KNOWLEDGE ENGINEERING

Building Cognitive Assistants for Evidence-Based Reasoning

This book presents a significant advancement in the theory and practice

of knowledge engineering, the discipline concerned with the development
of intelligent agents that use knowledge and reasoning to perform problem-
solving and decision-making tasks. It covers the main stages in the develop-
ment of a knowledge-based agent: understanding the application domain,
modeling problem solving in that domain, developing the ontology, learning
the reasoning rules, and testing the agent. The book focuses on a special
class of agents: cognitive assistants for evidence-based reasoning that learn
complex problem-solving expertise directly from human experts, support
experts, and nonexperts in problem solving and decision making, and teach
their problem-solving expertise to students.
A powerful learning agent shell, Disciple-EBR, is included with the book,
enabling students, practitioners, and researchers to develop cognitive
assistants rapidly in a wide variety of domains that require evidence-based
reasoning, including intelligence analysis, cybersecurity, law, forensics,
medicine, and education.

Gheorghe Tecuci (PhD, University of Paris-South, July 1988, and Poly-

technic Institute of Bucharest, December 1988) is Professor of Computer
Science and Director of the Learning Agents Center in the Volgenau School
of Engineering of George Mason University, Member of the Romanian
Academy, and former Chair of Artificial Intelligence in the Center for
Strategic Leadership of the U.S. Army War College.

Dorin Marcu (PhD, George Mason University, 2009) is Research Assistant

Professor, as well as Senior Software and Knowledge Engineer, in the
Learning Agents Center, Volgenau School of Engineering, George Mason
University.

Mihai Boicu (PhD, George Mason University, 2003) is Associate Professor

of Information Sciences and Technology, and Associate Director of the
Learning Agents Center, Volgenau School of Engineering, George Mason
University.

David A. Schum (PhD, Ohio State University, 1964) is Emeritus Professor

of Systems Engineering, Operations Research, and Law, as well as Chief
Scientist of the Learning Agents Center at George Mason University. He is
also Honorary Professor of Evidence Science at University College London.
Knowledge Engineering
Building Cognitive Assistants for
Evidence-Based Reasoning

GHEORGHE TECUCI
George Mason University

DORIN MARCU
George Mason University

MIHAI BOICU
George Mason University

DAVID A. SCHUM
George Mason University
One Liberty Plaza, New York, NY 10006

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of

education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781107122567
© Gheorghe Tecuci, Dorin Marcu, Mihai Boicu, and David A. Schum 2016

This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.

First published 2016

Printed in the United States of America by Sheridan Books, Inc.

A catalog record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: Tecuci, Gheorghe, author. | Marcu, Dorin, author. | Boicu, Mihai, author. |
Schum, David A., author.
Title: Knowledge engineering: building cognitive assistants for evidence-based reasoning /
Gheorghe Tecuci, George Mason University, Dorin Marcu, George Mason University,
Mihai Boicu, George Mason University, David A. Schum, George Mason University.
Description: New York NY : Cambridge University Press, 2016. | Includes bibliographical references and index.
Identifiers: LCCN 2015042941 | ISBN 9781107122567 (Hardback : alk. paper)
Subjects: LCSH: Expert systems (Computer science) | Intelligent agents (Computer software) | Machine learning |
Artificial intelligence | Knowledge, Theory of–Data processing.
Classification: LCC QA76.76.E95 T435 2016 | DDC 006.3/3–dc23 LC record available at
http://lccn.loc.gov/2015042941

ISBN 978-1-107-12256-7 Hardback

Cambridge University Press has no responsibility for the persistence or accuracy

of URLs for external or third-party Internet Web sites referred to in this publication
and does not guarantee that any content on such Web sites is, or will remain,
accurate or appropriate.
Contents

Preface page xv
Acknowledgments xxi
About the Authors xxiii

1 Introduction . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Understanding the World through Evidence-based
Reasoning 1
1.1.1 What Is Evidence? 1
1.1.2 Evidence, Data, and Information 1
1.1.3 Evidence and Fact 2
1.1.4 Evidence and Knowledge 2
1.1.5 Ubiquity of Evidence 5
1.2 Abductive Reasoning 5
1.2.1 From Aristotle to Peirce 5
1.2.2 Peirce and Sherlock Holmes on Abductive
Reasoning 6
1.3 Probabilistic Reasoning 9
1.3.1 Enumerative Probabilities: Obtained by Counting 9
1.3.1.1 Aleatory Probability 9
1.3.1.2 Relative Frequency and Statistics 9
1.3.2 Subjective Bayesian View of Probability 11
1.3.3 Belief Functions 13
1.3.4 Baconian Probability 16
1.3.4.1 Variative and Eliminative Inferences 16
1.3.4.2 Importance of Evidential Completeness 17
1.3.4.3 Baconian Probability of Boolean
Expressions 20
1.3.5 Fuzzy Probability 20
1.3.5.1 Fuzzy Force of Evidence 20
1.3.5.2 Fuzzy Probability of Boolean Expressions 21
1.3.5.3 On Verbal Assessments of Probabilities 22
1.3.6 A Summary of Uncertainty Methods and What
They Best Capture 23
1.4 Evidence-based Reasoning 25
1.4.1 Deduction, Induction, and Abduction 25
1.4.2 The Search for Knowledge 26
1.4.3 Evidence-based Reasoning Everywhere 27

13:46:00,
vi Contents

1.5 Artificial Intelligence 29

1.5.1 Intelligent Agents 30
1.5.2 Mixed-Initiative Reasoning 32
1.6 Knowledge Engineering 33
1.6.1 From Expert Systems to Knowledge-based Agents
and Cognitive Assistants 33
1.6.2 An Ontology of Problem-Solving Tasks 35
1.6.2.1 Analytic Tasks 36
1.6.2.2 Synthetic Tasks 36
1.6.3 Building Knowledge-based Agents 37
1.6.3.1 How Knowledge-based Agents Are Built
and Why It Is Hard 37
1.6.3.2 Teaching as an Alternative to
Programming: Disciple Agents 39
1.6.3.3 Disciple-EBR, Disciple-CD, and
TIACRITIS 40
1.7 Obtaining Disciple-EBR 41
1.8 Review Questions 42

2 Evidence-based Reasoning: Connecting the Dots . . . . 46

2.1 How Easy Is It to Connect the Dots? 46
2.1.1 How Many Kinds of Dots Are There? 47
2.1.2 Which Evidential Dots Can Be Believed? 48
2.1.3 Which Evidential Dots Should Be Considered? 50
2.1.4 Which Evidential Dots Should We Try to
Connect? 50
2.1.5 How to Connect Evidential Dots to Hypotheses? 52
2.1.6 What Do Our Dot Connections Mean? 54
2.2 Sample Evidence-based Reasoning Task: Intelligence
Analysis 56
2.2.1 Evidence in Search of Hypotheses 56
2.2.2 Hypotheses in Search of Evidence 58
2.2.3 Evidentiary Testing of Hypotheses 60
2.2.4 Completing the Analysis 62
2.3 Other Evidence-based Reasoning Tasks 64
2.3.1 Cyber Insider Threat Discovery and Analysis 64
2.3.2 Analysis of Wide-Area Motion Imagery 68
2.3.3 Inquiry-based Teaching and Learning in a Science
Classroom 70
2.3.3.1 Need for Inquiry-based Teaching and
Learning 70
2.3.3.2 Illustration of Inquiry-based Teaching
and Learning 71
2.3.3.3 Other Examples of Inquiry-based
Teaching and Learning 74
2.4 Hands On: Browsing an Argumentation 76
2.5 Project Assignment 1 81
2.6 Review Questions 81

13:46:00,
Contents vii

3 Methodologies and Tools for Agent Design and

Development . . . . . . . . . . . . . . . . . . . . . 83
3.1 A Conventional Design and Development Scenario 83
3.1.1 Conventional Design and Development Phases 83
3.1.2 Requirements Specification and Domain
Understanding 83
3.1.3 Ontology Design and Development 85
3.1.4 Development of the Problem-Solving Rules or
Methods 86
3.1.5 Verification, Validation, and Certification 87
3.2 Development Tools and Reusable Ontologies 88
3.2.1 Expert System Shells 88
3.2.2 Foundational and Utility Ontologies and Their
Reuse 89
3.2.3 Learning Agent Shells 90
3.2.4 Learning Agent Shell for Evidence-based
Reasoning 91
3.3 Agent Design and Development Using Learning
Technology 93
3.3.1 Requirements Specification and Domain
Understanding 93
3.3.2 Rapid Prototyping 93
3.3.3 Ontology Design and Development 100
3.3.4 Rule Learning and Ontology Refinement 101
3.3.5 Hierarchical Organization of the Knowledge
Repository 104
3.3.6 Learning-based Design and Development Phases 105
3.4 Hands On: Loading, Saving, and Closing Knowledge
Bases 107
3.5 Knowledge Base Guidelines 111
3.6 Project Assignment 2 111
3.7 Review Questions 112

4 Modeling the Problem-Solving Process . . . . . . . . . 113

4.1 Problem Solving through Analysis and Synthesis 113
4.2 Inquiry-driven Analysis and Synthesis 113
4.3 Inquiry-driven Analysis and Synthesis for Evidence-based
Reasoning 119
4.3.1 Hypothesis Reduction and Assessment Synthesis 119
4.3.2 Necessary and Sufficient Conditions 120
4.3.3 Sufficient Conditions and Scenarios 120
4.3.4 Indicators 121
4.4 Evidence-based Assessment 122
4.5 Hands On: Was the Cesium Stolen? 124
4.6 Hands On: Hypothesis Analysis and Evidence Search
and Representation 130
4.7 Believability Assessment 133
4.7.1 Tangible Evidence 133
4.7.2 Testimonial Evidence 135

13:46:00,
viii Contents

4.7.3 Missing Evidence 137

4.7.4 Authoritative Record 137
4.7.5 Mixed Evidence and Chains of Custody 138
4.8 Hands On: Believability Analysis 140
4.9 Drill-Down Analysis, Assumption-based Reasoning,
and What-If Scenarios 143
4.10 Hands On: Modeling, Formalization, and Pattern
Learning 144
4.11 Hands On: Analysis Based on Learned Patterns 146
4.12 Modeling Guidelines 147
4.13 Project Assignment 3 151
4.14 Review Questions 152

5 Ontologies . . . . . . . . . . . . . . . . . . . . . . 155
5.1 What Is an Ontology? 155
5.2 Concepts and Instances 156
5.3 Generalization Hierarchies 157
5.4 Object Features 158
5.5 Defining Features 158
5.6 Representation of N-ary Features 160
5.7 Transitivity 161
5.8 Inheritance 162
5.8.1 Default Inheritance 162
5.8.2 Multiple Inheritance 162
5.9 Concepts as Feature Values 163
5.10 Ontology Matching 164
5.11 Hands On: Browsing an Ontology 165
5.12 Project Assignment 4 168
5.13 Review Questions 168

6 Ontology Design and Development . . . . . . . . . . 174

6.1 Design and Development Methodology 174
6.2 Steps in Ontology Development 174
6.3 Domain Understanding and Concept Elicitation 176
6.3.1 Tutorial Session Delivered by the Expert 177
6.3.2 Ad-hoc List Created by the Expert 177
6.3.3 Book Index 177
6.3.4 Unstructured Interviews with the Expert 177
6.3.5 Structured Interviews with the Expert 177
6.3.6 Protocol Analysis (Think-Aloud Technique) 178
6.3.7 The Card-Sort Method 179
6.4 Modeling-based Ontology Specification 179
6.5 Hands On: Developing a Hierarchy of Concepts and
Instances 180
6.6 Guidelines for Developing Generalization Hierarchies 186
6.6.1 Well-Structured Hierarchies 186
6.6.2 Instance or Concept? 187
6.6.3 Specific Instance or Generic Instance? 188
6.6.4 Naming Conventions 188
6.6.5 Automatic Support 189

13:46:00,
Contents ix

6.7 Hands On: Developing a Hierarchy of Features 189

6.8 Hands On: Defining Instances and Their Features 192
6.9 Guidelines for Defining Features and Values 195
6.9.1 Concept or Feature? 195
6.9.2 Concept, Instance, or Constant? 196
6.9.3 Naming of Features 196
6.9.4 Automatic Support 197
6.10 Ontology Maintenance 197
6.11 Project Assignment 5 198
6.12 Review Questions 198

7 Reasoning with Ontologies and Rules . . . . . . . . . 202

7.1 Production System Architecture 202
7.2 Complex Ontology-based Concepts 203
7.3 Reduction and Synthesis Rules and the Inference Engine 204
7.4 Reduction and Synthesis Rules for Evidence-based
Hypotheses Analysis 206
7.5 Rule and Ontology Matching 207
7.6 Partially Learned Knowledge 212
7.6.1 Partially Learned Concepts 212
7.6.2 Partially Learned Features 213
7.6.3 Partially Learned Hypotheses 214
7.6.4 Partially Learned Rules 214
7.7 Reasoning with Partially Learned Knowledge 215
7.8 Review Questions 216

8 Learning for Knowledge-based Agents . . . . . . . . . 222

8.1 Introduction to Machine Learning 222
8.1.1 What Is Learning? 222
8.1.2 Inductive Learning from Examples 223
8.1.3 Explanation-based Learning 224
8.1.4 Learning by Analogy 225
8.1.5 Multistrategy Learning 226
8.2 Concepts 227
8.2.1 Concepts, Examples, and Exceptions 227
8.2.2 Examples and Exceptions of a Partially Learned
Concept 228
8.3 Generalization and Specialization Rules 229
8.3.1 Turning Constants into Variables 230
8.3.2 Turning Occurrences of a Variable into Different
Variables 230
8.3.3 Climbing the Generalization Hierarchies 231
8.3.4 Dropping Conditions 231
8.3.5 Extending Intervals 231
8.3.6 Extending Ordered Sets of Intervals 232
8.3.7 Extending Symbolic Probabilities 232
8.3.8 Extending Discrete Sets 232
8.3.9 Using Feature Definitions 233
8.3.10 Using Inference Rules 233

13:46:00,
x Contents

8.4 Types of Generalizations and Specializations 234

8.4.1 Definition of Generalization 234
8.4.2 Minimal Generalization 234
8.4.3 Minimal Specialization 235
8.4.4 Generalization of Two Concepts 236
8.4.5 Minimal Generalization of Two Concepts 236
8.4.6 Specialization of Two Concepts 237
8.4.7 Minimal Specialization of Two Concepts 237
8.5 Inductive Concept Learning from Examples 238
8.6 Learning with an Incomplete Representation Language 242
8.7 Formal Definition of Generalization 243
8.7.1 Formal Representation Language for Concepts 243
8.7.2 Term Generalization 245
8.7.3 Clause Generalization 245
8.7.4 BRU Generalization 246
8.7.5 Generalization of Concepts with Negations 247
8.7.6 Substitutions and the Generalization Rules 247
8.8 Review Questions 247

9 Rule Learning . . . . . . . . . . . . . . . . . . . . 252

9.1 Modeling, Learning, and Problem Solving 252
9.2 An Illustration of Rule Learning and Refinement 253
9.3 The Rule-Learning Problem 257
9.4 Overview of the Rule-Learning Method 258
9.5 Mixed-Initiative Example Understanding 260
9.5.1 What Is an Explanation of an Example? 260
9.5.2 Explanation Generation 262
9.6 Example Reformulation 264
9.7 Analogy-based Generalization 265
9.7.1 Analogical Problem Solving Based on
Explanation Similarity 265
9.7.2 Upper Bound Condition as a Maximally General
Analogy Criterion 266
9.7.3 Lower Bound Condition as a Minimally General
Analogy Criterion 268
9.8 Rule Generation and Analysis 270
9.9 Generalized Examples 270
9.10 Hypothesis Learning 271
9.11 Hands On: Rule and Hypotheses Learning 275
9.12 Explanation Generation Operations 279
9.12.1 Guiding Explanation Generation 279
9.12.2 Fixing Values 280
9.12.3 Explanations with Functions 280
9.12.4 Explanations with Comparisons 283
9.12.5 Hands On: Explanations with Functions and
Comparisons 285
9.13 Guidelines for Rule and Hypothesis Learning 285
9.14 Project Assignment 6 289
9.15 Review Questions 289

13:46:00,
Contents xi

10 Rule Refinement . . . . . . . . . . . . . . . . . . . 294

10.1 Incremental Rule Refinement 294
10.1.1 The Rule Refinement Problem 294
10.1.2 Overview of the Rule Refinement Method 295
10.1.3 Rule Refinement with Positive Examples 296
10.1.3.1 Illustration of Rule Refinement with a
Positive Example 296
10.1.3.2 The Method of Rule Refinement with
a Positive Example 298
10.1.3.3 Summary of Rule Refinement with a
Positive Example 300
10.1.4 Rule Refinement with Negative Examples 300
10.1.4.1 Illustration of Rule Refinement with
Except-When Conditions 300
10.1.4.2 The Method of Rule Refinement with
Except-When Conditions 305
10.1.4.3 Illustration of Rule Refinement
through Condition Specialization 305
10.1.4.4 The Method of Rule Refinement
through Condition Specialization 307
10.1.4.5 Summary of Rule Refinement with a
Negative Example 308
10.2 Learning with an Evolving Ontology 309
10.2.1 The Rule Regeneration Problem 309
10.2.2 On-Demand Rule Regeneration 310
10.2.3 Illustration of the Rule Regeneration Method 312
10.2.4 The Rule Regeneration Method 316
10.3 Hypothesis Refinement 316
10.4 Characterization of Rule Learning and Refinement 317
10.5 Hands On: Rule Refinement 319
10.6 Guidelines for Rule Refinement 321
10.7 Project Assignment 7 322
10.8 Review Questions 322

11 Abstraction of Reasoning . . . . . . . . . . . . . . . 329

11.1 Statement Abstraction 329
11.2 Reasoning Tree Abstraction 331
11.3 Reasoning Tree Browsing 331
11.4 Hands On: Abstraction of Reasoning 331
11.5 Abstraction Guideline 334
11.6 Project Assignment 8 335
11.7 Review Questions 335

12 Disciple Agents . . . . . . . . . . . . . . . . . . . . 338

12.1 Introduction 338
12.2 Disciple-WA: Military Engineering Planning 338
12.2.1 The Workaround Planning Problem 338
12.2.2 Modeling the Workaround Planning Process 341
12.2.3 Ontology Design and Development 343

13:46:00,
xii Contents

12.2.4 Rule Learning 345

12.2.5 Experimental Results 346
12.3 Disciple-COA: Course of Action Critiquing 348
12.3.1 The Course of Action Critiquing Problem 348
12.3.2 Modeling the COA Critiquing Process 351
12.3.3 Ontology Design and Development 352
12.3.4 Training the Disciple-COA Agent 355
12.3.5 Experimental Results 360
12.4 Disciple-COG: Center of Gravity Analysis 364
12.4.1 The Center of Gravity Analysis Problem 364
12.4.2 Overview of the Use of Disciple-COG 367
12.4.3 Ontology Design and Development 376
12.4.4 Script Development for Scenario Elicitation 376
12.4.5 Agent Teaching and Learning 380
12.4.6 Experimental Results 383
12.5 Disciple-VPT: Multi-Agent Collaborative Planning 387
12.5.1 Introduction 387
12.5.2 The Architecture of Disciple-VPT 388
12.5.3 The Emergency Response Planning Problem 389
12.5.4 The Disciple-VE Learning Agent Shell 390
12.5.5 Hierarchical Task Network Planning 394
12.5.6 Guidelines for HTN Planning 396
12.5.7 Integration of Planning and Inference 400
12.5.8 Teaching Disciple-VE to Perform Inference Tasks 403
12.5.9 Teaching Disciple-VE to Perform Planning Tasks 409
12.5.9.1 Why Learning Planning Rules Is
Difficult 409
12.5.9.2 Learning a Set of Correlated Planning
Rules 409
12.5.9.3 The Learning Problem and Method for
a Set of Correlated Planning Rules 413
12.5.9.4 Learning Correlated Planning Task
Reduction Rules 413
12.5.9.5 Learning Correlated Planning Task
Concretion Rules 414
12.5.9.6 Learning a Correlated Action
Concretion Rule 415
12.5.10 The Virtual Experts Library 416
12.5.11 Multidomain Collaborative Planning 420
12.5.12 Basic Virtual Planning Experts 421
12.5.13 Evaluation of Disciple-VPT 422
12.5.14 Final Remarks 422

13 Design Principles for Cognitive Assistants . . . . . . . 426

13.1 Learning-based Knowledge Engineering 426
13.2 Problem-Solving Paradigm for User–Agent
Collaboration 427
13.3 Multi-Agent and Multidomain Problem Solving 427
13.4 Knowledge Base Structuring for Knowledge Reuse 427
13.5 Integrated Teaching and Learning 428

13:46:00,
Contents xiii

13.6 Multistrategy Learning 428

13.7 Knowledge Adaptation 429
13.8 Mixed-Initiative Modeling, Learning, and Problem
Solving 429
13.9 Plausible Reasoning with Partially Learned Knowledge 430
13.10 User Tutoring in Problem Solving 430
13.11 Agent Architecture for Rapid Agent Development 430
13.12 Design Based on a Complete Agent Life Cycle 431

References 433
Appendixes 443
Summary: Knowledge Engineering Guidelines 443
Summary: Operations with Disciple-EBR 444
Summary: Hands-On Exercises 446
Index 447

13:46:00,
13:46:00,
Preface

BOOK PURPOSE

This is a book on knowledge engineering, the discipline concerned with the

development of intelligent agents that use knowledge and reasoning to
perform problem-solving and decision-making tasks. The book covers the
theory and practice of the main stages in the development of a knowledge-
based agent: understanding the application domain, modeling problem
solving in that domain, developing the ontology, learning the reasoning
rules, and testing the agent. However, it does this by focusing on a special
class of agents: cognitive assistants that learn complex problem-solving
expertise directly from human experts, support experts, and nonexperts in
problem solving and decision making and teach their problem-solving
expertise to students. These are learning agents that are taught by their
users in ways that are similar to how a student, an apprentice, or a new
collaborator is taught, through problem-solving examples and explanations
and by supervising and correcting their behavior. Because such agents
learn to replicate the problem-solving behavior of their users, we have
called them Disciple agents.
This book presents a significant advancement in the theory and practice
of knowledge engineering, where many tasks are performed by a typical
computer user and a learning agent, with only limited support from a
knowledge engineer. To simplify further the development of the cognitive
assistants by typical users, we have focused on the development of cogni-
tive assistants for evidence-based reasoning. Evidence-based reasoning is
at the core of many problem-solving and decision-making tasks in a wide
variety of domains, including intelligence analysis, cybersecurity, law,
forensics, medicine, physics, chemistry, history, archaeology, education,
and many others. Nevertheless, the last part of the book presents Disciple
agents for applications that did not involve evidence-based reasoning.
Because knowledge engineering is a practical activity, it is best learned
by doing. Therefore, this book presents the theory and methodology of
developing cognitive assistants in conjunction with a practical tool,
Disciple-EBR, a learning agent shell for evidence-based reasoning (EBR).
Consequently, each chapter typically contains a theoretical part presenting
general concepts and methods, a methodological part with guidelines on
the application of the methods, and a practical part on the actual employ-
ment of these methods with Disciple-EBR. It also includes project assign-
ments and review questions.

xv

13:47:03,
.001
xvi Preface

This book addresses issues of interest to a large spectrum of readers

from academia, research, and industry. We have used drafts of this book in
our computer science courses on knowledge engineering, expert systems,
and knowledge-based agents, at both undergraduate and graduate levels,
because it covers the theory and practice of the main stages in the devel-
opment of knowledge-based agents. We have also used some parts of the
book in introductory courses on artificial intelligence, and other parts in the
courses on knowledge acquisition and machine learning. These are all
examples of courses where this book could be used.
Researchers in knowledge engineering will find in this book an inte-
grated approach that advances the theory and practice in the field. We
believe that further research and development of this approach will enable
typical computer users to develop their own cognitive assistants without
any knowledge engineering assistance. Thus, non–computer scientists will
no longer be only users of generic programs developed by others (such as
word processors or Internet browsers), as they are today, but also agent
developers themselves. They will be able to train their personal assistants to
help them with their increasingly complex tasks in the knowledge society,
which should have a significant beneficial impact on their work and life.
Practitioners that develop various types of knowledge-based systems
will find in this book a detailed, yet intuitive, presentation of an agent
development methodology and tool, as well as several case studies of
developing intelligent agents that illustrate different types of agents that
are relevant to a wide variety of application domains. In fact, a comple-
mentary book, Intelligence Analysis as Discovery of Evidence, Hypotheses,
and Arguments: Connecting the Dots (Tecuci et al., 2016), presents the
practical application of Disciple-CD, an agent developed with Disciple-EBR
for intelligence analysis problems.

BOOK CONTENTS

Here is a route or map we will follow in the learning venture you will have
with the assistance of Disciple-EBR. Chapter 1 is a general introduction to
the topics that form the basis of this book. It starts with the problem
of understanding the world through evidence-based reasoning. It then
presents abductive reasoning, five different conceptions of probability
(enumerative, subjective Bayesian, Belief Functions, Baconian, and Fuzzy),
and how deductive, abductive, and inductive (probabilistic) reasoning are
used in evidence-based reasoning. After that, it introduces artificial intelli-
gence and intelligent agents, and the challenges of developing such agents
through conventional knowledge engineering. Afterward, it introduces the
development of agents through teaching and learning, which is the
approach presented in this book.
Chapter 2 is an overview of evidence-based reasoning, which is a focus
of this book. It starts with a discussion of the elements that make evidence-
based reasoning an astonishingly complex task. It then introduces a sys-
tematic approach that integrates abduction, deduction, and induction to
solve a typical evidence-based reasoning task, using intelligence analysis as
an example. Finally, it shows the application of the same approach to other

13:47:03,
.001
Preface xvii

evidence-based reasoning tasks in cybersecurity, geospatial intelligence,

and critical thinking education.
Chapter 3 is an overview of the main methodologies and tools for the
design and development of knowledge-based agents. It first presents the
conventional approach of developing such agents by a knowledge engineer
working with a subject matter expert. It then introduces different types of
agent shells, which are the typical tools for building agents, and discusses the
use of foundational and utility ontologies. The more advanced of these tools,
the learning agent shells, are at the basis of a new and more powerful
approach to agent design and development, an overview of which is pre-
sented in the second part of Chapter 3. This learning-based approach is
illustrated with the development of a cognitive assistant for assessing a
potential PhD advisor for a student, an example that is used in the following
chapters to present in detail each of the main agent development stages.
Chapter 4 presents the modeling of the problem-solving process
through analysis and synthesis, which is the most challenging task in the
development of a knowledge-based agent. This chapter starts with a more
general version of this process for any type of problems, which is used in
the Disciple agents presented in Chapter 12. After that, the chapter presents
an easier, customized version for evidence-based hypothesis analysis,
which is used in the chapters that follow. Chapter 4 also introduces an
ontology of evidence and presents the modeling of the believability assess-
ment for different types of evidence.
Chapters 5 and 6 present the representation of knowledge through
ontologies, as well as their design and development. They discuss the
representation of the objects from the agent’s application domain and their
relationships. These chapters also discuss the basic reasoning operations of
transitivity, inheritance, and matching. They cover several approaches to
concept elicitation, as well as a systematic approach to modeling-based
ontology specification. The chapters also address the complexity of ontol-
ogy maintenance.
Chapter 7 presents the agent’s reasoning based on ontologies and rules,
in the context of a production system architecture. It starts with defining
the representation of complex ontology-based concepts and the use of
these concepts in the reasoning rules. It defines the reduction and synthe-
ses rules in general, and the special case of these rules for evidence-based
reasoning. It then presents the process of problem solving through analysis
and synthesis, which is accomplished through rule and ontology matching.
This chapter also introduces the representation of and reasoning with
partially learned concepts, features, hypotheses, and rules.
Chapter 8 starts with a general introduction to machine learning and to
several learning strategies that are most relevant for knowledge engineer-
ing. It continues with the definition of the generalization and specializa-
tions of concepts through the use of inductive rules. After that, the chapter
defines the basic operations of minimal and maximal generalizations and
specialization of concepts, which are at the basis of rule learning and
refinement discussed in Chapters 9 and 10. The chapter ends with the
presentation of a formal definition of generalization.
Chapter 9 presents the mixed-initiative rule learning method that
enables an agent to learn a general rule from a specific example of a

13:47:03,
.001
xviii Preface

reasoning step. It starts with an overview of the integration of modeling,

learning, and problem solving and an illustration of this process. The
chapter then introduces the rule learning problem and method. After that,
it presents in detail the phases of rule learning, such as the mixed-initiative
understanding of the example, and its analogy-based generalizations,
which result in a minimal and a maximal generalization forming the
plausible version space of the rule. This chapter also discusses the learning
of rules involving functions and relational operators.
The refinement of a partially learned rule is presented in Chapter 10.
After introducing the incremental rule refinement problem, this chapter
presents the refinement of the rule based on positive examples, and then
the refinement based on negative examples. This may result in a very
complex rule characterized by a main applicability condition and several
“except-when” conditions, which capture the reasoning of a subject matter
expert. There are various refinement strategies, depending on the position
of the examples with respect to the bounds of the rule’s conditions and also
on the possible explanations of these examples, but in all cases, they
involve simple interactions with the subject matter expert. Changes in the
ontology require the learned rules to be updated, the corresponding
method being presented and illustrated in the second part of this chapter.
The chapter ends with a characterization of rule learning and refinement,
which enable a non–computer scientist to teach an agent.
The chapters dedicated to the individual phases of agent development
end with Chapter 11, which discusses the abstractions of individual hypoth-
eses and of the reasoning tree to facilitate its browsing, understanding, and
further development by the end-user.
As previously indicated, this book focuses on the development of agents
for evidence-based reasoning tasks. However, the presented theory and
methodology are also applicable to other types of agents. These agents have
been developed with learning agent shells representing previous imple-
mentations of the Disciple theory and methodology. Four of these agents
are presented in Chapter 12. Disciple-WA is an agent that uses expert
knowledge from military engineering manuals to develop alternative plans
of actions that a military unit can use to work around (WA) damages to a
transportation infrastructure, such as a damaged, destroyed, or mined
bridge, road, or tunnel. The goal is to find the quickest way for the military
unit to bypass the encountered obstacle. There were several cases where
the Disciple-WA agent generated better solutions than those of the human
expert who evaluated the developed systems, as well as cases where the
agent generated new solutions that this expert did not consider.
Disciple-COA is an agent trained by a military expert to critique courses
of action (COA) with respect to the principles of war and the tenets of Army
operations. It receives as input the description of a course of action and
returns a list of strengths and weaknesses of various levels, together with
their justifications. A remarkable feature of this agent is that it was judged
to exceed the performance of the subject matter experts that defined the
problems for its evaluation.
Disciple-COG is an agent that was trained to identify and assess the
center of gravity (COG) candidates of the opposing forces in a military
scenario. Correctly identifying the centers of gravity of the opposing forces

13:47:03,
.001
Preface xix

is of highest importance in any conflict, and Disciple-COG has been used

for many years in courses at the U.S. Army War College, as well as at other
military institutions, to teach students how to perform a strategic center-of-
gravity analysis of a scenario of interest.
The last agent described in Chapter 12 is Disciple-VPT (Virtual Planning
Team). Disciple-VPT consists of virtual planning experts that collaborate to
develop plans of actions requiring expertise from multiple domains. They
are assembled from an extensible library of such agents. The basic com-
ponent of Disciple-VPT is the Disciple-VE learning agent shell that can be
taught how to plan directly by a subject matter expert. Copies of the
Disciple-VE shells can be used by experts in different domains to rapidly
populate the library of virtual experts (VEs) of Disciple-VPT.
The learning-based approach to knowledge engineering presented in
this book illustrates the application of several design principles that are
useful in the development of cognitive assistants in general. Therefore, as a
conclusion, Chapter 13 summarizes these principles, which are illustrated
throughout this book.
The book also includes several appendices that summarize important
aspects from the chapters: the list of the knowledge-engineering guidelines
for each of the main stages of agent development, the list of operations of
Disciple-EBR, and the list of the hands-on exercises. Answers to selected
questions from each chapter are made available to the instructors.

BACKGROUND

The learning-based knowledge-engineering theory, methodology, and tools

presented in this book are the result of many years of research and experi-
mentation that produced increasingly more general and powerful versions.
During this evolution, one may identify four main stages. The first stage
corresponds to the PhD work of Gheorghe Tecuci presented in his thesis
“Disciple: A Theory, Methodology and System for Learning Expert Know-
ledge” (Thèse de Docteur en Science, Université de Paris-Sud, July 1988).
The main emphasis of that work was on rule learning. The developed
methods are among the first multistrategy approaches to learning that later
grew into the subfield of multistrategy machine learning (Tecuci, 1993;
Michalski and Tecuci, 1994). They also are among the first attempts to
integrate machine learning and knowledge acquisition (Tecuci et al., 1994;
Tecuci and Kodratoff, 1995). This work benefited from the collaboration of
Yves Kodratoff and the support of Mihai Drăgănescu. It was done with the
support of the Romanian Research Institute for Informatics, the Romanian
Academy, and the French National Center for Scientific Research.
The second stage in the evolution of the Disciple approach is presented
in the book Building Intelligent Agents: An Apprenticeship Multistrategy
Learning Theory, Methodology, Tool and Case Studies, published by Aca-
demic Press in 1998. It includes an improvement of Disciple’s multistrategy
learning methods and their extension with guided knowledge-elicitation
methods. It also includes ontology development tools and illustrations of
the application of the Disciple rule-learning approach to a variety of
domains, such as teaching of higher-order thinking skills in history and

xix

13:47:03,
.001
xx Preface

statistics, engineering design, and military simulation. This work benefited

from the collaboration of several of Tecuci’s PhD students, particularly
Thomas Dybala, Michael Hieb, Harry Keeling, and Kathryn Wright, and it
was partially supported by George Mason University, the National Science
Foundation, and the Defense Advanced Research Projects Agency.
The third stage in the evolution of the Disciple approach is represented
by the Disciple agents described in Chapter 12 of this book. This represents
a significant extension of the Disciple approach, with methods for problem
solving through analysis and synthesis, modeling of the problem solving
process, ontology development, and scenario elicitation. At this stage,
Disciple has become a complete, end-to-end agent development method-
ology that has been applied to develop powerful agents, such as Disciple-
COG, used for a period of ten years in courses at the U.S. Army War College
and at other institutions. This work also benefited from the collaboration of
Tecuci’s students, first of all Mihai Boicu and Dorin Marcu, and also
Michael Bowman, Vu Le, Cristina Boicu, Bogdan Stanescu, and Marcel
Barbulescu. This work was partially supported by George Mason University,
the Air Force Office of Scientific Research, the Air Force Research Labora-
tory, the Defense Advanced Research Projects Agency, the National Science
Foundation, and others.
Finally, the latest stage in the evolution of the Disciple approach is
represented by the rest of this book. A main conceptual advancement over
the previous stage consists in its extension with a theory of evidence-based
reasoning, greatly facilitated by the collaboration of David A. Schum. This
resulted in a very powerful theory, methodology, and tool for the develop-
ment of agents for complex evidence-based reasoning applications, such as
intelligence analysis. Two of such agents are TIACRITIS (Teaching Intelli-
gence Analysts Critical Thinking Skills) and Disciple-CD (Disciple cognitive
assistant for Connecting the Dots). This work was partially supported by
George Mason University and by several agencies of the U.S. government,
including the Department of Defense.

13:47:03,
.001
Acknowledgments

We are very grateful to the many individuals who, in various ways,

supported our research, including Kelcy Allwein, Cindy Ayers, Murray
Burke, Douglas Campbell, William Cleckner, Jerry Comello, John Donelan,
Jim Donlon, Susan Durham, Keri Eustis, Michael Fletcher, Erin Gibbens,
Lloyd Griffiths, David Gunning, Ben Hamilton, Sharon Hamilton, Robert
Herklotz, Phillip Hwang, Eric Jones, Alex Kilpatrick, David Luginbuhl, Joan
McIntyre, Jean-Michel Pomrade, Michelle Quirk, Peter Rocci, William
Rzepka, Kimberly Urban, Joan Vallancewhitacre, and Ben Wible.
We also want to thank Heather Bergman, who invited us to write this
book for the Cambridge University Press, as well as to senior editor Lauren
Cowles, who is a great professional to work with.

xxi

13:47:10,
13:47:10,
About the Authors

Gheorghe Tecuci (PhD, Université de Paris-Sud, July 1988, and Polytechnic

Institute of Bucharest, December 1988) is Professor of Computer Science
and Director of the Learning Agents Center in the Volgenau School of
Engineering of George Mason University, Member of the Romanian Acad-
emy, and former Chair of Artificial Intelligence in the Center for Strategic
Leadership of the U.S. Army War College. He has published around two
hundred papers, including eleven books, with contributions to artificial
intelligence, knowledge engineering, cognitive assistants, machine learn-
ing, evidence-based reasoning, and intelligence analysis. He has received
the U.S. Army Outstanding Civilian Service Medal (for “groundbreaking
contributions to the application of artificial intelligence to center of gravity
determination”) and the Innovative Application Award from the American
Association for Artificial Intelligence.

Dorin Marcu (PhD, George Mason University, 2009) is Research Assistant

Professor, as well as Senior Software and Knowledge Engineer, in the
Learning Agents Center, Volgenau School of Engineering, George Mason
University. He has published more than forty papers, including five books,
with contributions to adaptive user interfaces, mixed-initiative human–
computer interaction, and cognitive assistants. He has received the Innova-
tive Application Award from the American Association for Artificial
Intelligence.

Mihai Boicu (PhD, George Mason University, 2002) is Associate Professor

of Information Sciences and Technology and Associate Director of the
Learning Agents Center, in the Volgenau School of Engineering of George
Mason University. He has published over ninety papers, including five
books, with contributions to problem solving and multistrategy learning
in dynamic and evolving representation spaces, mixed-initiative inter-
action, multi-agent systems architecture, collaboration and coordination,
abstraction-based reasoning, knowledge representation, and knowledge
acquisition. He has received the Innovative Application Award from the
American Association for Artificial Intelligence.

David A. Schum (PhD, Ohio State University, 1964) is Emeritus Professor of

Systems Engineering, Operations Research, and Law, as well as Chief
Scientist of the Learning Agents Center at George Mason University, and
Honorary Professor of Evidence Science at University College London.

xxiii

13:47:11,
xxiv About the Authors

He has followed a career-long interest in the study of the properties,

uses, discovery, and marshaling of evidence in probabilistic reasoning.
Dr. Schum has published more than one hundred papers in a variety of
journals and eight books, including The Evidential Foundations of Probabil-
istic Reasoning, Analysis of Evidence, Evidence and Inference for the Intelli-
gence Analyst, and Probabilistic Analysis of the Sacco and Vanzetti Evidence,
being recognized as one of the founding fathers of the Science of Evidence.

13:47:11,
 1 Introduction

1.1 UNDERSTANDING THE WORLD THROUGH

EVIDENCE-BASED REASONING

We can try to understand the world in various ways, an obvious one being the employment of
empirical methods for gathering and analyzing various forms of evidence about phenomena,
events, and situations of interest to us. This will include work in all of the sciences, medicine,
law, intelligence analysis, history, political affairs, current events, and a variety of other contexts
too numerous to mention. In the sciences, this empirical work will involve both experimental
and nonexperimental methods. In some of these contexts, notably in the sciences, we are able
to devise mathematical and logical models that allow us to make inferences and predictions
about complex matters of interest to us. But in every case, our understanding rests on our
knowledge of the properties, uses, discovery, and marshaling of evidence. This is why we begin
this book with a careful consideration of reasoning based on evidence.

1.1.1 What Is Evidence?

You might think this question is unnecessary since everyone knows what evidence is. How-
ever, matters are not quite that simple, since the term evidence is not so easy to define and its
use often arouses controversy. One problem with the definition of evidence is that several
other terms are often used synonymously with it, when in fact there are distinctions to be made
among these terms that are not always apparent. Quite unnecessary controversy occurs since
some believe that the term evidence arises and has meaning only in the field of law.
Consulting a dictionary actually does not assist us much in defining the term. For
example, look at the Oxford English Dictionary under the term evidence and you will be led
in a circle; evidence is ultimately defined as being evidence.
A variety of terms are often used as synonyms for the term evidence: data, items of
information, facts, and knowledge. When examined carefully, there are some valid and
important distinctions to be made among these terms, as we will now discuss.

1.1.2 Evidence, Data, and Information

Consider the terms data and items of information.
Data are uninterpreted signals, raw observations, measurements, such as the number 6,
the color “red,” or the sequence of dots and lines “. . .–. . .”.

13:51:08,
.002
 2 Chapter 1. Introduction

Information is data equipped with meaning provided by a certain context, such as

“6 am,” “red traffic light,” “red tomato,” or the “S O S” emergency alert.
Untold trillions of data and items of information exist that will almost certainly never
become evidence in most inferences. Here’s an item of information for you: Professor
Schum has a long and steep driveway in front of his house that makes shoveling snow off
of it very difficult in the winter. Can you think of any situation in which this item of
information could become evidence? About the only matter in which this information
could become interesting evidence involves the question: Why did Schum and his wife,
Anne, ever purchase this house in the first place? As we will discuss, items of information
become evidence only when their relevance is established regarding some matter to be
proved or disproved.

1.1.3 Evidence and Fact

Now consider the term fact; there are some real troubles here as far as its relation to the
term evidence is concerned. How many times have you heard someone say, “I want all
the facts before I draw a conclusion or make a decision,” or, “I want to know the facts in
this matter”? The first question is easily answered: We will never have all the facts in any
matter of inferential interest. Answers to the second question require a bit of careful
thought. Here is an example of what is involved:
Suppose we are police investigators interviewing Bob, who is a witness of a car accident
that just happened at a particular intersection. Bob tells us that the Ford car did not stop at
the red light signal. Now we regard it as fact that Bob gave us this information: We all just
heard him give it to us. But whether the Ford car did not stop at the red light is only an
inference and is not a fact. This is precisely why we need to distinguish carefully between
an event and evidence for this event.
Here is what we have: Bob has given us evidence E*, saying that event E occurred,
where E is the event that the Ford car did not stop at the red light signal. Whether this
event E did occur or not is open to question and depends on Bob’s competence and
credibility. If we take it as fact that event E did occur, just because Bob said it did, we
would be overlooking the believability foundation for any inference we might make from
his evidence E*. Unfortunately, it so often happens that people regard the events reported
in evidence as being facts when they are not. Doing this suppresses all uncertainties we
may have about the source’s credibility and competence if the evidence is testimonial in
nature. We have exactly the same concerns about the credibility of tangible evidence. For
example, we have been given a tangible object or an image as evidence E* that we believe
reveals the occurrence of event E. But we must consider whether this object or image is
authentic and it is what we believe it to be. In any case, the events recorded in evidence
can be regarded as facts only if provided by perfectly credible sources, something we
almost never have. As another example, any information we find on the Internet should be
considered as only a claim by its source rather than as fact, that is, as evidence about a
potential fact rather than a fact.

1.1.4 Evidence and Knowledge

Now consider the term knowledge and its relation with evidence. Here is where things get
interesting and difficult. As you know, the field of epistemology is the study of knowledge,

13:51:08,
.002
1.1. Evidence-based Reasoning 3

what we believe it may be, and how we obtain it. Two questions we would normally ask
regarding what Bob just told us are as follows:

Does Bob really know what he just told us, that the Ford car did not stop at the red light
signal?

Do we ourselves then also know, based on Bob’s testimony, that the Ford car did not
stop at the red light signal?

Let’s consider the first question. For more than two millennia, some very learned people
have troubled over the question: What do we mean when we say that person A knows that
event B occurred? To apply this question to our source Bob, let’s make an assumption that
will simplify our answering this question. Let’s assume that Bob is a competent observer in
this matter. Suppose we have evidence that Bob was actually himself at the intersection
when the accident happened. This is a major element of Bob’s competence. Bob’s
credibility depends on different matters, as we will see.
Here is what a standard or conventional account says about whether Bob knows
that the car did not stop at the red light signal. First, here is a general statement of the
standard account of knowledge: Knowledge is justified true belief. Person knows that
event B occurred if:

Event B did occur [True];

got nondefective evidence that B occurred [Justified]; and

believed this evidence [Belief].

This standard analysis first says that event B must have occurred for to have knowledge
of its occurrence. This is what makes ’s belief true. If B did not occur, then could not
know that it occurred. Second, ’s getting nondefective evidence that B occurred is
actually where ’s competence arises. could not have gotten any evidence, defective
or nondefective, if was not where B could have occurred. Then, believed the evidence
received about the occurrence of event B, and was justified in having this belief by
obtaining nondefective evidence of B’s occurrence.
So, in the case involving Bob’s evidence, Bob knows that the Ford car did not stop at the
red light signal if:

The car did not stop at the red light signal,

Bob got nondefective evidence that the car did not stop at the red signal, and

Bob believed this evidence.

If all of these three things are true, we can state on this standard analysis that Bob knows
that the Ford car did not stop at the red light signal.
Before we proceed, we must acknowledge that this standard analysis has been very
controversial in fairly recent years and some philosophers claim to have found alleged
paradoxes and counterexamples associated with it. Other philosophers dispute these claims.
Most of the controversy here concerns the justification condition: What does it mean to say
that A is justified in believing that B occurred? In any case, we have found this standard
analysis very useful as a heuristic in our analyses of the credibility of testimonial evidence.
But now we have several matters to consider in answering the second question: Do we
ourselves also know, based on Bob’s testimony, that the Ford car did not stop at the red
light signal? The first and most obvious fact is that we do not know the extent to which any
of the three events just described in the standard analysis are true. We cannot get inside

13:51:08,
.002
4 Chapter 1. Introduction

Bob’s head to obtain necessary answers about these events. Starting at the bottom, we do
not know for sure that Bob believes what he just told us about the Ford car not stopping at
the red light signal. This is a matter of Bob’s veracity or truthfulness. We would not say that
Bob is being truthful if he told us something he did not believe.
Second, we do not know what sensory evidence Bob obtained on which to base his
belief and whether he based his belief at all on this evidence. Bob might have believed that
the Ford car did not stop at the red light signal either because he expected or desired this
to be true. This involves Bob’s objectivity as an observer. We would not say that Bob was
objective in this observation if he did not base his belief on the sensory evidence he
obtained in his observation.
Finally, even if we believe that Bob was an objective observer who based his belief
about the accident on sensory evidence, we do not know how good this evidence was.
Here we are obliged to consider Bob’s sensory sensitivities or accuracy in the conditions
under which Bob made his observations. Here we consider such obvious things as Bob’s
visual acuity. But there are many other considerations, such as, “Did Bob only get a
fleeting look at the accident when it happened?” “Is Bob color-blind?“ “Did he make this
observation during a storm?” and, “What time of day did he make this observation?” For a
variety of such reasons, Bob might simply have been mistaken in his observation: The light
signal was not red when the Ford car entered the intersection.
So, what it comes down to is that the extent of our knowledge about whether the Ford
car did not stop at the red light signal, based on Bob’s evidence, depends on these three
attributes of Bob’s credibility: his veracity, objectivity, and observational sensitivity. We
will have much more to say about assessing the credibility of sources of evidence, and how
Disciple-EBR can assist you in this difficult process, in Section 4.7 of this book.
Now, we return to our role as police investigators. Based on evidence we have about
Bob’s competence and credibility, suppose we believe that the event he reports did occur;
we believe that “E: The Ford car did not stop at the red light signal,” did occur. Now we
face the question: So what? Why is knowledge of event E of importance to us? Stated more
precisely: How is event E relevant in further inferences we must make? In our investigation
so far, we have other evidence besides Bob’s testimony. In particular, we observe a Toyota
car that has smashed into a light pole at this intersection, injuring the driver of the Toyota,
who was immediately taken to a hospital. In our minds, we form the tentative chain of
reasoning from Figure 1.1.
This sequence of events, E ➔ F ➔ G ➔ H, is a relevance argument or chain of reasoning
whose links represent sources of doubt interposed between the evidence E* and the
hypothesis H. An important thing to note is that some or all of these events may not be
true. Reducing our doubts or uncertainties regarding any of these events requires a variety

H: The driver of the Ford car bears the responsibility for the accident.

G: The driver of the Toyota swerved to avoid the Ford car and smashed into a light pole.

F: The driver of the Toyota car, having a green light at this intersection, saw the Ford car running the red light.

E: The Ford car did not stop at the red light signal at this intersection.

E*: Bob’s testimony that the Ford car did not stop at the red light signal at this intersection.

Figure 1.1. Tentative chain of reasoning.

13:51:08,
.002
 1.2. Abductive Reasoning 5

of additional evidence. The extent of our knowledge about the relative probabilities of our
final hypothesis depends on the believability of our evidence and on the defensibility
and strength of our relevance arguments, as discussed in Section 2.2. The whole point here
is that the relation between evidence and knowledge is not a simple one at all.

1.1.5 Ubiquity of Evidence

Finally, we must consider the controversy over the use of the term evidence instead of
the other terms we just examined. The mistake made by some people is to consider that
evidence concerns only objects, testimony, or other items introduced in a court trial. This
controversy and confusion has been recognized by eminent evidence scholars in the field
of law. For example, in his marvelous book Evidence, Proof, and Facts: A Book of Sources,
Professor Peter Murphy (2003, p. 1) notes the curious fact that the term evidence is so
commonly associated only with the field of law:

The word “evidence” is associated more often with lawyers and judicial trials
than with any other cross-section of society or form of activity. . . . In its
simplest sense, evidence may be defined as any factual datum which in some
manner assists in drawing conclusions, either favorable or unfavorable, to
some hypothesis whose proof or refutation is being attempted.

Murphy notes that this term is appropriate in any field in which conclusions are reached
from any relevant datum. Thus, physicians, scientists of any ilk, historians, and persons of
any other conceivable discipline, as well as ordinary persons, use evidence every day in
order to draw conclusions about matters of interest to them.
We believe there is a very good reason why many persons are so often tempted to
associate the term evidence only with the field of law. It happens that the Anglo-American
system of laws has provided us with by far the richest legacy of experience and scholarship
on evidence of any field known to us. This legacy has arisen as a result of the development
of the adversarial system for settling disputes and the gradual emergence of the jury
system, in which members of the jury deliberate on evidence provided by external
witnesses. This legacy has now been accumulating over at least the past six hundred years
(Anderson et al., 2005).
Evidence-based reasoning involves abductive, deductive, and inductive (probabilistic)
reasoning. The following sections briefly introduce them.

1.2 ABDUCTIVE REASONING

1.2.1 From Aristotle to Peirce

Throughout history, some of the greatest minds have tried to understand the world
through a process of discovery and testing of hypotheses based on evidence. We have
found the metaphor of an arch of knowledge to be very useful in summarizing the many
ideas expressed over the centuries concerning the generation of new thoughts and new
evidence. This metaphor comes from the work of the philosopher David Oldroyd in a
valuable work having this metaphor as its title (Oldroyd, 1986). Figure 1.2 shows this
metaphor applied in the context of science. Based upon existing records, it seems that

13:51:08,
.002
 6 Chapter 1. Introduction

Possible Hypotheses
or Explanations

Hypothesis
Generation
Hypothesis

Testing
? Deduction

Observations of New Observable

Events in Nature Phenomena

Figure 1.2. The “arch of knowledge” in science.

Aristotle (384 bc–322 bc) was the first to puzzle about the generation or discovery of new
ideas in science. From sensory observations, we generate possible explanations, in the
form of hypotheses, for these observations. It was never clear from Aristotle's work what
label should be placed on the upward, or discovery-related, arm of the arch in Figure 1.2.
By some accounts, Aristotle’s description of this act of generating hypotheses is called
"intuitive induction" (Cohen and Nagel, 1934; Kneale, 1949). The question mark on the
upward arm of the arch in Figure 1.2 simply indicates that there is still argument about
what this discovery-related arm should be called. By most accounts, the downward arm of
the arch concerns the deduction of new observable phenomena, assuming the truth of a
generated hypothesis (Schum, 2001b).
Over the millennia since Aristotle, many people have tried to give an account of the
process of discovering hypotheses and how this process differs from ones in which existing
hypotheses are justified. Galileo Galilei (1564–1642) thought that we “reason backward”
inductively to imagine causes (hypotheses) from observed events, and we reason deduct-
ively to test the hypotheses. A similar view was held by Isaac Newton (1642–1727), John
Locke (1632–1704), and William Whewell (1794–1866). Charles S. Peirce (1839–1914) was
the first to suggest that new ideas or hypotheses are generated through a different form of
reasoning, which he called abduction and associated with imaginative reasoning (Peirce,
1898; 1901). His views are very similar to those of Sherlock Holmes, the famous fictional
character of Conan Doyle (Schum, 1999).

1.2.2 Peirce and Sherlock Holmes on Abductive Reasoning

Until the time of Peirce, most persons interested in discovery and investigation supposed
that the discovery-related arms of the arch in Figure 1.2 involved some form of inductive
reasoning that proceeds from particulars, in the form of observations, to generalities,
in the form of hypotheses. But inductive reasoning is commonly associated with the
process of justifying or trying to prove existing hypotheses based on evidence. The
question remains: Where did these hypotheses come from? Pondering such matters,
Peirce relied on a figure of reasoning he found in the works of Aristotle. The reasoning
proceeds as follows:

If H were true, then E, F, and G would follow as a matter of course.

But E, F, and G have been observed.

Therefore, we have reason to believe that H might possibly be true.

As an illustration, let us assume that we observe E*: “Smoke in the East building” (E* being
evidence that event E occurred).

13:51:08,
.002
1.2. Abductive Reasoning 7

Based on our prior knowledge of contexts in which things like E: “Smoke in a building”
have occurred, we say: “Whenever something like H: ‘There is fire in a building’ has
occurred, then something like E: ‘Smoke in the building’ has also occurred.” Thus, there is
reason to suspect that H: “There is fire in the East building” may explain the occurrence of
the clue E*: “Smoke in the East building.” In other words, the clue E* points to H as a
possible explanation for its occurrence.
To summarize:

We observe smoke in the East building.

Fire causes smoke.

We hypothesize that there is a fire in the East building.

Peirce was unsure about what to call this form of reasoning. At various points in his work,
he called it “abduction,” “retroduction,” and even just “hypothesis” (Pierce, 1898; 1901).
The essential interpretation Peirce placed on the concept of abduction is illustrated in
Figure 1.3. He often used as a basis for his discussions of abduction the observation of an
anomaly in science. Let us suppose that we already have a collection of prior evidence in
some investigation and an existing collection of hypotheses H1, H2, . . . , Hn. To varying
degrees, these n hypotheses explain the evidence we have so far. But now we make an
observation E* that is embarrassing in the following way: We take E* seriously, but we cannot
explain it by any of the hypotheses we have generated so far. In other words, E* is an
anomaly. Vexed by this anomaly, we try to find an explanation for it. In some cases, often
much later when we are occupied by other things, we experience a “flash of insight” in which
it occurs to us that a new hypothesis Hn+1 could explain this anomaly E*. It is these “flashes of
insight” that Peirce associated with abduction. Asked at this moment to say exactly how Hn+1
explains E*, we may be unable to do so. However, further thought may produce a chain of
reasoning that plausibly connects Hn+1 and E*. The reasoning might go as follows:

I have evidence E* that event E happened.

If E did happen, then F might be true.

If F happened, then G might be true.

And if G happened, then Hn+1 might be true.

It is possible, of course, that the chain of reasoning might have started at the top with Hn+1
and ended at E*. This is why we have shown no direction on the links between E* and Hn+1
in Figure 1.3.

Existing hypotheses New hypothesis New hypothesis

{H1, H2, … , Hn} Hn+1 Hn+1

Reasoning G Later thought

stages not suggests a
immediately F plausible chain
Insight ! obvious of reasoning
E

Prior evidence bearing E*: Evidence not E*

on these hypotheses explainable by existing
hypotheses (an anomaly)
Figure 1.3. Peirce’s interpretation of abductive reasoning.

13:51:08,
.002
8 Chapter 1. Introduction

But our discovery-related activities are hardly over just because we have explained
this anomaly. Our new hypothesis Hn+1 would not be very appealing if it explained
only anomaly E*. Figure 1.4 shows the next steps in our use of this new hypothesis.
We first inquire about the extent to which it explains the prior evidence we collected
before we observed E*. An important test of the suitability of the new hypothesis Hn+1
involves asking how well this new hypothesis explains other observations we have
taken seriously. This new hypothesis would be especially valuable if it explains our
prior evidence better than any of our previously generated hypotheses. But there is
one other most important test of the adequacy of a new hypothesis Hn+1: How well does
this new hypothesis suggest new potentially observable evidence that our previous
hypotheses did not suggest? If Hn+1 would be true, then B, I, and K would also be true;
and if B would be true, then C would be true. Now if C would be true, then we would
need to observe D.
In the illustrations Peirce used, which are shown in Figures 1.3 and 1.4, we entered
the process of discovery at an intermediate point when we already had existing hypoth-
eses and evidence. In other contexts, we must of course consider abductive reasoning
from the beginning of an episode of fact investigation when we have no hypotheses and
no evidence bearing on them. Based on our initial observations, by this process of
abductive or insightful reasoning, we may generate initial guesses or hypotheses to
explain even the very first observations we make. Such hypotheses may of course be
vague, imprecise, or undifferentiated. Further observations and evidence we collect may
allow us to make an initial hypothesis more precise and may of course suggest entirely
new hypotheses.
It happens that at the very same time Peirce was writing about abductive reasoning,
insight, and discovery, across the Atlantic, Arthur Conan Doyle was exercising his
fictional character Sherlock Holmes in many mystery stories. At several points in the
Sherlock Holmes stories, Holmes describes to his colleague, Dr. Watson, his inferential
strategies during investigation. These strategies seem almost identical to the concept of
abductive reasoning described by Peirce. Holmes did not, of course, describe his investi-
gative reasoning as abductive. Instead, he said his reasoning was “backward," moving
from his observations to possible explanations for them. A very informative and enjoy-
able collection of papers on the connection between Peirce and Sherlock Holmes
appears in the work of Umberto Eco and Thomas Sebeok (1983). In spite of the similarity
of Peirce's and Holmes's (Conan Doyle's) views of discovery-related reasoning, there is
no evidence that Peirce and Conan Doyle ever shared ideas on the subject.

Hn+1 Hn+1 Hn+1

G
B I K
F
C J L
E How well does
D Observable
Hn+1 explain
E* prior evidence How well does Hn+1 suggest new
taken seriously kinds of observable evidence?
Figure 1.4. Putting an abduced hypothesis to work.

13:51:08,
.002
 1.3. Probabilistic Reasoning 9

1.3 PROBABILISTIC REASONING

A major trouble we all face in thinking about probability and uncertainty concerns the fact
that the necessity for probability calculations, estimations, or judgments arises in different
situations. In addition, there are many different attributes of our judgments that we would
like to capture in assessments of uncertainty we are obliged to make. There are situations
in which you can estimate probabilities of interest by counting things. But there are many
other situations in which we have uncertainty but will have nothing to count. These
situations involve events that are singular, unique, or one of a kind. In the following, we
will briefly discuss several alternative views of probability, starting with two views of
probability that involve processes in which we can obtain probabilities or estimates of
them by enumerative or counting processes.

1.3.1 Enumerative Probabilities: Obtained by Counting

1.3.1.1 Aleatory Probability
According to Laplace (1814, p. cv), “probability theory is nothing but common sense
reduced to calculation.” There are two conceptions of probability that involve counting
operations. The first is termed aleatory probability. This term has its roots in the Latin term
alea, meaning chance, game of chance, or devices such as dice involved in games of
chance. Games of chance have two important ground rules:

There is a finite number n(S) of possible outcomes

All outcomes in S are assumed to have equal probability

For example, in a game involving a pair of fair six-sided dice, where we roll and add the
two numbers showing up, there are thirty-six ways in which the numbers showing up will
have sums between two and twelve, inclusive. So, in this case, n(S) = 36. Suppose you wish
to determine the probability that you will roll a seven on a single throw of these dice. There
are exactly six ways in which this can happen. If E = “the sum of the numbers is seven,”
then n(E) = 6. The probability of E, P(E), is simply determined by dividing n(E) by n(S),
which in this example is P(E) = 6/36 = 1/6. So, aleatory probabilities are always determined
by dividing n(E) by n(S), whatever E and S are, as long as E is a subset of S.

1.3.1.2 Relative Frequency and Statistics

Another way of assessing probabilities involves the many situations in which aleatory
ground rules will not apply, but empirical methods are at hand to estimate probabilities.
These situations arise when we have replicable or repeatable processes in which we can
count the number of times events have occurred in the past. Suppose that, employing a
defensible method for gathering information about the number of times event E has
occurred, we determine the relative frequency of an occurrence of E by counting the
number of times E has occurred, n(E), and then dividing this number by N, where N is
the number of observations we have made, or the sample size we have taken. In this case,
the relative frequency of E, f(E), equals n(E)/N. You recognize that this is a statistical
process that can be performed in many situations, provided that we assume processes that
are replicable or repeatable. It is true, of course, that a relative frequency f(E) is just an

13:51:08,
.002
10 Chapter 1. Introduction

estimate of the true probability of E, P(E). The reason, of course, is that the number N of
observations we have made is always less than the total number of observations that could
be made. In some cases, there may be an infinite number of possible observations. If you
have had a course in probability theory, you will remember that there are several formal
statements, called the laws of large numbers, for showing how f(E) approaches P(E) when
N is made larger and larger.
Probability theory presents an interesting paradox. It has a very long history but a very
short past. There is abundant evidence that people as far back as Paleolithic times used
objects resembling dice either for gambling or, more likely, to foretell the future (David,
1962). But attempts to calculate probabilities date back only to the 1600s, and the first
attempt to develop a theory of mathematical probability dates back only to 1933 in the
work of A. N. Kolmogorov (1933). Kolmogorov was the first to put probability on an
axiomatic basis. The three basic axioms he proposed are the following ones:

Axiom 1: For any event E, P(E)  0.

Axiom 2: If an event is sure or certain to occur, which we label S, P(S) = 1.0.
Axiom 3: If two events, E and F, cannot occur together, or are mutually exclusive, the
probability that one or the other of these events occurring is the sum of their separate
probabilities. In symbols, P(E or F) = P(E) + P(F).

All Axiom 1 says is that probabilities are never negative. Axioms 1 and 2, taken together,
mean that probabilities are numbers between 0 and 1. An event having 0 probability is
commonly called an “impossible event.” Axiom 3 is called the additivity axiom, and it
holds for any number of mutually exclusive events.
Certain transformations of Kolmogorov’s probabilities are entirely permissible and are
often used. One common form involves odds. The odds of event E occurring to its not
occurring, which we label Odds(E, ¬E), is determined by Odds(E, ¬E) = P(E)/(1 – P(E)). For
any two mutually exclusive events E and F, the odds of E to F, Odds(E, F), are given by Odds
(E, F) = P(E)/P(F). Numerical odds scales range from zero to an unlimited upper value.
What is very interesting, but not always recognized, is that Kolmogorov had only
enumerative probability in mind when he settled on the preceding three axioms. He
makes this clear in his 1933 book and in his later writings (Kolmogorov, 1969). It is easily
shown that both aleatory probabilities and relative frequencies obey these three axioms.
But Kolmogorov went an important step further in defining conditional probabilities that
are necessary to show how the probability of an event may change as we learn new
information. He defined the probability of event E, given or conditional upon some other
event F, as P(E given F) = P(E and F)/P(F), assuming that P(F) is not zero. P(E given F) is
also written as P(E|F). He chose this particular definition since conditional probabilities, so
defined, will also obey the three axioms just mentioned. In other words, we do not need
any new axioms for conditional probabilities.
Now comes a very important concept you may have heard about. It is called Bayes’
rule and results directly from applying the definition of the conditional probability. From
P(E* and H) = P(H and E*), you obtain P(E*|H) P(H) = P(H|E*)P(E*). This can then be
written as shown in Figure 1.5.
This rule is named after the English clergyman, the Reverend Thomas Bayes (1702–1761),
who first saw the essentials of a rule for revising probabilities of hypotheses, based on new
evidence (Dale, 2003). He had written a paper describing his derivation and use of this rule
but he never published it; this paper was found in his desk after he died in 1761 by Richard

13:51:08,
.002
 1.3. Probabilistic Reasoning 11

Probability of given H Prior probability of hypothesis H

(Likelihood) (Prior)

Probability of H given
(Posterior)

Prior probability of e
(Normalizer)
Figure 1.5. The Bayes’ rule.

Price, the executor of Bayes’ will. Price realized the importance of Bayes’ paper and
recommended it for publication in the Transactions of the Royal Society, in which it
appeared in 1763. He rightly viewed Bayes’ rule as the first canon or rule for inductive or
probabilistic reasoning. Bayes’ rule follows directly from Kolmogorov’s three axioms and his
definition of a conditional probability, and is entirely uncontroversial as far as its derivation
is concerned. But this rule has always been a source of controversy on other grounds. The
reason is that it requires us to say how probable a hypothesis is before we have gathered
evidence that will possibly allow us to revise this probability. In short, we need prior
probabilities on hypotheses in order to revise them, when they become posterior
probabilities. Persons wedded to enumerative conceptions of probability say we can never
have prior probabilities of hypotheses, since in advance of data collection we have nothing
to count. Statisticians are still divided today about whether it makes sense to use Bayes’ rule
in statistical inferences. Some statisticians argue that initial prior probabilities could be
assessed only subjectively and that any subjective assessments have no place in any area
that calls itself scientific. Bayes’ rule says that if we are to talk about probability revisions in
our beliefs, based on evidence, we have to say where these beliefs were before we obtained
the evidence.
The Bayes’ rule is useful in practice because there are many cases where we have good
probability estimates for three of the four probabilities involved, and we can therefore
compute the fourth one (see, e.g., Question 1.9).
It is time for us to consider views of probability in situations where we will have nothing
to count, either a priori or anywhere else: the Subjective Bayesian view, Belief Functions,
Baconian probabilities, and Fuzzy probabilities. We provide a look at only the essentials of
these four views, focusing on what each one has to tell us about what the force or weight of
evidence on some hypothesis means. More extensive comparisons of these four views
appear in (Schum, 1994 [2001a], pp. 200–269).

1.3.2 Subjective Bayesian View of Probability

We refer to our first nonenumerative view as an epistemic view, since it assumes that
probabilities in any case are based on some kind of knowledge, whatever form it may take.
In short, probabilities are the result of informed judgments. Many statisticians now favor the
use of Bayes’ rule for combining subjective assessments of all the prior and likelihood
ingredients of Bayes’ rule. But what these persons require is that these assessments be entirely
consistent with Kolmogorov’s three axioms and his definition of conditional probabilities we

13:51:08,
.002
12 Chapter 1. Introduction

previously noted. Since Bayes’ rule rests on these axioms and definition, we must adhere to
them in order to say that our assessment process is coherent or consistent.
As we will show, the likelihoods and their ratios are the ingredients of Bayes’ rule that
concern the inferential force of evidence. Suppose we have two hypotheses, H and ¬H,
and a single item of evidence, E  , saying that event E occurred. What we are interested in
determining are the posterior probabilities: PðHjE  Þ and P ð¬HjE  Þ. Using the Bayes’ rule
from Figure 1.5, we can express these posterior probabilities as:

PðE  jH ÞPðH Þ
PðHjE  Þ ¼
P ðE  Þ

PðE  j¬H ÞPð¬H Þ

P ð¬HjE  Þ ¼
P ðE  Þ

The next step is to divide PðHjE  Þ by Pð¬HjE  Þ, which will produce three ratios; in the
process the term PðE  Þ will drop out. Here are the three ratios that result:

PðHjE  Þ PðH Þ P ðE  jH Þ
¼
Pð¬HjE  Þ Pð¬H Þ P ðE  j¬H Þ
P ðHjE  Þ
The left-hand ratio P ð¬HjE  Þ is called the posterior odds of H to ¬H, given evidence E  . In
symbols, we can express this ratio as OddsðH : ¬HjE  Þ. The first ratio on the right, PPðð¬H
HÞ
Þ, is
called the prior odds of H to ¬H. In symbols, we can express this ratio as OddsðH : ¬H Þ.

The remaining ratio on the right, PPððEE j¬H
jH Þ 
Þ, is called the likelihood ratio for evidence E ; we
give this ratio the symbol LE . In terms of these three ratios, Bayes' rule applied to this
situation can be written simply as follows:

OddsðH : ¬HjE  Þ ¼ OddsðH : ¬H ÞLE

This simple version of Bayes' rule is called the odds-likelihood ratio form. It is also called,
somewhat unkindly, “idiot’s Bayes.” If we divide both sides of this equation by the prior
odds, OddsðH : ¬H Þ, we observe that the likelihood ratio LE  is simply the ratio of posterior
odds to prior odds of H to ¬H. This likelihood ratio shows us how much, and in what
direction (toward H or ¬H), our evidence E  has caused us to change our beliefs from
what they were before we obtained evidence E  . In short, likelihood ratios grade the force
of evidence in Bayesian analyses.
Here is an example of how likelihoods and their ratios provide a method for grading the
force of an item of evidence on some hypothesis. This is an example of a situation
involving a singular evidence item where we have nothing to count. Suppose we are
interested in determining whether or not the Green state is supplying parts necessary for
the construction of shaped explosive devices to a certain insurgent militia group in the
neighboring Orange state. Thus we are considering two hypotheses:

H: “The Greens are supplying parts necessary for the construction of shaped explosive
devices.”

¬H: “The Greens are not supplying parts necessary for the construction of shaped
explosive devices.”

Suppose we believe, before we have any evidence, that the prior probability of H is
PðH Þ ¼ 0:20. Because we must obey the rules for enumerative probabilities, we must also

13:51:08,
.002
 1.3. Probabilistic Reasoning 13

say that Pð¬H Þ ¼ 0:80. This follows from the third axiom we discussed in Section 1.3.1. So,
our prior odds on H relative to ¬H have a value OddsðH : ¬H Þ ¼ PPðð¬H
HÞ 0:20 1
Þ ¼ 0:80 ¼ 4.
Suppose now that we receive the item of evidence E  : A member of the Green’s
military was captured less than one kilometer away from a location in Orange at which
parts necessary for the construction of these shaped explosives were found.
We ask ourselves how likely is this evidence E  if H were true, and how likely is this
evidence E  if H were not true. Suppose we say that PðE  jH Þ ¼ 0:80 and P ðE  j¬H Þ ¼ 0:10.
We are saying that this evidence is eight times more likely if H were true than if H were not

true. So, our likelihood ratio for evidence E  is LE ¼ PPððEE j¬H
jH Þ 0:8
Þ ¼ 0:1 ¼ 8.
We now have all the ingredients necessary in Bayes' rule to determine the posterior
odds and posterior probability of hypothesis H:

1
OddsðH : ¬HjE  Þ ¼ OddsðH : ¬H ÞLE ¼  8 ¼ 2:
4

This means that we now believe the posterior odds favoring H over ¬H are two to one. But
we started by believing that the prior odds of H to ¬H were one in four, so the evidence E 
changed our belief by a factor of 8.
We could just as easily express this inference in terms of probabilities. Our prior
probability of H was PðH Þ ¼ 0:20. But our posterior probability PðHjE  Þ ¼ 1þ2
2
¼ 23 ¼ 0:67.

So, in terms of probabilities, evidence E caused us to increase the probability of H by 0.47.
So, using this subjective Bayesian approach, we would be entitled to express the extent
of our uncertainty in an analysis using numerical probability assessments provided only
that they conform to Kolmogorov’s axioms.

1.3.3 Belief Functions

Both the enumerative and the subjective Bayesian interpretations of probability conform
to Kolmogorov’s three axioms. We asserted that these axioms rest on the investigation of
replicable or repeatable processes such as statistical analysis of the results obtained in a
sample of observations. But there are many reasons for wondering whether these three
axioms remain self-evident concerning subjective probability judgments we all make from
time to time involving unique events for which no enumerative process can be involved. In
a very influential work, the probabilist Professor Glenn Shafer pointed to an array of
difficulties associated with Axiom 3 concerning the additivity of enumerative probabilities
for mutually exclusive events (Shafer, 1976). In particular, Shafer asserts that this axiom
places various constraints on our judgments or beliefs about uncertainty that we may
not be willing to accept. Here it is only necessary to mention two of the difficulties
Shafer mentions:

Indecisions we routinely face concerning ambiguities in our evidence

Instances in which we encounter what historically has been called “pure” evidence

In so many instances, we may not be sure what evidence is telling us, and so we wish to be
able to withhold a portion of our beliefs and not commit it to any particular hypothesis or
possible conclusion. A very important element in what Shafer terms Belief Functions is that
the weight of evidence means the degree of support evidence provides to hypotheses we are
considering. Shafer allows that we can grade the degree of support s on a 0 to 1 scale

13:51:08,
.002
14 Chapter 1. Introduction

similar to the scale for Kolmogorov probabilities; but we can do things with support
assignment s that the Kolmogorov additivity Axiom 3 does not allow.
To illustrate, suppose we revisit the issue discussed in the previous section about
whether or not the Green state is supplying parts necessary for the construction of shaped
explosive devices to a certain insurgent militia group in the neighboring Orange state. At
some stage, we are required to state our beliefs about the extent to which the evidence
supports H or ¬H. Here is our assessment:

{H} {¬H} {H or ¬H}

s 0.5 0.3 0.2

What does this support assignment mean? We are saying that we believe the evidence
supports H exactly to degree s = 0.5, and that this evidence also supports ¬H exactly to
degree s = 0.3. But there is something about this evidence that makes us unsure about
whether it supports H or ¬H. So, we have left the balance of our s assignment, s = 0.2,
uncommitted among H or ¬H. In other words, we have withheld a portion of our beliefs
because we are not sure what some element of our evidence is telling us.
If we were required to obey Kolmogorov Axiom 3, we would not be allowed to be
indecisive in any way in stating our beliefs. Here is what our support assignment would
have to look like:

{H} {¬H}

s a 1a

In this case, we would be required to say that the evidence supports H to degree s = a,
and supports ¬H to degree s = (1 – a) in agreement with Axiom 3 since H and ¬H are
mutually exclusive and exhaustive. In short, Kolmogorov Axiom 3 does not permit us any
indecision in stating our beliefs; we must commit all of it to H and to ¬H. This, we believe,
would not be a faithful or accurate account of our beliefs.
But Shafer’s Belief Function approach allows us to cope with another difficulty
associated with Kolmogorov’s axioms. For centuries, it has been recognized that a
distinction is necessary between what has been termed mixed evidence and pure evi-
dence. Mixed evidence has some degree of probability under every hypothesis we are
considering. But pure evidence may support one hypothesis but say nothing at all about
other hypotheses. In other words, we may encounter evidence that we believe offers zero
support for some hypothesis. Here is another example involving our Green-Orange
situation. Suppose we encounter an item of evidence we believe supports H to a degree,
but we believe offers no support at all for ¬H. Here is our support assignment s for this
evidence:

{H} {¬H} {H or ¬H}

s 0.5 0 0.5

In this situation, we are saying that the evidence supports H to degree s = 0.5, but offers
no support at all to ¬H. The rest of our support we leave uncommitted between H and ¬H.
But now we have to examine what s = 0 for ¬H means; does it mean that ¬H could not be
supported by further evidence? The answer is no, and the reason why it is no allows us to

13:51:08,
.002
1.3. Probabilistic Reasoning 15

compare what ordinary probabilities mean in comparison with what support s means.
This comparison is shown in Figure 1.6.
The (a) scale in Figure 1.6, for conventional or Kolmogorov probabilities, has a lower
boundary with a meaning quite different from the meaning of this lower boundary on
Shafer’s support scale shown in (b). The value 0 in conventional probability refers to an
event judged to be impossible and one you completely disbelieve. But all 0 means on
Shafer’s s scale is lack of belief, not disbelief. This is very important, since we can go from
lack of belief to some belief as we gather more evidence. But we cannot go from disbelief
to some belief. On a conventional probability scale, a hypothesis once assigned the
probability value 0 can never be resuscitated by further evidence, regardless of how strong
it may be. But some hypothesis, assigned the value s = 0, can be revised upward since we
can go from lack of belief to some belief in this hypothesis when and if we have some
further evidence to support it. Thus, s allows us to account for pure evidence in ways that
ordinary probabilities cannot do. We will refer to this scale again in Section 1.3.4 when
discussing Baconian probability.
Consider the evidence in the dirty bomb example which will be discussed in Section 2.2.
We begin by listing the hypotheses we are considering at this moment:

H 1: A dirty bomb will be set off somewhere in the Washington, D.C., area.
¬H 1 : A dirty bomb will not be set off in the Washington, D.C., area (it might be set off
somewhere else or not at all).

In the Belief Functions approach, we have just specified what is called a frame of
discernment, in shorthand a frame F. What this frame F ¼ fH 1 ; ¬H 1 g shows is how we
are viewing our hypotheses right now. We might, on further evidence, wish to revise our
frame in any one of a variety of ways. For example, we might have evidence suggesting
other specific places where a dirty bomb might be set off, such as in Annapolis, Maryland,
or in Tysons Corner, Virginia. So our frame F in this case might be:

H 1: A dirty bomb will be set off in Washington, D.C.

H 2: A dirty bomb will be set off in Annapolis, Maryland.
H 3: A dirty bomb will be set off in Tysons Corner, Virginia.

All that is required in the Belief Functions approach is that the hypotheses in a frame be
mutually exclusive; they might or might not be exhaustive. The hypotheses are required to
0
be exhaustive in the Bayesian approach. So this revised frame F ¼ fH 1 ; H 2 ; H 3 g, as stated,
is not exhaustive. But we are assuming, for the moment at least, that these three

0 Conventional Probability 1
(a)
Disbelief or Impossible Certain or Complete Belief

0 Degree of Support or Belief 1

(b)
Lack of Support or Belief Total Support or Belief

0 Baconian Probability
(c)
Lack of Proof Proof
Figure 1.6. Different probability scales.

13:51:08,
.002
 16 Chapter 1. Introduction

hypotheses are mutually exclusive: The dirty bomb will be set off at exactly one of these
three locations. But, on further evidence, we might come to believe that dirty bombs will
be set off in both Washington, D.C., and in Tysons Corner, Virginia. We know that the
terrorists we are facing have a preference for simultaneous and coordinated attacks. So,
00
our revised frame F might be:

H 1: A dirty bomb will be set off in Washington, D.C., and in Tysons Corner, Virginia.
H 2: A dirty bomb will be set off in Annapolis, Maryland.

The point of all this so far is that the Belief Functions approach allows for the fact that our
hypotheses may mutate or change as a result of new evidence we obtain. This is a major
virtue of this approach to evidential reasoning.
The next thing we have to consider is the power set of the hypotheses in a frame. This
power set is simply the list of all possible combinations of the hypotheses in this frame.
When we have n hypotheses in F, there are 2n possible combinations of our hypotheses,
including all of them and none of them. For example, when F ¼{H 1 , ¬H 1 }, the power set
consists of {H 1 }, {¬H 1 }, {H 1 , ¬H 1 }, and ∅, where ∅ = the empty set (i.e., none of the
0
hypotheses). For F ¼ fH 1 ; H 2 ; H 3 g, as just defined, there are 23 = 8 possible combin-
ations: {H 1 }, {H 2 }, {H 3 }, {H 1 , H 2 }, {H 1 , H 3 }, {H 2 , H 3 ), {H 1 , H 2 , H 3 } and ∅. Now, here comes
an important point about support function s: The assigned values of s for any item or body
of evidence must sum to 1.0 across the power set of hypotheses in a frame. The only
restriction is that we must set s{∅} = 0. We cannot give any support to the set of none of
the hypotheses we are considering.
More details about the Belief Functions approach are provided in Schum (1994 [2001a],
pp. 222–243).

1.3.4 Baconian Probability

1.3.4.1 Variative and Eliminative Inferences
Here is a view of probabilistic reasoning that puts particular emphasis on a very important
matter not specifically addressed in any other view of probability. In this view, the
probability of a hypothesis depends on how much relevant and believable evidence we
have and on how complete is our coverage of existing evidence on matters we ourselves
have recognized as being relevant in the analysis at hand.
This Baconian view of probability rests on the work of Professor L. Jonathan Cohen
(1977, 1989). The label “Baconian” on this system of probability acknowledges the work
of Sir Francis Bacon (1561–1626), who revolutionized the process of inference in science.
Bacon argued that attempting to prove some hypothesis by gathering instances favorable
to it is mistaken, since all it would take to refute the generality of this hypothesis was
one unfavorable instance of it. What Bacon argued was that we ought to design research
with the objective of eliminating hypotheses. The hypothesis that best resists our
eliminative efforts is the one in which we should have the greatest confidence. As this
eliminative process proceeds, it is obvious that we should not keep performing the same
test over and over again. What we need is an array of different tests of our hypotheses.
The hypothesis that holds up under the most varied set of tests is the one having
the greatest probability of being correct. So, Baconian inferences are eliminative and
variative in nature.

13:51:08,
.002
1.3. Probabilistic Reasoning 17

Baconian probabilities have only ordinal properties and cannot be combined algebra-
ically in any way. The Baconian probability scale is shown as (c) in Figure 1.6, to be
compared with the conventional probability scale shown as (a) in Figure 1.6. On the
conventional probability scale, 0 means disproof; but on the Baconian scale, 0 simply
means lack of proof. A hypothesis now having zero Baconian probability can be revised
upward in probability as soon as we have some evidence for it. As noted, we cannot revise
upward in probability any hypothesis disproved, or having zero conventional probability.

1.3.4.2 Importance of Evidential Completeness

Figure 1.7 illustrates a major point of interest in the Baconian system. Professor Cohen
(1977, 1989) argued that in any evidential reasoning situation, we are always out on an
inferential limb that might be longer and weaker than we may believe it to be. Suppose
you have generated three hypotheses {H1, H2, and H3}. You have examined a body of
evidence and have used Bayes’ rule to combine the likelihoods for this evidence together
with stated prior probabilities. The result is that Bayes’ rule shows the posterior probability
of H3, in light of the evidence, to be 0.998, very close to certainty on the Kolmogorov
probability scale. Therefore, you confidently report your conclusion that H3 is true,
together with its very large posterior probability you have determined. A short time passes,
and you hear the depressing news that H3 is not true. What could have gone wrong? After
all, you performed an analysis that is highly respected by many persons.
A person having knowledge of Cohen’s Baconian probability (1977, 1989) arrives on the
scene of your distress and makes the following comments:

You gathered some evidence, fair enough, quite a bit of it, in fact. But, how
many relevant questions you can think of were not answered by the evidence
you had? Depending upon the number of these unanswered questions, you
were out on an inferential limb that was longer and weaker than you
imagined it to be (see Figure 1.7). If you believed that these unanswered
questions would supply evidence that also favored H3, you were misleading

Your conclusion
that H3 is true

Questions unanswered
by your existing evidence

P(H3 on evidence) = 0.998

(e.g., from Bayes’ rule)

Your existing
evidence

Figure 1.7. A Baconian inferential limb.

13:51:08,
.002
18 Chapter 1. Introduction

yourself since you did not obtain any answers to them. The posterior probabil-
ity you determined by itself is not a good indicator of the weight of evidence.
What makes better sense is to say the weight of evidence depends on the
amount of favorable evidence you have and how completely it covers matters
you said were relevant. In your analysis, you completely overlooked the infer-
ential importance of questions your existing evidence did not answer.

Apart from the Baconian system, no other probability view focuses on evidential com-
pleteness and the importance of taking into account questions recognized as being
relevant that remain unanswered by the evidence we do have. This is why Jonathan
Cohen’s Baconian system is so important (Cohen, 1977; 1989). What we do not take into
account in our analyses can hurt us very badly.
In many instances, such as reasoning in intelligence analysis, we frequently have to
make inferences about matters for which we have scant evidence, or no evidence at all. In
other instances in which there may be available evidence, we may have no time to search
for it or consider it carefully. In such cases, we are forced to make assumptions or
generalizations that license inferential steps. But this amounts to giving an assumption
or a generalization the benefit of the doubt (without supporting it in any way), to believing
as if some conclusion were true (absent any evidential support for it), or to taking
something for granted without testing it in any way. All of these situations involve the
suppression of uncertainties.
It happens that only the Baconian probability system provides any guidance about how
to proceed when we must give benefit of doubt, believe as if, or take things for granted.
The major reason is that it acknowledges what almost every logician says about the
necessity for asserting generalizations and supplying tests of them in evidential reasoning.
Search the Bayesian or Belief Functions literature, and you will find almost no discussion
of generalizations (assumptions) and ancillary tests of them. Suppose we are interested in
inferring F from E, that is, P(F|E). Bayes’ rule grinds to a halt when we have no basis for
assessing the likelihoods P(E|F) and P(E|¬F). Bayesians counter by saying that we will
always have some evidence on which to base these judgments. But they never say what
this evidence is in particular cases and how credible it might or might not be. The Belief
Functions approach comes closer by saying that we can assess the evidential support for a
body of evidence that may include both directly relevant and at least some ancillary
evidence (i.e., evidence about other evidence). Following is an account of the Baconian
license for giving an assumption or generalization benefit of doubt, believing as if it were
true, or taking it for granted, provided that we are willing to mention all of the uncertain-
ties we are suppressing when we do so. Stated another way, we must try to account for all
of the questions we can think of that remain unanswered by the absence, or very scant
amount, of evidence.
Here are the essentials of Cohen’s Baconian approach to reasoning based on little or no
ancillary evidence to either support or undermine a generalization (Cohen 1977; 1989).
The first step, of course, is to make sure the generalization is not a non sequitur, that is,
that it makes logical sense. In the simplest possible case, suppose we are interested in
inferring proposition or event F from proposition or event E. The generalization G in doing
so might read, “If E has occurred, then probably F has occurred.” We recognize this if-then
statement as an inductive generalization since it is hedged. Second, we consider various
tests of this generalization using relevant ancillary evidence. Third, we consider how many
evidential tests of this generalization there might be. Suppose we identify N such tests. The

13:51:08,
.002
1.3. Probabilistic Reasoning 19

best case would be when we perform all N of these tests and they all produce results
favorable to generalization G. But we must not overlook generalization G itself; we do so
by assigning it the value 1; so we have N + 1 things to consider. Now we are in a position to
show what happens in any possible case.
First, suppose we perform none of these N evidential tests. We could still proceed by
giving generalization G the benefit of the doubt and detach a belief that F occurred (or will
occur) just by invoking this generalization G regarding the linkage between events
E and F. So, when no evidential tests are performed, we are saying: “Let’s believe as if
F occurred based on E and generalization G.” This would amount to saying that the
Baconian probability of event F is B(F) = 1/(N + 1). This expression is never a ratio; all it
says is that we considered just one thing in our inference about F from E, namely just the
generalization G. We could also say, “Let’s take event F for granted and believe that
F occurred (or will occur) because E occurred, as our generalization G asserts.” However,
note that in doing so, we have left all N ancillary evidential questions unanswered.
This we represent by saying that our inference of F from E has involved only one of the
N + 1 considerations and so we have (N + 1 – 1) = N, the number of questions we have left
unanswered. As far as evidential completeness is concerned, this is when the evidence we
have is totally incomplete. But the Baconian system allows us to proceed anyway based on
giving a generalization the benefit of doubt. But our confidence in this result should
be very low.
Now suppose we have performed some number k of the N possible ancillary evidential
tests of generalization G, as asserted previously, and they were all passed. The Baconian
probability of F in this situation is given by B(F) = (k + 1)/(N +1). The difference between
the numerator and denominator in such an expression will always equal the number
of unanswered questions as far as the testing of G is concerned. In this case, we have
(N + 1) – (k + 1) = N – k questions that were unanswered in a test of generalization G. How
high our confidence is that F is true depends on how high k + 1 is as compared to N + 1.
But now suppose that not all answers to these k questions are favorable to generaliza-
tion G. Under what conditions are we entitled to detach a belief that event F occurred,
based on evidence E, generalization G, and the k tests of G? The answer requires a
subjective judgment by the analyst about whether the tests, on balance, favor or disfavor
G. When the number of the k tests disfavoring G exceeds the number of tests favoring G,
we might suppose that we would always detach a belief that event F did not occur, since
G has failed more tests than it survived. But this will not always be such an easy judgment
if the number of tests G passed were judged to be more important than the tests it failed to
pass. In any case, there are N – k tests that remain unanswered. Suppose that k is quite
large, but the number of tests favorable to G is only slightly larger than the number of tests
unfavorable to G. In such cases, the analyst might still give event F the benefit of the doubt,
or believe, at least tentatively, as if F occurred pending the possible acquisition of further
favorable tests of G. And in this case, the confidence of the analyst in this conclusion
should also be very low.
Whatever the basis for an assumption or a benefit of doubt judgment there is, one of
the most important things about the Baconian approach is that the analyst must be
prepared to give an account of the questions that remain unanswered in evidential tests
of possible conclusions. This will be especially important when analysts make assump-
tions, or more appropriately, give generalizations the benefit of doubt, draw as if conclu-
sions, or take certain events for granted. These are situations in which analysts are most
vulnerable and in which Baconian ideas are most helpful.

13:51:08,
.002
 20 Chapter 1. Introduction

1.3.4.3 Baconian Probability of Boolean Expressions

Some of the most important properties of Baconian probabilities concern their application
to Boolean combinations of propositions, such as hypotheses. Because the probabilities in
the Baconian system have only ordinal properties, we can say only that hypothesis H1 is
more likely than H2, but we cannot say how much more likely H1 is than H2. Also, in the
Baconian system, it is never necessary to assess subjective probabilities. In our saying that
H1 is more probable than H2, all we are saying is that there is more favorably relevant
evidence on H1 than there is on H2. What counts most in the Baconian system is the
completeness of our evidence and the extent to which we have questions that remain
unanswered by the evidence we have. Here are the three most important Baconian
properties of interest to us concerning intersections, unions, and negation.
Baconian Intersection: Suppose we have some events of interest such as events F, G,
and H. Suppose we have some favorably relevant evidence about each one of these events
and have also considered how complete the evidence is for these events. So we determine
that the Baconian probabilities (B) for these three events are B(F)  B(G)  B(H). Here’s
what these probabilities say: We have more favorably relevant and complete evidence for
event F than we do for event G, and more favorably relevant and complete evidence for
event G than we have for event H. So, asked what the Baconian probability is for their
intersection (F and G and H), we must say that B(F and G and H) = B(H). What this says is
that the Baconian probability of the intersection of these three events is equal to the
Baconian probability of the event with the least favorably relevant and complete evidence.
This is an example of the MIN rule for Baconian intersections. We might compare this with
the conventional probability of the intersection of these three events. Suppose that events
F, G, and H are independent events where P(F) = 0.8, P(G) = 0.6 and P(H) = 0.4. In this
case, P(F and G and H) = 0.8*0.6*0.4 = 0.192 < P(H) = 0.4. In the Baconian system, the
probability of a conjunction of events or propositions can never be less than that of the
event having the smallest Baconian probability.
Baconian Union: Now consider the same case involving events F, G, and H. Again,
suppose that B(F)  B(G)  B(H). Now what we wish to determine is the Baconian
probability B(F or G or H). In this case, B(F or G or H)  B(F), where B(F) is the largest
of the Baconian probability for the events we are considering. This is the MAX rule
for Baconian probability, and what it says is that the probability of a disjunction of events
is at least as large as the largest Baconian probability of any of the individual events.
Baconian Negation: Baconian negation is not complementary. The Baconian rule is
quite complex; here’s what it says: If we have A and ¬A, if B(A) > 0, then B(¬A) = 0. What
this means essentially is that we cannot commit beliefs simultaneously to two events that
cannot both occur.
What is quite interesting is that the Baconian treatment of conjunctions and disjunc-
tions is the same as in Zadeh’s Fuzzy probability system; namely, they both make use of
min-max rules for these connectives.
More details about Baconian probabilities are provided in Schum (1994 [2001a],
pp. 243–261).

1.3.5 Fuzzy Probability

1.3.5.1 Fuzzy Force of Evidence
One can also express the uncertainty about a conclusion reached by using words, such as
“likely,” “almost certain,” or “much less certain,” rather than numbers, as illustrated by the

13:51:08,
.002
1.3. Probabilistic Reasoning 21

following fragment from the letter sent by Albert Einstein to the United States President
Franklin D. Roosevelt, on the possibility of constructing nuclear bombs (Einstein, 1939):
. . . In the course of the last four months it has been made probable – through
the work of Joliot in France as well as Fermi and Szilárd in America – that it
may become possible to set up a nuclear chain reaction in a large mass of
uranium, by which vast amounts of power and large quantities of new
radium-like elements would be generated. Now it appears almost certain
that this could be achieved in the immediate future.
This new phenomenon would also lead to the construction of bombs, and it
is conceivable – though much less certain – that extremely powerful bombs of
a new type may thus be constructed. . . .
Verbal expressions of uncertainty are common in many areas. In the field of law, for
example, forensic standards of proof are always employed using words instead of
numbers. We all know about standards such as “beyond reasonable doubt” (in criminal
cases); “preponderance of evidence” (in civil cases); “clear and convincing evidence” (in
many Senate and congressional hearings); and “probable cause” (employed by magis-
trates to determine whether a person should be held in custody pending further hearings).
All the verbal examples just cited have a current name: They can be called Fuzzy
probabilities. Words are less precise than numbers. There is now extensive study of fuzzy
inference involving what has been termed approximate reasoning, which involves verbal
statements about things that are imprecisely stated. Here is an example of approximate
reasoning: “Since John believes he is overworked and underpaid, then he is probably not
very satisfied with his job.” We are indebted to Professor Lofti Zadeh (University of
California, Berkeley), and his many colleagues, for developing logics for dealing with fuzzy
statements, including Fuzzy probabilities (Zadeh, 1983; Negoita and Ralescu, 1975). In his
methods for relating verbal assessments of uncertainty with numerical equivalents, Zadeh
employed what he termed a possibility function, μ, to indicate ranges of numerical
probabilities a person might associate with a verbal expression of uncertainty. Zadeh
reasoned that a person might not be able to identify a single precise number he or she
would always associate with a verbal statement or Fuzzy probability. Here is an example of
a possibility function for the Fuzzy probability “very probable.”
Asked to grade what numerical probabilities might be associated with an analyst’s
Fuzzy probability of “very probable,” the analyst might respond as follows:
For me, “very probable” means a numerical probability of at least 0.75 and at
most 0.95. If it were any value above 0.95, I might use a stronger term, such
as “very, very probable.” I would further say that I would not use the term
“very probable” if I thought the probability was less than 0.75. In such cases,
I would weaken my verbal assessment. Finally, I think it is most possible (μ =
1.0) that my use of the verbal assessment “very probable” means something
that has about 0.85 of occurring. If the analyst decides that “very probable”
declines linearly on either side of μ = 1.0, we would have the possibility
function shown in Figure 1.8.

1.3.5.2 Fuzzy Probability of Boolean Expressions

As an example of using fuzzy probabilities, suppose we have three events, or propositions
A, B, and C. We consider the following Fuzzy probabilities (F) for these events, and we say
the following:

13:51:08,
.002
22 Chapter 1. Introduction

1.0

Very

Possibility
probable
0.5

0 0.25 0.50 0.75 1.0

Probability
Figure 1.8. Possibilities and Fuzzy probabilities.

Event A is very likely.

Event B is likely.

Event C is very unlikely.

We express this by saying that F(A) > F(B) > F(C).

Fuzzy Conjunction: The Fuzzy conjunction of several events is the minimum Fuzzy
probability of these events. For example, F(A and B and C) = F(C), which is the minimum
Fuzzy probability of these three events.
Fuzzy Disjunction: The Fuzzy disjunction of several events is the maximum Fuzzy
probability of these events. For example, F(A or B or C) = F(A), which is the maximum
Fuzzy probability of these three events.
Notice that both in the Baconian system and in the Fuzzy system we have MIN/MAX
rules for combining probabilities for complex events.
Fuzzy Negation: Fuzzy negation is complementary: F(A) = 1 – F(¬A).
More details about the Fuzzy probabilities are provided in Schum (1994 [2001a],
pp. 261–269).

1.3.5.3 On Verbal Assessments of Probabilities

Let us also consider the critics who sneer at verbal assessments of probabilities, saying that
only numerical assessments, conforming to the Kolmogorov axioms, are acceptable. As a
top-ranking analyst, you are asked by an equally high-ranking customer for the probability
of a crucial hypothesis HK. All the evidence in this case is for a one-of-a-kind event, so your
assessment is necessarily subjective. You tell the customer, “Sir, the probability of HK, on
our analysis is 78 percent.” The customer asks, “This is a very precise number. How did
you arrive at it, given the subjective nature of your assessment?” You reply, “Yes, sir, what
I really should have said was that my probability is between 73 percent and 83 percent,
and 78 percent seemed like a good figure to quote.” The customer then says, “But the
limits to the probability interval you quoted are also precise. How did you arrive at them?”
You might say, “Well, my lower limit is really between 70 percent and 76 percent and my
upper limit is between 80 percent and 86 percent.” Your customer says, “But these are also
precise numbers.” There is, as you see, an infinite regress of similar questions regarding
the basis for subjective numerical assessments.
There are many places to begin a defense of verbal or Fuzzy probability statements. The
most obvious one is law. All of the forensic standards of proof are given verbally: “beyond

13:51:08,
.002
 1.3. Probabilistic Reasoning 23

reasonable doubt,” “clear and convincing evidence,” “balance of probabilities,” “sufficient

evidence,” and “probable cause.” Over the centuries, attempts have been made to supply
numerical probability values and ranges for each of these standards, but none of them has
been successful. The reason, of course, is that every case is unique and rests upon many
subjective and imprecise judgments. Wigmore (1913, 1937) understood completely that
the catenated inferences in his Wigmorean networks were probabilistic in nature. Each of
the arrows in a chain of reasoning describes the force of one hypothesis on the next one,
such as E ➔ F. Wigmore graded the force of such linkages verbally, using such terms as
“strong force,” “weak force,” “provisional force,” and so on. Toulmin (1963) also used
fuzzy qualifiers in the probability statements of his system, which grounds the Rationale
analytical tool (van Gelder, 2007). There are many other examples of situations in which it
is difficult or impossible for people to find numerical equivalents for verbal probabilities
they assess. Intelligence analysis so often supplies very good examples in spite of what
Sherman Kent (1994) said some years ago. Indeed, using words is quite often necessary in
analyses based on masses of evidence that are so complex that they resist even the most
devoted attention to the construction of inference networks. Couple this with the fact that
different analysts might disagree substantially about what specific probability should be
assigned to a conclusion. In addition an analyst might assign a different probability to the
same conclusion, based on the same evidence, on different occasions. What this says is
that there will be inter-analyst and intra-analyst variation in the assessment of probabil-
ities. Words are less precise than numbers, so there will often be less disagreement about a
verbal or a Fuzzy probability.
We conclude this discussion by recalling what the well-known probabilist Professor
Glenn Shafer said years ago (Shafer, 1988):

Probability is more about structuring arguments than it is about numbers. All

probabilities rest upon arguments. If the arguments are faulty, the probabil-
ities, however determined, will make no sense.

1.3.6 A Summary of Uncertainty Methods and What They

Best Capture
Evidence-based reasoning is necessarily probabilistic in nature because our evidence
is always incomplete (we can look for more, if we have time), usually inconclusive (it
is consistent with the truth of more than one hypothesis or possible explanation),
frequently ambiguous (we cannot always determine exactly what the evidence is telling
us), commonly dissonant (some of it favors one hypothesis or possible explanation, but
other evidence favors other hypotheses), and with various degrees of believability shy
of perfection.
As illustrated in Table 1.1 and discussed in this section, each of the alternative views of
probability previously discussed captures best some of these characteristics of evidence,
but no single view captures best all of them. We include in this table just the four views
concerning nonenumerative situations. One can easily find many works on statistics,
frequentistic or Bayesian, in enumerative situations in which they can estimate probabil-
ities from observed relative frequencies.
The first entry in Table 1.1 lists a major strength that is exclusive to the Baconian
system: its concern about how much favorable evidence was taken into account in an

13:51:08,
.002
24 Chapter 1. Introduction

Table 1.1. A Summary of Nonenumerative Uncertainty Methods and What They Best
Capture

Subjective Belief
Major Strength Bayes Functions Baconian Fuzzy

Accounting for incompleteness of coverage

☑
of evidence

Coping with inconclusiveness in evidence ☑ ☑ ☑ ☑

Coping with ambiguities or imprecision

☑ ☑
in evidence and judgmental indecision

Coping with dissonance in evidence ☑ ☑ ☑ ☑

Coping with source believability issues ☑ ☑

analysis, and how completely this evidence covered matters judged relevant to conclusions
that could be reached. A major question this form of analysis allows us to address is the
extent to which questions that have not been answered by existing evidence could have
altered the conclusion being reached. It would be quite inappropriate to assume that
answers to the remaining unanswered questions would, if they were obtained, all favor the
conclusion that was being considered. This, of course, requires us to consider carefully
matters relevant to any conclusion that are not addressed by available evidence.
The second entry in Table 1.1 notes that all four of the probabilistic methods have
very good ways for dealing with the inconclusive nature of most evidence, but they do
so in different ways. The Subjective Bayesian does so by assessing nonzero likelihoods
for the evidence under every hypothesis being considered. Their relative sizes indicate
the force that the evidence is judged to have on each hypothesis. But the Belief
Functions advocate assigns numbers indicating the support evidence provides for
hypotheses or subsets of them. We should be quick to notice that Bayesian likelihoods
do not grade evidential support, since in Belief Functions one can say that an item of
evidence provides no support at all to some hypothesis. But a Bayesian likelihood of
zero under a particular hypothesis would mean that this hypothesis is impossible and
should be eliminated. Offering no support in Belief Functions does not entail that this
hypothesis is impossible, since some support for this hypothesis may be provided by
further evidence. The Baconian acknowledges the inconclusive nature of evidence by
assessing how completely, as well as how strongly, the evidence favors one hypothesis
over others. In Fuzzy probabilities, it would be quite appropriate to use words in
judging how an item or body of evidence bears on several hypotheses. For example,
one might say, “This evidence is indeed consistent with H1 and H2, but I believe it
strongly favors H1 over H2.”
The third entry in the table first acknowledges the Belief Functions and Fuzzy
concerns about ambiguities and imprecision in evidence. In the Belief Functions
approach, one is entitled to withhold belief for some hypotheses in the face of ambiguous
evidence. In such cases, one may not be able to decide upon the extent to which the
evidence may support any hypothesis being considered, or even whether the evidence

13:51:08,
.002
 1.4. Evidence-based Reasoning 25

supports any of them. Judgmental indecision is not allowed in the Bayesian system since
it assumes one can say precisely how strongly evidence judged relevant favors every
hypothesis being considered. Ambiguities in evidence may be commonly encountered.
The Fuzzy advocate will argue that ambiguities or imprecision in evidence hardly
justifies precise numerical judgments. In the face of fuzzy evidence, we can make only
fuzzy judgments of uncertainty.
The fourth entry in Table 1.1 shows that all four probability systems have very good
mechanisms for coping with dissonant evidence in which there are patterns of contradict-
ory and divergent evidence. Dissonant evidence is directionally inconsistent; some of it
will favor certain hypotheses and some of it will favor others. In resolving such inconsist-
encies, both the Bayesian and Belief Functions approaches will side with the evidence
having the strongest believability. The Bayesian approach to resolving contradictions is
especially interesting since it shows how “counting heads” is not the appropriate method
for resolving contradictions. In times past, “majority rule” was the governing principle.
Bayes’ rule shows that what matters is the aggregate believability on either side of a
contradiction. The Baconian approach also rests on the strength and aggregate believabil-
ity in matters of dissonance, but it also rests on how much evidence is available on either
side and upon the questions that remain unanswered. In Fuzzy terms, evidential disson-
ance, and how it might be resolved, can be indicated in verbal assessments of uncer-
tainty. In such instances, one might say, “We have dissonant evidence favoring both H1
and H2, but I believe the evidence favoring H1 predominates because of its very strong
believability.”
Row five in Table 1.1 concerns the vital matter of assessing the believability of
evidence. From considerable experience, we find that the Bayesian and Baconian
systems are especially important when they are combined. In many cases, these two
radically different schemes for assessing uncertainty are not at all antagonistic but are
entirely complementary. Let us consider a body of evidence about a human intelli-
gence (HUMINT) asset or informant. Ideas from the Baconian system allow us to ask,
“How much evidence do we have about this asset, and how many questions about
this asset remain unanswered?” Ideas from the Bayesian system allow us to ask, “How
strong is the evidence we do have about this asset?” (Schum, 1991; Schum and
Morris, 2007)

1.4 EVIDENCE-BASED REASONING

1.4.1 Deduction, Induction, and Abduction

These three types of inference involved in evidence-based reasoning may be intuitively
summarized as shown in the following.
Deduction shows that something is necessarily true:

A ➔ necessarily B
Socrates is a man ➔ necessarily Socrates is mortal

Induction shows that something is probably true:

A ➔ probably B
Julia was born in Switzerland ➔ probably Julia speaks German

13:51:08,
.002
 26 Chapter 1. Introduction

Abduction shows that something is possibly or plausibly true:

A ➔ possibly B
There is smoke in the East building ➔ possibly there is fire
in the East building

These types of inference are described more formally in the following.

Deductive Inference:

8x, U(x) ➔ V(x) Whenever U(x) is true, V(x) is also true

U(a1) U(a1) is true

Necessarily V(a1) Therefore V(a1) is necessarily true

Inductive Inference:

U(a1) and V(a1) When U(a1) was true, it was observed that V(a1) was also true
U(a2) and V(a2) When U(a2) was true, it was observed that V(a2) was also true

... ...

U(an) and V(an) When U(an) was true, it was observed that V(an) was also true

8x, U(x) ➔ Probably V(x) Therefore, whenever U(x) is true, V(x) is also probably true

Abductive Inference:

U(a1) ➔ V(a1) If U(a1) were true then V(a1) would follow as a matter of course
V(a1) V(a1) is true

Possibly U(a1) Therefore U(a1) is possibly true

1.4.2 The Search for Knowledge

We can extend Oldroyd’s Arch of Knowledge from Figure 1.2, as indicated in Figure 1.9, to
show how abduction, deduction, and induction are used in the search for knowledge.
They are at the basis of collaborative processes of evidence in search of hypotheses,
hypotheses in search of evidence, and evidentiary testing of hypotheses in a complex
dynamic world.
Through abductive reasoning (which shows that something is possibly true), we search
for hypotheses that explain our observations; through deductive reasoning (which shows
that something is necessarily true), we use our hypotheses to generate new lines of inquiry
and discover new evidence; and through inductive reasoning (which shows that something
is probably true), we test our hypotheses by evaluating our evidence. Now the discovery of
new evidence may lead to new hypotheses or the refinement of the existing ones. Also,
when there is more than one most likely hypothesis, we need to search for additional
evidence to determine which of them is actually the most likely. Therefore, the processes
of evidence in search of hypotheses, hypotheses in search of evidence, and evidentiary

13:51:08,
.002
 1.4. Evidence-based Reasoning 27

Explanatory Hypotheses Probability of Hypotheses

What hypotheses would What evidence is entailed What is the evidence-based

explain these observations? by each hypothesis? probability of each hypothesis?
Abduction: O  possibly H Deduction: H  necessarily E Induction: E  probably H

Observations New Evidence

Evidence in search Hypotheses in Evidentiary testing
of hypotheses search of evidence of hypotheses
Figure 1.9. The search for knowledge.

testing of hypotheses also take place in response to one another, as indicated by the
feedback loops from the bottom of Figure 1.9.

1.4.3 Evidence-based Reasoning Everywhere

As illustrated in Figure 1.10, evidence-based reasoning is at the core of many problem-
solving and decision-making tasks in a wide variety of domains, including physics,
chemistry, history, archaeology, medicine, law, forensics, intelligence analysis, cyberse-
curity, and many others. This is not surprising, because, as Jeremy Betham stated over
two centuries ago, “The field of evidence is no other than the field of knowledge”
(Betham, 1810, p. 5).
Scientists from various domains, such as physics, chemistry, or biology, may recognize
this as a formulation of the scientific method.
In medicine, a doctor makes observations with respect to a patient’s complaints and
attempts to generate possible diagnoses (hypotheses) that would explain them. He or
she then performs various medical tests that provide further evidence for or against
the various hypothesized illnesses. After that, the doctor uses the obtained evidence to
determine the most likely illness.
In law, an attorney makes observations in a criminal case and seeks to generate
hypotheses in the form of charges that seem possible in explaining these observations.
Then, assuming that a charge is justified, attempts are made to deduce further evidence
bearing on it. Finally, the obtained evidence is used to prove the charge.
In forensics, observations made at the site of an explosion in a power plant lead to the
formulation of several possible causes. Analysis of each possible cause leads to the discovery
of new evidence that eliminates or refines some of the causes, and may even suggest new
ones. This cycle continues until enough evidence is found to determine the most likely cause.
In intelligence analysis, an analyst formulates alternative hypotheses that would
explain the evidence about an event. Then the analyst puts each of the hypotheses at
work to guide him or her in the collection of additional evidence, which is used to assess
the probability of each hypothesis.
In cybersecurity, a suspicious connection to our computer from an external one triggers
the automatic generation of alternative threat and nonthreat hypotheses. Each generated

13:51:08,
.002
28 Chapter 1. Introduction

Natural Possible hypotheses or explanations Revised theory

Science

Evidence in search Hypotheses in Evidentiary tests

of hypotheses search of evidence of hypotheses

Observations of events in nature New observable phenomena

Medicine Possible illnesses Probabilities of illnesses

Evidence in search Hypotheses in Evidentiary tests

of hypotheses search of evidence of hypotheses

Patient’s complaints Test results

Law Possible charges or complaints Probabilities of charges

Evidence in search Hypotheses in Evidentiary tests

of hypotheses search of evidence of hypotheses

Observations during fact investigation New evidence

Forensics Possible causes Probabilities of causes

Evidence in search Hypotheses in Evidentiary tests

of hypotheses search of evidence of hypotheses

Observations at the site of incident New evidence

Intelligence Alternative hypotheses Probabilities of hypotheses

Analysis

Evidence in search Hypotheses in Evidentiary tests

of hypotheses search of evidence of hypotheses

Evidence New evidence

Cybersecurity Threat and nonthreat hypotheses Probabilities of hypotheses

Evidence in search Hypotheses in Evidentiary tests

of hypotheses search of evidence of hypotheses

Evidence of suspicious activity New evidence

Figure 1.10. Evidence-based reasoning everywhere.

13:51:08,
.002
 1.5. Artificial Intelligence 29

hypothesis guides the collection of additional evidence, which is used to assess the
probability of each hypothesis (Meckl et al., 2015).
The following, for instance, are different hypotheses one may be interested in assessing
based on evidence:

University U would be a good university for student S.

Professor P would be a good PhD Advisor for student S.

House H would be a good house to be bought by person P.

Country C will be a world leader in nonconventional energy sources within the next
decade.

Country C has nuclear weapons.

Person P was murdered by suspect S.

Patient P has illness I.

Building B has collapsed because of the use of low-quality materials.

Connection C is part of APT1 intrusion using malware M.

Evidence-based reasoning, however, is often highly complex, and the conclusions are
necessarily probabilistic in nature because our evidence is always incomplete, usually
inconclusive, frequently ambiguous, commonly dissonant, and has imperfect believability
(Schum, 1994 [2001a]; Tecuci et al., 2010b). Arguments requiring both imaginative and
critical reasoning, and involving all known types of inference (deduction, induction, and
abduction), are necessary in order to estimate the probability of the considered hypoth-
eses. Therefore, evidence-based reasoning can be best approached through the mixed-
initiative integration of human imagination and computer knowledge-based reasoning
(Tecuci et al., 2007a, 2007b), that is, by using knowledge-based intelligent agents for
evidence-based reasoning. Therefore, in the next section, we briefly review the field of
artificial intelligence.

1.5 ARTIFICIAL INTELLIGENCE

Artificial intelligence (AI) is the science and engineering domain concerned with the theory
and practice of developing systems that exhibit the characteristics we associate with intelli-
gence in human behavior, such as perception, natural language processing, problem solving
and planning, learning and adaptation, and acting on the environment. Its main scientific
goal is understanding the principles that enable intelligent behavior in humans, animals,
and artificial agents. This scientific goal directly supports several engineering goals, such
as developing intelligent agents, formalizing knowledge and mechanizing reasoning in all
areas of human endeavor, making working with computers as easy as working with
people, and developing human-machine systems that exploit the complementariness of
human and automated reasoning.
Artificial intelligence is a very broad interdisciplinary field that has roots in and
intersects with many domains, not only all the computing disciplines, but also mathemat-
ics, linguistics, psychology, neuroscience, mechanical engineering, statistics, economics,
control theory and cybernetics, philosophy, and many others. The field has adopted many
concepts and methods from these domains, but it has also contributed back.
While some of the developed systems, such as an expert system or a planning system,
can be characterized as pure applications of AI, most of the AI systems are developed as

13:51:08,
.002
 30 Chapter 1. Introduction

components of complex applications to which they add intelligence in various ways, for
instance, by enabling them to reason with knowledge, to process natural language, or to
learn and adapt.
Artificial intelligence researchers investigate powerful techniques in their quest for
realizing intelligent behavior. But these techniques are pervasive and are no longer
considered AI when they reach mainstream use. Examples include time-sharing, symbolic
programming languages (e.g., Lisp, Prolog, and Scheme), symbolic mathematics systems
(e.g., Mathematica), graphical user interfaces, computer games, object-oriented program-
ming, the personal computer, email, hypertext, and even the software agents. While this
tends to diminish the merits of AI, the field is continuously producing new results and, due
to its current level of maturity and the increased availability of cheap computational
power, it is a key technology in many of today's novel applications.

1.5.1 Intelligent Agents

It has become common to describe an AI system using the agent metaphor (Russell and
Norvig, 2010, pp. 34–63). An agent is a system that perceives its environment (which may be
the physical world, a user via a graphical user interface, a collection of other agents, the
Internet, or other complex environment); reasons to interpret perceptions, draw inferences,
solve problems, and determine actions; and acts upon that environment to realize a set of
goals or tasks for which it has been designed. An intelligent knowledge-based agent will
continuously improve its knowledge and performance through learning from input data,
from a user, from other agents, and/or from its own problem-solving experience. While
interacting with a human or some other agents, it may not blindly obey commands, but
may have the ability to modify requests, ask clarification questions, or even refuse to
satisfy certain requests. It can accept high-level requests indicating what the user wants
and can decide how to satisfy each request with some degree of independence or
autonomy, exhibiting goal-directed behavior and dynamically choosing which actions to
take and in what sequence. It can collaborate with users to improve the accomplishment
of their tasks or can carry out such tasks on their behalf, based on knowledge of their goals
or desires. It can monitor events or procedures for the users, can advise them on
performing various tasks, can train or teach them, or can help them collaborate (Tecuci,
1998, pp. 1–12).
Figure 1.11 shows the main components of a knowledge-based agent:

The knowledge base is a type of long-term memory that contains data structures
representing the objects from the application domain, general laws governing them,
and actions that can be performed with them.

The perceptual processing module implements methods to process natural language,
speech, and visual inputs.

The problem-solving engine implements general problem-solving methods that use the
knowledge from the knowledge base to interpret the input and provide an appropriate
output.

The learning engine implements learning methods for acquiring, extending, and refin-
ing the knowledge in the knowledge base.

The action processing module implements the agent’s actions upon that environment
aimed at realizing the goals or tasks for which it was designed (e.g., generation of

13:51:08,
.002
1.5. Artificial Intelligence 31

answers to questions, solutions of input problems, manipulation of objects, or naviga-

tion to a new position).

The reasoning area is a type of short-term memory where the actual reasoning
takes place.

An intelligent agent has an internal representation of its external environment that allows
it to reason about the environment by manipulating the elements of the representation.
For each relevant aspect of the environment, such as an object, a relation between objects,
a class of objects, a law, or an action, there is an expression in the agent’s knowledge base
that represents that aspect. For example, the left side of Figure 1.12 shows one way to
represent the simple world from the right side of Figure 1.12. The upper part is a
hierarchical representation of the objects and their relationships (an ontology). Under it
is a rule to be used for reasoning about these objects. This mapping between real entities
and their representations allows the agent to reason about the environment by manipu-
lating its internal representations and creating new ones. For example, by employing
natural deduction and its modus ponens rule, the agent may infer that cup1 is on table1.
The actual algorithm that implements natural deduction is part of the problem-solving
engine, while the actual reasoning is performed in the reasoning area (see Figure 1.11).
Since such an agent integrates many of the intelligent behaviors that we observe in humans,
it is also called a cognitive agent or system (Langley, 2012).
Most of the current AI agents, however, will not have all the components from
Figure 1.11, or some of the components will have very limited functionality. For example,
a user may speak with an automated agent (representing her Internet service provider)
that will guide her in troubleshooting the Internet connection. This agent may have

Problem-Solving Perceptual sensory

Engine Processing input
Environment

Knowledge Reasoning
Base Area
Action action
Learning Engine output
Processing

Figure 1.11. Notional architecture of an intelligent agent.

Model of the domain Applicaon domain

Ontology object Simple world

subconcept of

vessel publicaon furniture

subconcept of subconcept of subconcept of
Represents
cup book table
instance of instance of instance of
on book1 on table1
cup1
If an object is on top of another object that is
Rule itself on top of a third object, then the first
" x,y,z Î object, (on x y) & (on y z)  (on x z) object is on top of the third object.

Figure 1.12. An ontology fragment and a reasoning rule representing a simple agent world.

13:51:08,
.002
 32 Chapter 1. Introduction

advanced speech, natural language, and reasoning capabilities, but no visual or learning
capabilities. A natural language interface to a database may have only natural language
processing capabilities, while a face recognition system may have only learning and visual
perception capabilities.

1.5.2 Mixed-Initiative Reasoning

Table 1.2 lists some of the complementary abilities of humans and computer agents.
Humans are slow, sloppy, forgetful, implicit, and subjective, but have common sense
and intuition, and may find creative solutions in new situations. In contrast, computer
agents are fast, rigorous, precise, explicit, and objective, but they lack common sense and
the ability to deal with new situations (Turoff, 2007). Moreover, in contrast to an agent,
a human has a very limited attention span and can analyze only a small number of
alternatives at a time (Pomerol and Adam, 2006). That is why we have claimed that the
evidence-based reasoning tasks are best performed in a problem-solving and decision-
making environment that synergistically integrates the complementary capabilities of
humans and computer agents, taking advantage of their relative strengths to compensate
for each other’s weaknesses. Such an integration becomes even more important in face of
the globalization and the rapid evolution toward the knowledge economies (Filip, 2001;
Rooney et al., 2005). Indeed, these developments present additional challenges to decision
makers who need to cope with dynamic and increasingly complex situations and make
good decisions in face of an overwhelming amount of incomplete, uncertain, and mostly
irrelevant information.
Mixed-initiative reasoning is a type of collaboration between humans and computer
agents that mirrors the flexible collaboration between people. It is based on an efficient,
natural interleaving of contributions by people and agents that is determined by their
relative knowledge and skills and the problem-solving context, rather than by fixed roles,
enabling each participant to contribute what it does best, at the appropriate moment
(Horvitz, 1999; Tecuci et al., 2007a, 2007b).

Table 1.2. Complementary Computational Abilities of Humans and

Computer Agents

Humans are Computer agents are

slow fast
sloppy rigorous
forgetful precise
implicit explicit
subjective objective
but but
have common sense lack common sense
have intuition lack intuition
may find creative solutions in have poor ability to deal with
new situations new situations

13:51:08,
.002
 1.6. Knowledge Engineering 33

Mixed-initiative systems can either accomplish goals unachievable by the component

agents, assuming they work independently, or they can achieve the same goals more effectively.
An example of a good problem-solving and decision-making environment is one where
the human acts as the orchestrator of the reasoning process, guiding the high-level
exploration, while the agent implements this guidance by taking into account the human’s
preferred problem-solving strategies, assumptions, and biases (Tecuci et al., 2007c). In
such an environment, the agent is an extension of the reasoning capabilities of the human,
much like a calculator is an extension of the computational capabilities of an accountant.
The emphasis is on enhancing human’s creativity (Filip, 1989), relying on the human to
take the most critical decisions and only to critique and correct the more routine ones that
are proposed by the agent (Tecuci et al., 2007c).
Developing such an environment requires an automatic approach to problem solving
that is very natural and easy to understand. Moreover, the human and the agent should
collaborate in a natural way, similarly to how humans collaborate, as opposed to the usual
human–computer interaction that is inflexible and mostly unidirectional. Also, because
most of the complex decisions are based on incomplete and imperfect information, the
decision-making environment should allow the investigation of what-if scenarios, where
the decision maker can make various assumptions about a situation.
In the following section, we discuss the development of knowledge-based agents that
could be used for mixed-initiative reasoning.

1.6 KNOWLEDGE ENGINEERING

Knowledge engineering is the area of artificial intelligence that is concerned with the design,
development, and maintenance of agents that use knowledge and reasoning to perform
problem solving and decision making tasks.
Knowledge engineering is a central discipline in the Knowledge Society, the society
where knowledge is the primary production resource instead of capital and labor (Druker,
1993). Currently, human societies are rapidly evolving toward knowledge societies and an
Integrated Global Knowledge Society because of the development of the information
technologies, the Internet, the World Wide Web, and the Semantic Web that no longer
restrict knowledge societies to geographic proximity and that facilitate the sharing, archiv-
ing, retrieving, and processing of knowledge (Schreiber et al., 2000; David and Foray, 2003;
UNESCO, 2005). Moreover, the Semantic Web, an extension of the World Wide Web in
which Web content is expressed both in a natural form for humans, and in a format that can
be understood by software agents, is becoming the main infrastructure for the Knowledge
Society, allowing knowledge-based agents to automatically find, integrate, process, and
share information (Allemang and Hendler, 2011; W3C, 2015).

1.6.1 From Expert Systems to Knowledge-based Agents and

Cognitive Assistants
Expert systems are among the most successful applications of artificial intelligence.
According to Edward Feigenbaum (1982, p. 1), their founding father, “An ‘expert system’
is an intelligent computer program that uses knowledge and inference procedures to solve
problems that are difficult enough to require significant human expertise for their solution.

13:51:08,
.002
34 Chapter 1. Introduction

The knowledge necessary to perform at such a level, plus the inference procedures used, can
be thought of as a model of the expertise of the best practitioners in that field.”
Two early and very influential expert systems were DENDRAL (Buchanan and Feigen-
baum, 1978) and MYCIN (Buchanan and Shortliffe, 1984). DENDRAL, an expert system for
organic chemistry, analyzed mass spectral data and inferred a complete structural hypoth-
esis of a molecule. MYCIN, a medical expert system, produced diagnoses of infectious
diseases and advised the physician on antibiotic therapies for treating them.
Expert systems and knowledge-based systems are often used as synonyms since all expert
systems are knowledge-based systems. However, not all the knowledge-based systems
are expert systems, such as the Watson natural language question-answering system
(Ferrucci et al., 2010), and the Siri personal assistant (2011).
Continuous advances in artificial intelligence, particularly with respect to knowledge
representation and reasoning, learning, and natural language processing, are reflected in
more and more powerful and useful knowledge-based systems that, as discussed Section
1.5.1, are now more commonly called knowledge-based agents, or simply intelligent agents.
In this book, we are primarily concerned with a very important and newer class of
intelligent agents, namely cognitive assistants, which have the following capabilities:

Learn complex problem-solving expertise directly from human experts

Assist nonexpert users in solving problems requiring subject matter expertise

Assist human experts in complex problem solving and decision making

Teach problem solving and decision making to students

Expert systems, knowledge-based agents, and cognitive assistants are used in business,
science, engineering, manufacturing, military, intelligence, and many other areas (Durkin,
1994; Giarratano and Riley, 1994; Tecuci, 1998; Tecuci et al., 2001; 2008b). They are everywhere.
The following are examples of such successful systems.
Digital Equipment Corporation’s R1 (McDermott, 1982), which helped configure orders
for new computers, is considered the first successful commercial system. By 1986, it was
saving the company an estimated $40 million a year.
Intuit’s TurboTax, an American tax preparation software package initially developed by
Michael A. Chipman of Chipsoft in the mid-1980s (Forbes, 2013), helps you fill in your
taxes with the maximal deductions, according to the law.
The Defense Advanced Research Projects Agency’s (DARPA) DARP logistics planning
system was used during the Persian Gulf crisis of 1991 (Cross and Walker, 1994). It
planned with up to fifty thousand vehicles, cargo, and people, and reportedly more than
paid back DARPA’s thirty-year investment in artificial intelligence.
IBM’s Deep Blue chess playing system defeated Gary Kasparov, the chess world
champion, in 1997 (Goodman and Keene, 1997).
Disciple-COG agent for center of gravity analysis helped senior military officers from
the U.S. Army War College to learn how to identify the centers of gravity of the opposing
forces in complex war scenarios (Tecuci et al., 2002a; 2002b; 2008b).
NASA’s planning and scheduling systems helped plan and control the operations of
NASA’s spacecraft. For example, MAPGEN, a mixed-initiative planner, was deployed as a
mission-critical component of the ground operations system for the Mars Exploration
Rover mission (Bresina and Morris, 2007).
IBM’s Watson natural language question-answering system defeated the best human
players at the quiz show Jeopardy in 2011 (Ferrucci et al., 2010).

13:51:08,
.002
 1.6. Knowledge Engineering 35

Apple’s Siri, running on a cell phone, answers questions, makes recommendations on

nearby places of interest, and provides directions (2011).
TIACRITIS (Tecuci et al., 2011a, 2011b) and Disciple-CD (Tecuci et al., 2014) are
intelligent agents for evidence-based hypothesis analysis that help analysts in analyzing
complex hypotheses and teach new analysts.
In the following section, we present an overview of the types of tasks that can be
performed by expert systems, knowledge-based agents, and cognitive assistants.

1.6.2 An Ontology of Problem-Solving Tasks

Immanuel Kant, in his Critique of Pure Reason (Kant, 1781), considered that the two main
reasoning operations are analysis and synthesis. Analysis comes from the Greek word
analyein, which means “to break up.” It is a reasoning operation by which we break down
a system or problem into parts or components to better understand or solve it. Synthesis
also comes from a Greek word, syntithenai, which means “to put together.” It is the
complementary reasoning operation by which we combine system or solution compon-
ents to form a coherent system or solution.
Analysis and synthesis can be used as organizing principles for an ontology of problem-
solving tasks for knowledge-based agents, as was done by Clancey (1985), Breuker and
Wielinga (1989), and Schreiber et al. (2000, pp. 123–166). A fragment of such an ontology is
presented in Figure 1.13.
An analytic task is one that takes as input some data about a system or object and
produces a characterization of it as output. For example, in diagnosis, one analyzes the
symptoms of a malfunctioning system in order to determine their cause.

knowledge-
intensive task

analytic task synthetic task

classification prediction design assignment

critiquing monitoring planning debugging

configuration
intelligence analysis interpretation design scheduling

diagnosis
computer
configuration
medical mechanical
diagnosis diagnosis

Figure 1.13. Ontology of knowledge-intensive problem-solving tasks.

13:51:08,
.002
36 Chapter 1. Introduction

A synthetic task is one which takes as input the requirements of an object or system and
produces the corresponding object or system, as, for instance, in designing a car based on
given specifications.
Under each type of analytic or synthetic task, one may consider more specialized
versions of that task. For example, special cases of diagnosis are medical diagnosis and
mechanical diagnosis. Special cases of mechanical diagnosis are car diagnosis, airplane
diagnosis, and so on.
The importance of identifying such types of tasks is that one may create general models
for solving them (e.g., a general model of diagnosis) that could guide the development of
specialized systems (e.g., a system to diagnose Toyota cars).
Schreiber et al. (2000, pp. 123–166) present abstract problem-solving models for many
of the task types in Figure 1.13. These abstract models can provide initial guidance when
developing knowledge-based agents to perform such tasks.

1.6.2.1 Analytic Tasks

The following is a brief characterization with examples of the analytic tasks from Figure 1.13.
Classification means determining the class to which an object belongs as, for instance,
determining that a plant is a rose based on its characteristics. Credit card companies, for
instance, use classification software learned from examples that classify applicants into
various categories corresponding to the type of card to be approved for them (if any).
Critiquing means expressing judgments about something according to certain standards.
For example, one may identify strengths and weaknesses in a military course of action by
considering the principles of war and the tenets of Army operations and determining how
well each principle or tenet is illustrated by that course of action, as done by the Disciple-
COA agent (Tecuci et al., 2001), which will be presented in Section 12.3.
Interpretation means inferring situation description from sensory data as, for example,
interpreting gauge readings in a chemical process plant to infer the status of the process.
Monitoring means comparing observations of a dynamic system or process with the expected
outcomes, to identify changes in its state and take appropriate actions. Examples include
monitoring instrument readings in a nuclear reactor to detect accident conditions, or monitoring
a patient in an intensive care unit based on the data from the monitoring equipment.
Prediction means inferring likely consequences of given situations, such as predicting
the damage to crops from some type of insect, estimating global oil demand from the
current geopolitical world situation, or forecasting the weather.
Diagnosis means inferring system malfunctions from observables, such as determining
the disease of a patient from the observed symptoms, locating faults in electrical circuits,
finding defective components in the cooling system of nuclear reactors, or diagnosing the
faults in an electrical network.
Intelligence analysis means analyzing information to estimate the probability of various
hypotheses, such as estimating the probability that Al Qaeda has nuclear weapons.
Examples of systems for intelligence analysis are Disciple-LTA (Tecuci et al., 2005a;
2008a), TIACRITIS (Tecuci et al., 2011a, 2011b), and Disciple-CD (Tecuci et al., 2014).

1.6.2.2 Synthetic Tasks

The following is a brief characterization with examples of the synthetic tasks from
Figure 1.13.

13:51:08,
.002
 1.6. Knowledge Engineering 37

Design means configuring objects under constraints, such as designing an elevator with
a certain capacity and speed, as done by the SALT system (Marcus, 1988), or designing a
computer system with certain memory, speed, and graphical processing characteristics,
based on a set of predefined components.
Planning means finding a set of actions that achieve a certain goal, such as determining
the actions that need to be performed in order to repair a bridge, as done by the Disciple-
WA agent (Tecuci et al., 2000), which will be discussed in Section 12.2. A more complex
example is collaborative emergency response planning, illustrated by Disciple-VPT
(Tecuci et al., 2008c), which is presented in Section 12.5.
Scheduling means allocating sequences of activities or jobs to resources or machines on
which they can be executed, such as scheduling the lectures in the classrooms of a
university, or scheduling the sequence of operations needed to produce an object on
the available machines in a factory.
Debugging means prescribing remedies for malfunctions, such as determining how to
tune a computer system to reduce a particular type of performance problem.
Assignment means creating a partial mapping between two sets of objects, such as
allocating offices to employees in a company or allocating airplanes to gates in an airport.

1.6.3 Building Knowledge-based Agents

1.6.3.1 How Knowledge-based Agents Are Built and Why It Is Hard
As shown in the right-hand side of Figure 1.14, the basic components of a knowledge-
based agent are the problem-solving engine and the knowledge base.
The problem-solving engine implements a general method of solving the input prob-
lems based on the knowledge from the knowledge base. An example of such a general
method is problem reduction and solution synthesis, which will be discussed in detail in
Chapter 4. As illustrated in the right side of Figure 1.15, this method consists in solving a
problem, such as P1, by successively reducing it, from the top down, to simpler and
simpler problems; finding the solutions of the simplest problems; and successively com-
bining these solutions, from the bottom up, into the solution of the initial problem
(i.e., S1).
Knowledge of the actual problems to solve and how they can be solved depends on the
expertise domain. This knowledge is represented into the knowledge base of the system
by using different representation formalisms. The left and middle parts of Figure 1.15
illustrate the representation of this knowledge by using an ontology of concepts and a set
of rules expressed with these concepts. The ontology describes the types of objects in
the application domain, as well as the relationships between them. Some of the rules are

Subject Knowledge Knowledge-based

Matter Expert Engineer Agent
Problem-Solving
Programming
Dialogue Engine

Knowledge
Base

Results

Figure 1.14. Conventional approach to building a knowledge-based agent.

13:51:08,
.002
38 Chapter 1. Introduction

Ontology Rules Reasoning tree

IF the problem to solve is P1g

person IF the problem to solve is P1g P1 S1
Condition
IF the problem to solve is P1g
Condition
IFExcept-
the problem to solve is P1g
… Condition When Condition
IFExcept-
the problem When to solve is P1g
Condition
faculty member student … Condition
IFExcept-
the problem to solve is P1g
Except-When WhenCondition
Condition
IF…the problem
Condition to solve is P1g P11 S11 … Pn1 S1n
IFExcept-
Except-When
Condition
…the
WhenCondition
problem Condition
to solve is P1g
graduate THEN Except-When
Except-
solve
Condition itsWhen Condition
Condition
subproblems
…Except- When Condition
student THEN Except-When
P
…Condition
solve
1 … Png its
1 Condition
subproblems
Except-
…
1g
THENExcept-When
solve
P1g1 … its PWhen Condition
1 subproblems
Condition
professor Except- When
ng Condition
THEN Except-When
Psolve
1 …P its Condition
1 subproblems
instructor undergraduate … 1g
Except-When ng
Condition
student THEN 1 … its
Psolve P 1 subproblems
ng P21 S21 … Pm
2 2
Sm
THEN Except-When
1g Condition
Psolve
1 …P its1 subproblems
ng1
assistant full THEN solve P1g1g
1 … its Png subproblems
professor professor THEN Psolve 1 …P its1 subproblems
PhD student 1g1 ng
1
P1g … Png
associate
professor MS student BS student P13 S31 … P3 S3
p p

Figure 1.15. Knowledge and reasoning based on problem reduction and solution synthesis.

if-then structures that indicate the conditions under which a general complex problem
(such as P1g) can be reduced to simpler problems. Other rules indicate how the solutions
of simpler problems can be combined into the solution of the more complex problem.
These rules are applied to generate the reasoning tree from the right part of Figure 1.15.
Figure 1.14 illustrates the conventional approach to building a knowledge-based agent.
A knowledge engineer interviews the subject matter expert to understand how the expert
reasons and solves problems, identifies the knowledge used by the expert, and then
represents it into the agent's knowledge base. For instance, the knowledge engineer may
represent the knowledge acquired from the expert as an ontology of concepts and a set of
reasoning rules expressed with these concepts, like those from Figure 1.15. Then the agent
is used to solve typical problems, and the subject matter expert analyzes the generated
solutions (e.g., the reasoning tree from the right side of Figure 1.15), and often the
knowledge base itself, to identify errors. Referring to the identified errors, the knowledge
engineer corrects the knowledge base.
After more than two decades of work on expert systems, Edward Feigenbaum (1993)
characterized the knowledge-based technology as a tiger in a cage:

The systems offer remarkable cost savings; some dramatically “hot selling”
products; great return-on-investment; speedup of professional work by
factors of ten to several hundred; improved quality of human decision making
(often reducing errors to zero); and the preservation and “publishing” of
knowledge assets of a firm. . . . These stories of successful applications,
repeated a thousand fold around the world, show that knowledge-based
technology is a tiger. Rarely does a technology arise that offers such a wide
range of important benefits of this magnitude. Yet as the technology moved
through the phase of early adoption to general industry adoption, the
response has been cautious, slow, and “linear” (rather than exponential).

The main reason for this less than exponential growth of expert systems lies in the
difficulty of capturing and representing the knowledge of the subject matter expert in
the system’s knowledge base. This long, difficult, and error-prone process is known as
the “knowledge acquisition bottleneck” of the system development process. But why is

13:51:08,
.002
1.6. Knowledge Engineering 39

knowledge acquisition so difficult? First, much of an expert’s problem-solving knowledge

is in the form of tacit knowledge, which is difficult to make explicit (Nonaka and Krogh,
2009). Second, when describing their knowledge, experts use natural language, visual
representations, and common sense, often omitting essential details that are implicit in
human communication. This is very different from the way in which knowledge has to be
represented into the knowledge base, which is formal, precise, and sufficiently complete.
Finally, the knowledge engineer needs domain training in order to understand properly
the expert’s problem-solving knowledge, and this takes time and effort.

1.6.3.2 Teaching as an Alternative to Programming: Disciple Agents

In an interesting way, the difficulty of building knowledge-based systems confirms Tur-
ing’s intuition that building an intelligent machine by programming is too difficult and
that a simpler way might be to teach a learning agent: “Instead of trying to produce a
program to simulate the adult mind, why not rather try to produce one which simulates the
child's? If this were then subjected to an appropriate course of education one would obtain
the adult brain” (Turing, 1950).
The Disciple learning and reasoning agents, which will be presented in detail in this
book, represent one realization of Turing’s intuition. A subject matter expert can teach a
Disciple agent how to solve problems in ways that are similar to how the expert would
teach a student or a collaborator (Tecuci, 1988; Tecuci, 1998; Boicu et al., 2000). For
example, the expert will show the agent examples of how to solve specific problems,
helping it to understand the solutions, and will supervise and correct the problem-
solving behavior of the agent. The agent will learn from the expert by generalizing the
examples and building its knowledge base. In essence, this creates a synergism between
the expert who has the knowledge to be formalized and the agent that knows how to
formalize it. This approach is based on methods for mixed-initiative problem solving,
integrated teaching and learning, and multistrategy learning. In mixed-initiative prob-
lem solving (Tecuci et al., 2007a, 2007b), the expert solves the more creative parts of the
problem and the agent solves the more routine ones. In integrated teaching and learning
(Tecuci and Kodratoff, 1995; Tecuci, 1998), for example, the agent helps the expert to
teach it by asking relevant questions, and the expert helps the agent to learn by providing
examples, hints, and explanations. In multistrategy learning (Tecuci, 1993; Michalski and
Tecuci, 1994), the agent integrates multiple learning strategies, such as learning from
examples, learning from explanations, and learning by analogy, to learn from the expert
how to solve problems.
Disciple agents have been developed in many domains, including intelligence analysis
(Tecuci et al., 2011a, 2011b; 2014; 2015), modeling of the behavior of violent extremists,
military center of gravity determination (Tecuci et al., 2002b; 2008b), course of action
critiquing (Tecuci et al, 2001), planning (Tecuci et al, 2000; 2008c), financial services, and
education (Tecuci and Keeling, 1999).
As illustrated in Figure 1.16, learning agents such as the Disciple agents contribute to a
new age in the software systems development process. In the age of mainframe computers,
the software systems were both built and used by computer science experts. In the current
age of personal computers, such systems are still being built by computer science experts,
but many of them (such as text processors, email programs, and Internet browsers) are
regularly used by persons who have no formal computer science education. We think
that this trend will continue, with the next age being that of cognitive assistants on the

13:51:08,
.002
40 Chapter 1. Introduction

Soware systems
Soware systems developed developed and used by
by computer experts and persons who are not
used by persons who are not computer experts
Soware systems computer experts
developed and used
by computer experts Cognive
Personal Assistants
Computers
Mainframe
Computers

Semanc
Web

Figure 1.16. Evolution of software development and use.

Semantic Web, where typical computer users will be able to both develop and use special
types of software agents.
The learning agent technology illustrated by the Disciple approach attempts to
change the way the knowledge-based agents are built, from “being programmed” by a
knowledge engineer to “being taught” by a user who does not have prior knowledge
engineering or computer science experience. This approach will allow typical computer
users, who are not trained knowledge engineers, to build by themselves cognitive
assistants. Thus, non–computer scientists will no longer be only users of generic pro-
grams developed by others (such as word processors or Internet browsers), as they are
today, but also agent developers themselves. They will be able to train their cognitive
assistants to help them with their increasingly complex tasks in the Knowledge Society,
which should have a significant beneficial impact on their work and life. This goal is
consistent with the Semantic Web vision of enabling typical users to author Web content
that can be understood by automated agents (Allemang and Hendler, 2011; W3C, 2015).
Bill Gates has also stressed the great potential and importance of software assistants
(Simonite, 2013).
Because the subject matter expert teaches a Disciple agent similarly to how the expert
would teach a student, through explicit examples and explanations, a trained Disciple
agent can be used as an assistant by a student, learning from the agent’s explicit reasoning.
Alternatively, Disciple may behave as a tutoring system, guiding the student through a
series of lessons and exercises. Educational Disciple agents have been developed for
intelligence analysis (Tecuci et al., 2011a, 2011b) and for center of gravity determination
(Tecuci et al., 2008b). Thus the Disciple agents also contribute to advancing “Personalized
Learning,” which is one of the fourteen Grand Challenges for the Twenty-first Century
identified by the U.S. National Academy of Engineering (NAE, 2008).

1.6.3.3 Disciple-EBR, Disciple-CD, and TIACRITIS

The Disciple learning agents theory and technology has been continuously developed over
many years, with significant stages presented in a series of PhD theses (Tecuci, 1988;
Dybala, 1996; Hieb, 1996; Keeling, 1998; Boicu M., 2002; Bowman, 2002; Boicu C., 2006;

13:51:08,
.002
 1.7. Obtaining Disciple-EBR 41

Le 2008; Marcu 2009), and several books (e.g., Tecuci, 1998; Tecuci et al., 2008b). Some of
the most representative implementations of this evolving theory and technology are
discussed in Chapter 12. The rest of this book, however, focuses on the most recent
advances of this theory and technology that enables the development of Disciple agents
for evidence-based reasoning (EBR) tasks such as those introduced in Section 1.4.3.
The corresponding agent development environment is called Disciple-EBR, which can
be used by a subject matter expert, with support from a knowledge engineer, to develop a
knowledge-based agent incorporating his or her expertise.
Disciple-EBR (the Disciple learning agent shell for evidence-based reasoning) will be
used throughout this book to explain knowledge engineering concepts, principles, and
methods using a hands-on approach. It will also be the software environment used in the
agent design and development project.
There is also a reduced version of Disciple-EBR, called Disciple-CD (the Disciple
cognitive assistant for “Connecting the Dots”). This version was created for the end-user
who has no knowledge engineering experience and receives no support from a
knowledge engineer (Tecuci et al., 2014). Therefore, when using Disciple-CD, the user
does not have access to any Disciple-EBR module that may require any kind of
knowledge engineering support, such as Ontology Development, Rule Learning, or
Rule Refinement.
We have written a book for intelligence analysis courses, titled Intelligence Analysis as
Discovery of Evidence, Hypotheses, and Arguments: Connecting the Dots (Tecuci et al.,
2016), which uses Disciple-CD. This is because Disciple-CD incorporates a significant
amount of knowledge about evidence and its properties, uses, and discovery to help the
students acquire the knowledge, skills, and abilities involved in discovering and processing
of evidence and in drawing defensible and persuasive conclusions from it, by employing
an effective learning-by-doing approach. The students can practice and learn how to link
evidence to hypotheses through abductive, deductive, and inductive reasoning that estab-
lish the basic credentials of evidence: its relevance, believability or credibility, and infer-
ential force or weight. They can experiment with “what-if” scenarios and can study the
influence of various assumptions on the final result of analysis. So, their learning experi-
ence will be a joint venture involving the intelligence analysis book together with their
interaction with Disciple-CD.
Disciple-CD is a significant improvement over an earlier system that we have
developed for intelligence analysis, called TIACRITIS (Teaching Intelligence Analysts
Critical Thinking Skills), and it subsumes all the reasoning and learning capabilities
of TIACRITIS that have been described in several papers (Tecuci et al., 2010b;
2011a; 2011b).

1.7 OBTAINING DISCIPLE-EBR

Disciple-EBR (Disciple, for short) is a learning agent shell for evidence-based reasoning. It
is a research prototype implemented in Java and tested on PC. Disciple-EBR is a stand-
alone system that needs to be installed on the user’s computer.
For installation requirements and to download the system, visit http://lac.gmu.edu/
KEBook/Disciple-EBR/. At this address, you will also find instructions on how to install
and uninstall Disciple-EBR, a section with frequently asked questions (FAQs), and a
section that allows users to submit error reports to the developers of the system.

13:51:08,
.002
 42 Chapter 1. Introduction

1.8 REVIEW QUESTIONS

1.1. Consider the following illustrations of the concepts data, information, and
knowledge:
Data: the color red.
Information: red tomato.
Knowledge: If the tomato is red, then it is ripe.
Data: the sequence of dots and lines “. . .–. . .”
Information: the “S O S” emergency alert.
Knowledge: If there is an emergency alert, then start rescue operations.
Provide two other illustrations of these concepts.

1.2. What is evidence?

1.3. Provide an example of an item of evidence.

1.4. Why does evidence differ from data or items of information?

1.5. Give an example of a fact F and of evidence about F. In general, what is the
difference between a fact and evidence about that fact?

1.6. Formulate a hypothesis. Indicate an item of evidence that favors this hypothesis, an
item of evidence that disfavors this hypothesis, and an item of information that is
not evidence for this hypothesis.

1.7. What is deduction? Provide an example of deductive reasoning.

1.8. What is abduction? Give an example of abductive reasoning. Provide other explan-
ations or hypotheses that are less plausible. Specify a context where one of these
alternative explanatory hypotheses would actually be more plausible.

1.9. A doctor knows that the disease hepatitis causes the patient to have yellow eyes
90 percent of the time. The doctor also knows that the probability that a patient has
hepatitis is one in one hundred thousand, and the probability that any patient has
yellow eyes is one in ten thousand. What is the probability that a patient with
yellow eyes has hepatitis?

1.10. Suppose that in answering a multiple-choice test question with five choices, a
student either knows the answer, with probability p, or she guesses it with prob-
ability 1 – p. Assume that the probability of answering a question correctly is 1 for a
student who knows the answer. If the student guesses the answer, she chooses one
of the options with equal probability. What is the probability that a student knew
the answer, given that she answered it correctly? What is this probability in the case
of a true-false test question?

1.11. Consider a hypothesis H and its negation ¬H. Suppose you believe, before you
have any evidence, that the prior probability of H is PðH Þ ¼ 0:30. Now you receive
an item of evidence E  and ask yourself how likely is this evidence E  if H were true,
and how likely is this evidence E  , if H were not true. Suppose you say that
P ðE  jH Þ ¼ 0:70 and P ðE  j¬H Þ ¼ 0:10. What are your prior odds OddsðH : ¬H Þ,

13:51:08,
.002
1.8. Review Questions 43

and how have these odds changed as a result of the evidence E  ? What is the
posterior probability PðHjE  Þ?

1.12. Suppose in Question 1.11 you said that the prior probability of H is P ðH Þ ¼ 0:20
and the posterior probability P ðHjE  Þ ¼ 0:95. What would be the force of evidence
E  (i.e., the likelihood ratio LE*) that is implied by these assessments you
have made?

1.13. Think back to the very first time you were ever tutored about probability, what it
means, and how it is determined. What were you told about these matters? Then
describe your present views about these probability matters.

1.14. As we noted, the subjective Bayesian view of probability lets us assess prob-
abilities for singular, unique, or one-of-a-kind events, provided that our
assessed probabilities obey the three Kolmogorov axioms we discussed
regarding enumerative probabilities. First, is there any way of showing that
these axioms for enumerative probabilities also form the basis for ideal or
optimal probability assessments in the nonenumerative case? Second, can
this really be the rational basis for all probability assessments based on
evidence?

1.15. Show how Bayes’ rule supplies no method for incorporating “pure evidence” as
does the Belief Function system.

1.16. Provide an example showing how an analyst’s numerical assessment of a probabil-

ity applied to a conclusion can invite criticism.

1.17. What is induction? Provide an example of inductive reasoning.

1.18. Consider the following statements:

(a) All the beans from this bag are white.
(b) These beans are from this bag.
(c) These beans are white.
We will arrange these statements in three different ways and you will have to
identify what type of inference it represents (Walton, 2004).

(1) Premise One: (a) All the beans from this bag are white.
Premise Two: (b) These beans are from this bag.
Conclusion: (c) These beans are white.

(2) Premise One: (b) These beans are from this bag.
Premise Two: (c) These beans are white.
Conclusion: (a) All the beans from this bag are white.

(3) Premise One: (a) All the beans from this bag are white.
Premise Two: (c) These beans are white.
Conclusion: (b) These beans are from this bag.

13:51:08,
.002
44 Chapter 1. Introduction

1.19. Consider the following statements:

(a) These persons are French nationals.

(b) All the French nationals speak French.
(c) These persons speak French.
Arrange these statements in three different ways so that the reasoning is the
indicated one (using a word in the conclusion that is characteristic to that type of
reasoning):

Deductive Inference

Premise One:
Premise Two:
Conclusion:

Inductive Inference

Premise One:
Premise Two:
Conclusion:

Abductive Inference

Premise One:
Premise Two:
Conclusion:

1.20. Provide examples of evidence-based reasoning in law, medicine, geography, cyber-

security, and intelligence analysis.

1.21. Give an example of an observation and of several hypotheses that would explain it.

1.22. Define artificial intelligence.

1.23. What is an intelligent agent?

1.24. Describe the generic architecture of an intelligent agent and the role of each main
component.

1.25. Which are two main types of knowledge often found in the knowledge base of
an agent?

1.26. What are some of the complementary abilities of humans and computer agents?

1.27. What would be a good mixed-initiative environment for problem solving and
decision making? What are some key requirements for such an environment?

1.28. How do assumptions enable mixed-initiative problem solving? How do they enable
problem solving in the context of incomplete information?

1.29. What is a knowledge-based system?

13:51:08,
.002
1.8. Review Questions 45

1.30. Which are some other examples of the analytic tasks introduced in Section
1.6.2.1?

1.31. Which are some other examples of the synthetic tasks introduced in Section
1.6.2.2?

1.32. Why is building a knowledge-based system difficult?

13:51:08,
.002
 Evidence-based Reasoning:
2 Connecting the Dots

In Section 1.4.3, we have briefly introduced evidence-based reasoning in various domains

(see Figure 1.10, p. 28). In this section, we start with discussing the complexity of evidence-
based reasoning by using the “connecting the dots” metaphor. Then we discuss in more
detail evidence-based reasoning in a representative EBR domain, intelligence analysis. We
conclude this section with other examples of evidence-based reasoning. Then the
following chapters will address the development of such systems and of knowledge-
based agents in general.

2.1 HOW EASY IS IT TO CONNECT THE DOTS?

The “connecting the dots” metaphor seems appropriate for characterizing evidence-based
reasoning. This metaphor may have gained its current popularity following the terrorist
attacks in New York City and Washington, D.C., on September 11, 2001. It was frequently
said that the intelligence services did not connect the dots appropriately in order to have
possibly prevented the catastrophes that occurred. Since then, we have seen and heard
this metaphor applied in the news media to inferences in a very wide array of contexts, in
addition to the intelligence, including legal, military, and business contexts. For example,
we have seen it applied to allegedly faulty medical diagnoses; to allegedly faulty conclu-
sions in historical studies; to allegedly faulty or unpopular governmental decisions; and in
discussions involving the conclusions reached by competing politicians. What is also true
is that the commentators on television and radio, or the sources of written accounts of
inferential failures, never tell us what they mean by the phrase “connecting the dots.”
A natural explanation is that they have never even considered what this phrase means
and what it might involve.
But we have made a detailed study of what “connecting the dots” entails. We have
found this metaphor very useful, and quite intuitive, in illustrating the extraordinary
complexity of the evidential and inferential reasoning required in the contexts we have
mentioned. Listening or seeing some media accounts of this process may lead one to
believe that it resembles the simple tasks we performed as children when, if we connected
some collection of numbered dots correctly, a figure of Santa Claus, or some other familiar
figure, would emerge. Our belief is that critics employing this metaphor in criticizing
intelligence analysts and others have very little awareness of how astonishingly difficult the
process of connecting unnumbered dots can be in so many contexts (Schum, 1987).
A natural place to begin our examination is by trying to define what is meant by the
metaphor "connecting the dots" when it is applied to evidence-based reasoning tasks:

46

13:50:24,
.003
 2.1. How Easy Is It to Connect the Dots? 47

“Connecting the Dots” refers to the task of marshaling thoughts and evidence in the
generation or discovery of productive hypotheses and new evidence, and in the construction
of defensible and persuasive arguments on hypotheses we believe to be most favored by the
evidence we have gathered and evaluated.
The following represents an account of seven complexities in the process of “connect-
ing the dots.”

2.1.1 How Many Kinds of Dots Are There?

It is so easy to assume that the only kind of dot to be connected concerns details in the
observable information or data we collect that may eventually be considered as evi-
dence in some analysis. We might refer to these dots as being evidential dots. Sherlock
Holmes had another term for the details in observations he made, calling them trifles.
As he told Dr. Watson: “You know my method, it is based on the observance of
trifles.” A related problem here is that most items of evidence may contain many
details, dots, or trifles, some of which are interesting and others not. What this means
is the information must be carefully parsed in order to observe its significant evidential
dots. Not all data or items of information we have will ever become evidence in an
analysis task.
Consider the bombing during the Boston Marathon that took place on April 15,
2013. Many images were taken during this event. One is a widely televised videotape
of two young men, one walking closely behind the other, both carrying black back-
packs. This is the evidential dot shown in the bottom-left of Figure 2.1. Why should we
be interested in this evidence dot? Because it suggests ideas or hypotheses of what
might have actually happened. Consider our ideas or thoughts concerning the rele-
vance of the backpack dot just described. We have other evidence that the two bombs
that were set off were small enough to be carried in backpacks. This allows the
inference that the backpacks carried by the two young men might have contained
explosive devices and that they should be considered as suspects in the bombing.
A further inference is that these two men were the ones who actually detonated the
two bombs.

These two men were the ones who

Hypothesis actually detonated the two bombs.

They should be considered

as suspects in the bombing.

Idea Dots The backpacks carried by the

two young men might have
contained explosive devices.

Videotape of two young The two bombs that were

Evidential men, one walking closely set off were small enough
Dots behind the other, both to be carried in backpacks.
carrying black backpacks.

Figure 2.1. Types of dots to be connected: evidence, ideas, and hypotheses.

13:50:24,
.003
 48 Chapter 2. Evidence-based Reasoning: Connecting the Dots

Thus, the second type of dot concerns ideas we have about how some evidential dots
are connected to matters we are trying to prove or disprove. We commonly refer to the
matters to be proved or disproved as hypotheses. Hypotheses commonly refer to possible
alternative conclusions we could entertain about matters of interest in an analysis. The
other dots, which we call idea dots, come in the form of links in chains of reasoning or
arguments we construct to link evidential dots to hypotheses. Of course, hypotheses are
also ideas. Each of these idea dots refers to sources of uncertainty or doubt we believe to
be interposed between our evidence and our hypotheses. This is precisely where imagina-
tive reasoning is involved. The essential task for the analyst is to imagine what evidential
dots mean as far as hypotheses or possible conclusions are concerned. Careful critical
reasoning is then required to check on the logical coherence of sequences of idea dots in
our arguments or chains of reasoning. In other words, does the meaning we have attached
to sequences of idea dots make logical sense?

2.1.2 Which Evidential Dots Can Be Believed?

The next problem we discuss is one of the most important, challenging, and interesting
problems raised in any area of analysis. From some source, a sensor of some sort, or from
a person, we obtain an evidential dot saying that a certain event has occurred. Just because
this source says that this event occurred does not entail that it did occur. So, as discussed
in Section 1.1.3, what is vitally necessary is to distinguish between evidence of an event and
the event itself. We adopt the following notational device to make this distinction:

 E represents the actual occurrence of event E

 E*i represents the reported occurrence of event E from source .

So, a basic inference we encounter is whether or not E did occur based on our evidence
E*i. Clearly, this inference rests upon what we know about the believability of source .
There are some real challenges here in discussing the believability of source . Section 4.7
of this book is devoted to the task of assessing the believability of the sources of our
evidence. As we will see, Disciple-EBR (as well as Disciple-CD and TIACRITIS, introduced
in Section 1.6.3.3) already knows much about this crucial task.
But there are even distinctions to be made in what we have called evidential dots. Some
of these dots arise from objects we obtain or from sensors that supply us with records or
images of various sorts. So one major kind of evidential dot involves what we can call
tangible evidence that we can observe for ourselves to see what events it may reveal. In
many other cases, we have no such tangible evidence but must rely upon the reports of
human sources who allegedly have made observations of events of interest to us. Their
reports to us come in the form of testimonial evidence or assertions about what they have
observed. Therefore, an evidential dot E*i can be one of the following types:

 Tangible evidence, such as objects of various kinds, or sensor records such as those
obtained by signals intelligence (SIGINT), imagery intelligence (IMINT), measurement
and signature intelligence (MASINT), and other possible sources
 Testimonial evidence obtained from human sources (HUMINT)

The origin of one of the greatest challenges in assessing the believability of evidence
is that we must ask different questions about the sources of tangible evidence than

13:50:24,
.003
2.1. How Easy Is It to Connect the Dots? 49

those we ask about the sources of testimonial evidence. Stated another way, the
believability attributes of tangible evidence are different from the believability attri-
butes of testimonial evidence. Consider again the evidential dot concerning the two
men carrying backpacks. This is an example of tangible evidence. We can all examine
this videotape to our heart’s content to see what events it might reveal. The most
important attribute of tangible evidence is its authenticity: Is this evidential dot what
it is claimed to be? The FBI claims that this videotape was recorded on April 15, 2013,
on Boyleston Street in Boston, Massachusetts, where the bombings occurred, and
recorded before the bombings occurred. Our imaginations are excited by this claim
and lead to questions such as those that would certainly arise in the minds of defense
attorneys during trial. Was this videotape actually recorded on April 15, 2013? Maybe it
was recorded on a different date. If it was recorded on April 15, 2013, was it recorded
before the bombings occurred? Perhaps it was recorded after the bombings occurred.
And, was this videotape actually recorded on Boyleston Street in Boston, Massachu-
setts? It may have been recorded on a different street in Boston, or perhaps on a street
in a different city.
But there is another difficulty that is not always recognized that can cause endless
trouble. While, in the case of tangible evidence, believability and credibility may be
considered as equivalent terms, human sources of evidence have another characteristic
apart from credibility; this characteristic involves their competence. As we discuss in
Section 4.7.2, the credibility and competence characteristics of human sources must not
be confused; to do so invites inferential catastrophes, as we will illustrate. The questions
required to assess human source competence are different from those required to assess
human source credibility. Competence requires answers to questions concerning the
source’s actual access to, and understanding of, the evidence he or she reports. Credibil-
ity assessment for a testimonial source requires answers to questions concerning the
veracity, objectivity, and observational sensitivity or accuracy of the source. Disciple-EBR
knows the credibility-related questions to ask of tangible evidence and the competence-
and credibility-related questions to ask of HUMINT sources.
There is no better way of illustrating the importance of evidence believability assess-
ments than to show how such assessments form the very foundation for all arguments we
make from evidence to possible conclusions. In many situations, people will mistakenly
base inferences on the assumption that an event E has occurred just because we have
evidence E*i from source . This amounts to the suppression of any uncertainty we have
about the believability of source (whatever this source might be). In Figure 2.2 is a

{H, not H}

{G, not G}
Relevance
links in the
argument
{F, not F}

Believability {E, not E}

foundation of
the argument Evidence E*i

Figure 2.2. The believability foundation for an argument.

13:50:24,
.003
 50 Chapter 2. Evidence-based Reasoning: Connecting the Dots

simple example illustrating this believability foundation; it will also allow us to introduce
the next problem in connecting the dots.
What this figure shows is an argument from evidence E*i as to whether or not hypoth-
esis H is true. As shown, the very first stage in this argument concerns an inference about
whether or not event E actually occurred. This is precisely where we consider whatever
evidence we may have about the believability of source . We may have considerable
uncertainty about whether or not event E occurred. All subsequent links in this argument
concern the relevance of event E on hypothesis H. As we noted in Figure 2.1, these
relevance links connect the idea dots we discussed. As Figure 2.2 shows, each idea dot is
a source of uncertainty associated with the logical connection between whether or not
event E did occur and whether or not H is true. Consideration of these relevance links is
our next problem in connecting the dots.

2.1.3 Which Evidential Dots Should Be Considered?

In all of the contexts we have considered, there is usually no shortage of potential
evidential dots. In fact, in many of these contexts, persons drawing conclusions about
matters of importance are swamped with information or data. This situation is currently
being called the “big data problem.” Here we begin to consider vital matters concerning
the discovery-related or investigative tasks and the imaginative or creative reasoning these
tasks involve. Unfortunately, in many situations people or organizations try to collect
everything in the hope of finding something useful in an inference task. This wasteful
practice is one reason why the big data problem exists, since only a minute fraction of the
information collected will be relevant in any inference of concern. In our work, we have
paid great attention to the process of discovery that necessarily takes place in a world that
keeps changing all the while we are trying to understand parts of it of interest to us in our
inference tasks. As will be discussed in Section 2.2, this is an ongoing, seamless activity in
which we have evidence in search of hypotheses, hypotheses in search of evidence, and
the testing of hypotheses all going on at the same time. Hypotheses you entertain,
questions you ask, particular evidence items, and your accumulated experience all allow
you to examine which evidential dots to consider. Part of our objectives here is to make the
process of discovery more efficient. As was discussed in Section 1.4.2, these discovery tasks
involve mixtures of three different forms of reasoning: deduction, induction (probabilistic),
and abduction (imaginative, creative, or insightful) reasoning. These forms of reasoning
provide the bases for our idea dots.

2.1.4 Which Evidential Dots Should We Try to Connect?

Here comes a matter of great complexity. It usually happens that hypotheses we entertain
are generated from observations we have made involving potential evidential dots. On
limited occasions, we can generate a hypothesis from a single evidential dot. For example,
in a criminal investigation finding a fingerprint will suggest a possible suspect in the case.
But in most cases, it takes consideration of combinations of evidential dots in order to
generate plausible and useful hypotheses, as illustrated in the following example based on
accounts given in Time Magazine and the Washington Post. From European sources came
word that terrorists of Middle Eastern origin would make new attempts to destroy the
World Trade Center, this time using airliners. Many threats are received every day, most of

13:50:24,
.003
2.1. How Easy Is It to Connect the Dots? 51

which come to nothing. However, from several civilian flying schools in the United States
came word (to the FBI) that persons from the Middle East were taking flying lessons,
paying for them in cash, and wanting only to learn how to steer and navigate heavy aircraft
but not how to make takeoffs and landings in these aircraft. By itself, this information,
though admittedly strange, may not have seemed very important. But, taken together,
these two items of information might have caused even Inspector Lestrade (the rather
incompetent police investigator in Sherlock Holmes stories) to generate the hypothesis
that there would be attacks on the World Trade Center using hijacked airliners. The
hijackers would not need to learn how to make takeoffs; the aircrafts’ regular pilots would
do this. There would be no need for the hijackers to know how to land aircraft, since no
landings were intended, only crashes into the World Trade Center and the Pentagon. Why
were these two crucial items of information not considered together? The answer seems to
be that they were not shared among relevant agencies. Information not shared cannot
be considered jointly, with the result that their joint inferential impact could never have
been assessed. For all time, this may become the best (or worst) example of failure to
consider evidence items together. Even Sherlock Holmes would perhaps not have inferred
what happened on September 11, 2001, if he had not been given these two items of
information together.
The problem, however, is that here we encounter a combinatorial explosion, since the
number of possible combinations of two or more evidential dots is exponentially related to
the number of evidential dots we are considering. Suppose we consider having some
number N of evidential dots. We ask the question: How many combinations C of two or
more evidential dots are there when we have N evidential dots? The answer is given by the
following expression: C = 2N – (N + 1). This expression by itself does not reveal how quickly
this combinatorial explosion takes place. Here are a few examples showing how quickly
C mounts up with increases in N:

 For N = 10, C = 1013

 For N = 25, C = 33,554,406
 For N = 50, C = 1.13  1015
 For N = 100, C = 1.27  1030

There are several important messages in this combinatorial analysis for evidence-based
reasoning. The first concerns the size of N, the number of potential evidential dots that
might be connected. Given the array of sensing devices and human observers available,
the number N of potential evidential dots is as large as you wish to make it. In most
analyses, N would certainly be greater than one hundred and would increase as time
passes. Remember that we live in a nonstationary world in which things change and we
find out about new things all the time. So, in most cases, even if we had access to the
world’s fastest computer, we could not possibly examine all possible evidential dot
combinations, even when N is quite small.
Second, trying to examine all possible evidential dot combinations would be the act of
looking through everything with the hope of finding something. This would be a silly thing to
do, even if it were possible. The reason, of course, is that most of the dot combinations
would tell us nothing at all. What we are looking for are combinations of evidential dots that
interact or are dependent in ways that suggest new hypotheses or possible conclusions. If we
would examine these dots separately or independently, we would not perceive these new
possibilities. A tragic real-life example is what happened on September 11, 2001.

13:50:24,
.003
 52 Chapter 2. Evidence-based Reasoning: Connecting the Dots

Figure 2.3 is an abstract example involving four numbered evidential dots. The
numbers might indicate the order in which we obtained them. In part (a) of the figure,
we show an instance where these four dots have been examined separately or independ-
ently, in which case they tell us nothing interesting. Then someone notices that, taken
together, these four dots combine to suggest a new hypothesis HK that no one has thought
about before, as shown in part (b) of the figure. What we have here is a case of evidential
synergism in which two or more evidence items mean something quite different when they
are examined jointly than they would mean if examined separately or independently. Here
we come to one of the most interesting and crucial evidence subtleties or complexities that
have, quite frankly, led to intelligence failures in the past: failure to identify and exploit
evidential synergisms. We will address this matter in other problems we mention concern-
ing connecting the dots.
It might be said that the act of looking through everything in the hope of finding
something is the equivalent of giving yourself a prefrontal lobotomy, meaning that you
are ignoring any imaginative capability you naturally have concerning which evidential
dot combinations to look for in your analytic problem area. What is absolutely crucial
in selecting dot combinations to examine is an analyst’s experience and imaginative
reasoning capabilities. What we should like to have is a conceptual “magnet” that we
could direct at a base of evidential dots that would “attract” interesting and important
dot combinations.

2.1.5 How to Connect Evidential Dots to Hypotheses?

As will be discussed in Section 4.4, all evidence has three major credentials or properties:
relevance, believability or credibility, and inferential force or weight. No evidence ever
comes to us with these three credentials already attached; they must be established by
defensible and persuasive arguments linking the evidence to the hypotheses we are
considering. As we will see, relevance answers the question: So what? How is this datum
or information item linked to something we are trying to prove or disprove? If such
relevance linkage cannot be established, this datum is irrelevant or useless. As discussed
previously, believability answers the question: Can we believe what this evidence is telling
us? The force or weight credential asks: How strong is this evidence in favoring or
disfavoring the hypothesis? This is where probability enters our picture, since, for very
good reasons, the force or weight of evidence is always graded in probabilistic terms.
A relevance argument is precisely where the idea dots become so important. Consider-
ing an item of information, an analyst must imagine how this item could be linked to some
hypothesis being considered before it could become an item of evidence. These idea dots

Examined separately, these

(a)
dots tell us nothing.
17 103 246 300
300

Examined jointly, these dots

(b) 103
246 suggest a new hypothesis Hk.

17

Figure 2.3. Evidential synergism.

13:50:24,
.003
2.1. How Easy Is It to Connect the Dots? 53

forming this linkage come in the form of propositions or statements indicating possible
sources of doubt or uncertainty in the imagined linkage between the item of information
and hypotheses being considered. For a simple example, look again at Figure 2.2 (p. 49),
where we show a connection between evidence E*i and hypothesis H. An analyst has an
item of information from source concerning the occurrence of event E that sounds very
interesting. This analyst attempts to show how event E, if it did occur, would be relevant in
an inference about whether hypothesis H is true or not. So the analyst forms the following
chain of reasoning involving idea dots. The analyst says, “If event E were true, this would
allow us to infer that event F might be true, and if F were true, this would allow us to infer
that event G might be true. Finally, if event G were true, this would make hypothesis
H more probable.” If this chain of reasoning is defensible, the analyst has established the
relevance of evidence E*i on hypothesis H.
In forming this argument, the analyst wisely begins with the believability foundation for
this whole argument: Did event E really occur just because source says it did? Also notice
in Figure 2.2 that we have indicated the uncertainty associated with each idea dot in this
argument. For example, the analyst only infers from E that F might have occurred and so
we note that we must consider F and ‘not F’ as possibilities. The same is true for the other
idea dot G and for the hypothesis dot H.
There are several important things to note about relevance arguments; the first concerns
their defense. Suppose the argument in Figure 2.2 was constructed by analyst . shows this
argument to analyst , who can have an assortment of quibbles about this argument. Suppose
says, “You cannot infer F directly from E; you need another step here involving event
K. From E you can infer that K occurred, and then if K occurred, then you can infer F.” Now
comes analyst , who also listens to ’s argument. says, “I think your whole argument is
wrong. I see a different reasoning route from E to hypothesis H. From E we can infer event R,
and from R we can infer event S, and from S we can infer T, which will show that hypothesis
H is less probable.” Whether or not there is any final agreement about the relevance of
evidence E*i, analyst has performed a real service by making the argument openly and
available for discourse and criticism by colleagues. There are several important messages here.
First, there is no such thing as a uniquely correct argument from evidence to hypoth-
eses. What we all try to avoid are disconnects or non sequiturs in the arguments we
construct. But even when we have an argument that has no disconnects, someone may be
able to come up with a better argument. Second, we have considered only the simplest
possible situation in which we used just a single item of potential evidence. But intelli-
gence analysis and other evidence-based reasoning tasks are based on masses of evidence
of many different kinds and from an array of different sources. In this case, we are obliged
to consider multiple lines of argument that can be connected in different ways. It is
customary to call these complex arguments inference networks.
From years of experience teaching law students to construct defensible and persuasive
arguments from evidence, we have found that most of them often experience difficulty in
constructing arguments from single items of evidence; they quickly become overwhelmed
when they are confronted with argument construction involving masses of evidence.
But they gain much assistance in such tasks by learning about argument construction
methods devised nearly a hundred years ago by a world-class evidence scholar named
John H. Wigmore (1863–1943). Wigmore (1913, 1937) was the very first person to study
carefully what today we call inference networks. We will encounter Wigmore’s work in
several places in our discussions, and you will see that Disciple-EBR employs elements of
Wigmore’s methods of argument construction.

13:50:24,
.003
 54 Chapter 2. Evidence-based Reasoning: Connecting the Dots

There is also a message here for critics, such as newswriters and the talking heads on
television. These critics always have an advantage never available to practicing intelligence
analysts. Namely, they know how things turned out or what actually happened in some
previously investigated matter. In the absence of clairvoyance, analysts studying a problem
will never know for sure, or be able to predict with absolute certainty, what will happen in
the future. A natural question to ask these critics is, “What arguments would you have
constructed if all you knew was what the analysts had when they made their assessments?”
This would be a very difficult question for them to answer fairly, even if they were given
access to the classified evidence the analysts may have known at the time.

2.1.6 What Do Our Dot Connections Mean?

The previous topic concerns efforts designed to establish the defensibility of complex
arguments. But what do these arguments mean to persons for whom these arguments
are being constructed? This question raises matters concerning how persuasive are our
arguments when they are taken all together. Our view is that the persuasiveness of an
argument structure depends, in large part, upon the nature of the probabilities we assess
and combine in making our arguments and in stating our major conclusions.
Here we consider the direction and force of our arguments based on the combined
evidence we have considered. Direction refers to the hypothesis we believe our evidence
favors most. Force means how strongly we believe the evidence favors this hypothesis over
alternative hypotheses we have considered. There are two uncontroversial statements we
can make about the force or weight of evidence. The first is that the force or weight of
evidence has vector-like properties. What this means is that evidence points us in the
direction of certain hypotheses or possible conclusions with varying degrees of strength.
The second is that the force or weight of evidence is always graded in probabilistic terms,
indicating our uncertainties or doubts about what the evidence means in terms of its
inferential direction and force. But beyond these two statements, controversies begin
to arise.
Before we consider assorted controversies, it is advisable to consider where our uncer-
tainties or doubts come from in the conclusions we reach from evidence. Have a look once
again at Figure 2.2, involving a simple example based on a single item of evidence. Our
evidence here was E*i, from source , saying that event E occurred. We ask the question:
How strongly does this evidence E*i favor hypothesis H over not H? As we discussed, this
argument was indicated by what we termed idea dots, each one indicating what the analyst
constructing this argument believed to be sources of doubt or uncertainty associated with
the argument from the evidence to the hypothesis. As you see, there are two major origins
of uncertainty: those associated with the believability of source , and those associated
with links in the analyst’s relevance argument. So, the force of evidence E*i on hypotheses
H and not H depends on how much uncertainty exists in this entire argument involving
each one of its believability and relevance links. The interesting message here is that
the evidence force or weight credential depends on its other two credentials: believability
and relevance.
In the simple example just discussed, there are four major origins of uncertainty, one
associated with believability and three associated with relevance. But this is the easiest
possible situation since it involves only one item of evidence. Think of how many sources
of uncertainty there might be when we have a mass of evidence together with multiple

13:50:24,
.003
2.1. How Easy Is It to Connect the Dots? 55

complex and possibly interrelated arguments. The mind boggles at the enormity of the
task of assessing the force or weight of a mass of evidence commonly encountered in
intelligence analysis when we have some untold numbers of sources of believability and
relevance uncertainties to assess and combine (Schum, 1987). We are certain that critics of
intelligence analysts have never considered how many evidential and idea dots there
would be to connect.
So, the question remains: How do we assess and combine the assorted uncertainties in
complex arguments in intelligence analysis and in any other context in which we have the
task of trying to make sense out of masses of evidence? Here is where controversies arise.
The problem is that there are several quite different views among probabilists about what
the force or weight of evidence means and how it should be assessed and combined across
evidence in either simple or complex arguments: Bayesian, Belief Functions, Baconian,
and Fuzzy (Schum, 1994[2001a]). Each of these views has something interesting to say,
but no one view says it all, as discussed in Section 1.3.6.
Later in this book, we will discuss how Disciple-EBR allows you to assess and combine
probabilistic judgments in situations in which many such judgments are required. There is
further difficulty as far as judgments of the weight or force of evidence are concerned.
Analysts, or teams of analysts, may agree about the construction of an argument but
disagree, often vigorously, about the extent and direction of the force or weight this
argument reveals. There may be strong disagreements about the believability of sources
of evidence or about the strength of relevance linkages. These disagreements can be
resolved only when arguments are made carefully and are openly revealed so that they
can be tested by colleagues. A major mission of Disciple-EBR is to allow you to construct
arguments carefully and critically and encourage you to share them with colleagues so that
they can be critically examined.
There is one final matter of interest in making sense out of masses of evidence and
complex arguments. Careful and detailed argument construction might seem a very
laborious task, no matter how necessary it is. Now consider the task of revealing the
conclusions resulting from an analysis to some policy-making “customer” who has deci-
sions to make that rest in no small part on the results of an intelligence analysis. What this
customer will probably not wish to see is a detailed inference network analysis that
displays all of the dots that have been connected and the uncertainties that have been
assessed and combined in the process. A fair guess is that this customer will wish to have a
narrative account or a story about what the analysis predicts or explains. In some cases,
customers will require only short and not extensive narratives. This person may say, “Just
tell me the conclusions you have reached and briefly why you have reached them.” So the
question may be asked: Why go to all the trouble to construct defensible and persuasive
arguments when our customers may not wish to see their details?
There is a very good answer to the question just raised. Your narrative account of an
analysis must be appropriately anchored on the evidence you have. What you wish to be able
to tell is a story that you believe contains some truth; that is, it is not just a good story. The
virtue of careful and critical argument construction is that it will allow you to anchor your
narrative not only on your imagination, but also on the care you have taken to subject your
analysis to critical examination. There is no telling what questions you might be asked about
your analysis. Rigor in constructing your arguments from your evidence is the best protec-
tion you have in dealing with customers and other critics who might have entirely different
views regarding the conclusions you have reached. Disciple-EBR is designed to allow you
and others to evaluate critically the arguments you have constructed.

13:50:24,
.003
 56 Chapter 2. Evidence-based Reasoning: Connecting the Dots

2.2 SAMPLE EVIDENCE-BASED REASONING TASK:

INTELLIGENCE ANALYSIS

The purpose of intelligence analysis is to answer questions arising in the decision-making

process by analyzing evidence about the world, such as, “Does Al Qaeda have nuclear
weapons?” or, “Will the United States be the world leader in nonconventional energy
sources within the next decade?” This is done by determining the probabilities of alterna-
tive (hypothesized) answers, based on evidence, and by selecting the most likely answer.
As discussed in Section 1.4.2, intelligence analysis, like other evidence-based reasoning
tasks, can be viewed as ceaseless discovery of evidence, hypotheses, and arguments in
a nonstationary world, involving collaborative processes of evidence in search of hypoth-
eses, hypotheses in search of evidence, and evidentiary testing of hypotheses (see
Figure 1.9, p. 27). Since these processes are generally very complex and involve both
imaginative and critical reasoning, they can be best approached through the synergistic
integration of the analyst’s imaginative reasoning and the computer’s knowledge-based
critical reasoning, as was discussed in Section 1.5.2, and will be illustrated with the use of
the Disciple-EBR cognitive assistant.
The following sections illustrate this systematic approach to intelligence analysis by
using a specific example of anticipatory analysis. At the same time, this example intro-
duces the main concepts related to evidence and inference.

2.2.1 Evidence in Search of Hypotheses

Consider that you are an intelligence analyst and you read in today’s Washington Post an
article that concerns how safely radioactive materials are stored in this general area.
Willard, the investigative reporter and author of this piece, begins by noting how the
storage of nuclear and radioactive materials is so frequently haphazard in other countries
and wonders how carefully these materials are guarded here in the United States, particu-
larly in this general area. In the process of his investigations, the reporter notes his
discovery that a canister containing cesium-137 has gone missing from the XYZ Company
in Maryland, just three days ago. The XYZ Company manufactures devices for sterilizing
medical equipment and uses cesium-137 in these devices along with other radioactive
materials. This piece arouses your curiosity because of your concern about terrorists
planting dirty bombs in our cities. The question is: What hypotheses would explain this
observation? You experience a flash of insight that a dirty bomb may be set off in the
Washington, D.C., area (see Figure 2.4).
However, no matter how imaginative or important this hypothesis is, no one will take it
seriously unless you are able to justify it. So you develop the chain of abductive inferences
shown in Table 2.1 and in Figure 2.5.
The chain of inferences from Table 2.1 and Figure 2.5 shows clearly the possibility that
a dirty bomb will be set off in the Washington, D.C., area. Can you then conclude that this
will actually happen? No, because there are many other hypotheses that may explain this
evidence, as shown in Figure 2.6 and discussed in the following paragraphs.
Just because there is evidence that the cesium-137 canister is missing does not mean
that it is indeed missing. At issue here is the believability of Willard, the source of this
information. What if this Willard is mistaken or deceptive? Thus an alternative hypothesis
is that the cesium-137 canister is not missing.

13:50:24,
.003
2.2. Sample Evidence-based Reasoning Task 57

Table 2.1 Abductive Reasoning Steps Justifying a Hypothesis

There is evidence that the cesium-137 canister is missing (E*).

Therefore, it is possible that the cesium-137 canister is indeed missing (H1).
Therefore, it is possible that the cesium-137 canister was stolen (H2).
Therefore, it is possible that the cesium-137 canister was stolen
by someone associated with a terrorist organization (H3).
Therefore, it is possible that the terrorist organization will use
the cesium-137 canister to construct a dirty bomb (H4).
Therefore, it is possible that the dirty bomb will be set off
in the Washington, D.C., area (H5).

H: A dirty bomb
will be set off in
the Washington,
D.C., area
Insight

E*: Article on
cesium-137
canister
missing

What hypotheses would

explain this observation?
Figure 2.4. Hypothesis generation through imaginative reasoning.

But let us assume that the cesium-137 canister is indeed missing. Then it is possible
that it was stolen. But it is also possible that it was misplaced, or maybe it was used in a
project at the XYZ Company without being checked out from the warehouse?
However, let us assume that the cesium-137 canister was indeed stolen. It is then
possible that it might have been stolen by a terrorist organization, but it is also possible
that it might have been stolen by a competitor or by an employee, and so on.
This is the process of evidence in search of hypotheses, shown in the left side of Figure 1.9
(p. 27). You cannot conclude that a dirty bomb will be set off in the Washington, D.C., area
(i.e., hypothesis H5) until you consider all the alternative hypotheses and show that those
on the chain from E* to H5 are actually more likely than their alternatives. But to analyze all
these alternative hypotheses and make such an assessment, you need additional evidence.
How can you get it? As represented in the middle of Figure 1.9, you put each hypothesis to
work to guide you in the collection of additional evidence. This process is discussed in the
next section.

13:50:24,
.003
 58 Chapter 2. Evidence-based Reasoning: Connecting the Dots

H5: A dirty bomb

will be set off in
the Washington,
D.C., area

H4: build dirty

bomb

H3: stolen

possibly Q
by terrorist

Abduction
organization

H2: stolen
P

H1: missing

E*: Article on
cesium-137
canister
missing

What hypotheses would explain this observation?

Evidence in search of hypotheses
Figure 2.5. Justification of the generated hypothesis.

2.2.2 Hypotheses in Search of Evidence

Let us first consider the hypothesis “H1: missing” from the bottom of Figure 2.6, shown
as “H1: cesium-137 canister is missing from warehouse,” in the top-left of Figure 2.7. The
question is: Assuming that this hypothesis is true, what other things should be observable?
Which are the necessary and sufficient conditions for an object to be missing from a
warehouse? It was in the warehouse, it is no longer there, and no one has checked it out.
This suggests the decomposition of the hypothesis H1 into three simpler hypotheses, as
shown in the left part of Figure 2.7. This clearly indicates that you should look for evidence
that indeed the cesium-137 canister was in the warehouse, that it is no longer there, and
that no one has checked it out. That is, by putting hypothesis H1 to work, you were guided
to perform the collection tasks from Table 2.2, represented in Figure 2.7 by the gray circles.
Guided by the evidence collection tasks in Table 2.2, you contact Ralph, the super-
visor of the XYZ warehouse, who provides the information shown in Table 2.3 and in
Figure 2.7.
When you are given testimonial information, or descriptions of tangible items, the
information might contain very many details, dots, or trifles. Some of the details might be
interesting and relevant evidence, and others not. What you always have to do is to parse
the information to extract the information that you believe is relevant in the inference task

13:50:24,
.003
2.2. Sample Evidence-based Reasoning Task 59

H5: A dirty bomb will H’5: A dirty bomb

be set off in the will be set off in
Washington, D.C., area the New York area

H4: build dirty H’4: build unshielded

bomb radiation sources

H3: stolen
H’3: stolen by H”3: stolen by
possibly Q

by terrorist
competitor employee
Abduction

organization

H2: stolen H’2: misplaced H”2: used in project

P

H1: missing H’1: not missing

E*: Article on
cesium-137
canister
missing

What hypotheses would explain this observation?

Evidence in search of hypotheses
Figure 2.6. Competing hypotheses explaining an item of evidence.

Hypothesis-driven collection Multi-INT fusion

H1: cesium-137 canister is
H1 very likely
missing from warehouse

Deduction Induction
H13: was not H11: H12: H13:
H11: H12:
checked out almost very very
was in is not in
from certain likely likely
warehouse warehouse
warehouse
inferential force
very likely
E*
relevance believability
Ralph, the supervisor of the warehouse, reports that certain very likely
the cesium-137 canister is registered as being in the warehouse
and that no one at the XYZ Company had checked it out,
but it is not located anywhere in the hazardous materials locker.
He also indicates that the lock on the hazardous materials locker
appears to have been forced.

Figure 2.7. Hypothesis-driven evidence collection and hypothesis testing.

13:50:24,
.003
 60 Chapter 2. Evidence-based Reasoning: Connecting the Dots

Table 2.2 Evidence Collection Tasks Obtained from the Analysis in Figure 2.7

Collection Task1: Look for evidence that the cesium-137 canister was in the XYZ warehouse before
being reported as missing.
Collection Task2: Look for evidence that the cesium-137 canister is no longer in the XYZ
warehouse.
Collection Task3: Look for evidence that the cesium-137 canister was not checked out from the
XYZ warehouse.

Table 2.3 Information Obtained through the Collection Tasks in Table 2.2

INFO-002-Ralph: Ralph, the supervisor of the warehouse, reports that the cesium-137 canister is
registered as being in the warehouse and that no one at the XYZ Company had checked it out, but it is
not located anywhere in the hazardous materials locker. He also indicates that the lock on the
hazardous materials locker appears to have been forced.

Table 2.4 Dots or Items of Evidence Obtained from Willard and Ralph

E001-Willard: Willard’s report in the Washington Post that a canister containing cesium-137 was
missing from the XYZ warehouse in Baltimore, MD.
E002-Ralph: Ralph’s testimony that the cesium-137 canister is registered as being in the XYZ
warehouse.
E003-Ralph: Ralph’s testimony that no one at the XYZ Company had checked out the cesium-137
canister.
E004-Ralph: Ralph’s testimony that the canister is not located anywhere in the hazardous materials
locker.
E005-Ralph: Ralph’s testimony that the lock on the hazardous materials locker appears to have
been forced.

at hand. Consider, for example, the information provided by Willard in his Washington
Post article. You parse it to extract the relevant information represented as E001-Willard in
Table 2.4. Similarly, Ralph’s testimony from Table 2.3 provides you with several dots or
items of evidence that are relevant to assessing the hypotheses from Figure 2.7. These
items of evidence are represented in Table 2.4.
This is the process of hypotheses in search of evidence that guides you in collecting new
evidence. The next step now is to assess the probability of hypothesis H1 based on the
collected evidence, as represented in the right-hand side of Figure 1.9 (p. 27), and
discussed in the next section.

2.2.3 Evidentiary Testing of Hypotheses

Having identified evidence relevant to the leaf hypotheses in Figure 2.7, the next step is to
use it in order to assess these hypotheses. The assessments of the hypotheses will be done

13:50:24,
.003
2.2. Sample Evidence-based Reasoning Task 61

by using probabilities that are expressed in words rather than in numbers. In particular,
we will use the ordered symbolic probability scale from Table 2.5. This is based on a
combination of ideas from the Baconian and Fuzzy probability systems (Schum, 1994
[2001a], pp. 243–269). As in the Baconian system, “no support” for a hypothesis means that
we have no basis to consider that the hypothesis might be true. However, we may later
find evidence that may make us believe that the hypothesis is “very likely,” for instance.
To assess the hypotheses, you first need to attach each item of evidence to the
hypothesis to which it is relevant, as shown in the right side of Figure 2.7. Then you need
to establish the relevance and the believability of each item of evidence, which will result in
the inferential force of that item of evidence on the corresponding hypothesis, as illustrated
in the right side of Figure 2.7 and explained in the following.
So let us consider the hypothesis “H13: cesium-137 canister was not checked out from
the warehouse” and the item of evidence “E003-Ralph: Ralph’s testimony that no one at
the XYZ Company had checked out the cesium-137 canister.”
Relevance answers the question: So what? How does E003-Ralph bear on the hypothesis
H13 that you are trying to prove or disprove? If you believe what E003-Ralph is telling us,
then H13 is “certain.”
Believability answers the question: To what extent can you believe what E003-Ralph is
telling you? Let us assume this to be “very likely.”
Inferential force or weight answers the question: How strong is E003-Ralph in favoring
H13? Obviously, an item of evidence that is not relevant to the considered hypothesis will
have no inferential force on it and will not convince you that the hypothesis is true. An
item of evidence that is not believable will have no inferential force either. Only an item
of evidence that is both very relevant and very believable will make you believe that
the hypothesis is true. In general, the inferential force of an item of evidence (such as
E003-Ralph) on a hypothesis (such as H13) is the minimum of its relevance and its
believability. You can therefore conclude that, based on E003-Ralph, the probability of
the hypothesis H13 is “very likely” (i.e., the minimum of “certain” and “very likely”), as
shown in Figure 2.7.
Notice in Figure 2.7 that there are two items of evidence that are relevant to the
hypothesis H12. In this case, the probability of H12 is the result of the combined (max-
imum) inferential force of these two items of evidence.
Once you have the assessments of the hypotheses H11, H12, and H13, the assessment of
the hypothesis H1 is obtained as their minimum, because these three subhypotheses are
necessary and sufficient conditions for H1. Therefore, all need to be true in order for H1 to
be true, and H1 is as weak as its weakest component.
Thus, as shown at the top-right side of Figure 2.7, you conclude that it is “very likely”
that the cesium-137 canister is missing from the warehouse.
Notice that this is a process of multi-intelligence fusion since, in general, the assessment
of a hypothesis involves fusing different types of evidence.
Figure 2.8 summarizes the preceding analysis, which is an illustration of the general
framework from Figure 1.9 (p. 27).

Table 2.5 Ordered Symbolic Probability Scale

no support < likely < very likely < almost certain < certain

13:50:24,
.003
 62 Chapter 2. Evidence-based Reasoning: Connecting the Dots

H5: A dirty bomb

will be set off in Hypothesis-driven collection Multi-INT fusion
the Washington,
H1: cesium-137
DC, area
canister is missing H1 very likely
H4: build dirty from warehouse
bomb
Abduction Deduction Induction
H13: was not H11: H12: H13:
H3: stolen H’3: stolen H”3: stolen H11: H12:
checked out almost very very
by terrorist by by was in is not in
from certain likely likely
organization competitor employee warehouse warehouse
warehouse

H’2:mis- H”2: used

H2: stolen E
placed in project

H1: missing Ralph, the supervisor of the warehouse, reports that

very likely the cesium-137 canister is registered as being in the warehouse
and that no one at the XYZ Company had checked it out,
E*: Article on but it is not located anywhere in the hazardous materials locker.
cesi
cesium-137 He also indicates that the lock on the hazardous materials locker
canister appears to have been forced.
missing
W
What hypotheses would Assuming that this hypothesis is true, What is the probability of the hypothesis
explain this observation? what other things should be observable? based on the available evidence?
Evidence in search Hypotheses in search Evidentiary testing
of hypotheses of evidence of hypotheses
P possibly Q P necessarily Q P probably Q

Figure 2.8. An illustration of the general framework from Figure 1.9 (p. 27).

Now that you have concluded “H1: missing,” you repeat this process for the upper
hypotheses (i.e., H2: stolen, H’2: misplaced, and H”2: used in project), as will be discussed
in the next section.

2.2.4 Completing the Analysis

Let us first consider the hypothesis “H2: stolen.” You need to put this hypothesis to work to
guide you in collecting relevant evidence for its analysis. During your investigation of the
security camera of the warehouse, you discover a video segment showing a person loading
a container into a U-Haul panel truck (E007-SecurityCamera). This new item of evidence,
together with Ralph’s testimony that the lock on the hazardous materials locker appears
to have been forced (E005-Ralph in Table 2.4), suggests the following scenario of how
the cesium-137 canister might have been stolen (see Figure 2.9): The truck entered the
company, the canister was stolen from the locker, the canister was loaded into the truck, and
the truck left with the canister.
Such scenarios have enormous heuristic value in advancing the investigation because
they consist of mixtures of what is taken to be factual and what is conjectural. Conjecture is
necessary in order to fill in natural gaps left by the absence of existing evidence. Each such
conjecture opens up new avenues of investigation, and the discovery of additional evi-
dence, if the scenario turns out to be true. For instance, the first hypothesized action from
the scenario (“Truck entered company”) leads you to check the record of the security
guard, which shows that a panel truck bearing Maryland license plate number MDC-578
was in the XYZ parking area the day before the discovery that the cesium-137 canister was
missing (E008-GuardReport).

13:50:24,
.003
2.2. Sample Evidence-based Reasoning Task 63

H2: cesium-137 very

canister was stolen likely

min
Scenario: The truck entered the company, the canister was stolen from the locker,
the canister was loaded into the truck, and the truck left with the canister.

H: Truck H: Truck left almost

almost
entered with cesium- certain
certain
company 137 canister

H: cesium-137 H: cesium-137
very very
canister stolen canister loaded
likely likely
from locker into truck
min

H1: cesium-137 very H: Warehouse almost

canister missing likely locker was certain
from warehouse forced
max

Figure 2.9. Another example of hypothesis-driven evidence collection and hypothesis testing.

The second hypothesized action in the scenario (i.e., “cesium-137 canister stolen from
locker”) is further decomposed into two hypotheses. The first one was already analyzed:
“It is very likely that the cesium-137 canister is missing from the warehouse.” The second
subhypothesis (“Warehouse locker was forced”) is supported both by Ralph’s testimony
(i.e., E005-Ralph in Table 2.4) and by the professional locksmith Clyde who was asked to
examine it (E006-Clyde: Professional locksmith Clyde testimony that the lock has been
forced, but it was a clumsy job).
After continuing the process for the remaining hypothesized actions in the scenario and
fusing all the discovered evidence, you and Disciple-EBR conclude that it is “very likely”
that the cesium-137 canister was stolen.
You repeat the same process for the other two competing hypotheses, “H’2: misplaced,”
and “H”2: used in project.” However, you find no evidence that the cesium-137 canister
might have been misplaced. Moreover, you find disfavoring evidence for the second com-
peting hypothesis: Grace, the Vice President for Operations at XYZ, tells us that no one at the
XYZ Company had checked out the canister for work on any project (E014-Grace).
Thus you conclude that the canister-137 was stolen and you continue the analysis with
investigating the next level up of competing hypotheses: “H3: stolen by terrorist organiza-
tion”; “H’3: stolen by competitor”; and “H”3: stolen by employee.” Of course, at any point,
the discovery of new evidence may lead you to refine your hypotheses, define new
hypotheses, or eliminate existing hypotheses.
This example is not as simple as it may be inferred from this presentation. It is the
methodology that guides you and makes it look simple. Many things can and will indeed
go wrong. But the computational theory of evidence-based reasoning and Disciple-EBR

13:50:24,
.003
 64 Chapter 2. Evidence-based Reasoning: Connecting the Dots

provide you the means to deal with them. Based on evidence, you come up with some
hypotheses, but then you cannot find evidence to support any of them. So you need to
come up with other hypotheses, and you should always consider alternative hypotheses.
The deduction-based decomposition approach guides you on how to look for evidence,
but your knowledge and imagination also play a crucial role. As illustrated here, you
imagined a scenario where the cesium-137 canister was stolen with a truck. But let us now
assume that you did not find supporting evidence for this scenario. Should you conclude
that the cesium-137 canister was not stolen? No, because this was just one scenario. If you
can prove it, you have an assessment of your hypothesis. However, if you cannot prove it,
there still may be another scenario on how the cesium-137 canister might have been
stolen. Maybe the cesium-137 canister was stolen by someone working at the XYZ
Company. Maybe it was stolen by Ralph, the administrator of the warehouse. The import-
ant thing is that each such scenario opens a new line of investigation and a new way to
prove the hypothesis.
Having established that the cesium-137 canister was stolen, you would further like to
determine by whom and for what purpose. If it is for building and setting off a dirty bomb,
you would like to know who will do this; where exactly in the Washington, D.C., area will
the bomb be set off; precisely when this action will happen; what form of dirty bomb will
be used; and how powerful it will be. These are very hard questions that the computational
theory of evidence-based reasoning (as well as its current implementation in Disciple-
EBR) will help you answer.
One major challenge in performing such an analysis is the development of argumenta-
tion structures. An advantage of using an advanced tool, such as Disciple-EBR, is that it
can learn reasoning rules from the user to greatly facilitate and improve the analysis of
similar hypotheses, as will be shown in the next chapters of this book.
In conclusion, the computational theory of evidence-based reasoning presented in
this volume, as well as its current implementation in Disciple-EBR, provides a framework
for integrating the art and science of evidence-based reasoning, to cope with its aston-
ishing complexity.
More details about intelligence analysis are presented in our book Intelligence Analy-
sis as Discovery of Evidence, Hypotheses, and Arguments: Connecting the Dots (Tecuci
et al., 2016). Other examples of applications of evidence-based reasoning are presented
in the next section.

2.3 OTHER EVIDENCE-BASED REASONING TASKS

2.3.1 Cyber Insider Threat Discovery and Analysis

Cyber insider threats are persons who operate inside an organization and use legitimate
access and normal tactics to accomplish abnormal and malicious cyber-missions, such as
to perform data reconnaissance, collection, and exfiltration, or to create vulnerabilities for
cyber-attacks by outsiders. These persons represent a major national security concern, a
critical concern for businesses that need to protect their intellectual property, as well as a
huge privacy concern.
In the following, we will present a hypothetical agent for cyber-insider threat discovery
and analysis. This example was developed with the assistance of Professor Angelos Stavrou

13:50:24,
.003
2.3. Other Evidence-based Reasoning Tasks 65

from George Mason University. We will assume that we have a set of monitoring agents
that perform persistent surveillance of a computer network and host systems. They
include login monitors, file system monitors, internal network monitors, port monitors,
outside network monitors, and others. These monitoring agents are constantly looking for
indicators and warnings of insider missions.
Let us further assume that the evidence collection agents have detected a record inside
network logs involving an instance of denied access from the device with the Internet
Protocol address IP1 to the device with the address IP2, at time T. This is evidence E* at the
bottom of Figure 2.10. While this denied service access might be a normal event of an
accidental access to a shared resource generated by legitimate browsing on a local host, it
can also be an indication of an improper attempt to access a shared network resource.
Therefore, the question is: What insider missions might explain this observation?
By means of abductive reasoning, which shows that something is possibly true, the
analysis agent may formulate the chain of explanatory hypotheses from the left side of
Figure 2.10:
It is possible that the observed denied access from IP1 to IP2 is part of a sequence of
attempted network service accesses from IP1 to other IPs (hypothesis H11). It is further
possible that this is part of a network scan for files (hypothesis H21). It is possible that this
network scan is in fact a malicious attempt to discover network shared files (hypothesis
H31), part of malicious covert reconnaissance (hypothesis H41), which may itself be part
of a covert reconnaissance, collection, and exfiltration mission (hypothesis H51).
As one can notice, these hypotheses are very vague at this point. Moreover, for each of
these hypotheses there are alternative hypotheses, as shown in the right-hand side of
Figure 2.10. For example, the denied access may be part of a single isolated attempt
(hypothesis H12). However, even in the case where we have established that the denied
service access is part of a sequence of accesses (hypothesis H11), it is still possible that this

H51: Covert reconnaissance, H52: Covert reconnaissance

collection, and exfiltration for remote vulnerabilities

H41: Covert H42: Use for covert exfiltration

reconnaissance using stepping stone

H31: Non-account H32: Account owner

owner scan for files scan for files

H21: Network H22: Policy/user

scan for files access changed

H11: Sequence H12: Single isolated

of accesses attempt

E*: Log record of

denied service access
Figure 2.10. Evidence in search of insider missions (abductive reasoning).

13:50:24,
.003
66 Chapter 2. Evidence-based Reasoning: Connecting the Dots

sequence of accesses is due to recent policy changes that affected the user’s access to
specific services or objects (hypothesis H22).
What the agent needs to do is to test each of these alternative hypotheses, starting
from bottom up, to make sure that, if an insider mission is actually being performed, it is
promptly detected. Each of the bottom-level alternative hypotheses (i.e., H11 and H12)
is put to work to guide the collection of relevant evidence (see Figure 2.11). The
discovered evidence may lead to the refinement of the hypotheses, including the
possible formulation of new hypotheses, and these refined hypotheses may lead to
new evidence. Next, the discovered evidence is used to assess the probabilities of the
bottom-level alternative hypotheses.
Assuming that the most likely hypothesis was determined to be H11, this process
continues with the next level up of alternative hypotheses (i.e., H21 and H22), using them
to collect evidence, and assessing which of them is most likely, as illustrated in Figure 2.12.
Now, since H21 was assessed as being “very likely,” it is possible that H31 is true. But it is
also possible that H32 is true. The right-hand side of Figure 2.13 illustrates the process of
using the hypothesis H31 in order to guide the collection of evidence to test it:

H11: Sequence H11: very

of accesses likely

H11: Sequence H12: Single

of accesses isolated attempt H12: Single H12: no
isolated attempt support
E*: Log record
of denied
service access
Figure 2.11. Evidence collection and assessment of the bottom-level hypotheses in Figure 2.10.

H21: Network scan for H21: very

shared files from IP1, likely
between T1 and T2

Search for logs of H:

network scan for H:
very likely
H21: Network shared files, from IP1, likely
scan for shared between T1 and T2
files from IP1,
between H22: Policy/user
T1 and T2 access changed Search
for …
Search for
H11: Sequence of user activity Search for host
accesses to network between T1 machine logs
services for several and T2 for with file sharing
systems from IP1, the user(s) request from IP1
between T1 and T2 assigned to IP1
(very likely)

E*: Log record

of denied … H22: Policy/user H22: no
access changed support
service access

Figure 2.12. Evidence collection and assessment for the next level up of alternative hypotheses.

13:50:24,
.003
2.3. Other Evidence-based Reasoning Tasks 67

H31: Non-account owner

on IP1 scanned the H31: very
network for shared files, likely
between T1 and T2
Computer C1 has IP1 address

Non-account owner accessed Network scan for

an account on computer C1 shared files from IP1,
between T1 and T2 between T1 and T2

Scenario: Physical access of C1 in conference room CR1

A person from CR1,

H31: Non-account H32: Account owner accessed someone’s
owner on IP1 on IP1 scanned H32: Account
account on computer C1, owner on IP1
scanned network network for shared between T1 and T2
for shared files files between scanned network
for shared H32: no
between T1 and T2 T1 and T2 Search for persons present support
files between
in conference room CR1
T1 and T2
between T1 and T2
H21: Network scan
for shared files from Search for persons who
IP1, between T1 and T2 entered conference
(very likely) room CR1 before T2

H11: Sequence of accesses

to network services for
several systems from IP1,
between T1 and T2
(very likely) Search for persons Search for persons Search for persons
who entered CR1 who entered CR1 who entered CR1
before T2, based on before T2, based on before T2, based on
E*: Log record of denied scheduled meetings outside surveillance
door logs
access to network service participants video camera VC1
from IP1 to IP2 at time T

Figure 2.13. Evidence collection and assessment of higher-level alternative hypotheses.

If “H31: Non-account owner on IP1 scanned the network for shared files, between
T1 and T2” were true
Then the following subhypotheses would also be true:
“Non-account owner accessed an account on computer C1 between T1 and T2”
“Network scan for shared files from IP1, between T1 and T2”

To collect evidence for the first subhypothesis, we need to consider possible scenarios for
a non-account owner to access computer C1. The scenario illustrated in Figure 2.13 is a
physical access to C1 in the conference room CR1 where C1 is located. Another possible
scenario is a virtual access to C1.
As discussed in Section 2.2.4, such scenarios have enormous heuristic value in advan-
cing the investigation. In this case, for example, we are guided toward searching for
persons who were present in CR1 between T1 and T2. As indicated at the bottom of
Figure 2.13, there are several possible strategies to look for such persons:

Search for persons who entered CR1 before T2, based on door logs.
Search for persons who entered CR1 before T2, based on scheduled meetings participants.
Search for persons who entered CR1 before T2, based on outside surveillance video
camera VC1.

Notice that these are precise queries that can be answered very fast. Notice also that, in this
particular case, they involve all-source (non-computer) evidence. It is very important to be
able to use both computer and non-computer evidence to discover cyber insider threats.
This process will continue until the top-level hypothesis is assessed, as illustrated in
Figure 2.14.

13:50:24,
.003
 68 Chapter 2. Evidence-based Reasoning: Connecting the Dots

H51: P1, non-account

owner on IP1 H52: P1, non-account
performed covert owner on IP1 performed H51: P1, non-account owner
reconnaissance, covert reconnaissance on IP1 performed covert
collection, and for remote reconnaissance, collection, H51: likely
exfiltration vulnerabilities and exfiltration

H: H: H:
H41: P1, non-account Covert recon- Covert Covert
likely likely likely
owner on IP1 performed naissance collection exfiltration
reconnaissance H: no
between T1 and T2 support
(very likely) Covert H: H: no
Cs of IPs
scan (Cs) likely support
H31: Non-account owner of network Cs of processes H: likely
Cs of host H: likely
on IP1 scanned network
for shared files Cs of directory H:
between T1 and T2 Network almost
(very likely) scan data certain
for files temporary
email
file shares trash
H21: Network scan for H: H: very H: no
shared files from IP1, almost likely support
accounts
between T1 and T2 Sequence certain
(very likely) of accesses printers
remote
H11: Sequence of accesses access web services
to network services for
E*
several systems from IP1,
between T1 and T2
(very likely)
H52: P1, non-account owner on IP1 H52: no
E*: Log record of denied performed covert reconnaissance support
access to network service for remote vulnerabilities
from IP1 to IP2 at time T

Figure 2.14. Evidence collection and assessment of top-level alternative hypotheses.

2.3.2 Analysis of Wide-Area Motion Imagery

Capabilities exist today to persistently monitor fixed geographic locations (such as conflict
areas) as wide as 100 km2, for long periods of time, using electro-optic sensors (see
Figure 2.15). This leads to the collection of huge amounts of data to be used either in
real-time analysis or in forensic analysis. During real-time analysis, analysts attempt to
discover impeding threat events (e.g., ambush, kidnapping, rocket launch, false check-
point, suicide bomber, and improvised explosive devices [IEDs]) in time to react. During
forensic analysis, the analysts backtrack from such an event (e.g., an ambush) in order to
discover the participants, possible related locations and events, and the specific move-
ment patterns (Desai, 2009). The problem, however, is that the manual analysis of these
huge amounts of data would require thousands of analysts. Thus the use of a cognitive
assistant, such as Disciple-EBR, is very helpful.
Let us consider an analyst who performs real-time analysis of the wide area motion
imagery of a region characterized by insurgency operations. Road work at a highway
intersection suggests the hypothesis that there is an ambush threat to the U.S. forces.
Table 2.6 shows a chain of abductive inferences leading to this hypothesis. The corres-
ponding intermediary hypotheses and their alternatives are also shown in the left-hand
side of Figure 2.16. All these hypotheses would need to be analyzed.
The analyst uses Disciple-EBR to analyze these hypotheses. Because Disciple-EBR was
previously taught how to analyze such hypotheses, it automatically decomposes them into
increasingly simpler hypotheses that guide the collection of relevant evidence. This is
illustrated in the middle of Figure 2.16 with the top hypothesis “Hk: Ambush threat to U.S.
forces at Al Batha highway junction.” Thus, as indicated in the middle of Figure 2.16,
an ambush threat at a certain location requires that location to be good for ambush

13:50:24,
.003
2.3. Other Evidence-based Reasoning Tasks 69

Table 2.6 Evidence in Search of Hypotheses through Abductive Reasoning

E*i: There is evidence of road work at 1:17 am at the Al Batha highway junction.
➔ Ei: It is possible that there is indeed road work at the Al Batha highway junction.
➔ Ha: It is possible that the road work is for blocking the road.
➔ Hc: It is possible that there is ambush preparation at the Al Batha highway junction.
➔ Hk: It is possible that there is an ambush threat at the Al Batha highway junction.

Figure 2.15. Wide-area motion imagery (background image reprinted from www.turbophoto.com/
Free-Stock-Images/Images/Aerial%20City%20View.jpg).

(Hb: Ambush location), and there should also be some observable ambush preparation
activities (Hc: Ambush preparation). Further on, to be a good location for ambush requires
the corresponding route to be used by the U.S. forces (Hd: Blue route), and there should also
be cover at that location (He: Cover). This directly guides the analyst to check whether the U.S.
forces are using that route. It also guides the analyst to analyze images of that location for the
existence of cover. Having obtained the corresponding evidence, the analyst assesses its
support of the corresponding subhypotheses, and Disciple-EBR automatically aggregates
these assessments, concluding that it is almost certain that the location is good for ambush.
The rest of the analysis is developed in a similar way. The “Hc: Ambush preparation”
activity is automatically decomposed into three simpler activities: “Hf: Deployment,”
“Ha: Road blocking,” and “Hp: Move to cover.” Further on, “Hf: Deployment” is decom-
posed into “Hg: Vehicle deployment” and “Hm: Insurgent vehicle,” the last one being
further decomposed into two subhypotheses. All these simpler subhypotheses guide the
collection of corresponding relevant evidence which, in this example, is found and
evaluated, leading Disciple-EBR to infer that the top-level hypothesis is very likely.
Let us now consider that the analyst does not perform real-time analysis, but forensic
analysis. The ambush has already taken place (hence a third subhypothesis, “Hq: Ambush
execution,” of the top-level hypothesis), and the goal is to trace back the wide-area motion

13:50:24,
.003
 70 Chapter 2. Evidence-based Reasoning: Connecting the Dots

Hk: Ambush Hk1: Ambush Hk: Ambush threat to Hk: It is very likely that there is an
threat deception U.S. forces at Al Batha ambush threat to U.S. forces at Al
highway junction Batha highway junction

Hc: Hc1: Hc2:

Ambush Traffic Traffic Hc:
Hb: Ambush Hq:
preparation disruption deviation Ambush preparation Ambush Hb Hc Hq
location execution
Hf:
Deploy-
ment Ha: Hp: Hp
Hd He Hf
Ha: Ha1: Road Move
Hd: He: Hs: Ht:
Road Road blocking to Ha
Blue Cover Hs Ht
blocking repair cover Heavy Assault
route fire
Hm:
Insurgent Hm
vehicle
Ei
Ei: Road ¬ Ei: No Ei:
work road work Hg: Hn: Ho: Potential Hn Ho
Road Hg
Vehicle Vehicle Insurgent work items of
deployment departed facility evidence
from
E*i:Evidence facility
of road work E*i

Abductive reasoning Deductive reasoning Inductive reasoning

P  possibly Q P  necessarily Q P  probably Q
Evidence in search Hypotheses in search Evidentiary testing
of hypotheses of evidence of hypotheses

Figure 2.16. Another illustration of the general reasoning framework from Figure 1.9 (p. 27).

imagery in order to identify the participants together with the related locations and events.
The analyst and Disciple-EBR develop a similar analysis tree, as discussed previously,
which leads to the following hypotheses from the bottom of Figure 2.16: “Hn: Vehicle
departed from facility” and “Ho: Insurgent facility.” Thus, forensic analysis leads to the
discovery of the facility from which the insurgents have departed, identifying it as an
insurgent facility.
The same approach can also be used as a basis for the development of collaborative
autonomous agents engaged in persistent surveillance and interpretation of unconstrained
dynamic environments, continuously generating and testing hypotheses about the state of
the world. Consider, for example, the use of such agents in counterinsurgency operations
with the mission to automatically discover threat activities, such as IEDs, suicide bombers,
rocket launches, kidnappings, or ambushes. Discovery by sensor agents of road work at a
location that is often used by the U.S. forces leads to the hypothesis that there is an
ambush threat at that location. This hypothesis is then automatically decomposed into
simpler and simpler hypotheses, as discussed previously, guiding the agents to discover
additional evidence. Then the ambush threat hypothesis is automatically assessed and an
alert is issued if its probability is above a certain threshold.

2.3.3 Inquiry-based Teaching and Learning in a Science Classroom

2.3.3.1 Need for Inquiry-based Teaching and Learning
Significant progress has been made in K–12 science education with the development of
the National Science Education Standards (NRC, 1996). These standards call for inquiry-
based teaching and learning that, according to the standards, “refers to the diverse ways in
which scientists study the natural world and propose explanations based on the evidence
derived from their work. Inquiry also refers to the activities of students in which they

13:50:24,
.003
2.3. Other Evidence-based Reasoning Tasks 71

develop knowledge and understanding of scientific ideas, as well as an understanding of

how scientists study the natural world” (NRC, 1996, p. 23). “The research base on learning
and on effective learning environments makes a strong case for inquiry-based
approaches” (NRC, 2000, p. 128). The standards indicate that “teaching through inquiry
is effective” (NRC, 2000, p. 126) and that the appropriate use of inquiry by the teachers
“can have a powerful influence on their students’ science learning” (NRC, 2000, p. 128).
However, “for students to understand inquiry and use it to learn science, their teachers
need to be well-versed in inquiry and inquiry-based methods. Yet most teachers have not
had opportunities to learn science through inquiry or to conduct scientific inquiries
themselves. Nor do many teachers have the understanding and skills they need to use
inquiry thoughtfully and appropriately in their classrooms” (NRC 2000, p. 87). “Currently,
K–12 science education in the United States . . . does not provide students with engaging
opportunities to experience how science is actually done” (NRC, 2011, p. 1). “Instruction
throughout K–12 education is likely to develop science proficiency if it provides students
with opportunities for a range of scientific activities and scientific thinking, including, but
not limited to: inquiry and investigation, collection and analysis of evidence, logical
reasoning, and communication and application of information” (NRC, 2010, p. 137).
In this section, we will illustrate the use of a hypothetical Disciple-EBR agent in the
science classroom. We will call this hypothetical agent Inquirer, to emphasize its purpose.

2.3.3.2 Illustration of Inquiry-based Teaching and Learning

The illustration presented in this section follows closely the introductory example of
inquiry in a science classroom from Inquiry and the National Science Education Standards
(NRC, 2000, pp. 5–11), modified only to show a hypothetical use of Inquirer.
Several of the students in Mrs. Graham’s science class were excited when they returned
to their room after recess one fall day. They pulled their teacher over to a window, pointed
outside, and said, ‘We noticed something about the trees on the playground. The left tree
had lost all its leaves while the right tree had lush, green leaves. Why are those trees
different? They used to look the same, didn’t they?’ Mrs. Graham didn’t know the answer.
Mrs. Graham knew that her class was scheduled to study plants later in the year, and
this was an opportunity for them to investigate questions about plant growth that they had
originated and thus were especially motivated to answer. Although she was uncertain
about where her students’ questions would lead, Mrs. Graham chose to take the risk of
letting her students pursue investigations with Inquirer’s assistance and her guidance.
“Let’s make a list of hypotheses that might explain what’s happening to the left tree.” They
came up with a list of competing explanatory hypotheses, including the following ones
(shown also in the top-left of Figure 2.17):

H1: The left tree has lost its leaves because there is too much water at its root.
H2: The left tree has lost its leaves because it is older than the right tree.
H3: The left tree has lost its leaves because it is ill.

She then invited each student to pick one explanatory hypothesis, which led to several
groups: a “water” group, an “age” group, an “illness” group, and so on. She asked each
group to use the Inquirer assistant in order to plan and conduct a simple investigation to
test their preferred hypothesis.
For the next three weeks, science periods were set aside for each group to carry out
its investigation. Each group used the Inquirer assistant to conduct its investigation,

13:50:24,
.003
72 Chapter 2. Evidence-based Reasoning: Connecting the Dots

H1: The H2: The H3: The H1: The left tree H1: It is very likely that
left tree left tree left tree has lost its leaves the left tree has lost
has lost has lost has lost … because there is its leaves because
its leaves its leaves its leaves too much water there is too much
because because it because at its root. water at its root.
there is is older it is ill.
too much than the
water at right tree. H1a: H1b: Too H1a: It is very H1b: It is
its root. There is much likely that almost
too much water at a there is too certain that
water at tree’s root much water too much
the root causes it at the root water at a
of the left to lose its of the left tree’s root
tree. leaves tree. causes it
and die. to lose its
leaves
and die.
E*: Evidence that
the left tree has
lost its leaves.

Figure 2.17. Systematic inquiry with a cognitive assistant (based on NRC, 2000, pp. 5–11).

discovering a variety of sources with information about characteristics of trees, their

life cycles, and their environments. For example, the water group entered the following
hypothesis into Inquirer:

H1: The left tree has lost its leaves because there is too much water at its root.

The group’s members reasoned that, if this hypothesis were true, then two simpler
subhypotheses need to be true:

H1a: There is too much water at the root of the left tree.
H1b: Too much water at a tree’s root causes it to lose its leaves and die.

Therefore, they decomposed H1 into H1a and H1b by entering them into Inquirer (see the
middle part of Figure 2.17) and decided to assess them based on evidence. As a result,
Inquirer guided the students to look for both favoring and disfavoring evidence for each of
these two subhypotheses.
To collect relevant evidence for H1a, the students decided to look at the ground around
the two trees every hour that they could. They took turns on making individual observa-
tions, and since some of them lived near the school, their observations continued after
school hours and on weekends. Even though they missed some hourly observations, they
had sufficient data that they introduced into Inquirer as evidence E*1 favoring H1a, because
their observations confirmed the presence of excessive water at the root of the tree. As a
result, Inquirer extended the analysis of H1a from the middle of Figure 2.17 with the blue
tree shown in Figure 2.18, asking the students to assess the relevance of E1 (the event
indicated by the evidence E*1) with respect to H1a, as well as the believability of the
evidence E*1.
Inquirer reminded the students that relevance answers the question: So what? May E1
change my belief in the truthfulness of H1a? The students’ answer was: Assuming that E*1
is believable, it is very likely that there is too much water at the root of the left tree. They
did this by selecting one value from the following list of probabilistic assessments
displayed by Inquirer: {no support, likely, very likely, almost certain, certain}. They also
justified their choice with the fact that during their observations, the tree was standing in
water, which means that it is very likely that there is too much water there.

13:50:24,
.003
2.3. Other Evidence-based Reasoning Tasks 73

H1a: There is too much water very likely

at the root of the left tree.

Inferential force of E*1

Relevance of E1 to H1a very likely Believability of E*1 almost certain

Figure 2.18. Evidence-based assessment of an elementary hypothesis.

H1b: It is almost certain that too much water at

a tree’s root causes it to lose its leaves and die.

Inferential force of evidence

Based on favoring evidence, There is no disfavoring
H1b is almost certain evidence for H1b

Inferential force of favoring evidence

Based on E*1, Based on E*2, Based on E*3, H1b

H1b is likely H1b is likely is almost certain

Inferential force of E*1 Inferential force of E*2 Inferential force of E*3

Relevance Believability Relevance Believability Relevance Believability

of E1 to H1b of E*1 is of E2 to of E*2 is of E3 to of E*3 is
is likely almost certain H1b is likely very likely H1b is certain almost certain

Figure 2.19. Inferential force of evidence on an elementary hypothesis.

Inquirer also reminded them that believability answers the question: Can we believe
what E*1 is telling us? Here the students’ answer was: Believability of E*1 is almost certain,
since a few data points were missing and, on rare occasions, the left tree was not standing
in water.
Based on the students’ assessments, Inquirer determined the inferential force of E*1 on
H1a, as shown by the green reasoning tree in Figure 2.18: Based on E*1, it is very likely that
there is too much water at the root of the left tree. Inquirer explained the students that
inferential force answers the question: How strong is E*1 in favoring H1a? An item of
evidence, such as E*1, will make us believe that the hypothesis H1a is true if and only if
E1* is both highly relevant and highly believable. Therefore, the inferential force of E*1 on
H1a was computed as the minimum of the relevance of E1 (very likely) and the believability
of E*1 (almost certain), the minimum of which is very likely.
The students agreed that E1* also supports the hypothesis “H1b: Too much water at a
tree’s root causes it to lose its leaves and die.” They assessed the relevance of E1 as likely
(because E1 is only one instance of this phenomenon), and the believability of E1* as
almost certain, leading Inquirer to assess the inferential force of E1* on H1b as likely, which
is the minimum of the two (see the left part of Figure 2.19).
One of the students recalled that several months ago the leaves on one of his mother’s
geraniums had begun to turn yellow. She told him that the geranium was getting too much
water. This item of information was represented in Inquirer as item of evidence E*2
favoring the hypothesis H1b. The students agreed to assess E2‘s relevance as likely (because
geranium is a different type of plant) and E*2‘s believability as very likely (because although
the mother has experience with plants, she is not a professional), leading Inquirer to

13:50:24,
.003
74 Chapter 2. Evidence-based Reasoning: Connecting the Dots

compute E*2‘s inferential force on H1b as likely (see the bottom-middle part of Figure 2.19).
Additionally, Mrs. Graham gave the group a pamphlet from a local nursery entitled Growing
Healthy Plants. The water group read the pamphlet and found that when plant roots are
surrounded by water, they cannot take in air from the space around the roots and they
essentially “drown.” This item of information was represented in Inquirer as item of
evidence E*3 favoring the hypothesis H1b. The students agreed to assess E3‘s relevance as
certain (since it, in fact, asserted the hypothesis) and the believability of E*3 as almost certain
(because this information is from a highly credible expert), leading Inquirer to compute E*3‘s
inferential force on H1b as almost certain (see the bottom-right part of Figure 2.19). Addition-
ally, Inquirer computed the inferential force of all favoring evidence (i.e., E*1, E*2, and E*3)
on H1b as almost certain, by taking the maximum of them. This is also the probability of H1b,
because no disfavoring evidence was found. However, if any disfavoring evidence would
have been found, then the Inquirer would have needed to determine whether, on balance,
the totality of evidence favors or disfavors H1b, and to what degree.
Having assessed the probability of H1a as very likely and that of H1b as almost certain,
the students and Inquirer inferred that the probability of their top-level hypothesis H1 is
the minimum of the two because both are required to infer H1 (see the top-right part of
Figure 2.17). Finally, Inquirer automatically generated a report describing the analysis
logic, citing sources of data used and the manner in which the analysis was performed.
The report was further edited by the water group before being presented to the class,
together with the reports of the other teams.
As different groups presented and compared their analyses, the class learned that some
evidence – such as that from the group investigating whether the trees were different – did
not explain the observations. The results of other investigations, such as the idea that the
trees could have a disease, partly supported the observations. But the explanation that
seemed most reasonable to the students, that fit all the observations and conformed with
what they had learned from other sources, was “too much water.” After their three weeks
of work, the class was satisfied that together the students had found a reasonable answer
to their question.

2.3.3.3 Other Examples of Inquiry-based Teaching and Learning

The preceding scenario illustrates one form of classroom inquiry where the investigations are
free-ranging explorations of unexplained phenomena. However, depending on the educa-
tional goals for students, other forms of inquiry can be much more structured, allowing
Inquirer to provide even more guidance and assistance, as discussed in the following.
According to the National Science Education Standards, “Inquiry is a multifaceted
activity that involves making observations; posing questions; examining books and other
sources of information to see what is already known; planning investigations; reviewing
what is already known in light of experimental evidence; using tools to gather, analyze,
and interpret data; proposing answers, explanations, and predictions; and communicating
the results. Inquiry requires identification of assumptions, use of critical and logical
thinking, and consideration of alternative explanations” (NRC, 1996, p. 23). “Developing
the ability to understand and engage in this kind of activity requires direct experience and
continued practice with the processes of inquiry. Students do not come to understand
inquiry simply by learning words such as ‘hypothesis’ and ‘inference’ or by memorizing
procedures such as ‘the steps of the scientific method.’ They must experience inquiry
directly to gain a deep understanding of its characteristics” (NRC 2000, p. 14).

13:50:24,
.003
2.3. Other Evidence-based Reasoning Tasks 75

The Inquirer cognitive assistant illustrates a computational view of the inquiry process
as ceaseless discovery of evidence, hypotheses, and arguments, through evidence in
search of hypotheses, hypotheses in search of evidence, and evidentiary testing of hypoth-
eses. This allows students hands-on experiences with this process, as illustrated in the
previous section and described in a more general way in Section 1.4.2. The hypothetical
Inquirer cognitive assistant incorporates this general model of inquiry, together with a
significant amount of knowledge about the properties, uses, discovery, and marshaling of
evidence. This allows it to be used as a general tool supporting the learning of those
inquiry-based practices that all sciences share, as advocated by the National Research
Council (NRC) framework for K–12 science education (NRC, 2011). This can be done
through a sequence of hands-on exercises in biology, chemistry, or physics, allowing the
students to experience the same inquiry-based scientific practices in multiple domains.
For example, in one exercise, Inquirer will teach the students how a complex hypothesis
is decomposed into simpler hypotheses, and how the assessments of the simpler hypoth-
eses are combined into the assessment of the top-level hypothesis, as was illustrated
in Figure 2.17.
In another exercise, Inquirer will teach the students how to assess the relevance and
the believability of evidence. This case study will provide both a decomposition tree (like
the one in the middle of Figure 2.17, with the elementary hypotheses H1a and H1b) and a
set of items of information, some relevant to the considered hypotheses and some
irrelevant. The students will be asked to determine which item of information is
relevant to which elementary hypothesis, and whether it is favoring or disfavoring
evidence. They will also be asked to assess and justify the relevance and the believability
of each item of evidence. After completing their analysis, the teacher and Inquirer
will provide additional information, asking the students to update their analysis in the
light of the new evidence. Finally, the students will present, compare, and debate their
analyses in class.
In yet another exercise, Inquirer will provide an analysis tree like the one from the
middle of Figure 2.17 but no items of information, asking the students to look for relevant
evidence (e.g., by searching the Internet or by performing various experiments) and to
complete the analysis.
In a more complex exercise, Inquirer will present a scenario with some unusual
characteristics, like the two trees previously discussed. The students will be asked to
formulate competing hypotheses that may explain the surprising observations, use the
formulated hypotheses to collect evidence, and use the evidence to assess each hypoth-
esis. Then they will compare their analyses of the competing hypotheses in terms of the
evidence used and assessments made, and will select the most likely hypothesis. Inquirer
will assist the students in this process by guiding them in decomposing hypotheses,
searching for evidence, assessing the elementary hypotheses, combining the assessments
of the simpler hypotheses, comparing the analyses of the competing hypotheses, and
producing an analysis report.
Inquirer can also provide many opportunities for collaborative work, in addition to
those that we have illustrated. For example, a complex hypothesis will be decomposed
into simpler hypotheses, each assessed by a different student. Then the results obtained
by different students will be combined to produce the assessment of the complex
hypothesis. Or different students will analyze the same hypothesis. Then they will
compare and debate their analysis and evidence and work together toward producing
a consensus analysis.

13:50:24,
.003
 76 Chapter 2. Evidence-based Reasoning: Connecting the Dots

Consistent with the writers of the National Science Education Standards, who “treated
inquiry as both a learning goal and as a teaching method” (NRC 2000, p. 18), Inquirer is
envisioned as both a teaching tool for teachers and as a learning assistant for students.
For example, the teacher will demonstrate some of these exercises in class. Other exercises
will be performed by the students, under the guidance of the teacher and with the
assistance of Inquirer.

2.4 HANDS ON: BROWSING AN ARGUMENTATION

The use of the various modules of Disciple-EBR will be introduced with the help of case
studies with associated instructions that will provide detailed guidance. We will illustrate
this process by running the case study stored in the knowledge base called “01-Browse-
Argumentation.” This case study concerns the hypothesis “The cesium-137 canister is
missing from the XYZ warehouse,” which is part of the analysis example discussed in
Section 2.2.1. This case study has three objectives:

 Learning how to run a case study

 Learning how to browse a reasoning tree or argumentation
 Understanding the process of hypothesis analysis through problem reduction and
solution synthesis

Figure 2.20 shows an example of an argumentation or reasoning tree in the interface of the
Reasoner. The left panel shows an abstract view of the entire reasoning tree. It consists of
brief names for the main hypotheses and their assessments, if determined. You can expand
(or show the decomposition of) a hypothesis by clicking on it or on the plus sign (+) on its
left. To collapse a decomposition, click on the minus sign (–). You can expand or
collapse the entire tree under a hypothesis by right-clicking on it and selecting Expand or
Collapse.
When you click on a hypothesis in the left panel, the right panel shows the detailed
description of the reasoning step abstracted in the left panel (see Figure 2.20). If you click
on [HIDE SOLUTIONS] at the top of the window, the agent will no longer display the
solutions/assessments of the hypotheses in the right panel. You may show the solutions
again by clicking on [SHOW SOLUTIONS].
The detailed argumentation from the right panel may be a single decomposition or a
deeper tree. In both cases, the leaves of the tree in the right panel are the detailed
descriptions of the subhypotheses of the hypothesis selected in the left panel (see the
arrows in Figure 2.20 that indicate these correspondences).
A detailed description shows the entire name of a hypothesis. It also shows the
question/answer pair that justifies the decomposition of the hypothesis into subhypoth-
eses. The detailed descriptions with solutions also show the synthesis functions that were
used to obtain those solutions from children solutions in the tree. These functions will be
discussed in Section 4.3.
Notice that, in the actual interface of Disciple-EBR, some of the words appear in bright
blue while others appear in dark blue (this distinction is not clearly shown in this book).
The bright blue words are names of specific entities or instances, such as “cesium-137
canister.” The dark blue words correspond to more general notions or concepts, such as
“evidence.” If you click on such a (blue) word in the right panel, the agent automatically

13:50:24,
.003
.003
13:50:24,

Figure 2.20. Abstract and detailed views of an argumentation.

78 Chapter 2. Evidence-based Reasoning: Connecting the Dots

switches to the Description module and displays its description. To view the reasoning tree
again, just click on Reasoner on the top of the window.
Figure 2.21 shows another part of the reasoning tree from Figure 2.20 to be browsed
in this case study. Notice that some of the solutions in the left panel have a yellow
background. This indicates that they are assessments or assumptions made by the user,
as will be discussed in Sections 4.4 and 4.9.
In this case study, you will practice the aforementioned operations. You will first select
the hypothesis “The cesium-137 canister is missing from the XYZ warehouse.” Then you will
browse its analysis tree to see how it is decomposed into simpler hypotheses and how the
assessments of these simpler hypotheses are composed. You will visualize both detailed
descriptions of these decomposition and synthesis operations, as well as abstract ones,
including an abstract view of the entire tree. Then you will visualize the descriptions of the
concepts and instances to which the analysis tree refers. Start by following the instructions
described in Operation 2.1 and illustrated in Figure 2.22.

Operation 2.1. Run a case study

 Start Disciple-EBR.
 In the System Workspace Manager, click on the knowledge base icon containing the
plus sign (+).
 The Knowledge Manager window opens, showing all the knowledge bases from the
repository.
 Click on the plus sign (+) of the case study domain knowledge base to be run. This will
display one or several scenario knowledge bases.
 Click on the scenario knowledge base corresponding to the case study to be run.
 Click on the Select button. This will both load the case study knowledge bases in
memory (i.e., the domain KB and the scenario KB) and select them as the current ones
to work with. Because they are loaded in memory, their names will be shown in bold in
the Knowledge Manager window.
 Follow the instructions at the bottom of the displayed window and run the case study.
After running the case study, you have to close the corresponding knowledge base, as
instructed in the following steps.
 Close all the workspaces open on the current knowledge base (case study) by selecting
one of them and clicking on the minus sign (–) to the right of the knowledge base icon
containing the plus sign (+).
 Click on the knowledge base icon containing the plus sign (+) situated at the right of
the workspace selector (see Figure 3.23).
 The Knowledge Manager window will be opened, showing all the knowledge bases.
 Click on the Scen node corresponding to the case study knowledge base that
was run.
 Click on the Close button in the right side of the window.
 Click on the X button to close the Knowledge Manager window.

This case study has also illustrated the following basic operations:

Operation 2.2. Browse the analysis of a hypothesis

 In the Hypothesis module, select a hypothesis.
 The Reasoner module will be automatically selected, showing the corresponding
reasoning tree.

13:50:24,
.003
.003
13:50:24,

Figure 2.21. Views of another part the reasoning tree from Figure 2.20.
80 Chapter 2. Evidence-based Reasoning: Connecting the Dots

1. Click on this icon to load a KB

2. System displays all KBs

3. Click on + to expand the KB to be loaded

5. Click on Select

4. Click on Scen

6. Perform the case study by following the displayed instructions

Figure 2.22. Running a case study.

 To browse the entire reasoning tree, step by step, click on the hypotheses in the left
panel, or one of the + and – signs preceding them.
 To expand or collapse the entire subtree of a hypothesis, right-click on it and select the
corresponding action.
 To view the detailed description of an abstract decomposition of a hypothesis in the left
panel, click on the hypothesis, and the detailed decomposition will be displayed in the
right panel.
 To browse a detailed reasoning tree in the right panel click on the + and – signs.

Operation 2.3. End a case study

 Select the Scenario workspace (repository\KB-name\Scen)
 From the System menu, select Save All to save the knowledge bases.

13:50:24,
.003
 2.6. Review Questions 81

 Close all the workspaces open on the current knowledge bases (case studies) by
clicking on the minus sign (–) to the right of the knowledge base icon containing the
plus sign (+).
 Close the opened knowledge bases corresponding to the case study by following the
instructions from Operation 3.3.

2.5 PROJECT ASSIGNMENT 1

This is the first of a sequence of assignments in which you will develop a knowledge-based
agent for analyzing hypotheses in a domain of interest to you. Since you will develop the
knowledge base of the agent, you will have to consider a familiar domain. Select a domain
and illustrate the process of evidence-based reasoning with a diagram such as those in
Figures 2.8, 2.14, 2.26, or 2.27. That is, specify an item of evidence, an abductive reasoning
chain from that evidence to a hypothesis of interest, alternative hypotheses to those from
the reasoning chain, and the analysis of one of these hypotheses.

2.6 REVIEW QUESTIONS

2.1. Consider the intelligence analysis problem from Section 2.2.1. What might constitute
an item of evidence for the hypothesis that the cesium-137 canister was misplaced?

2.2. What might by an alternative hypothesis for “H5: A dirty bomb will be set off in the
Washington, D.C., area”?

2.3. A terrorist incident occurred two weeks ago in an American city involving consider-
able destruction and some loss of life. After an investigation, two foreign terrorist
groups have been identified as possible initiators of this terrorist action: Group
A and Group B. Which are some hypotheses we could entertain about this event?

2.4. Consider the hypothesis that Group A from Country W was involved in a recent
terrorist incident in an American city. What evidence we might find concerning this
hypothesis?

2.5. Sometimes we have evidence in search of hypotheses or possible explanations. For

example, consider the dog-tag containing the name of one of our soldiers who has
been missing since the end of our conflict with Country X. This tag was allegedly
given to a recent visitor in Country X, who then gave it to us. One possibility is that
this soldier is still being held as a prisoner in Country X. What are some other
possibilities?

2.6. Sometimes we have hypotheses in search of evidence. Suppose our hypothesis is

that Person X was involved in a recent terrorist incident in an American city. So far,
all we have is evidence that he was at the scene of the incident an hour before it
happened. If this hypothesis were true, what other kinds of evidence about X might
we want to look for?

2.7. Consider the hypothesis that the leadership of Country A is planning an armed
conflict against Country B. You have just obtained a report that says that there has

13:50:24,
.003
82 Chapter 2. Evidence-based Reasoning: Connecting the Dots

just been an attempt on the life of the president of Country B by an unknown

assailant. Why is this report, if believable, relevant evidence on the hypothesis that
the leadership of Country A is planning an armed conflict against Country B?

2.8. Defendant Dave is accused of shooting a victim Vic. When Dave was arrested
sometime after the shooting he was carrying a 32 caliber Colt automatic pistol.
Let H be the hypothesis that it was Dave who shot Vic. A witness named Frank
appears and says he saw Dave fire a pistol at the scene of the crime when it
occurred; that’s all Frank can tell us. Construct a simple chain of reasoning that
connects Frank’s report to the hypothesis that it was Dave who shot Vic.

2.9. Consider the situation from Question 2.8. The chain of reasoning that connects
Frank’s report to the hypothesis that it was Dave who shot Vic shows only the
possibility of this hypothesis being true. What are some alternative hypotheses?

2.10. Consider again the situation from Questions 2.8 and 2.9. In order to prove the
hypothesis that it was Dave who shot Vic, we need additional evidence. As dis-
cussed in Section 2.2.2, we need to put this hypothesis to work to guide us in
collecting new evidence. Decompose this hypothesis into simpler hypotheses,
as was illustrated by the blue trees in Figures 2.8 and 2.9, in order to discover
new evidence.

2.11. Our investigation described in Questions 2.8, 2.9, and 2.10 has led to the discovery
of additional evidence. By itself, each evidence item is hardly conclusive that Dave
was the one who shot Vic. Someone else might have been using Dave’s Colt
automatic. But Frank’s testimony, along with the fact that Dave was carrying his
weapon and with the ballistics evidence, puts additional heat on Dave. Analyze the
hypothesis that it was Dave who shot Vic, based on all these items of evidence, as
was illustrated by the green trees in Figures 2.8 and 2.9. In Chapter 4, we will
discuss more rigorous methods for making such probabilistic assessments. In this
exercise, just use your common sense.

2.12. A car bomb was set off in front of a power substation in Washington, D.C., on
November 25. The building was damaged but, fortunately, no one was injured.
From the car’s identification plate, which survived, it was learned that the car
belonged to Budget Car Rental Agency. From information provided by Budget, it
was learned that the car was last rented on November 24 by a man named
M. Construct an argument from this evidence to the hypothesis that Person
M was involved in this car-bombing incident.

2.13. Consider again the situation from Question 2.12, and suppose that we have deter-
mined that the evidence that M rented a car on November 24 is believable. We
want now to assess whether M drove the car on November 25. For this, we need
additional evidence. As discussed in Section 2.2.2, we need to put this hypothesis to
work to guide us in collecting new evidence. Decompose this hypothesis into
simpler hypotheses, as was illustrated by the blue trees in Figures 2.8 and 2.9, in
order to discover new evidence.

13:50:24,
.003
 Methodologies and Tools for Agent
3 Design and Development

3.1 A CONVENTIONAL DESIGN AND DEVELOPMENT

SCENARIO

3.1.1 Conventional Design and Development Phases

Table 3.1 shows the main development phases for a knowledge-based agent. As in the
case of developing a typical software system, there are feedback loops between all these
phases. We will illustrate these phases with a conventional scenario of developing a
knowledge-based agent. This scenario is an adaptation of the one described in Buchanan
et al. (1983).

3.1.2 Requirements Specification and Domain Understanding

The development process starts with identifying a problem that may be solved by develop-
ing a knowledge-based agent. Table 3.2 shows an example of such a problem.
A knowledge engineer, who is a person specialized in developing knowledge-based
agents, is assigned the job of building an agent that incorporates the expertise of subject
matter experts, who are people familiar with the detection, monitoring, and containment
of the spills of hazardous materials. The goal of the system is to assist its users in
performing such tasks. The knowledge engineer has to identify a subject matter expert
from whom to capture the domain expertise and represent it into the agent’s knowledge
base, as was briefly discussed in Section 1.6.3.1.

Table 3.1 Main Phases in the Development of a Knowledge-based Agent

1. Requirements specification (specifying the types of problems to be solved or hypotheses to be

analyzed, and the agent to be built).
2. Domain understanding.
3. Ontology design and development.
4. Development of problem-solving rules or methods.
5. Refinement of the knowledge base (ontology and rules/methods).
6. Verification, validation, and certification of the agent.

83

13:44:04,
.004
84 Chapter 3. Methodologies and Tools

Table 3.2 Problem to Be Solved with a Knowledge-based Agent (from Buchanan et al.,
p. 132).

The director of the Oak Ridge National Lab (ORNL) faces a problem. Environmental Protection
Agency (EPA) regulations forbid the discharge of quantities of oil or hazardous chemicals into or
upon waters of the United States when this discharge violates specified quality standards. ORNL
has approximately two thousand buildings on a two-hundred-square-mile government
reservation, with ninety-three discharge sites entering White Oak Creek. Oil and hazardous
chemicals are stored and used extensively at ORNL. The problem is to detect, monitor,
and contain spills of these materials, and this problem may be solved with a
knowledge-based agent.

Table 3.3 Specification of the Actual Problem to Be Solved (from Buchanan et al., p.133).

When an accidental inland spill of an oil or chemical occurs, an emergency situation may exist,
depending on the properties and the quantity of the substance released, the location of the
substance, and whether or not the substance enters a body of water.
The observer of a spill should:
1. Characterize the spill and the probable hazards.
2. Contain the spill material.
3. Locate the source of the spill and stop any further release.
4. Notify the Department of Environmental Management.

What issues may concern the subject matter expert? First of all, the expert may be
concerned that once his or her expertise is represented into the agent, the organization
may no longer need him or her because the job can be performed by the agent. Replacing
human experts was a bad and generally inaccurate way of promoting expert systems.
Usually, the knowledge-based agents and even the expert systems are used by experts in
order to better and more efficiently solve problems from their areas of expertise. They are
also used by people who need the expertise but do not have access to a human expert, or
the expert would be too expensive.
What are some examples of knowledge-based agents? Think, for instance, of any tax-
preparation software. Is it a knowledge-based agent? What about the software systems that
help us with various legal problems, such as creating a will? They all are based on large
amounts of subject matter expertise that are represented in their knowledge bases.
Once the subject matter expert is identified and agrees to work on this project, the
knowledge engineer and the expert have a series of meetings to better define the actual
problem to be solved, which is shown in Table 3.3.
The knowledge engineer has many meetings with the subject matter expert to elicit his
or her knowledge on how to solve the specified problem. There are several knowledge
elicitation methods that can be employed, as discussed later in Section 6.3. Table 3.4
illustrates the unstructured interview, where the questions of the knowledge engineer and
the responses of the expert are open-ended.

13:44:04,
.004
 3.1. Conventional Design and Development Scenario 85

Table 3.4 Unstructured Interview to Elicit Subject Matter Expertise

(from Buchanan et al., p. 134)

KE: Suppose you were told that a spill had been detected in White Oak Creek one mile
before it enters White Oak Lake. What would you do to contain the spill?
SME: That depends on a number of factors. I would need to find the source in order to
prevent the possibility of further contamination, probably by checking drains and manholes for
signs of the spill material. And it helps to know what the spilled material is.
KE: How can you tell what it is?
SME: Sometimes you can tell what the substance is by its smell. Sometimes you can
tell by its color, but that's not always reliable since dyes are used a lot nowadays. Oil,
however, floats on the surface and forms a silvery film, while acids dissolve completely in
the water. Once you discover the type of material spilled, you can eliminate any
building that either doesn’t store the material at all or doesn’t store enough of it to account
for the spill.

Table 3.5 Identification of the Basic Concepts and Features Employed by the
Subject Matter Expert

KE: Suppose you were told that a spill had been detected in White Oak Creek one mile
before it enters White Oak Lake. What would you do to contain the spill?
SME: That depends on a number of factors. I would need to find the source in order
to prevent the possibility of further contamination, probably by checking drains
and manholes for signs of the spill material. And it helps to know what the spilled
material is.
KE: How can you tell what it is?
SME: Sometimes you can tell what the substance is by its smell. Sometimes you can
tell by its color, but that's not always reliable since dyes are used a lot nowadays. Oil, however,
floats on the surface and forms a silvery film, while acids dissolve completely in the water. Once
you discover the type of material spilled, you can eliminate any building that either doesn’t store
the material at all or doesn’t store enough of it to account for the spill.

3.1.3 Ontology Design and Development

The main goal of the initial interview sessions is to identify the basic concepts and features
from the application domain. Some of these domain concepts and features are underlined
in Table 3.5.
The identified concepts, instances, and features are used to design and develop the
ontology of the system to be built. A fragment of the developed concept and instance
hierarchy is shown in Figure 3.1.
The nodes in light blue (such as “building 3023”) represent specific objects or
instances. The nodes in dark blue (such as “building”) represent sets of instances and
are called concepts. The instances and concepts may have features that are represented
in green, as for instance, “spill-1 has as type acid” and “spill-1 has as odor vinegar odor.”

13:44:04,
.004
 86 Chapter 3. Methodologies and Tools

object

building subconcept of
odor
appearance
instance of
building 3024 subconcept of
source spill
building 3023 instance of no film
instance of
silver film subconcept of
s6-1 s6-2 spill-1
pungent odor
location substance has as type
subconcept of has as odor
subconcept of
vinegar odor
drain manhole water acid
oil
instance of instance of subconcept of subconcept of
d6-1 d6-2 m6-1 m6-2
diesel oil gasoline sulfuric acid acetic acid

hydraulic oil hydrochloric acid

Figure 3.1. Hierarchy of domain concepts and instances.

feature

has as type has as odor

has as location has as color
has as amount

Figure 3.2. Features used to describe the instances and concepts.

The features are also represented hierarchically, as shown in Figure 3.2. Ontologies are
discussed in detail in Chapter 5.

3.1.4 Development of the Problem-Solving Rules or Methods

Once an initial ontology is developed, the knowledge engineer and the subject matter
expert use the component concepts and features to represent the expert’s methods of
determining the spill material as a set of rules. Consider, for example, the reasoning of
the subject matter expert (SME) from the top of Table 3.6. What IF-THEN rules may
represent this reasoning?
In an iterative process, the knowledge engineer represents the expert’s reasoning as
rules, asks the expert to critique them, and correspondingly updates the rules, as illus-
trated in Table 3.7. As would be expected, this is quite a long and difficult process, for the
reasons discussed in Section 1.6.3.1.

13:44:04,
.004
 3.1. Conventional Design and Development Scenario 87

Table 3.6 Representing the Expert’s Reasoning as If-Then Rules

SME: Sometimes you can tell what the substance is by its smell. Sometimes you can tell by its color,
but that’s not always reliable since dyes are used a lot nowadays. Oil, however, floats on the
surface and forms a silvery film, while acid dissolves completely in the water.
IF the spill . . .
THEN the substance of the spill is oil
IF the spill . . .
THEN the substance of the spill is acid

Table 3.7 Iterative Process of Rules Development and Refinement (Based on Buchanan et al.,
p.138)

KE: Here are some rules I think capture your explanation about determining the substance of the
spill. What do you think?
IF the spill does not dissolve in water
and the spill forms a silvery film
THEN the substance of the spill is oil
IF the spill dissolves in water
and the spill does not form a film
THEN the substance of the spill is acid

SME: Uh-huh (long pause). Yes, that begins to capture it. Of course, if the substance is silver nitrate,
it will dissolve only partially in water.
KE: I see. Rather than talking about a substance dissolving or not dissolving in water, we should talk
about its solubility, which we may consider as being high, moderate, or low. Let’s add that
information to the knowledge base and see what it looks like.
IF the solubility of the spill is low
and the spill forms a silvery film
THEN the substance of the spill is oil
IF the solubility of the spill is moderate
THEN the substance of the spill is silver-nitrate

SME: If the solubility of the spill is moderate, I would be about 60 percent

sure that the substance of the spill is silver-nitrate.
KE: Okay, we will represent this information in the rule.
IF the solubility of the spill is moderate
THEN the substance of the spill is silver-nitrate with certainty 0.6

3.1.5 Verification, Validation, and Certification

Once the ontology and the rules are defined and represented into the knowledge base of
the agent, the knowledge engineer and the subject matter expert can run the prototype
agent and analyze its solutions, which are generated by the chaining of the rules.

13:44:04,
.004
 88 Chapter 3. Methodologies and Tools

Testing of the agent involves three types of activity: verification, validation, and certifi-
cation (O’Keefe et al., 1987; Awad, 1996).
In essence, verification attempts to answer the question: Are we building the agent right?
Its goal is to test the consistency and the completeness of the agent with respect to its initial
specification. For example, in the case of a rule-based agent, one would check the rules to
identify various types of errors, such as the existence of rules that are redundant, conflicting,
subsumed, circular, dead-end, missing, unreachable, or with unnecessary IF conditions.
Validation, on the other hand, attempts to answer the question: Are we building the
right agent? In essence, this activity checks whether the agent meets the user’s needs and
requirements.
Finally, certification is a written guarantee that the agent complies with its specified
requirements and is acceptable for operational use.

3.2 DEVELOPMENT TOOLS AND REUSABLE ONTOLOGIES

Various types of tools can be used to develop a knowledge-based agent. We will briefly
discuss three different types: expert system shells, learning agent shells, and learning agent
shells for evidence-based reasoning. We will also discuss the reuse of knowledge in the
development of a knowledge-based agent.

3.2.1 Expert System Shells

These tools exploit the architectural separation in a knowledge-based system between the
knowledge base and the inference engine (see the right-hand side of Figure 1.14, p. 37).
Given this separation, two knowledge-based agents, one for diagnosing ships and the
other for diagnosing airplanes, may potentially use the same inference engine for diagno-
sis. The difference between the two agents will be in the content of their knowledge bases
(e.g., ship parts versus airplane parts).
An expert system shell is a tool that consists of an inference engine for a certain class of
tasks (e.g., planning, design, diagnosis, monitoring, prediction, interpretation, etc.); a repre-
sentation formalism in which the knowledge base can be encoded; and mechanisms for
acquiring, verifying, and revising knowledge expressed in that formalism.
If the inference engine of an expert system shell is adequate for a certain expertise task
(e.g., planning), then the process of building an expert system or agent for that type of
tasks is, in principle, reduced to the building of the knowledge base.
Different expert system shells trade generality (i.e., their domain of applicability)
against power (i.e., the assistance given to the expert and the knowledge engineer in the
development process), covering a large spectrum.
At the generality end of the spectrum are very general shells that can be used to build
almost any type of expert system, but they do not provide too much assistance in building
a specific system. Examples of such shell are:

 OPS (Cooper and Wogrin, 1988), which has a general rule engine
 CLIPS (Giarratano and Riley, 1994), which also has a general rule engine
 CYC (Lenat, 1995; CYC, 2008; 2016), a very large knowledge base with ontologies
covering many domains, and with several rule engines
 EXPECT (Gil and Paris, 1995; EXPECT 2015), a shell that enables the acquisition of
problem-solving knowledge both from knowledge engineers and from end-users

13:44:04,
.004
 3.2. Development Tools and Reusable Ontologies 89

 JESS (Friedman-Hill, 2003; JESS, 2016), which is a version of CLIPS with a Java-based
rule engine
 CommonKADS (Schreiber et al., 2000), which is a general methodology with support-
ing tools for the development of knowledge-based systems
 Jena, which is a toolkit for developing applications for the Semantic Web (Jena, 2012)
 Pellet, an ontology (OWL2) reasoner that can be used to develop knowledge-based
applications for the Semantic Web (Pellet, 2012)
 Protégé (Musen, 1989; Protégé, 2015), an ontology editor and knowledge base frame-
work, also used to develop knowledge-based applications for the Semantic Web
 TopBraid Composer (Allemang and Hendler, 2011; TopBraid Composer, 2012) Ontology
Development Tool for Semantic Web applications

At the power end of the spectrum are shells that employ much more specific problem-
solving methods, such the propose-and-revise design method used in SALT (Marcus, 1988)
to design elevators. The knowledge for such a system can be elicited by simply filling in
forms, which are then automatically converted into rules. Thus the shell provides signifi-
cant assistance in building the system, but the type of systems for which it can be used is
much more limited.
In between these two types of shells are the shells applicable to a certain type of
problems (such as planning, or diagnosis, or design). A representative example is EMYCIN
(van Melle et al., 1981), a general rule-based shell for medical diagnosis.

3.2.2 Foundational and Utility Ontologies and Their Reuse

The use of an expert system shell facilitates the development of an expert system because
it reduces the process of developing the system to that of developing the knowledge base
for the provided inference engine. However, it is this development of the knowledge base
that is the most challenging part of developing the system.
Despite the fact that building the knowledge base is such a difficult problem, historic-
ally knowledge bases were very rarely reused, primarily for two reasons. First, the know-
ledge in the knowledge base is usually very specific to a particular domain and type of
problems and cannot be applied directly to a different application area. Second, even if the
knowledge base of a system is directly relevant to the new area, its reuse by a system with a
different knowledge representation is likely to be very difficult because of the differences
between the knowledge models of the two systems.
Progress in knowledge engineering has changed this situation. A new architectural
separation has emerged at the level of the knowledge base, which is structured into two
main components: an ontology that defines the concepts of the application domain, and a
set of problem-solving rules or methods expressed with these concepts, as was illustrated
in Section 1.6.3.1. The ontology is the more general part, being characteristic to an entire
application domain, such as military or medicine. In the military domain, for example, the
ontology includes descriptions of military units and of military equipment. These descrip-
tions are most likely needed in almost any specific military application and can therefore
be reused. The rules are the more specific part of the knowledge base, corresponding to a
certain type of application. For example, there may be rules for an agent, such as Disciple-
COA, that assists a commander in critiquing courses of action, or rules for an agent, such
as Disciple-WA, that assists in planning the repair of damaged bridges or roads. These
rules offer much fewer opportunities for reuse, if any.

13:44:04,
.004
 90 Chapter 3. Methodologies and Tools

As a result of terminological standardization to facilitate automatic processing of infor-

mation, particularly in the context of the Semantic Web, many domain and general-purpose
ontologies have been developed. The general-purpose (domain-independent) ontologies
are also called upper, foundational, or universal ontologies because they provide high-level,
domain-independent concepts and relationships that can be included in the top part of a
domain ontology (Obrst et al., 2012). Examples of such ontologies are Cyc/OpenCyc (2016;
Lenat, 1995), Suggested Upper Merged Ontology (SUMO) (Pease, 2011), WordNet (2012;
Fellbaum 1988), Descriptive Ontology for Linguistic and Cognitive Engineering (DOLCE,
2012), Basic Formal Ontology (BFO, 2012), Object-centered High-level Reference (OCHRE)
(Schneider, 2003), General Formal Ontology (GFO, 2012), Unified Medical Language System
(UMLS) (Humphreys and Lindberg, 1993), and Unified Foundational Ontology (UFO)
(Guizzardi and Wagner, 2005a, 2005b).
There are also utility ontologies (Obrst et al., 2012) that include representations of
commonly used concepts, such as persons, social roles, and organizations (Masolo et al.,
2004; FOAF, 2012), temporal concepts (Hobbs and Pan, 2004; Pan and Hobbs, 2004; 2012),
and geospatial concepts (Ressler et al., 2010; Geonames, 2012).
The Open Knowledge Base Connectivity (OKBC) protocol (OKBC, 2008) has been defined
to facilitate knowledge sharing and reuse (Chaudhri et al., 1998). OKBC is a standard for
accessing knowledge bases stored in different frame-representation systems. It provides a set
of operations for a generic interface to such systems. As a result, OKBC servers for various
frame-based systems, such as Ontolingua (Ontolingua 1997; 2008; Farquhar et al., 1997) and
Loom (Loom, 1999; MacGregor, 1991), have been developed. These servers are repositories of
reusable ontologies and domain theories and can be accessed using the OKBC protocol.
Additionally, there are many tools that can query existing ontologies on the Semantic
Web or simply import them into the knowledge base to be built, including Jena (2012),
OWLIM (2012), Pellet (2012), Protégé (2015), and TopBraid Composer (2012).

3.2.3 Learning Agent Shells

While part of the ontology of the agent to be developed can be imported from previously
developed ontologies, the reasoning rules of the agent are application-specific and cannot
be reused. An alternative approach to rules development is to employ a learning agent that
is able to learn the rules. In addition to containing the knowledge base and the inference
engine, the architecture of a learning agent includes a learning engine consisting of the
programs that create and update the data structures in the knowledge base. The learning
agent may learn from a variety of information sources in the environment. It may learn from
its user or from other agents, either by being directly instructed by them or just by observing
and imitating their behavior. It may learn from a repository of information (such as a
database), or it may learn from its own experience. Building a practical autonomous
learning agent that can acquire and update its knowledge by itself is not yet practical, except
for very simple problems, such as classification. Therefore, a more practical approach is to
develop an interactive learning agent that can interact with an expert. Such an agent can
perform many of the functions of the knowledge engineer. It allows the human expert to
communicate expertise in a way familiar to him or her and is responsible for building,
updating, and reorganizing the knowledge base. We call such a tool a learning agent shell.
A learning agent shell is an advanced tool for building knowledge-based agents. It
contains a general problem-solving engine, a learning engine, and a general knowledge

13:44:04,
.004
 3.2. Development Tools and Reusable Ontologies 91

base structured into an ontology and a set of rules (see Figure 3.3). Building a knowledge-
based agent for a specific application consists of customizing the shell for that application
and developing the knowledge base. The learning engine facilitates the building of the
knowledge base by subject matter experts and knowledge engineers.
Examples of learning agent shells are Disciple-LAS (Tecuci et al., 1999), Disciple-COA
(Tecuci et al., 2001), and Disciple-COG/RKF (Tecuci et al., 2005b), the last two being
presented in Section 12.4.

3.2.4 Learning Agent Shell for Evidence-based Reasoning

As discussed in Section 1.4.3, tasks in many domains, such as law, intelligence analysis,
cybersecurity, forensics, medicine, physics, chemistry, history, or archaeology, involve
evidence-based reasoning. All these tasks use general evidence-based reasoning concepts
and rules. Examples of general concepts are different types of evidence, such as tangible
evidence and testimonial evidence. An example of a general rule is to assess the credibility of
a human source of information by assessing his or her veracity, objectivity, and observational
sensitivity. Since all this knowledge is domain-independent, it makes sense to develop a
general knowledge base for evidence-based reasoning that can be reused each time we are
developing an agent that also needs to perform this kind of reasoning. To facilitate this
process, the knowledge base of the agent is structured into a hierarchy of knowledge bases,
with the knowledge base for evidence-based reasoning at top of the hierarchy.
A learning agent shell for evidence-based reasoning is a learning agent shell that contains
a hierarchy of knowledge bases the top of which is a knowledge base for evidence-based
reasoning (see Figure 3.4). As will be illustrated in the following sections, building an agent
for a specific evidence-based reasoning application consists in extending the knowledge
base with domain-specific knowledge through learning from a subject matter expert.
An example of a learning agent shell for evidence-based reasoning is Disciple-EBR, the
architecture of which is shown in the center of Figure 3.5. It includes multiple modules for
problem solving, learning, tutoring, evidence-based reasoning, mixed-initiative inter-
action, as well as a hierarchically organized knowledge base with domain-independent
knowledge for evidence-based reasoning at the top of the knowledge hierarchy. The
Disciple-EBR shell can learn complex problem-solving expertise directly from human
experts, and in doing so it evolves into a cognitive assistant that can support experts and
nonexperts in problem solving and can teach expert problem-solving to students.
The outside hexagon in Figure 3.5 summarizes a possible life cycle of a Disciple
cognitive assistant for evidence-based reasoning. The first stage is shell customization,

Problem
Interface

Solving
Ontology
+ Rules
Learning

Figure 3.3. The overall architecture of a learning agent shell.

13:44:04,
.004
92 Chapter 3. Methodologies and Tools

Problem EBR KB
Solving

Interface
Domain KB Domain KB Domain KB

Scenario KB Scenario KB
Learning
KB = Ontology + Rules

Figure 3.4. The overall architecture of a learning agent shell for evidence-based reasoning.

1. Shell
Customization
Disciple Developer and Knowledge Engineer
Disciple
Learning Agent Multi-
Disciple Ontology
Shell Strategy
Development
Learning
Modules
Modules

EBR KB

Mixed- Tutoring
Initiative Domain KB Domain KB Domain KB
Modules
Interaction
Scenario KB Scenario KB

Knowledge Base
Management
End
-
Use
r

Problem- Evidence-
Solving Specific
Modules Specialized Modules
Agent
End-User

4. Field Use

Figure 3.5. Life cycle of a Disciple agent for evidence-based reasoning.

where, based on the specification of the type of problems to be solved and the agent to be
built, the developer and the knowledge engineer may decide that some extensions of the
Disciple shell may be necessary or useful. It is through such successive extensions during
the development of Disciple agents for various applications that the current version of the
Disciple shell for evidence-based reasoning problems (which includes the EBR knowledge
base) has emerged.
The next stage is agent teaching by the subject matter expert and the knowledge
engineer, supported by the agent itself, which simplifies and speeds up the knowledge
base development process (Tecuci et al., 2001; 2002b; 2005b). Once an operational agent is
developed, it is used for the education and training of the end-users, possibly in a
classroom environment.
The fourth stage is field use, where copies of the developed agent support users in their
operational environments. During this stage, an agent assists its user both in solving

13:44:04,
.004
 3.3. System Design and Development 93

problems and in collaborating with other users and their cognitive assistants. At the same
time, it continuously learns from this problem-solving experience by employing a form of
nondisruptive learning. In essence, it learns new rules from examples. However, because
there is no learning assistance from the user, the learned rules will not include a formal
applicability condition. It is during the next stage of after action review, when the user and
the agent analyze past problem-solving episodes, that the formal applicability conditions
are learned based on the accumulated examples.
In time, each cognitive assistant extends its knowledge with expertise acquired from its
user. This results in different agents and creates the opportunity to develop a more
competent agent by integrating the knowledge of all these agents. This can be accom-
plished by a knowledge engineer, with assistance from a subject matter expert, during the
next stage of knowledge integration. The result is an improved agent that may be used in a
new iteration of a spiral process of development and use.

3.3 AGENT DESIGN AND DEVELOPMENT USING LEARNING

TECHNOLOGY

In Section 3.1, we briefly illustrated a conventional scenario for building knowledge-based

agents. Now that we have reviewed several learning-based agent building tools, we can
illustrate the development of a knowledge-based agent using learning technology. While the
main design and development phases are, in essence, still those from Table 3.1, the way they
are performed is different. The tasks of the knowledge engineer are reduced because part of
them will be performed by the subject matter expert (e.g., explaining to the agent shell the
sequence of reasoning steps to solve a problem), and part by the agent shell (e.g., learning rules
from example reasoning steps rather than having them defined by the knowledge engineer).

3.3.1 Requirements Specification and Domain Understanding

Let us consider the development of an agent that will assist a PhD student in assessing a
potential PhD advisor, a case study that will be used throughout this book.
The agent should receive a hypothesis, such as “John Doe would be a good PhD advisor
for Bob Sharp,” and should return the probability that this hypothesis is true, such as “very
likely,” together with an easy-to-understand reasoning that has led this assessment.
To build such an agent, one needs first to understand the expertise domain. In this case, a
simple search on the Internet will lead to the discovery of many papers written on this subject
(see Figure 3.6), which may supplement the expertise of a specific subject matter expert.
Reading some of these papers, you will discover that assessing a PhD advisor is a very
complex task. Many questions need to be addressed whose answers need to be aggregated
to evaluate a potential PhD advisor. A few of these questions, identified by the knowledge
engineer and the subject matter expert (in such a case, an experienced professor), are
shown in Table 3.8.

3.3.2 Rapid Prototyping

The next agent development phase is to develop rapidly a prototype that can be validated
by the end-users.

13:44:04,
.004
94 Chapter 3. Methodologies and Tools

Figure 3.6. Understanding the expertise domain.

Table 3.8 Some Relevant Questions to Consider When Assessing a Potential PhD Advisor

(1) What is the reputation of the PhD advisor within the professional community at large?
(2) Does the advisor have many publications?
(3) Is his or her work cited?
(4) What is the opinion of the peers of this PhD advisor?
(5) What do the students think about this PhD advisor?
(6) Is the PhD advisor likely to remain on the faculty for the duration of your degree program?
(7) What is the placement record of the students of this PhD advisor? Where do they get jobs?
(8) Is the PhD advisor expert in your areas of interest?
(9) Does the PhD advisor publish with students?
(10) Does the PhD advisor have a research group or merely a string of individual students?
(11) Is the PhD advisor’s research work funded?

An analysis of the questions in Table 3.8 shows that some of them point to necessary
conditions that need to be satisfied by the PhD advisor, while others refer to various
desirable qualities. Which questions from Table 3.8 point to necessary conditions? The
answers to questions (6) and (8) need to be “yes” in order to further consider a potential
PhD advisor.
Now let us consider the desirable qualities of a PhD advisor revealed by the other
questions in Table 3.8. Some of these qualities seem to be more closely related than
others. It would be useful to organize them in classes of quality criteria. Could you identify
a class of related criteria? Questions (2), (3), (4), and (11) all characterize aspects of the
professional reputation of the advisor.
What might be other classes of related criteria suggested by the questions in Table 3.8?
Questions (7) and (9) characterize the results of the students of the PhD advisor, while
questions (5) and (10) characterize their learning experience.

13:44:04,
.004
3.3. System Design and Development 95

Based on these observations, we can develop the ontology of criteria from

Figure 3.7.
Each of the criteria from the right side corresponds to one of the questions in Table 3.8.
They are components of the higher-order criteria shown in the middle of the figure, which
are all components of a top-level criterion that characterizes the quality of the PhD
advisor. This is what is called a part-of hierarchy. All these individual criteria are all
instances of the “criterion” concept.
The preceding analysis suggests that, in order to assess a PhD advisor, one needs
to check that the advisor satisfies the necessary conditions and to assess his or her
advisor qualities.
The knowledge engineer and the subject matter expert need to develop a formal,
yet intuitive way of representing the assessment logic. This has to be natural enough,
such that subject matter experts who do not have knowledge engineering experience
are able to express how to solve different problems by themselves, with no or
limited support from knowledge engineers. But the assessment logic also needs to
be formal enough so that an agent can learn general rules from such problem-solving
examples.
A general problem-solving paradigm, called problem reduction and solution synthesis,
which has been illustrated in Section 2.2 and will be discussed in detail in Chapter 4,
satisfies both these requirements. It will be again illustrated in this case.
To clarify the reduction logic, the knowledge engineer and the subject matter expert
consider a particular hypothesis:

John Doe would be a good PhD advisor for Bob Sharp.

criterion

instance of

Q11 research funding criterion

has as criterion

Q2, 3, 4, 11 Q2 research publications criterion

professional
reputation criterion Q3
citations criterion
Q4
peer opinion criterion
has as criterion

has as criterion

PhD advisor Q7, 9

quality criterion Q9 publications with advisor criterion
quality of student
results criterion Q7
employers of graduates criterion
has as criterion

Q5, 10 Q10 research group status criterion

student learning
experience criterion Q5
student opinion criterion

Figure 3.7. Sample criteria ontology for assessing a PhD advisor.

13:44:04,
.004
96 Chapter 3. Methodologies and Tools

They express the hypothesis in natural language and select the phrases that may be
different for other similar hypotheses, such as the names of the advisor and student.
The selected phrases will appear in blue, guiding the agent to learn a general hypothesis
pattern:

?O1 would be a good PhD advisor for ?O2.

This top-level hypothesis will be successively reduced to simpler and simpler hypotheses,
guided by questions and answers, as shown in Figure 3.8 and discussed in this section.

John Doe would be a good PhD advisor for Bob Sharp.

Which are the necessary conditions?
Bob Sharp should be interested in an area of expertise of John Doe who should stay on
the faculty of George Mason University for the duration of the PhD dissertation of Bob
Sharp, and should have the qualities of a good PhD advisor.

Therefore, the initial hypothesis can be reduced to three simpler hypotheses:

Bob Sharp is interested in an area of expertise of John Doe.

John Doe will stay on the faculty of George Mason University for the duration of the PhD
dissertation of Bob Sharp.
John Doe would be a good PhD advisor with respect to the PhD advisor quality criterion.

John Doe would be a good PhD advisor for Bob Sharp.

Which are the necessary conditions?

Bob Sharp should be interested in an area of expertise of John Doe, who should
stay on the faculty of George Mason University for the duration of the PhD
dissertation of Bob Sharp and should have the qualities of a good PhD advisor.

Bob Sharp is John Doe will stay on the faculty John Doe would be a
interested in an of George Mason University for good PhD advisor with
area of expertise the duration of the PhD respect to the PhD
of John Doe. dissertation of Bob Sharp. advisor quality criterion.

Is Bob Sharp interested in an Which are the necessary quality criteria for a good PhD advisor?
area of expertise of John Doe?
professional reputation criterion, quality of student results criterion,
Yes, Artificial Intelligence. and student learning experience criterion.

It is certain that Bob Sharp John Doe would be a good John Doe would be a John Doe would be a good
is interested in an area of PhD advisor with respect good PhD advisor with PhD advisor with respect
expertise of John Doe. to the professional respect to the quality of to the student learning
reputation criterion. student results criterion. experience criterion.

Which is a professional reputation criterion?

research funding criterion

John Doe would be a good PhD

advisor with respect to the
research funding criterion.

Figure 3.8. Reduction logic for assessing a specific hypothesis.

13:44:04,
.004
3.3. System Design and Development 97

The reductions of the subhypotheses continues in the same way, until solutions are
obtained for them:

Bob Sharp is interested in an area of expertise of John Doe.

Is Bob Sharp interested in an area of expertise of John Doe?
Yes, Artificial Intelligence.

Therefore, one may conclude:

It is certain that Bob Sharp is interested in an area of expertise of John Doe.

Consider now the third subhypothesis of the initial hypothesis:

John Doe would be a good PhD advisor with respect to the PhD advisor quality criterion.

Its reduction is also guided by a question/answer pair:

Which are the necessary quality criteria for a good PhD advisor?
professional reputation criterion, quality of student results criterion, and student learning
experience criterion.

Therefore, the preceding hypothesis can be reduced to three simpler hypotheses:

John Doe would be a good PhD advisor with respect to the professional reputation
criterion.
John Doe would be a good PhD advisor with respect to the quality of student results
criterion.
John Doe would be a good PhD advisor with respect to the student learning experience
criterion.

Each of these subhypotheses can now be reduced to simpler hypotheses, each corres-
ponding to one of the elementary criteria from the right side of Figure 3.7 (e.g.,
research funding criterion). Since each of these reductions reduces a criterion to a
subcriterion, the agent could be asked to learn a general reduction pattern, as shown
in Figure 3.9.
Why is pattern learning useful? One reason is that the pattern can be applied
to reduce a criterion to its subcriteria, as shown in Figure 3.10. Additionally, as will
be illustrated later, the pattern will evolve into a rule that will automatically generate all
the reductions of criteria to their sub-criteria. If, instead of learning a pattern and

John Doe would be a good PhD ?O1 would be a good PhD

advisor with respect to the advisor with respect to the
professional reputation criterion. ?O2.
Pattern
learning
Which is a professional reputation criterion? Which is a ?O2?
research funding criterion ?O3

John Doe would be a good PhD ?O1 would be a good

advisor with respect to the PhD advisor with respect to
research funding criterion. the ?O3.

Figure 3.9. Pattern learning.

13:44:04,
.004
98 Chapter 3. Methodologies and Tools

John Doe would be a good PhD advisor with

respect to the PhD advisor quality criterion.

Which are the necessary quality criteria for a good PhD advisor?
professional reputation criterion, quality of student results criterion,
and student learning experience criterion.

John Doe would be a good John Doe would be a good John Doe would be a good
?O1 would be a good PhD advisor with respect to PhD advisor with respect PhD advisor with respect to
PhD advisor with the professional reputation to the quality of student the student learning
respect to the ?O2. criterion. results criterion. experience criterion.
Pattern instantiation
Which is a ?O2? Which is a quality of student results criterion? Which is a quality of student results criterion?
?O3 publications with advisor criterion employers of graduates criterion

?O1 would be a good John Doe would be a good PhD John Doe would be a good PhD
PhD advisor with advisor with respect to the advisor with respect to the
respect to the ?O3. publications with advisor criterion. employers of graduates criterion.

Figure 3.10. Uniform modeling through the instantiation of a learned pattern.

applying it, the user would manually define these reductions, then any syntactic differ-
ences between these reductions would lead to the learning of different rules. These
rules would only be superficially different, leading to an inefficient and difficult to
maintain agent.
After the top-level criterion (i.e., PhD advisor quality criterion) is reduced to a set of
elementary criteria, specific knowledge and evidence about the advisor need to be used
to evaluate John Doe with respect to each such elementary criterion. For example, the
following hypothesis will be evaluated based on favoring and disfavoring evidence from
John Doe’s peers:

John Doe would be a good PhD advisor with respect to the peer opinion criterion.

A learning agent shell for evidence-based reasoning already knows how to assess such
hypotheses based on evidence.
Through this process, the initial hypothesis is reduced to elementary hypotheses for
which assessments are made. Then these assessments are successively combined, from
bottom-up, until the assessment of the initial hypothesis is obtained, as illustrated in
Figure 3.11.
Notice at the bottom-right side of Figure 3.11 the assessments corresponding to the
subcriteria of the quality of student results criterion:

It is likely that John Doe would be a good PhD advisor with respect to the publications
with advisor criterion.
It is very likely that John Doe would be a good PhD advisor with respect to the employers
of graduates criterion.

These assessments are combined by taking their maximum, leading to an evaluation of

“very likely” for the quality of student results criterion:

It is very likely that John Doe would be a good PhD advisor with respect to the quality of
student results criterion.

13:44:04,
.004
3.3. System Design and Development 99

John Doe would be a good PhD advisor for Bob Sharp.

very likely

min
Which are the necessary conditions?
Bob Sharp should be interested in an area of expertise of John Doe, who should stay on
the faculty of George Mason University for the duration of the PhD dissertation of Bob
Sharp and should have the qualities of a good PhD advisor.

Bob Sharp is John Doe will stay on the faculty John Doe would be a
interested in an of George Mason University for good PhD advisor with
area of expertise of the duration of the PhD respect to the PhD
John Doe. dissertation of Bob Sharp. advisor quality criterion.
certain almost certain very likely

min
Is Bob Sharp interested in an Which are the necessary quality criteria for a good PhD advisor?
area of expertise of John Doe?
professional reputation criterion, quality of student results criterion,
Yes, Artificial Intelligence. and student learning experience criterion.

It is certain that Bob Sharp John Doe would be a good John Doe would be a John Doe would be a good
is interested in an area of PhD advisor with respect to good PhD advisor with PhD advisor with respect to
expertise of John Doe. the professional reputation respect to the quality of the student learning
criterion. student results criterion. experience criterion.
very likely very likely almost certain

max
Which is a quality of student results criterion? Which is a quality of student results criterion?
publications with advisor criterion employers of graduates criterion

John Doe would be a good PhD John Doe would be a good PhD
advisor with respect to the advisor with respect to the
publications with advisor criterion. employers of graduates criterion.
likely very likely

Figure 3.11. Reduction and synthesis tree for assessing a specific hypothesis.

Then this assessment is combined with the assessments corresponding to the other major
criteria (very likely for the professional reputation criterion, and almost certain for the student
learning experience criterion), through a minimum function (because they are necessary
conditions), to obtain the assessment very likely for the PhD advisor quality criterion.
Finally, consider the assessments of the three subhypotheses of the top-level hypothesis:

It is certain that Bob Sharp is interested in an area of expertise of John Doe.

It is almost certain that John Doe will stay on the faculty of George Mason University for
the duration of the PhD dissertation of Bob Sharp.
It is very likely that John Doe would be a good PhD advisor with respect to the PhD
advisor quality criterion.

These assessments are combined by taking their minimum, leading to the following
assessment of the initial hypothesis:

It is very likely that John Doe would be a good PhD advisor for Bob Sharp.

Could you justify the preceding solution synthesis function? We used minimum because
each of the three subhypotheses of the initial hypothesis corresponds to a necessary
condition. If any of them has a low probability, we would like this to be reflected in the
overall evaluation.

13:44:04,
.004
 100 Chapter 3. Methodologies and Tools

Notice that, at this point, the knowledge engineer and the subject matter expert have
completely modeled the assessment of the specific hypothesis considered. This is the most
creative and the most challenging part of developing the agent. Once such a model for
assessing hypotheses (or solving problems, in general) is clarified, the agent can be rapidly
prototyped by modeling a set of typical hypotheses. The rest of the agent development
process consists of developing its knowledge base so that the agent can automatically
assess other hypotheses. The knowledge base will consist of an ontology of domain
concepts and relationships and of problem/hypothesis reduction and solution synthesis
rules, as was discussed in Section 1.6.3.1 and illustrated in Figure 1.15 (p. 38). As will be
discussed in the following, the way the preceding assessments were modeled will greatly
facilitate this process.

3.3.3 Ontology Design and Development

We want the developed agent to generate by itself reasoning trees like that shown in
Figure 3.11. What knowledge does it need in its ontology to perform this kind of reasoning?
Consider, for example, the reasoning step from the bottom-left of Figure 3.8, shown
again in the left-hand side of Figure 3.12. What knowledge does the agent need in order to
answer the question from this reasoning step? It needs the knowledge from the right-hand
side of Figure 3.12.
But this is just an example. We want the agent to be able to answer similar questions,
corresponding to similar hypotheses. Therefore, the ontological knowledge from the right-
hand side of Figure 3.12 is just a specification for the ontological knowledge needed by
the agent.
What other concepts and instances should we add to the ontology, based on the
specification in Figure 3.12? We would obviously need to consider other areas of expertise,
since we want to develop a general advisor assistant, capable of assessing advisors from
different disciplines. Thus the ontology can be expanded as shown in Figure 3.13.
The knowledge engineer and the subject matter expert will consider all the reasoning
steps from the developed reasoning trees and will correspondingly develop the ontology of
the agent.
What resources could be used to develop the ontology? Obvious resources are the
many foundational and utility ontologies for the Semantic Web that were discussed in
Section 3.2.2.

Bob Sharp is interested in

area of expertise
an area of expertise of
John Doe. Bob Sharp
instance of
Is Bob Sharp interested in an is interested in
area of expertise of John Doe? Artificial Intelligence
Yes, Artificial Intelligence.
is expert in

It is certain that Bob Sharp John Doe

is interested in an area of
expertise of John Doe.

Figure 3.12. Modeling-based ontology specification.

13:44:04,
.004
 3.3. System Design and Development 101

The goal of this phase is to develop an ontology that is as complete as possible. This will
enable the agent to learn reasoning rules based on the concepts and the features from the
ontology, as will be briefly illustrated in the following.

3.3.4 Rule Learning and Ontology Refinement

During this phase, the subject matter expert will interact with the learning agent shell to
explain to it a specific reasoning tree. From each reasoning step, the agent will learn a
general reasoning rule, as shown in Figure 3.14.
Figure 3.15 illustrates the learning of a general hypothesis reduction rule from a specific
hypothesis reduction step. Notice that the specific instances from the example (e.g., Bob
Sharp or John Doe) are generalized to variables (i.e., ?O1, ?O2). In essence, the rule

area of expertise

subconcept of

Mathematics Computer Science Biology

instance of
Bob Sharp
is interested in
Artificial Intelligence Software Engineering
is expert in
John Doe

Figure 3.13. Expanded ontology based on the specification from Figure 3.12.

John Doe would be a good

Learned rule
PhD advisor for Bob Sharp.

Which are the necessary conditions?

Bob Sharp should be interested in an area of expertise of John Doe, who should
stay on the faculty of George Mason University for the duration of the PhD
dissertation of Bob Sharp and should have the qualities of a good PhD advisor.

John Doe will stay on the faculty of John Doe would be a

Bob Sharp is interested George Mason University for the good PhD advisor with
in an area of expertise duration of the PhD dissertation of respect to the PhD
of John Doe. Bob Sharp. advisor quality criterion. Learned rule

Is Bob Sharp interested in an Which are the necessary quality criteria for a good PhD advisor?
area of expertise of John Doe?
professional reputation criterion, quality of student results criterion,
Yes, Artificial Intelligence. and student learning experience criterion.

It is certain that Bob Sharp is John Doe would be a good John Doe would be a good John Doe would be a good
interested in an area of PhD advisor with respect to PhD advisor with respect PhD advisor with respect to
expertise of John Doe. the professional reputation to the quality of student the student learning
criterion. results criterion. experience criterion.

Which is a professional
reputation criterion?
research funding criterion

John Doe would be a

good PhD advisor with
respect to the research
funding criterion.

Figure 3.14. Rules learning.

13:44:04,
.004
102 Chapter 3. Methodologies and Tools

Bob Sharp is
interested an area of
expertise of John Doe.

Is Bob Sharp interested in an

area of expertise of John Doe?
Yes, Artificial Intelligence.

It is certain that Bob Sharp

is interested in an area of
expertise of John Doe.

Figure 3.15. Learning a reduction rule from a reduction step.

indicates the conditions that need to be satisfied by these variables so that the IF
hypothesis can be assessed as indicated in the example. For example, ?O1 should be a
PhD student or possibly a person (the agent does not yet know precisely what concept to
apply because the rule is only partially learned), ?O1 should be interested in ?O3, and ?O3
should be Artificial Intelligence or possibly any area of expertise.
The way the question and its answer from the reduction step are formulated is very
important for learning. What could the agent learn if the answer were simply “yes”? The
agent would only be able to learn the fact that “Bob Sharp is interested in an area of
expertise of John Doe.” By providing an explanation of why this fact is true (“Yes, Artificial
Intelligence” meaning: “Yes, because Bob Sharp is interested in Artificial Intelligence which is
an area of expertise of John Doe”), we help the agent to learn a general rule where it will
check that the student ?O1 is interested in some area ?O3, which is an area of expertise of
the advisor ?O2. This is precisely the condition of the rule that can be easily verified
because this type of knowledge was represented in the ontology, as discussed previously
and shown in Figure 3.12.
What is the difference between the pattern learning illustrated in Figure 3.9 and the rule
learning illustrated in Figure 3.15? The difference is in the formal applicability condition of
the rule, which restricts the possible values of the rule variables and allows the automatic
application of the rule in situations where the condition is satisfied. A learned pattern,
such as that from Figure 3.9, cannot be automatically applied because the agent does not
know how to instantiate its variables correctly. Therefore, its application, during the
modeling phase, is controlled by the user, who selects the instances of the variables.
A remarkable capability of the agent is that it learns a general rule, like the one in
Figure 3.15, from a single example rather than requiring the rule to be manually developed
by the knowledge engineer and the subject matter expert. Rule learning will be discussed
in detail in Chapter 9.

13:44:04,
.004
3.3. System Design and Development 103

As indicated in the preceding, the rule in Figure 3.15 is only partially learned, because
instead of an exact applicability condition, it contains an upper and a lower bound for this
condition. The upper bound condition (represented as the larger ellipse from Figure 3.15)
corresponds to the most general generalization of the example (represented as the point
from the center of the two ellipses) in the context of the agent’s ontology, which is used as
a generalization hierarchy for learning. The lower bound condition (represented as the
smaller ellipse) corresponds to the least general generalization of the example.
The next phase is to refine the learned rules and, at the same time, test the agent with
new hypotheses. Therefore, the subject matter expert will formulate new hypotheses,
for example:

Dan Smith would be a good PhD advisor for Bob Sharp.

Using the learned rules, the agent will automatically generate the reasoning tree from
Figure 3.16. Notice that, in this case, the area of common interest/expertise of Dan Smith
and Bob Sharp is Information Security. The expert will have to check each reasoning step.
Those that are correct represent new positive examples that are used to generalize the
lower bound conditions of the corresponding rules. Those that are incorrect are used as
negative examples. The expert will interact with the agent to explain to it why a reasoning
step is incorrect, and the agent will correspondingly specialize the upper bound condition
of the rule, or both conditions. During the rule refinement process, the rule’s conditions
will converge toward one another and toward the exact applicability condition. During this
process, the ontology may also be extended, for instance to include the new concepts used
to explain the agent’s error. Rule refinement will be discussed in detail in Chapter 10.

Dan Smith would be a good PhD advisor for Bob Sharp.

Which are the necessary conditions?

Bob Sharp should be interested in an area of expertise of Dan Smith, who should
stay on the faculty of George Mason University for the duration of the PhD
dissertation of Bob Sharp and should have the qualities of a good PhD advisor.

Bob Sharp is Dan Smith will stay on the faculty Dan Smith would be a
interested in an of George Mason University for good PhD advisor with
area of expertise the duration of the PhD respect to the PhD advisor
of Dan Smith. dissertation of Bob Sharp. quality criterion.

Is Bob Sharp interested in an Which are the necessary quality criteria for a good PhD advisor?
area of expertise of Dan Smith?
professional reputation criterion, quality of student results criterion,
Yes, Information Security. and student learning experience criterion.

It is certain that Bob Sharp Dan Smith would be a good Dan Smith would be a good Dan Smith would be a good
is interested in an area of PhD advisor with respect to PhD advisor with respect to PhD advisor with respect to
expertise of Dan Smith. the professional reputation the quality of student the student learning
criterion. results criterion. experience criterion.

Which is a professional reputation criterion?

research funding criterion

Dan Smith would be a good PhD advisor with

respect to the research funding criterion.

Figure 3.16. Reasoning tree automatically generated by the agent.

13:44:04,
.004
 104 Chapter 3. Methodologies and Tools

3.3.5 Hierarchical Organization of the Knowledge Repository

Figure 3.17 illustrates the organization of a knowledge repository developed with the
Disciple-EBR learning agent shell, which was trained to assess different types of hypoth-
eses in different domains.
The top of the knowledge repository is the Shared KB, which contains general know-
ledge for evidence-based reasoning applicable in all the domains.
Under the Shared KB are Domain KBs, each corresponding to a different application
domain. One of these domain KBs will correspond to PhD advisor assessment. This PhD
advisor domain KB will contain general knowledge for assessing a PhD advisor, such as the
criteria for assessing a PhD advisor shown in Figure 3.7, and the learned reasoning rules. This
knowledge is, in principle, applicable for assessing any PhD advisor from any university.
Let us assume that we are using this agent to assess PhD advisors at three different
universities: George Mason University, Old Dominion University, and the University of
Virginia. How many scenario knowledge bases do we need and what knowledge will be
stored in each of them?
For each university there is at least one Scenario KB containing knowledge specific to
that university, such as individual professors and their characteristics.
When we need to assess PhD advisors at a specific university, the agent will use only the
corresponding Scenario KB, the Domain KB for PhD advisor assessment, and the Shared
KB for evidence-based reasoning.
The left-hand side of Figure 3.17 shows a reasoning step in the assessment of John Doe.
In which knowledge base will the following elements be represented: John Doe, Bob Sharp,
Artificial Intelligence, and certain?
John Doe and Bob Sharp are individuals at a specific university and are therefore
represented as specific instances in the Scenario KB corresponding to that university.
Artificial Intelligence is an instance to be used in reasoning in the context of any
university and is therefore represented as a generic instance in the Domain KB corres-
ponding to the assessment of PhD advisors.

Bob Sharp is Shared KB Domain-independent

interested in an Evidence-based concepts, generic
area of expertise reasoning knowledge
kn instances, and rules
of John Doe.

Is Bob Sharp interested

in an area of expertise of
Domain-
John Doe?
generic Domaini KB Domain KB dependent
Yes, Artificial Intelligence. concepts,
instance Domain knowledge … Domain knowledge
generic
adv
(PhD advisor) (Insider threat)
It is certain that Bob instances,
Sharp is interested in and rules
the area of expertise
of John Doe.
specific
Scenario KB Domain-
instance i KB
Scenario
dependent
Scenario knowledge … Scenario knowledge
specific
(George Mason Univ.) (Univ. of Virginia)
instances
Figure 3.17. Organization of the knowledge repository of a Disciple-EBR agent.

13:44:04,
.004
 3.3. System Design and Development 105

Finally, certain is a constant to be used in evidence-based reasoning in any application

domain and is therefore represented in the Shared KB.

3.3.6 Learning-based Design and Development Phases

Figure 3.18 summarizes the main phases of agent development when using learning
technology. The first phase is agent specification, during which a knowledge engineer and
a subject matter expert define the types of hypotheses to assess (or problems to be solved)
by the agent. For example, the hypotheses might be to assess potential PhD advisors.
The second phase is rapid prototyping, where the knowledge engineer supports the
subject matter expert to develop reasoning trees for specific but representative hypotheses
(or problems). An example of a reasoning tree is the one from Figure 3.8. In this reasoning
tree, a complex hypothesis is assessed as illustrated in Figure 3.11 by:

Agent PhD advisor

specification assessment

P1 S1
P1 S1
QuestionP1 S1
Question
Answer
Question
Answer
Question
Answer
Question

Rapid
Answer
Question
Answer
Answer IF the problem to solve is P1g
P1 S1 P 1 S1 IF the problem to solve is P1g
1 1 n n IF the problem to solve is P1g
THEN solve its sub-problemsIF the problem to solve is P1g
P11 S11 Pn1 S1n
prototyping
THEN solve its sub-problems
P11 S11 Pn1 S1n 1 … P
P1g 1
THEN solve its sub-problems
Question
Png
1 … P 1
THEN solve its sub-problems
1g Png
1 … P 1
Question
Answer Examples: E1, … , Ek 1g Png
1 … P1
ng
Question
Answer Examples: E1, … , Ek 1g
Question
Answer Examples: E1, … , Ek
Question Examples: E1, … , Ek
Answer
Question
Answer
P21 S2 Answer
2 2
1 Pm Sm
P21 S21 2
Pm 2
Sm
P21 S21 2
Pm 2
Sm

actor

Ontology person
organization

development university employee student

graduate educational
faculty member staff member company
student organization

instructor professor PhD advisor PhD student college university

actor

IF the problem to solve is P1g

IF the problem to solve is P1g person organization
IF the problem to solve is P1g
Condition IF the problem to solve is P1g
Condition IF the problem to solve is P1g

Rule learning and Except-WhenCondition

…
…
Condition
Except-When Condition
Condition
Except-When Condition
Except-When…Condition
IF the problem to solve is P1g
Condition
Except-WhenCondition
Condition
university employee student

Except-When…Condition Except-When Condition

ontology refinement
educational
Except-When…Condition Except-When Condition
…Condition faculty member staff member organization
company
THEN solve its sub-problems Except-When
Except-When Condition graduate
1 …solve
THEN
P1g
its sub-problems
P11ng solve Except-When Condition
THEN its sub-problems student
P1g … THEN
P11ng solve its sub-problems
P1g … THEN
P 1ng solve
1
its sub-problems
P1g … THEN
P11ng solve its sub-problems instructor professor PhD advisor
P1g … P11ng 1
P1g … P ng college university

assistant associate full

MS student PhD student
professor professor professor

P1 S1
P1 S1
QuestionP1 S1
Question
Answer IF the problem to solve is P1g
Question
Answer IF the problem to solve is P1g
Question
Answer
Question IF the problem to solve is P1g
IF the problem to solve is P1g

Agent use and

Answer
Question THEN solve its sub-problems
Answer THEN solve its sub-problems
Answer
1 … P
P1g 1
THEN solve its sub-problems
P1 S1 P 1 S1 Png
1 … P 1
THEN solve its sub-problems
1 1 n n 1g Png
1 … P 1
Examples: E1, … , Ek 1g Png
1 … P1
P11 S11 Pn1 S1n
personalization
ng
Examples: E1, … , Ek 1g
1 P1 S1 Pn1 S1n Examples: E1, … , Ek
Examples: E1, … , Ek
Question 1
Question
Answer
Question
Answer
Question
Answer
Question
Answer
Question
Answer
P2 S2 Answer
2
1 P1 S2 m m
P21 S21 2
Pm 2
Sm
P21 S21 2
Pm 2
Sm
actor

person
IF the problem to solve is P1g organization
IF the problem to solve is P1g
IF the problem to solve is P1g
Condition IF the problem to solve is P1g
Condition IF the problem to solve is P1g
Condition IF the problem to solve is P1g university employee student
Except-When Condition Condition IF the problem to solve is P1g
IF the problem to solve is P1g
Agent
… Except-When
… Except-When
Condition Condition
Condition Condition educational
… Except-When
Except-When Condition
… Except-When
Condition Condition
Condition faculty member staff member organization
company
Except-When Condition Condition
… Except-When
Except-When Condition Condition graduate undergraduate
… Except-When Condition

optimization
THEN solve its sub-problemsExcept-When Condition
… Except-When Condition
student student
THEN solve its sub-problems Except-When Condition
…
1 … P1
P1g THEN solve its sub-problemsExcept-When Condition
Png
1 … P 1
THEN solve its sub-problemsExcept-When Condition instructor professor PhD advisor
1g Png
1 … P 1
THEN solve its sub-problemsExcept-When Condition
1g Png
1 … P 1
THEN solve its sub-problems BS student college university
1g Png
1 … P 1
THEN solve its sub-problems
1g Png
1 … P 1
THEN solve its sub-problems
1g Png
1 … P
1g
1
ng
1 1 assistant associate full
P1g … Png professor professor professor
MS student PhD student

Figure 3.18. Main phases of agent development when using learning technology.

13:44:04,
.004
106 Chapter 3. Methodologies and Tools

 Successively reducing it, from the top down, to simpler and simpler hypotheses
 Assessing the simplest hypotheses
 Successively combining, from the bottom up, the assessments of the subhypotheses,
until the assessment of the top-level hypothesis is obtained

During this phase, the agent also learns reduction patterns.

The next phase is that of ontology development. The guiding question is: What are
the domain concepts, relationships, and instances that would enable the agent to
automatically generate the reasoning trees developed during rapid prototyping, as well
as similar ones? From each reasoning step, the knowledge engineer and the subject
matter expert identify the ontology elements mentioned in it, as illustrated in Figure 3.12.
Such semantic network fragments represent a specification of the needed ontology.
In particular, this fragment suggests the need for the ontology fragment shown in
Figure 3.13. Based on such specifications, and using the ontology development tools
of the Disciple-EBR shell, the knowledge engineer develops an ontology that is suffi-
ciently complete. As part of ontology development, a knowledge engineer may reuse
concepts and relationships from previously developed ontologies, including those on
the Semantic Web.
The next phase in agent development is that of rule learning and ontology refinement.
From each hypothesis reduction step of a reasoning tree developed during rapid proto-
typing, the agent will learn a general hypothesis reduction rule by using the ontology as a
generalization hierarchy. This was illustrated in Figures 3.14 and 3.15. During this phase,
the expert teaches the agent to assess other similar hypotheses. The expert instantiates a
learned hypothesis pattern, such as “Dan Smith would be a good PhD advisor for Bob
Sharp,” and the agent automatically generates a reasoning tree, by applying the learned
rules, as was illustrated in Figure 3.16. The expert then critiques the agent’s reasoning,
guiding it in refining the rules.
Now the assistant is ready for the typical end-user, as part of the next phase, agent use
and personalization. Typically, the user will specify the hypothesis to assess by simply
instantiating a corresponding pattern. Then the agent will automatically generate a reduc-
tion tree like the one in Figure 3.16 by automatically applying the learned reduction rules.
This tree reduces the top-level hypothesis to elementary hypotheses to be directly assessed
based on knowledge and evidence. The assessments of these elementary hypotheses are
then automatically combined, from the bottom up, until the probability of the initial
hypothesis is obtained.
The end-user may also assess an entirely new hypothesis, but in this case he or she also
needs to indicate its decomposition, either to elementary hypotheses or to known hypotheses.
So when does the additional learning take place? It may be the case that the agent does
not know how to decompose a specific hypothesis. In such a case, the user can indicate
the desired decomposition, which will be used in the current analysis. As a result, the
agent will automatically learn a decomposition pattern, such as the one from the right-
hand side of Figure 3.9.
As mentioned previously, an important difference between a learned rule and a learned
pattern concerns its use in problem solving. A learned rule is always automatically applied,
while a pattern will be proposed to the user who may decide to select it and instantiate it
appropriately. If this is done, the corresponding instances will be added to the pattern.
Thus this type of learning is much simpler for the end-user, who only needs to specify, in
natural language, the decomposition of a hypothesis into simpler hypotheses.

13:44:04,
.004
 3.4. Hands On 107

Periodically, the agent can undergo an optimization phase, which is the last phase in
Figure 3.18. During this phase, the knowledge engineer and the subject matter expert will
review the patterns learned from the end-user, will learn corresponding rules from them,
and will correspondingly refine the ontology. The current version of the Disciple-EBR shell
reapplies its rule and ontology learning methods to do this. However, improved methods
can be used when the pattern has more than one set of instances, because each represents
a different example of the rule to be learned. This is part of future research.
Figure 3.19 compares the conventional knowledge engineering process of developing a
knowledge-based agent (which was discussed in Section 3.1) with the process discussed in
this section, which is based on the learning agent technology.
The top part of Figure 3.19 shows the complex knowledge engineering activities that are
required to build the knowledge base. The knowledge engineer and the subject matter
expert have to develop a model of the application domain that will make explicit the way
the subject matter expert assesses hypotheses. Then the knowledge engineer has to
develop the ontology. He or she also needs to define general hypotheses decomposition
rules and to debug them, with the help of the subject matter expert.
As shown at the bottom of Figure 3.19, each such activity is replaced with an equivalent
activity that is performed by the subject matter expert and the agent, with limited assistance
from the knowledge engineer. The knowledge engineer still needs to help the subject matter
expert to define a formal model of how to assess hypotheses and to develop the ontology.
After that, the subject matter expert will teach the agent how to assess hypotheses, through
examples and explanations, and the agent will learn and refine the rules by itself.
The next chapters discuss each phase of this process in much more detail.

3.4 HANDS ON: LOADING, SAVING, AND CLOSING

KNOWLEDGE BASES

The knowledge bases developed with Disciple-EBR are located in the repository folder,
which is inside the installation folder. As shown in Figure 3.17, the knowledge bases are

Subject
ct Dialogue
Conventional Knowledge
Knowledge Matter
er
rtt
Expert Engineer
Engineering Programming
Model Define Verify and
Understand Develop
problem reasoning update rules
domain ontology
solving rules and ontology

D
Develop Define concepts Provide and Analyze
Explain
reasoning with hierarchical explain agent’s
errors
model organization examples solutions
Knowledge
Engineer Assist in Extend ontology Learn Learn
Refine
(support role)) solution based on modeled reasoning ontology
rules
development solutions rules elements
Learning-based Subject Mixed-Initiative
Knowledge Matter Dialogue
COGNITIVE ASSISTANT
Engineering Expert

Figure 3.19. Conventional knowledge engineering versus learning-based knowledge engineering.

13:44:04,
.004
108 Chapter 3. Methodologies and Tools

organized hierarchically, with the knowledge base for evidence-based reasoning at the top
of the hierarchy. The user cannot change this knowledge base, whose knowledge elements
are inherited in the domain and scenario knowledge bases. From the user’s point of view,
each knowledge base consists of a top-level domain part (which contains knowledge
common to several applications or scenarios in a domain) and one scenario part (con-
taining knowledge specific to a particular application or scenario). As illustrated in
Figure 3.17, there can be more than one scenario under a domain. In such a case, the
domain and each of the scenarios correspond to a different knowledge base. Loading,
saving, or closing a scenario will automatically load, save, or close both the scenario part
and the corresponding domain part of the knowledge base.
Loading and selecting a knowledge base are described in Operation 3.1 and illustrated
in Figure 3.20.

Operation 3.1. Load and select a knowledge base

 In the System Workspace Manager, click on the knowledge base icon containing the
plus sign (+).
 The Knowledge Manager window opens, showing all the knowledge bases from the
repository.
 Click on the plus sign (+) of the domain knowledge base to be loaded, to display all its
scenario knowledge bases.
 Click on the scenario knowledge base to be loaded.
 Click on the Select button. This both loads the scenario and domain KBs and selects
them as the current ones to work with. Their names will be shown in bold in the
Knowledge Manager window.

1. Click on this icon to load a KB

2. System displays all KBs

3. Click on + to expand the KB to be loaded

5. Click on Select

4. Click on Scen

Figure 3.20. Loading and selecting a knowledge base.

13:44:04,
.004
3.4. Hands On 109

Once a knowledge base is selected, you can invoke different modules of Disciple-EBR to
use it. Each module is accessible in a specific workspace. As illustrated in Figure 3.21,
there are three workspaces:

Evidence workspace: “Evidence: repository\KB-name\Scen”

Scenario workspace: “repository\KB-name\Scen”
Domain workspace: “repository\KB-name”

The user can switch between the workspaces to use the corresponding modules. For
example, you must switch to the Evidence workspace to work with evidence items. Then,
to save the knowledge base, you must switch to the Scenario workspace.
The user should work with only one set of the three workspaces at a time (correspond-
ing to the same KB). Therefore, you should close the workspaces corresponding to a
knowledge base (by clicking on the icon with the minus sign [–]) before opening the
workspaces corresponding to another knowledge base.
The steps to save all the knowledge bases loaded in memory are described in Operation
3.2 and illustrated in Figure 3.22.

Operation 3.2. Save all the knowledge bases

 Select the Scenario workspace (“repository\KB-name\Scen”).
 Select the System menu.
 Select Save All.

You may click on this icon to close a workspace

Scenario workspace

Domain workspace
Evidence workspace

Figure 3.21. Workspaces in Disciple-EBR.

. Select Scenario workspace

2. Scenario workspace

3. Select System
Save All

Figure 3.22. Saving all the knowledge bases.

13:44:04,
.004
110 Chapter 3. Methodologies and Tools

It is highly recommended to have only one knowledge base loaded in memory.

Therefore, before loading a new knowledge base, you should close all the opened
ones by following the instructions described in Operation 3.3 and illustrated in
Figure 3.23.

Operation 3.3. Close a knowledge base

 Click on the knowledge base icon containing the plus sign (+) situated at the right of
the workspace selector.
 The Knowledge Manager window opens, showing all the knowledge bases.
 Click on the scenario knowledge base to be closed.
 Click on the Close button in the right side of the window.
 Click on the X button to close the Knowledge Manager window.

The predefined knowledge base “00-Reference-KB” contains general knowledge for

evidence-based reasoning, an empty domain, and an empty scenario. Users can create
their knowledge bases as renamed copies of “00-Reference-KB,” as indicated in Operation
3.4. They should never work with “00-Reference-KB,” which should be kept as a reference
knowledge base.

Operation 3.4. Create a user knowledge base

 Open the “repository” folder from the “installation” directory in Windows Explorer.
 Make a copy of the entire “00-Reference-KB” folder and give it a new name.
 Use the Disciple-EBR modules to develop the newly created knowledge base.

. Click on this icon

2. System displays all KBs

3. Click on the Scen KB to be closed

5. Click on the x button to close the window

4. With Scen selected, click on Close

Figure 3.23. Closing a knowledge base.

13:44:04,
.004
 3.6. Project Assignment 2 111

3.5 KNOWLEDGE BASE GUIDELINES

The following are knowledge engineering guidelines for knowledge base development.

Guideline 3.1. Work with only one knowledge base loaded in memory
To maintain the performance of the Disciple-EBR modules, work with only one knowledge
base loaded in memory. Therefore, close all the knowledge bases before loading a new one.
If several knowledge bases are loaded, work only with the set of three workspaces
corresponding to the knowledge base you are currently using. Close all the other
workspaces.

Guideline 3.2. Create a knowledge base and save successive versions

A user can introduce errors into the knowledge base while developing it. When Disciple-
EBR attempts to load a knowledge base, it makes a rigorous check of its correctness and
will not load the knowledge base if it contains errors. It is therefore important to save
successive versions of the knowledge base being developed in order to fall back to a
previous version in case the latest one contains errors.
Create your knowledge base by making a copy of “00-Reference-KB” and giving it a new
name, for instance “WA.” From now on, you will work only with “WA,” but as you develop it,
you will have to save successive copies of it with different names, as explained in the following.
Suppose that you have developed “WA” to contain part of the modeling for a hypothesis.
Save “WA” in Disciple-EBR, then using Windows Explorer make a copy of it, and rename the
copy as “WA-1m.” Continue working with “WA” and expand the modeling of the hypothesis.
Save and make a copy of “WA.” Then rename the copy as “WA-2m,” and so on.
Through such a process, you will save a sequence of knowledge bases: “WA-1m,” “WA-
2m,” “WA-3o,” “WA-4o,” and so on, each corresponding to a given stage in your develop-
ment of “WA.” In this way, if your “WA” knowledge base is damaged for any reason, you can
always resume from the most recently saved version, as illustrated in the following scenario:

 “WA” has errors, and the most recently saved version is “WA-4o.”
 Delete “WA” in Windows Explorer, copy “WA-4o,” and rename this copy as “WA.”
 Continue with the development of “WA.”

3.6 PROJECT ASSIGNMENT 2

Finalize the project team, specify the type of hypotheses to be analyzed by the agent to
be developed, and study the application domain. Prepare a short presentation of the
following:

 The application domain of your agent and why your agent is important.
 A bibliography containing the expertise domain and your current familiarity with the
domain, keeping in mind that you should choose a domain where you already are or
could become an expert without investing a significant amount of time.
 Three examples of hypotheses and a probability for each.

13:44:04,
.004
 112 Chapter 3. Methodologies and Tools

3.7 REVIEW QUESTIONS

3.1. Which are the main phases in the development of a knowledge-based agent?

3.2. Which are the required qualifications of a knowledge engineer?

3.3. Which are the required qualifications of a subject matter expert?

3.4. Briefly define “verification,” “validation,” and “certification.”

3.5. Consider the fact that a knowledge-based agent may need to have hundreds or
thousands of rules. What can be said about the difficulty of defining and refining
these rules through the conventional process discussed in Section 3.1.4?

3.6. Use the scenario from Section 3.1 to illustrate the different difficulties of building a
knowledge-based agent discussed in Section 1.6.3.1.

3.7. What is an expert system shell?

3.8. Which are different types of expert system shells?

3.9. What is a learning agent shell?

3.10. What is a learning agent shell for evidence-based reasoning?

3.11. What is the organization of the knowledge repository of a learning agent shell for
evidence-based reasoning?

3.12. What is the difference between a specific instance and a generic instance? Provide
an example of each.

3.13. Are there any mistakes in the reasoning step from Figure 3.24 with respect to the
goal of teaching the agent? If the answer is yes, explain and indicate corrections.

3.14. Which are the main stages of developing a knowledge-based agent using learning
agent technology?

3.15. Compare the manual knowledge engineering process of developing a knowledge-

based agent, as described in Section 3.1, with the process using learning agent
technology, described in Section 3.3.

John Doe will stay on the faculty of

George Mason University for the duration
of the PhD dissertation of Bob Sharp.

Is John Doe likely to stay on the faculty of George Mason

University for the duration of the PhD dissertation of Bob Sharp?
Yes

It is almost certain that John Doe is likely to stay

on the faculty of George Mason University for the
duration of the PhD dissertation of Bob Sharp.

Figure 3.24. Reduction step.

13:44:04,
.004
 Modeling the Problem-Solving
4 Process

4.1 PROBLEM SOLVING THROUGH ANALYSIS AND SYNTHESIS

Analysis and synthesis, introduced in Section 1.6.2, form the basis of a general divide-and-
conquer problem-solving strategy that can be applied to a wide variety of problems. The
general idea, illustrated in Figure 4.1, is to decompose or reduce a complex problem P1 to
n simpler problems P11, P12, . . . , P1n, which represent its components. If we can then find
the solutions S11, S12, . . . , S1n of these subproblems, then these solutions can be combined
into the solution S1 of the problem P1.
If any of the subproblems is still too complex, it can be approached in a similar way, by
successively decomposing or reducing it to simpler problems, until one obtains problems
whose solutions are known, as illustrated in Figure 4.2.
Figures 4.3 and 4.4 illustrate the application of this divide-and-conquer approach to
solve a symbolic integration problem. Notice that each reduction and synthesis operation
is justified by a specific symbolic integration operator.

4.2 INQUIRY-DRIVEN ANALYSIS AND SYNTHESIS

Cognitive assistants require the use of problem-solving paradigms that are both natural
enough for their human users and formal enough to be automatically executed by the
agents. Inquiry-driven analysis and synthesis comprise such a problem-solving paradigm
where the reduction and synthesis operations are guided by corresponding questions and
answers. The typical questions are those from Rudyard Kipling’s well-known poem “I Keep
Six Honest . . .”

P1 S 1

P11 S11 P12 S12 … P1n S1n

Figure 4.1. Problem reduction (decomposition) and solutions synthesis (composition).

113

13:52:06,
.005
114 Chapter 4. Modeling the Problem-Solving Process

I keep six honest serving-men

(They taught me all I knew);
Their names are What and Why and When
And How and Where and Who.

A complex problem P1 is P1 S1
solved by:
• Successively reducing
it to simpler and
simpler problems;
P11 S11 P12 S12 … P1n S1n
• Finding the solutions
of the simplest
problems;
• Successively
combining these P1n1 S1n1 … P1nm S1nm
solutions, from
bottom up, until the
solution of the initial
problem is obtained
(synthesized). P1n11 S1n11 … P1n1q S1n1q

Figure 4.2. Problem solving through analysis and synthesis.

Figure 4.3. Reduction of a symbolic integration problem.

13:52:06,
.005
4.2. Inquiry-Driven Analysis and Synthesis 115

Figure 4.4. Symbolic integration through analysis and synthesis.

Problem 1

Question Q Question Q
Answer A Answer B

Problem 2 Problem 3 Problem 4 Problem 5

Figure 4.5. Inquiry-driven problem reduction.

We have already illustrated this paradigm in Sections 2.2 and 3.3.2. To better understand
it, let us consider the simple abstract example from Figure 4.5. To solve Problem 1, one
asks Question Q related to some aspect of Problem 1. Let us assume that there are two
answers to Q: Answer A and Answer B. For example, the question, “Which is a sub-
criterion of the quality of student results criterion?” has two answers, “publications with
advisor criterion” and “employers of graduates criterion.”
Let us further assume that Answer A leads to the reduction of Problem 1 to two simpler
problems, Problem 2 and Problem 3. Similarly, Answer B leads to the reduction of
Problem 1 to the other simpler problems, Problem 4 and Problem 5.
Let us now assume that we have obtained the solutions of these four subproblems. How
do we combine them to obtain the solution of Problem 1? As shown in Figure 4.6, first the

13:52:06,
.005
116 Chapter 4. Modeling the Problem-Solving Process

solutions to Problem 2 and Problem 3 are combined to obtain a solution to Problem

1 corresponding to Answer A (called Solution A). Also the solutions to Problem 4 and
Problem 5 are combined to obtain a solution to Problem 1 corresponding to Answer
B (called Solution B). Each of these two synthesis (or composition) operations is called a
reduction-level synthesis because it corresponds to a specific reduction of the top-level
problem. Second, we need to combine Solution A with Solution B into the final solution of
Problem 1. This synthesis operation is called problem-level synthesis because it corres-
ponds to all reductions of Problem 1.
Figures 4.7 and 4.8 illustrate inquiry-driven analysis and synthesis in the context of the
military center of gravity (COG) determination problem (Tecuci et al., 2005b; 2008b),
which will be discussed in Section 12.4. Figure 4.7 shows the reduction of a problem to
three simpler problems, guided by a question/answer pair.
Figure 4.8 illustrates the composition of the solutions of the three subproblems
into the solution of the problem. One aspect to notice is that the reduction-level
synthesis operation is guided by a question/answer pair, while the problem-level syn-
thesis operation is not. Thus the use of questions and answers is actually optional.

Problem 1
Solution
of Problem 1
Problem-
level
Question Q synthesis
Question Q
Answer A Answer B
Solution A Solution B
of Problem 1 of Problem 1

Reduction-
level
synthesis
Problem 2 Problem 3 Problem 4 Problem 5
Solution Solution Solution Solution
of Problem 2 of Problem 3 of Problem 4 of Problem 5

Figure 4.6. A more detailed view of the analysis and synthesis process.

Test whether President Roosevelt has the critical capability to maintain support.

Which are the critical requirements for President Roosevelt to maintain support?
President Roosevelt needs means to secure support from the government, means to
secure support from the military, and means to secure support from the people.

Test whether President Roosevelt Test whether President Roosevelt Test whether President Roosevelt
has means to secure support from has means to secure support from has means to secure support from
the government. the military. the people.

Figure 4.7. Illustration of a reduction operation in the COG domain.

13:52:06,
.005
4.2. Inquiry-Driven Analysis and Synthesis 117

Test whether President Roosevelt has the critical capability to maintain support.
President Roosevelt has the critical capability to maintain support because President
Roosevelt has means to secure support from the government, has means to secure
support from the military, and has means to secure support from the people.

critical capability is
an emerging property
Which are the critical requirements for President President Roosevelt has the critical capability to
Roosevelt to maintain support? President Roosevelt maintain support because President Roosevelt
needs means to secure support from the government, has means to secure support from the government,
means to secure support from the military, and means has means to secure support from the military, and
to secure support from the people. has means to secure support from the people.

Does President Roosevelt have the critical capability to maintain support?

Yes, because President Roosevelt has all the needed critical requirements.

Test whether President Roosevelt has Test whether President Roosevelt Test whether President Roosevelt
means to secure support from the has means to secure support from has means to secure support from
government. the military. the people.
President Roosevelt has means to President Roosevelt has means to President Roosevelt has means to
secure support from the government. secure support from the military. secure support from the people.

Figure 4.8. Illustration of reduction and synthesis operations in the COG domain.

Moreover, it is assumed that the questions guiding the synthesis operations may have
only one answer, which typically indicates how to combine the solutions. Allowing more
questions and more answers in the synthesis tree would lead to a combinatorial explo-
sion of solutions.
Another interesting aspect is that the three leaf solutions in Figure 4.8 are about the
means of President Roosevelt, while their composition is about a capability. Thus this
illustrates how synthesis operations may lead to emerging properties.
A third aspect to notice is how the reduction-level composition is actually performed. In
the example from Figure 4.8, the solutions to combine are:

President Roosevelt has means to secure support from the government.

President Roosevelt has means to secure support from the military.
President Roosevelt has means to secure support from the people.

The synthesized solution is obtained by concatenating substrings from these solutions

because, as indicated in corresponding question/answer pair, President Roosevelt has all
the needed critical requirements:

President Roosevelt has the critical capability to maintain support because President
Roosevelt has means to secure support from the government, has means to secure
support from the military, and has means to secure support from the people.

The preceding synthesis operation in the interface of Disciple-COG is shown in

Figure 12.28 (p. 373) from Section 12.4.2.

13:52:06,
.005
118 Chapter 4. Modeling the Problem-Solving Process

Figure 4.9 shows another example of reduction and synthesis in the COG domain. In
this case, the solutions to combine are:

PM Mussolini has means to secure support from the government.

PM Mussolini has means to secure support from the military.
PM Mussolini does not have means to secure support from the people.

In this case, the synthesized solution is no longer obtained by concatenating substrings

from these solutions because, as indicated in corresponding question/answer pair, PM
Mussolini does not have all the needed critical requirements:

PM Mussolini does not have the critical capability to maintain support because PM
Mussolini does not have means to secure support from the people.

Additional examples of solution synthesis from the COG domain are presented in
Figures 12.27 (p. 372), 12.28 (p. 373), 12.29 (p. 373), and 12.30 (p. 374) from Section 12.4.2.
As suggested by the preceding examples, there are many ways in which solutions may
be combined.
One last important aspect related to problem solving through analysis and synthesis is
that the solutions of the elementary problems may be obtained by applying any other type of
reasoning strategy. This enables the solving of problems through a multistrategy approach.
Chapter 12 presents Disciple cognitive assistants for different types of tasks, illustrating
the use of the inquiry-driven analysis and synthesis in different domains. Section 12.2
discusses this problem-solving paradigm in the context of military engineering planning.
Section 12.3 discusses it in the context of course of action critiquing. Section 12.4 discusses
it in the context of center of gravity analysis, and Section 12.5 discusses it in the context of
collaborative emergency response planning.

Test whether PM Mussolini has the critical capability to maintain support.

PM Mussolini does not have the critical capability to maintain support because
PM Mussolini does not have means to secure support from the people.

Which are the critical requirements for PM Mussolini to PM Mussolini does not have the critical
maintain support? PM Mussolini needs means to secure capability to maintain support because PM
support from the government, means to secure support from Mussolini does not have means to secure
the military, and means to secure support from the people. support from the people.

Does PM Mussolini have the critical capability to maintain support?

No, because PM Mussolini does not have all the needed critical requirements.

Test whether PM Mussolini has Test whether PM Mussolini has Test whether PM Mussolini has
means to secure support from the means to secure support from means to secure support from the
government. the military. people.
PM Mussolini has means to secure PM Mussolini has means to PM Mussolini does not have means
support from the government. secure support from the military. to secure support from the people.

Figure 4.9. Another illustration of reduction and synthesis operations in the COG domain.

13:52:06,
.005
 4.3. Evidence-based Reasoning 119

4.3 INQUIRY-DRIVEN ANALYSIS AND SYNTHESIS FOR

EVIDENCE-BASED REASONING

4.3.1 Hypothesis Reduction and Assessment Synthesis

In this section, we discuss the specialization of the inquiry-driven analysis and synthesis
paradigm for evidence-based reasoning where one assesses the probability of hypotheses
based on evidence, as was illustrated in Section 2.2. In this case, a complex hypothesis is
assessed by:

 Successively reducing it, from the top down, to simpler and simpler hypotheses
(guided by introspective questions and answers).
 Assessing the simplest hypotheses based on evidence.
 Successively combining, from the bottom up, the assessments of the simpler hypoth-
eses, until the assessment of the top-level hypothesis is obtained.

Figure 4.10 shows a possible analysis of the hypothesis that Country X has nuclear
weapons.

Country X has nuclear weapons.

likely

max

What are the necessary and sufficient conditions?

Reasons, desires, and capabilities.
likely
min

Country X has reasons to Country X desires Country X has the capability

have nuclear weapons. nuclear weapons. to obtain nuclear weapons.
almost certain very likely likely

max

Which is a strategy to obtain Which is a strategy to obtain

nuclear weapons? Build nuclear weapons? Buy
likely no support
max max

Country X can build Country X can buy

nuclear weapons. nuclear weapons.
likely no support
max

Which is an indicator?
Capability to produce enriched uranium.
likely
very likely indicator

Country X can produce

enriched uranium.
likely

Figure 4.10. An example of different types of reductions and corresponding synthesis functions.

13:52:06,
.005
 120 Chapter 4. Modeling the Problem-Solving Process

As pointed out by Alvin Toffler (1984, p. xi):

One of the most highly developed skills in contemporary Western civilization

is dissection; the split-up of problems into their smallest possible compon-
ents. We are good at it. So good, we often forget to put the pieces back
together again.

Evidence-based reasoning offers an opportunity to simplify the more complex synthesis

process. Indeed, in this case, the solution of a hypothesis may no longer need to be a complex
phrase or expression, but just its probability of being true, given the available evidence. The
actual solution synthesis function depends on the type of reduction strategy used.
In the following section, we will review different types of reduction strategies and the
corresponding synthesis functions.

4.3.2 Necessary and Sufficient Conditions

Ideally, a hypothesis would be reduced to several subhypotheses that would represent
necessary and sufficient conditions. That is, the hypothesis is true if and only if all the
subhypotheses are true. An example is the top-level reduction from Figure 4.10, where the
top-level hypothesis is reduced to three subhypotheses. Let us assume that we have
obtained the following assessments of these subhypotheses: almost certain, very likely,
and likely, respectively. Then the assessment of the top hypothesis, corresponding to this
necessary and sufficient condition, is the minimum of the three assessments (i.e., likely),
because each of the three subhypotheses would need to be true to ensure that the top-
level hypothesis is true. This value and the minimum (min) function that produced it are
associated with the question/answer pair.
In general, as will be illustrated later in this chapter, there may be more than one
strategy to reduce a hypothesis to simpler hypotheses, each resulting in a possibly different
assessment. In such a case, the assessment of the hypothesis should be taken as the
maximum of all these possible assessments. In this particular example, since we have only
one strategy, the assessment of the top-level hypothesis is max(likely) = likely.

4.3.3 Sufficient Conditions and Scenarios

Many times it is not easy or even possible to identify necessary and sufficient conditions to
reduce a hypothesis. In such a case, a second-best reduction would be a sufficient
condition. This means that if the subhypotheses are true, then the hypothesis is true.
But, as we have discussed previously, there may be more than one sufficient condition for
a hypothesis. For example, the middle part of Figure 4.10 shows two possible strategies for
Country X to obtain nuclear weapons: It can build them, or it can buy them. Each strategy
has to be assessed and the maximum assessment represents the assessment of the
hypothesis that Country X has the capability to obtain nuclear weapons.
A special type of a sufficient condition for a hypothesis is a scenario in which the
hypothesis would be true, such as the one illustrated in Figure 2.9 (p. 63) from Section
2.2.4. But, as we have discussed in Section 2.2.4, there may be multiple alternative
scenarios. Figure 4.11 shows an abstract example where there are two alternative scenarios
for Hypothesis 1 to be true. Scenario 1 consists of action 2 and action 3. For this scenario to
have happened, both these actions should have happened. Therefore, we combine their

13:52:06,
.005
 4.3. Evidence-based Reasoning 121

Hypothesis 1

Assessment of
Hypothesis 1

max

Which is a possible scenario? Which is a possible scenario?

Scenario 1 Scenario 2

Assessment 1 of Hypothesis 1 Assessment 2 of Hypothesis 1

min min

Hypothesis (action) 2 Hypothesis (action) 3 Hypothesis (action) 4 Hypothesis (action) 5

Assessment of Hypothesis 2 Assessment of Hypothesis 3 Assessment of Hypothesis 4 Assessment of Hypothesis 5

Figure 4.11. Reductions and syntheses corresponding to two sufficient conditions (scenarios).

corresponding assessments with a minimum function, as shown at the bottom-left of

Figure 4.11. Hypothesis 1, however, would be true if either of the two scenarios would
happen. Therefore, we combine the assessments corresponding to the two scenarios
through a maximum function.

4.3.4 Indicators
Many times when we are assessing a hypothesis, we have only indicators. For example, as
shown at the bottom part of Figure 4.10, having the capability to produce enriched uranium
is an indicator that a country can build nuclear weapons. An indicator is, however, weaker
than a sufficient condition. If we determine that a sufficient condition is satisfied (e.g., a
scenario has actually happened), we may conclude that the hypothesis is true. But we
cannot draw such a conclusion just because we have discovered an indicator. However, we
may be more or less inclined to conclude that the hypothesis is true, based on the relevance
(strength) of the indicator. Therefore, given the symbolic probabilities from Table 2.5, we
distinguish between three types of indicators of different relevance (strength): “likely indica-
tor,” “very likely indicator,” and “almost certain indicator.”
A “likely indicator” is one that, if discovered to be true, would lead to the conclusion
that the considered hypothesis is likely. Similarly, a “very likely indicator” would lead to the
conclusion that the hypothesis is very likely, and an “almost certain indicator” would lead to
the conclusion that the hypothesis is almost certain.
In the example from the bottom part of Figure 4.10 it is likely that Country X can
produce enriched uranium, and this is a very likely indicator that Country X can build
nuclear weapons. Therefore, we can conclude that the probability of the hypothesis
that Country X can build nuclear weapons is likely, the minimum between likely (the
probability of the indicator) and very likely (the strength of the indicator).
In general, the probability of a hypothesis H based on an indicator I is the minimum
between the probability of the indicator and the relevance (strength) of the indicator (which
could be likely, very likely, or almost certain).

13:52:06,
.005
 122 Chapter 4. Modeling the Problem-Solving Process

It makes no sense to consider the type “certain indicator,” because this would be a
sufficient condition. Similarly, it makes no sense to consider the type “no support indica-
tor,” because this would not be an indicator.
As an abstract example, Figure 4.12 shows a hypothesis that has two likely indicators,
A and B, if only one of them is observed. However, if both of them are observed, they
synergize to become an almost certain indicator.
As a concrete example, consider Person , who has been under surveillance in
connection with terrorist activities. We suspect that will attempt to leave the country
in a short while. Three days ago, we received information that sold his car. Today, we
received information that he closed his account at his bank. Each of these is only a likely
indicator of the hypothesis that plans to leave the country. He could be planning to buy
a new car, or he could be dissatisfied with his bank. But, taken together, these two
indicators suggest that it is almost certain that is planning to leave the country.
Coming back to the abstract example in Figure 4.12, let us assume that indicator A is
almost certain and indicator B is very likely. In such a case, the assessment of Hypothesis
1, based only on indicator A, is minimum(almost certain, likely) = likely. Similarly, the
assessment of Hypothesis 1, based only on indicator B, is minimum(very likely, likely) =
likely. But the assessment of Hypothesis 1, based on both indicators A and B, is
minimum(minimum(almost certain, very likely), almost certain) = very likely. Also, the
assessment of Hypothesis 1 based on all the indicators is the maximum of all the
individual assessments (i.e., very likely), because these are three alternative solutions
for Hypothesis 1.

4.4 EVIDENCE-BASED ASSESSMENT

Now we discuss the assessment of the leaf hypotheses of the argumentation structure,
based on the identified relevant evidence. Let us consider an abstract example where the
leaf hypothesis to be directly assessed based on evidence is Q (see Figure 4.13).
We begin by discussing how to assess the probability of hypothesis Q based only on one
item of favoring evidence Ek* (see the bottom of Figure 4.13). First notice that we call
this likeliness of Q, and not likelihood, because in classic probability theory, likelihood is

Hypothesis 1
very likely

max

Which is an indicator? Which is an indicator? Which is an indicator?

Indicator A Indicator B Indicators A and B
likely likely very likely

likely indicator likely indicator almost certain indicator

Indicator A Indicator B Indicator A Indicator B

almost certain very likely almost certain very likely

Figure 4.12. Hypothesis assessment based on indicators.

13:52:06,
.005
4.4. Evidence-based Assessment 123

Q very likely

on balance
Inferential force of evidence on Q

Q based only on almost Qc based only on

likely
favoring evidence certain disfavoring evidence
max
Inferential force of favoring evidence on Q
Q based almost Q based
likely
only on Ei* certain only on Ek*
min
Inferential force of Ek* on Q
How likely is Q based only on Ek*?

Relevance of Ek to Q likely Believability of Ek* very likely

How likely is Q, based only on Ek* How likely is it that Ek* is true?
and assuming that Ek* is true?

Figure 4.13. The relevance, believability, and inferential force of evidence.

P(Ek*|Q), while here we are interested in P(Q|Ek*), the posterior probability of Q given Ek*.
To assess Q based only on Ek*, there are three judgments to be made by answering
three questions:

The relevance question is: How likely is Q, based only on Ek* and assuming that Ek* is
true? If Ek* tends to favor Q, then our answer should be one of the values from likely
to certain. If Ek* is not relevant to Q, then our answer should be no support, because
Ek* provides no support for the truthfulness of Q. Finally, if Ek* tends to disfavor Q,
then it tends to favor the complement of Q, that is, Qc. Therefore, it should be used
as favoring evidence for Qc, as discussed later in this section.
The believability question is: How likely is it that Ek* is true? Here the answer should be
one of the values from no support to certain. The maximal value, certain, means
that we are sure that the event Ek reported in Ek* did indeed happen. The minimal
value, no support, means that Ek* provides us no reason to believe that the event
Ek reported in Ek* did happen. For example, we believe that the source of Ek* has
lied to us.
The inferential force or weight question is: How likely is Q based only on Ek*? The agent
automatically computes this answer as the minimum of the relevance and believ-
ability answers. What is the justification for this? Because to believe that Q is true
based only on Ek*, Ek* should be both relevant to Q and believable.

When we assess a hypothesis Q, we may have several items of evidence, some favoring Q
and some disfavoring Q. The agent uses the favoring evidence to assess the probability of
Q and the disfavoring evidence to assess the probability of Qc. As mentioned previously,
because the disfavoring evidence for Q is favoring evidence for Qc, the assessment process
for Qc is similar to the assessment for Q.
When we have several items of favoring evidence, we evaluate Q based on each of them
(as was explained previously), and then we compose the obtained results. This is illus-
trated in Figure 4.13, where the assessment of Q based only on Ei* (almost certain) is

13:52:06,
.005
 124 Chapter 4. Modeling the Problem-Solving Process

composed with the assessment of Q based only on Ek* (likely), through the maximum
function, to obtain the assessment of Q based only on favoring evidence (almost certain). In
this case, the use of the maximum function is justified because it is enough to have one
item of evidence that is both very relevant and very believable to persuade us that the
hypothesis Q is true.
Let us assume that Qc based only on disfavoring evidence is likely. How should we
combine this with the assessment of Q based only on favoring evidence? As illustrated at
the top of Figure 4.13, the agent uses an on-balance judgment: Because Q is almost certain
and Qc is likely, it concludes that, based on all available evidence, Q is very likely.
In general, as indicated in the right and upper side of Table 4.1, if the assessment of Qc
(based on disfavoring evidence for Q) is higher than or equal to the assessment of Q
(based on favoring evidence), then we conclude that, based on all the available evidence,
there is no support for Q. If, on the other hand, the assessment of Q is strictly greater than
the assessment of Qc, then the assessment of Q is decreased, depending on the actual
assessment of Qc (see the left and lower side of Table 4.1).
One important aspect to notice is that the direct assessment of hypotheses based on
favoring and disfavoring evidence is done automatically by the agent, once the user
assesses the relevance and the believability of evidence.
Another important aspect to notice is that the evaluation of upper-level hypotheses (such
as those from Figure 4.10) requires the user to indicate what function to use when
composing the assessments of their direct subhypotheses. This was discussed in Section 4.3.

4.5 HANDS ON: WAS THE CESIUM STOLEN?

To illustrate further the divide-and-conquer approach to hypothesis analysis, let us

continue with the cesium example introduced in Section 2.2, where we have already

Table 4.1 An “On-Balance” Synthesis Function

c
Q based Q based only on disfavoring evidence
on all no
likely
very almost
certain
evidence support likely certain
Q based only on favoring evidence

no no no no no no
support support support support support support

no no no no
likely likely
support support support support

very very no no no
likely
likely likely support support support

almost almost very no no

likely
certain certain likely support support

almost very no
certain certain likely
certain likely support

13:52:06,
.005
4.5. Hands On: Was the Cesium Stolen? 125

established that the cesium-137 canister is missing (see Figure 2.8 on p. 62). The next step
is to consider the competing hypotheses:

H2: The cesium-137 canister was stolen.

H2’: The cesium-137 canister was misplaced.
H2”: The cesium-137 canister is used in a project without being checked out from the
XYZ warehouse.

We have to put each of these hypotheses to work, to guide the collection of relevant
evidence. In Section 2.2.4, we have already discussed, at a conceptual level, the collec-
tion of evidence for hypothesis H2. Table 4.2 shows the result of our information
collection efforts.
The collected information from Table 4.2 suggests that the cesium-137 canister was
stolen with the panel truck having Maryland license MDC-578. This has led to the
development of the analysis tree in Figure 2.9 (p. 63). In this case study, you are going
to actually perform this analysis. You have to identify the “dots” in the information from

Table 4.2 Additional Information on the Missing Cesium-137 Canister

INFO-003-Clyde: We talked to a professional locksmith named Clyde, who said that

the lock had been forced, but it was a clumsy job.
INFO-004-SecurityCamera: The security camera of the XYZ warehouse contains a
video segment showing a person loading a container into a U-Haul panel truck.
INFO-005-Guard: There is a security perimeter around the XYZ warehouse and
employee parking area having just one gate that is controlled by a guard. On the day before the
missing canister was observed, the security guard, Sam, recorded that a panel truck having
Maryland license plate MDC-578 was granted entry at 4:45 pm just before the XYZ closing hour
at 5:00 pm. The driver of this vehicle showed the guard a manifest containing items being
delivered to the XYZ warehouse. This manifest contained a list of packing materials allegedly
ordered by the XYZ Company. The vehicle was allowed to enter the parking area. At 8:30 pm,
this same vehicle was allowed to exit the parking area. A different guard was on duty in the
evenings and noticed that his records showed that this vehicle had been permitted entry, and so
he allowed the vehicle to exit the parking area.
INFO-006-TRUXINC: Maryland DOT’s record indicates that the panel truck carrying
the license plate number MDC-578 is registered in the name of a truck-rental company called
TRUXINC, located in Silver Spring, MD. The manager of this agency showed records indicating
that this truck was rented to a person who gave his name as Omer Riley, having as his
listed address 6176 Williams Ave. in Silver Spring. The truck was rented on the day before
Willard’s discovery about the missing cesium-137, and it was returned the day after he made
the discovery.
INFO-007-SilverSpring: Silver Spring city record according to which there is no
residence at 6176 Williams Ave. in Silver Spring, MD.
INFO-008-InvestigativeRecord: An examination of the panel truck rented by Omer
Riley, using a Geiger counter, revealed minute traces of cesium-137.
INFO-009-Grace: Grace, the Vice President for Operations at XYZ, tells us that no one
at the XYZ Company had checked out the canister for work on any project the XYZ Company was
working on at the time. She says that the XYZ Company had other projects involving hazardous
materials, but none that involved the use of cesium-137.

13:52:06,
.005
126 Chapter 4. Modeling the Problem-Solving Process

Table 4.2, which are fragments representing relevant items of evidence for the leaf
hypotheses in Figure 2.9. These dots are presented in Table 4.3.
This case study has several objectives:

 Learning to associate with a hypothesis in an argument the evidence that is relevant

to it
 Learning to evaluate the relevance and the believability of evidence
 Learning to select synthesis functions
 Better understanding the process of evaluating the probability or likeliness of a hypoth-
esis based on the available evidence

When you associate an item of evidence with a hypothesis, the agent automatically
generates a decomposition tree like the one in Figure 4.14. The bottom part of Figure 4.14
shows the abstraction of the tree that is automatically generated by the agent when you
indicate that the item of evidence E005-Ralph favors the leaf hypothesis “The XYZ hazardous
material locker was forced.”
The agent also automatically generates the reduction from the top of Figure 4.14, where
the leaf hypothesis, “The XYZ hazardous material locker was forced,” is reduced to the
elementary hypothesis with the name, “The XYZ hazardous material locker was forced,” to be
directly assessed based on evidence. Although these two hypotheses are composed of the
same words, internally they are different, the latter being an instance introduced in the
agent’s ontology. This elementary hypothesis corresponds to the hypothesis Q in
Figure 4.13. The agent decomposes this hypothesis as shown in the bottom part of
Figure 4.14, which corresponds to the tree in Figure 4.13 except that there is only one

Table 4.3 Dots from Table 4.2

E006-Clyde: Locksmith Clyde’s report that the lock was forced.

E007-SecurityCamera: Video segment of the security camera of the XYZ warehouse,
showing a person loading a container into a U-Haul panel truck.
E008-GuardReport: The record, made by Sam, security guard at the XYZ Company,
that a panel truck bearing Maryland license plate number MDC-578 was in the XYZ parking area on
the day before Willard’s discovery about the missing cesium-137 canister.
E009-MDDOTRecord: Maryland DOT’s record that the truck bearing license plate
number MDC-578 is registered in the name of the TRUXINC Company in Silver Spring, MD.
E010-TRUXINCRecord1: TRUXINC’s record that the truck bearing MD license plate
number MDC-578 was rented to a man who gave his name as Omer Riley on the day before
Willard’s discovery of the missing cesium-137 canister.
E011-TRUXINCRecord2: TRUXINC’s record that Omer Riley gave his address as
6176 Williams Ave.
E012-SilverSpringRecord: Silver Spring city record according to which there is no
residence at 6176 Williams Ave. in Silver Spring, MD.
E013-InvestigativeRecord: Investigative record that traces of cesium-137 were found
in the truck bearing license plate number MDC-578.
E014-Grace: Grace, the Vice President for Operations at XYZ, tells us that no one at
the XYZ Company had checked out the canister for work on any project.

13:52:06,
.005
4.5. Hands On: Was the Cesium Stolen? 127

Figure 4.14. Evidence-based assessment of an elementary hypothesis.

item of favoring evidence, namely E005-Ralph. After that, you have to assess the relevance
of this item of evidence to the considered hypothesis (e.g., likely), as well as its believability
(e.g., very likely), and the agent automatically composes them, from the bottom up, to
obtain the assessment of the leaf hypothesis. When you add additional items of evidence
as either favoring or disfavoring evidence, the agent extends the reasoning tree from
Figure 4.14 as indicated in Figure 4.13.
Figure 4.15 illustrates the selection of a synthesis function indicating how to evaluate
the probability of a node based on the probability of its children. You have to right-click on
the node (but not on any word in blue), select New Solution with. . ., and then select the
function from the displayed list.
Now you can perform the case study. Start Disciple-EBR, select the case study know-
ledge base “02-Evidence-based-Analysis/Scen,” and proceed as indicated in the instruc-
tions from the bottom of the opened window.
This case study illustrates several basic hypothesis analysis operations described in the
following.

Operation 4.1. Associate evidence to hypotheses

 In the Evidence workspace, click on the Evidence menu at the top of the window.
 Notice the four modes of operations from the top part of the left panel. Because the
selected one is [AVAILABLE EVIDENCE], the left panel shows the current evidence from
the knowledge base.
 In the left panel, click on the item of evidence you would like to associate with a leaf
hypothesis from the current argumentation. As a result, the upper part of the right
panel shows the main characteristics of this item of evidence, followed by all the leaf
hypotheses in the analysis tree (see Figure 4.16). You will have to decide whether the
selected item of evidence favors or disfavors any of the hypotheses under the Irrelevant
to label, and indicate this by clicking on [FAVORS] or [DISFAVORS] following that
hypothesis.
 Clicking on [FAVORS] or [DISFAVORS] automatically creates an elementary hypoth-
esis to be assessed based on evidence, and moves it under the Favors (or Disfavors)

13:52:06,
.005
128 Chapter 4. Modeling the Problem-Solving Process

Figure 4.15. Selecting the synthesis function for a node.

Figure 4.16. Associating an evidence item to a hypothesis.

label. Clicking on [REMOVE] will restore the leaf hypothesis under the Irrelevant
to label.
 To associate another evidence item to a hypothesis, click on it in the left panel and
repeat the preceding operations.
 To return to the Reasoner module, click on [REASONING] following the hypothesis.

13:52:06,
.005
4.5. Hands On: Was the Cesium Stolen? 129

Operation 4.2. Update the name of an elementary hypothesis

 In the right panel of the Reasoner module, right-click on a hypothesis that was
reduced to an elementary hypothesis to be assessed based on evidence, and select
Improve Phrasing.
 In the opened editor, update the phrasing of the hypothesis and then click outside
the box.
 Notice that both this hypothesis and the corresponding elementary hypothesis have
been updated accordingly.

Operation 4.3. Assess evidence

 In the Reasoner module, in the left panel, click on the name of the item of evidence to
assess. You may need to right-click on the top hypothesis and select Expand, to make
the evidence item visible. As a result, the right panel shows the decomposition of
evidence assessment into a relevance assessment (the left leaf) and a believability
assessment (the right leaf), as illustrated in Figure 4.17.
 If the right panel does not show the solutions of the hypotheses, then click on [SHOW
SOLUTIONS] at the top of the panel.
 In the right panel, right-click on the left (relevance) leaf and select New Assumption.
As a result, the agent proposes the default solution (e.g., certain).
 If necessary, click on the default solution (the underlined text) and, from the displayed
list, select (double-click) the appropriate value.
 In the right panel, right-click on the right (believability) leaf and select New Assump-
tion (as illustrated in Figure 4.17). As a result, the agent proposes the default solution.
 If necessary, click on the default solution and, from the displayed list, select the
appropriate value.
 The agent automatically determines the inferential force of the item of evidence.

Operation 4.4. Select the synthesis function

 In the Reasoner module, in the right panel, right-click on the node for which you have
to select the synthesis function, select New Solution with. . ., and then select the
function from the displayed list (see Figure 4.18).
 To select a function for a node, all its children must have solutions.

Figure 4.17. Assessing an item of evidence.

13:52:06,
.005
 130 Chapter 4. Modeling the Problem-Solving Process

Figure 4.18. Selecting a synthesis function.

Figure 4.19. Evidence collection guidance for a selected hypothesis.

4.6 HANDS ON: HYPOTHESIS ANALYSIS AND EVIDENCE

SEARCH AND REPRESENTATION

The objective of this case study is to learn how to use Disciple-EBR to analyze hypotheses based
on evidence retrieved from the Internet, by associating search criteria with elementary hypoth-
eses, invoking various search engines (such as Google, Yahoo!, or Bing), identifying relevant
information, extracting evidence from it, and using the evidence to evaluate the hypotheses.
This case study concerns the hypothesis that the United States will be a global leader in
wind power within the next decade.
To search for evidence that is relevant to a leaf hypothesis, the agent guides you to
associate search criteria with it and to invoke various search engines on the Internet.
Figure 4.19 shows the corresponding interface of the Evidence module. Because the
[COLLECTION GUIDANCE] mode is selected in the left panel, it shows all the leaf hypotheses
and their current evidential support. If you click on one of these hypotheses, such as
“United States imports huge quantities of oil,” it displays this hypothesis in the right panel,
enabling you to define search criteria for it. You just need to click on the [NEW] button
following the Search criterion label, and the agent will open an editor in which you can
enter the search criterion.
Figure 4.20 shows two defined search criteria: “oil import by United States” and “top oil
importing countries.” You can now invoke Bing, Google, or Yahoo! with any one of these
criteria to search for relevant evidence on the Internet. This will open a new window with
the results of the search, as shown in Figure 4.21.

13:52:06,
.005
4.6. Hands On: Hypothesis Analysis 131

Figure 4.20. Defined search criteria for a selected hypothesis.

Figure 4.21. Searching relevant evidence on the Internet.

You have to browse the retrieved documents shown in Figure 4.21 and determine
whether any of them contains information that is relevant to the hypothesis that the
United States imports huge quantities of oil. Such a document is the second one, whose
content is shown in Figure 4.22.
You can now define one or several items of evidence with information copied from the
retrieved document, as illustrated in Figure 4.23. In the left panel of the Evidence module,
you switch the selection mode to [AVAILABLE EVIDENCE] and then click on [NEW]. As a
result, the right panel displays a partial name for the evidence E001- to be completed by
you. You then have to click on the [EDIT] button, which opens an editor where you can
copy the description of this item of evidence from the retrieved document. The result is
shown in the right panel of Figure 4.23.
You can define additional characteristics of this item of evidence, such as its type (as
will be discussed in Section 4.7), and you should indicate whether this item of evidence
favors or disfavors the hypothesis that the United States imports huge quantities of oil, as
explained previously.

13:52:06,
.005
132 Chapter 4. Modeling the Problem-Solving Process

Figure 4.22. Selected document providing relevant information.

Figure 4.23. Defining an item of evidence.

In this case study, you will first select the hypothesis, ”United States will be a global
leader in wind power within the next decade,” and then you will browse its analysis tree to
see how it is reduced to simpler hypotheses that you have to assess by searching evidence
on the Internet. You will associate specific search criteria with the leaf hypotheses, invoke
specific search engines with those criteria, identify relevant Web information, define
evidence from this information, associate evidence with the corresponding hypotheses,
and evaluate its relevance and believability, with the goal of assessing the probability of the
top-level hypothesis.
Start Disciple-EBR, select the case study knowledge base “03-Evidence-Search/Scen,”
and proceed as indicated in the instructions from the bottom of the opened window.
This case study illustrates the following hypothesis analysis operation:

Operation 4.5. Associate search criteria with hypotheses

 In the Evidence workspace, click on the Evidence menu and then click on [COLLECTION
GUIDANCE]. The left panel shows the leaf hypotheses and their evidential support.

13:52:06,
.005
 4.7. Believability Assessment 133

 In the left panel, select a hypothesis.

 In the right panel, after Search criterion, click on [NEW] to define a new criterion.
 Type the search criterion and click on [SAVE].
 You may define additional criteria by repeating the preceding two steps.
 Select one of the search criteria by clicking on it.
 After Search with, click on one of the available search engines (i.e., [BING], [GOOGLE],
[YAHOO]) to search the Internet with the selected criterion.
 Browse the documents returned by the search engine, select the relevant ones,
and define items of evidence based on them, as indicated in Operation 4.6, later in
this chapter.

4.7 BELIEVABILITY ASSESSMENT

In the previous sections, we have discussed and illustrated how you may directly assess
the believability of an item of evidence. However, the Disciple-EBR agent has a significant
amount of knowledge about the various types of evidence and its believability credentials,
enabling you to perform a much deeper believability analysis, as will be discussed in this
section. You may wish to perform such a detailed believability analysis for those items of
evidence that are critical to the final result of the analysis. We will start with presenting a
classification or ontology of evidence.
Attempts to categorize evidence in terms of its substance or content would be a fruitless
task, the essential reason being that the substance or content of evidence is virtually
unlimited. What we have termed a substance-blind classification of evidence refers to a
classification of recurrent forms and combinations of evidence, based not on substance or
content, but on the inferential properties of evidence (Schum, 1994 [2001a], pp. 114–130;
Schum, 2011). In what follows, we identify specific attributes of the believability of various
recurrent types of evidence without regard to their substance or content.
Here is an important question you are asked to answer regarding the individual
kinds of evidence you have: How do you stand in relation to this item of evidence? Can
you examine it for yourself to see what events it might reveal? If you can, we say that the
evidence is tangible in nature. But suppose instead you must rely upon other persons
to tell you about events of interest. Their reports to you about these events are
examples of testimonial evidence. Figure 4.24 shows a substance-blind classification
of evidence based on its believability credentials. This classification is discussed in
the following sections.

4.7.1 Tangible Evidence

There is an assortment of tangible items you might encounter. Both imagery intelligence
(IMINT) and signals intelligence (SIGINT) provide various kinds of sensor records
and images that can be examined. Measurement and signature intelligence (MASINT)
and technical intelligence (TECHINT) provide various objects, such as soil samples and
weapons, that can be examined. Communications intelligence (COMINT) can provide
audio recordings of communications that can be overheard and translated if the commu-
nication has occurred in a foreign language. Documents, tabled measurements, charts,
maps, and diagrams or plans of various kinds are also tangible evidence.

13:52:06,
.005
134 Chapter 4. Modeling the Problem-Solving Process

evidence

tangible testimonial missing authoritative

evidence evidence evidence record

real demonstrative unequivocal equivocal

tangible tangible testimonial testimonial
evidence evidence evidence evidence

unequivocal unequivocal testimonial completely probabilistically

testimonial evidence testimonial evidence equivocal equivocal
based upon direct evidence obtained based on testimonial testimonial
observation at second hand opinion evidence evidence

Figure 4.24. Substance-blind classification of evidence.

There are two different kinds of tangible evidence: real tangible evidence and demon-
strative tangible evidence (Lempert et al., 2000, pp. 1146–1148). Real tangible evidence is
an actual thing and has only one major believability attribute: authenticity. Is this object
what it is represented as being or is claimed to be? There are as many ways of generating
deceptive and inauthentic evidence as there are persons wishing to generate it. Docu-
ments or written communications may be faked, captured weapons may have been
tampered with, and photographs may have been altered in various ways. One problem
is that it usually requires considerable expertise to detect inauthentic evidence.
Demonstrative tangible evidence does not concern things themselves but only repre-
sentations or illustrations of these things. Examples include diagrams, maps, scale models,
statistical or other tabled measurements, and sensor images or records of various sorts
such as IMINT, SIGINT, and COMINT. Demonstrative tangible evidence has three believ-
ability attributes. The first concerns its authenticity. For example, suppose we obtain a
hand-drawn map from a captured insurgent showing the locations of various groups in his
insurgency organization. Has this map been deliberately contrived to mislead our military
forces, or is it a genuine representation of the location of these insurgency groups?
The second believability attribute is accuracy of the representation provided by the
demonstrative tangible item. The accuracy question concerns the extent to which the
device that produced the representation of the real tangible item had a degree of sensitiv-
ity (resolving power or accuracy) that allows us to tell what events were observed. We
would be as concerned about the accuracy of the hand-drawn map allegedly showing
insurgent groups locations as we would about the accuracy of a sensor in detecting traces
of some physical occurrence. Different sensors have different resolving power that also
depends on various settings of their physical parameters (e.g., the settings of a camera).
The third major attribute, reliability, is especially relevant to various forms of sensors
that provide us with many forms of demonstrative tangible evidence. A system, sensor,
or test of any kind is reliable to the extent that the results it provides are repeatable or

13:52:06,
.005
 4.7. Believability Assessment 135

consistent. You say that a sensing device is reliable if it provides the same image or report
on successive occasions on which this device is used.
The left side of Figure 4.25 shows how the agent assesses the believability of an item of
demonstrative tangible evidence Ei* as the minimum of its authenticity, accuracy, and
reliability.
Here are additional examples involving evidence that is tangible and that you can
examine personally to see what events it reveals.
Have a look at evidence item E009-MDDOTRecord in Table 4.3 (p. 126). The Maryland
DOT record, in the form of a tangible document, could be given to the analyst to verify
that the vehicle carrying MD license plate number MDC-578 is registered in the name of
the TRUXINC Company in Silver Spring, Maryland.
Now consider evidence item E008-GuardReport in Table 4.3. Here we have a document
in the form of a log showing that the truck bearing license plate number MDC-578 exited
the XYZ parking lot at 8:30 pm on the day in question. This tangible item could also be
made available to analysts investigating this matter.

4.7.2 Testimonial Evidence

For testimonial evidence, we have two basic sources of uncertainty: competence and
credibility. This is one reason why it is more appropriate to talk about the believability of
testimonial evidence, which is a broader concept that includes both competence and
credibility considerations. The first question to ask related to competence is whether this
source actually made the observation the source claims to have made or had access to the
information the source reports. The second competence question concerns whether this
source understood what was being observed well enough to provide us with an intelligible
account of what was observed. Thus competence involves access and understandability.
Assessments of human source credibility require consideration of entirely different
attributes: veracity (or truthfulness), objectivity, and observational sensitivity under the
conditions of observation (Schum, 1989). Here is an account of why these are the major
attributes of testimonial credibility. First, is this source telling us about an event this source
believes to have occurred? This source would be untruthful if he or she did not believe the
reported event actually occurred. So, this question involves the source's veracity. The
second question involves the source's objectivity. The question is: Did this source base a
belief on sensory evidence received during an observation, or did this source believe
the reported event occurred because this source either expected or wished it to occur?

believability of Ei* believability of Ek*

min min
i
Source’s Source’s
competence credibility
authenticity of Ei* reliability of Ei* min min

Source’s veracity Source’s objectivity

accuracy of Ei* Source’s access Source’s understandability
Source’s observational sensitivity

Figure 4.25. Assessing the believability of evidence with Disciple-EBR.

13:52:06,
.005
136 Chapter 4. Modeling the Problem-Solving Process

An objective observer is one who bases a belief on the sensory evidence instead of desires
or expectations. Finally, if the source did base a belief on sensory evidence, how good was
this evidence? This involves information about the source's relevant sensory capabilities
and the conditions under which a relevant observation was made.
As indicated in Figure 4.24, there are several types of testimonial evidence. If the source
does not hedge or equivocate about what the source observed (i.e., the source reports that
he or she is certain that the event did occur), then we have unequivocal testimonial
evidence. If, however, the source hedges or equivocate in any way (e.g., "I'm fairly sure
that E occurred"), then we have equivocal testimonial evidence. The first question we
would ask a source of unequivocal testimonial evidence is: How did you obtain information
about what you have just reported? It seems that this source has three possible answers to
this question. The first answer is, "I made a direct observation myself.” In this case, we have
unequivocal testimonial evidence based upon direct observation. The second possible
answer is, "I did not observe this event myself but heard about its occurrence (or
nonoccurrence) from another person." Here we have a case of second hand or hearsay
evidence, called unequivocal testimonial evidence obtained at second hand. A third answer
is possible: "I did not observe event E myself nor did I hear about it from another source.
But I did observe events C and D and inferred from them that event E definitely occurred."
This is called testimonial evidence based on opinion, and it requires some very difficult
questions. The first concerns the source's credibility as far as his or her observation of
events C and D; the second involves our examination of whether we ourselves would infer
E based on events C and D. This matter involves our assessment of the source's reasoning
ability. It might well be the case that we do not question this source's credibility in
observing events C and D, but we question the conclusion that the source has drawn
from his or her observations that event E occurred. We would also question the certainty
with which the source has reported the opinion that E occurred. Despite the source’s
conclusion that “event E definitely occurred," and because of many sources of uncertainty,
we should consider that testimonial evidence based on opinion is a type of equivocal
testimonial evidence.
There are two other types of equivocal testimonial evidence. The first we call completely
equivocal testimonial evidence. Asked whether event E occurred or did not, our source
says, "I don't know," or, "I can't remember."
But there is another way a source of HUMINT can equivocate: The source can provide
probabilistically equivocal testimonial evidence in various ways: "I'm 60 percent sure that
event E happened”; or "I'm fairly sure that E occurred”; or, "It is very likely that
E occurred." We could look upon this particular probabilistic equivocation as an assess-
ment by the source of the source’s own observational sensitivity.
The right side of Figure 4.25 shows how a Disciple-EBR agent assesses the believability
of an item of testimonial evidence based upon direct observation Ek* by a source, as the
minimum of the source’s competence and credibility. The source’s competence is
assessed as the minimum of the source’s access and understandability, while the source’s
credibility is assessed as the minimum of the source’s veracity, objectivity, and observa-
tional sensitivity.
Here are some examples involving testimonial evidence from human sources that is not
hedged or qualified in any away.
Evidence item E014-Grace in Table 4.3 (p. 126) is Grace’s testimony that no one at the
XYZ Company had checked out the canister for work on any project. Grace states this
unequivocally. You should also note that she has given negative evidence saying the

13:52:06,
.005
 4.7. Believability Assessment 137

cesium-137 was not being used by the XYZ Company. This negative evidence is very
important, because it strengthens our inference that the cesium-137 canister was stolen.
E006-Clyde in Table 4.3 is unequivocal testimonial evidence. It represents positive evidence.
Here are some examples involving testimonial evidence given by human sources who
equivocate or hedge in what they tell us.
Consider the evidence item E005-Ralph in Table 2.4 (p. 60). Here Ralph hedges a bit by
saying that the lock on the hazardous materials storage area appears to have been forced.
He cannot say for sure that the lock had been forced, so he hedges in what he tells us.
In new evidence regarding the dirty bomb example, suppose we have a source code-
named “Yasmin.” She tells us that she knew a man in Saudi Arabia named Omar al-
Massari. Yasmin says she is “quite sure” that Omar spent two years “somewhere” in
Afghanistan “sometime” in the years 1998 to 2000.

4.7.3 Missing Evidence

To say that evidence is missing entails that we must have had some basis for expecting we
could obtain it. There are some important sources of uncertainty as far as missing
evidence is concerned. In certain situations, missing evidence can itself be evidence.
Consider some form of tangible evidence, such as a document, that we have been unable
to obtain. There are several reasons for our inability to find it, some of which are more
important than others. First, it is possible that this tangible item never existed in the first
place; our expectation that it existed was wrong. Second, the tangible item exists, but we
have simply been looking in the wrong places for it. Third, the tangible item existed at one
time but has been destroyed or misplaced. Fourth, the tangible item exists, but someone is
keeping it from us. This fourth consideration has some very important inferential implica-
tions, including denial and possibly deception. An adverse inference can be drawn from
someone's failure to produce evidence.
We should not confuse negative evidence with missing evidence. To adopt a common
phrase, “evidence of absence (negative evidence) is not the same as absence of evidence
(missing evidence).” Entirely different conclusions can be drawn from evidence that an
event did not occur than can be drawn from our failure to find evidence. We are obliged to
ask different questions in these two situations.
Consider our discussion on the cesium-137 canister. Upon further investigation, we
identify the person who rented the truck as Omar al-Massari, alias Omer Riley. We tell him
that we wish to see his laptop computer. We are, of course, interested in what it might
reveal about the terrorists with whom he may be associating. He refuses to tell us where it
is. This we referred to as the nonproduction of evidence.

4.7.4 Authoritative Record

This final category of evidence would never oblige an analyst to assess its believability.
Tabled information of various sorts such as tide tables, celestial tables, tables of physical or
mathematical results such as probabilities associated with statistical calculations, and
many other tables of information we would accept as being believable provided that we
used these tables correctly. For example, we would not be obliged to prove that tempera-
tures in Iraq can be around 120 degrees Fahrenheit during summer months or that the
population of Baghdad is greater than that of Basra.

13:52:06,
.005
 138 Chapter 4. Modeling the Problem-Solving Process

4.7.5 Mixed Evidence and Chains of Custody

We have just described a categorization of individual items of evidence. But there are
situations in which individual items can reveal various mixtures of the types of evidence
shown in Figure 4.24. One example is testimonial evidence about tangible evidence where
a source describes a weapon observed at a scene of a crime. Another example is a tangible
document containing a testimonial assertion based on other alleged tangible evidence.
Figure 4.26, for example, shows how one would need to assess the believability of tangible
evidence about testimonial evidence.
Here is an example of a mixture of two or more items of tangible evidence; it is called a
passport. A passport is a tangible document alleging the existence of other tangible
documents recording the place of birth and country of origin of the holder of the passport.
In other words, a passport sets up a paper trail certifying the identity of the holder of the
passport. In addition to needing to check the authenticity of the passport itself, we are
also interested in the authenticity of all the other tangible documents on which this
passport is based.
Here is another mixture of forms of evidence, this time recording a mixture of tangible
and testimonial evidence. We return to our asset “Yasmin,” who has given us further
evidence about Omar al-Massari in our cesium-137 example. Suppose we have a tangible
document recording Yasmin’s account of her past experience with Omar al-Massari. This
document records Yasmin’s testimony about having seen a document, detailing plans for
constructing weapons of various sorts, that was in Omar al-Massari’s possession. As far as
believability issues are concerned, we first have the authenticity of the transcription of her
testimony to consider. Yasmin speaks only in Arabic, so we wonder how adequate the
translation of her testimony has been. Also, we have concerns about Yasmin’s competence
and credibility to consider in her recorded testimony. Finally, we have further interest in
the authenticity of the document she allegedly saw in Omar al-Massari’s possession.
But the believability analysis of an item of evidence can be even more complicated. For
example, very rarely, if ever, has an analyst access to the original evidence. Most often,
what is being analyzed is an item of evidence that has undergone a series of transform-
ations through a chain of custody (Schum et al., 2009). Here we have borrowed an

believability of Ei*

min

Source’s
believability
authenticity of Ei* accuracy of Ei*
min

Source’s Source’s
competence credibility
min min

Source’s Source’s
Source’s Source’s veracity objectivity
access understandability
Source’s
observational sensitivity

Figure 4.26. Believability analysis of tangible evidence about testimonial evidence.

13:52:06,
.005
4.7. Believability Assessment 139

important concept from the field of law, where a chain of custody refers to the persons or
devices having access to the original evidence, the time at which they had such access, and
what they did to the original evidence when they had access to it. These chains of custody
add three major sources of uncertainty for intelligence analysts to consider, all of which
are associated with the persons in the chains of custody, whose competence and credibil-
ity need to be considered. The first and most important question involves authenticity:
Is the evidence received by the analyst exactly what the initial evidence said, and is it
complete? The other questions involve assessing the reliability and accuracy of the
processes used to produce the evidence if it is tangible in nature or also used to take
various actions on the evidence in a chain of custody, whether the evidence is tangible or
testimonial. As an illustration, consider the situation from Figure 4.27. We have an item of
testimonial HUMINT coming from a foreign national whose code name is “Wallflower,”
who does not speak English. Wallflower gives his report to the case officer Bob. This
report is recorded by Bob and then translated by Husam. Then Wallflower’s translated
report is transmitted to the report’s officer Marsha, who edits it and transmits it to the
analyst Clyde, who evaluates it.
Figure 4.28 shows how a Disciple-EBR agent may determine the believability of the
evidence received by the analyst. A more detailed discussion is provided in Schum
et al. (2009).
The case officer might have intentionally overlooked details in his recording of Wall-
flower’s report. Thus, as shown at the bottom of Figure 4.28, the believability of the
recorded testimony of Wallflower is the minimum between the believability of Wallflower
and the believability of the recording. Then Husam, the translator, may have intentionally
altered or deleted parts of this report. Thus, the believability of the translated recording is
the minimum between the believability of the recorded testimony and the believability of
the translation by Husam. Then Marsha, the report’s officer, might have altered or deleted
parts of the translated report of Wallflower’s testimony in her editing of it, and so on.

Wallflower’s
received Wallflower’s
testimony about reported Wallflower’s
translated Wallflower’s
Emir Z. in English. testimony about
testimony about recorded
Emir Z. in English.
Emir Z. in English. testimony about Wallflower’s
Emir Z. in Farsi. testimony about
E005-Emir- Emir Z. in Farsi.
Iran E004-Marsha-
report E003-Husam-
translation E002-Bob-
recording E001-Wallflower-
testimony

recording
translation
Clyde editing
transmission
Wallflower
• Competence
• Credibility
Bob
Marsha Husam A.
Marsha • Competence ce
e
• Competence • Competence
• Competence ce
e • Veracity Sony Recorder
• Credibility • Credibility
• Veracity SN 247 • Fidelity
• Fidelity • Reliability
• Reliability
• Security

Figure 4.27. The chain of custody of the Wallflower’s testimony.

13:52:06,
.005
 140 Chapter 4. Modeling the Problem-Solving Process

Believability of transmitted, edited,

translated, recorded testimony of Wallflower
min
Believability of Believability of
edited translation transmission by Marsha
min
Believability of Believability of
translated recording editing by Marsha
min
Believability of Believability of
recorded testimony translation by Husam
min
Believability of Believability of
Wallflower recording by Bob

Figure 4.28. Assessing the believability of Wallflower’s testimony.

The result of these actions is that the analyst receiving this evidence almost certainly
did not receive an authentic and complete account of it, nor did he receive a good account
of its reliability and accuracy. What Clyde received was the transmitted, edited, translated,
and recorded testimony of Wallflower. Although the information to make such an analysis
may not be available, the analyst should adjust the confidence in his conclusion in
recognition of these uncertainties.

4.8 HANDS ON: BELIEVABILITY ANALYSIS

This case study, which continues the analysis from Section 4.5 with the analysis of the
hypothesis, “The cesium-137 canister is used in a project without being checked-out from
the XYZ warehouse,” has two main objectives:

 Learning to define a more detail representation of an item of evidence

 Better understanding the process of believability analysis

In Section 4.6, we have presented how you can define an item of evidence, and Figure 4.23
(p. 132) shows the definition of E001-US-top-oil-importer with type evidence. You can
specify the type by clicking on the [CHANGE] button. Figure 4.29, for instance, shows the
definition of E014-Grace. After you click on the [CHANGE] button, the agent displays the
various evidence types from the right panel. You just need to click on the [SELECT] button
following the correct type, which in this case is unequivocal testimonial evidence based upon
direct observation.
Once you have selected the type of E014-Grace, the agent displays it after the label Type
and asks for its source, which is Grace (see Figure 4.30).
As shown in Figure 4.30, we have also indicated that this item of evidence disfavors the
hypothesis “The missing cesium-137 canister is used in a project at the XYZ company.” As a
result, the agent introduced it into the analysis tree and generated a more detailed analysis
of its believability, which is shown in Figure 4.31.
You can now perform a more detailed believability analysis, as illustrated in Figure 4.32,
where we have assessed the competence, veracity, objectivity, and observational sensitiv-
ity of Grace, and the agent has automatically determined her believability.
In this case study, you will practice the preceding operations. You will first select the
hypothesis, “The cesium-137 canister is used in a project without being checked out from

13:52:06,
.005
4.8. Hands On: Believability Analysis 141

Figure 4.29. Selecting the type of evidence.

Figure 4.30. Definition of an item of evidence.

13:52:06,
.005
142 Chapter 4. Modeling the Problem-Solving Process

Figure 4.31. Decomposition of the believability assessment for an item of testimonial evidence.

Figure 4.32. More detailed believability analysis.

the XYZ warehouse.” Then you will browse its analysis to see how it is reduced to simpler
hypotheses that need to be assessed based on the evidence. After that, you will represent a
new item of evidence, will associate it with the hypothesis to which it is relevant, assess its
relevance, evaluate its believability by assessing its credentials, and browse the resulting
analysis tree.

13:52:06,
.005
 4.9. Drill-Down Analysis 143

Start Disciple-EBR, select the case study knowledge base “04-Believability-Analysis/

Scen,” and proceed as indicated in the instructions from the bottom of the opened window.
This case study illustrates the general operation of defining an item of evidence,
summarized as follows.

Operation 4.6. Define an item of evidence

 In the Evidence workspace, click on the Evidence menu at the top of the window.
 Notice the four modes of operations from the top part of the left panel. Because the
selected one is [AVAILABLE EVIDENCE], the left panel shows the current evidence (if any)
from the knowledge base.
 In the left panel, click on [NEW]. The right panel now shows a partially defined item of
evidence, such as E002-. You will complete the definition of this item of evidence.
 Complete the name E. . .- at the top of the right panel and click on [SAVE].
 Click on [EDIT] for Description, click inside the pane and type the description of the item
of evidence.
 Click on [SAVE].
 You may now provide additional information about the item of evidence (as indicated
in the following steps) or define additional items of evidence (by repeating the
preceding steps).
 After “Type: evidence,” click on [CHANGE] to specify the type of this item of evidence.
 Inspect the different evidence types and click on [SELECT] following the type corres-
ponding to the current item of evidence.
 Provide the additional, type-related information, requested by the system (e.g., the
source in the case of a testimonial item of evidence).

4.9 DRILL-DOWN ANALYSIS, ASSUMPTION-BASED

REASONING, AND WHAT-IF SCENARIOS

An important feature of the Disciple-EBR agent is that it allows you to perform analyses at
different levels of detail. What this means is that a hypothesis may be reduced to many levels
of subhypotheses or just a few levels that are then assessed based on relevant evidence. The
same applies to assessing the believability of evidence. You may directly assess it, as was
illustrated in Figure 4.14 (p. 127), where the believability of E005-Ralph was assessed as very
likely. But if an item of evidence has an important influence on the analysis, then you may
wish to perform a deeper believability analysis, as was illustrated in Figure 4.32, where the
user assessed lower-level believability credentials. The user could have drilled even deeper
to assess the source’s access and understandability instead of his or her competence.
It may also happen that you do not have the time or the evidence to assess a
subhypothesis, in which case you may make various assumptions with respect to its
probability. Consider, for example, the analysis from the case study in Section 4.5, partially
shown in Figure 4.15 (p. 128) and the four subhypotheses of the top-level hypothesis. The
first three of these subhypotheses have been analyzed as discussed in the previous
sections. However, for the last subhypothesis, you have made the following assumption:

It is certain that the MDC-578 truck left with the cesium-137 canister.

Assumptions are distinguished from system-computed assessments by the fact that the
assumed probabilities have a yellow background.

13:52:06,
.005
 144 Chapter 4. Modeling the Problem-Solving Process

You may provide justifications for the assumptions made. You may also experiment
with various what-if scenarios, where you make different assumptions to determine their
influence on the final result of the analysis.
Thus the agent gives you the flexibility of performing the analysis that makes the best
use of your time constraints and available evidence.
The Disciple-EBR shell includes a customized modeling assistant to model the hypoth-
esis analysis process. The following two case studies demonstrate its use.

4.10 HANDS ON: MODELING, FORMALIZATION,

AND PATTERN LEARNING

The objective of this case study is to learn how to use Disciple-EBR to model the analysis of
a hypothesis. More specifically, you will learn how to:

 Specify a new hypothesis

 Specify a question/answer pair that suggests how the hypothesis can be reduced to
simpler hypotheses
 Specify the subhypotheses suggested by the question/answer pair
 Select ontology names to be used in hypotheses, questions, and answers
 Convert a hypothesis to an elementary solution (assessment)
 Formalize a reasoning tree or a part of it to learn reduction patterns
 Convert formalized nodes back to modeling to further update them

This case study will guide you through the process of defining and analyzing a hypothesis
by using, as an example, the following hypothesis: “CS580 is a potential course for Mike
Rice.” You will first define the reduction tree shown in Figure 4.33. Then you will formalize
it and specify the synthesis functions.
Start Disciple-EBR, select the case study knowledge base “05-Modeling-Learning/Scen,”
and proceed as indicated in the instructions from the bottom of the opened window.

Figure 4.33. Hypotheses reduction tree.

13:52:06,
.005
4.10. Hands On: Modeling, Formalization 145

This case study illustrates several important operations, which are described in the
following.

Operation 4.7. Specify a new hypothesis

 In the Evidence workspace, click on the Hypothesis menu at the top of the window.
 Click on [NEW].
 At the top of the window, click on [NEW EXAMPLE]. The Reasoner module is automatic-
ally selected.
 Double-click on the red-border box and define the hypothesis: "CS580 is a potential
course for Mike Rice.”
 Define instances and concepts as indicated in Operation 4.8.
 Click outside the editing box when finished.

Operation 4.8. Define instances and constants

 In the Evidence workspace, in the Reasoner module, while editing a node in the
reasoning tree, select the text representing the instance or constant, right-click on it,
and select the corresponding type, as illustrated in Figure 4.34.

Operation 4.9. Specify question/answer nodes, subhypotheses,

and solutions
 In the Evidence workspace, in the right panel of the Reasoner module, right-click on
the node under which the new node will be defined and select Generate Suggestions.
 If only a generic suggestion is made, double-click on that node and write the
desired text.
 If several suggestions are made, right-click on one of them and select Accept Suggestion.

Operation 4.10. Insert an intermediary hypothesis

 In the Evidence workspace, in the right panel of the Reasoner module, right-click on
the hypothesis node above which an intermediary hypothesis is to be inserted and
select New Hypothesis Above.

Operation 4.11. Move a hypothesis to the left or right

 In the Evidence workspace, in the right panel of the Reasoner module, right-click on a
subhypothesis that has siblings, and select Move Left or Move Right to move it to the
left of its left sibling or to the right of its right sibling.

Operation 4.12. Delete question/answer nodes, subhypotheses,

and solutions
 When a node is deleted, the entire subtree under it is also deleted.
 In the Evidence workspace, in the right panel of the Reasoner module, right-click on
the node to be deleted and select Remove Node.

Figure 4.34. Defining Mike Rice as a specific instance.

13:52:06,
.005
 146 Chapter 4. Modeling the Problem-Solving Process

Operation 4.13. Convert between modeled and formalized

hypotheses and solutions/assessments
 In the Evidence workspace, in the right panel of the Reasoner module, right-click on
the node to be converted and select the desired and applicable conversion option, for
example, Convert to Modeling (to convert a formalized node), Learn Hypothesis
Pattern (to formalize a modeled hypothesis and learn a hypothesis pattern), Learn
Tree Patterns (to formalize the entire tree under a hypothesis and learn reduction
patterns), Change to Assessment (to convert a hypothesis into an assessment), or
Change to Hypothesis (to convert an assessment into a hypothesis to be assessed).

Operation 4.14. Define an assumption

 In the Evidence workspace, in the right panel of the Reasoner module, right-click on a
formalized or learned hypothesis node and select New Assumption.
 If several assumption patterns are proposed, select the desired one from the
displayed list.
 If necessary, change the automatically selected probability value by clicking on it and
selecting another one from the displayed list.

Operation 4.15. Define an assumption with justification

 Define an assumption following the steps from Operation 4.14.
 Click on the Assumption menu at the top of the window.
 Click on [MODIFY].
 Click on the underlined space following Justification and write your justification in the
opened editor. Then click outside the box.
 Click on [SAVE] and then on the Reasoner menu.

Operation 4.16. Delete an assumption

 In the Evidence workspace, in the right panel of the Reasoner module, right-click on
the hypothesis node with the assumption to be deleted and select Delete Assumption.

4.11 HANDS ON: ANALYSIS BASED ON LEARNED PATTERNS

The objective of this case study is to learn how to use Disciple-EBR to model the analysis of
a hypothesis by reusing learned patterns. More specifically, you will learn how to:

 Specify a new hypothesis by instantiating a learned pattern

 Specify the reduction of a hypothesis by reusing a learned reduction pattern
 Instantiate variables in a reduction
 Understand how the solution composition functions from the employed patterns are
automatically applied

You will first define the hypothesis by selecting an existing pattern and instantiating it to:
“CS681 is a potential course for Dan Bolt.” Then you will successively reduce it to simpler
hypotheses by reusing learned patterns. This will include the instantiation of variables
from the learned patterns.
Start Disciple-EBR, select the case study knowledge base “06-Analysis-Reuse/Scen” and
proceed as indicated in the instructions from the bottom of the opened window.
This case study illustrates several important operations described in the following.

13:52:06,
.005
 4.12. Modeling Guidelines 147

Operation 4.17. Specify a new hypothesis by instantiating a pattern

 In the Evidence workspace, click on the Hypothesis menu at the top of the window.
 Click on [NEW].
 Click on the pattern to instantiate and notice that each pattern variable is replaced
with “. . .”
 Click on each “. . .” and, in the text field that appears under it, write the desired value
and press the Enter key.
 Select an answer from those proposed by the system.
 After all the values have been defined, click on [CREATE].

Operation 4.18. Specify the reduction of a hypothesis by reusing

a learned pattern
 In the Evidence workspace, in the right panel of the Reasoner module, right-click on
a formalized hypothesis node and select Generate Suggestions.
 If the knowledge base has applicable learned patterns, it will propose them together
with a generic suggestion.
 If a generated pattern to be selected contains variables, click on each of them and
select the desired values.
 When the pattern to be selected is completely instantiated, right-click on its Q/A node
and select Accept Suggestion.

4.12 MODELING GUIDELINES

The following are several knowledge engineering guidelines for modeling the reasoning
process. In general, we will refer to hypotheses in these guidelines, although the guidelines
are applicable to problems as well. To make this clearer, Guideline 4.1 uses the form “problem/
hypothesis,” and we will illustrate it with planning problems. However, the rest of the
guidelines refer only to “hypothesis,” although “hypothesis” may be replaced with “problem.”

Guideline 4.1. Structure the modeling process based on

the agent’s specification
A main result of the agent specification phase (see Table 3.1, p. 83) is the identification of
the types of problems to be solved by the envisioned agent or the types of hypotheses to be
analyzed. The entire modeling process can be structured based on the types or classes of
these problems or hypotheses, as indicated Table 4.4.

Table 4.4 General Structure of the Modeling Process

 Partition the domain into classes of problems/hypotheses.

 Select representative problems/hypotheses for each class.
 Model one class at a time.
 Model one example solution at a time.
 Organize the top-level part of the reasoning tree to identify the class of the problem/hypothesis.

13:52:06,
.005
148 Chapter 4. Modeling the Problem-Solving Process

As an illustration, consider developing a workaround military planning agent that

needs to determine the actions to be performed in order to work around damage to
transportation infrastructures, such a tunnels, bridges, or roads (Tecuci et al., 2000). This
agent, called Disciple-WA, will be presented in Section 12.2. Figure 4.35 shows a possible
organization of the top-level part of the reasoning tree of this agent, which identifies the
class of the current problem to be solved.

Guideline 4.2. Define reduction trees in natural language using

simple questions
Table 4.5 shows a recommended sequence of steps to be followed when developing the
reduction tree for a specific hypothesis.

Guideline 4.3. Identify the specific instances, the generic instances,

and the constants
After defining each hypothesis and question/answer pair, identify the specific instances,
the generic instances, and the constants, such as “certain” or “5.” The agent will
automatically add all the instances under a temporary concept called “user instance,”
as shown in Figure 4.36. The concepts will be identified later as part of ontology
development.

Site 103:cross-section

Workaround Planning System

Damage 200: destroyed bridge

Esmates the best plan for a military unit to work Near approach
(Right approach)
Bridge/River
gap = 25 meters
Far approach
(Left approach)
Site 108 Site 104
around damage to a transportaon infrastructure,
such as a damaged bridge or road.
Left bank
Right bank
Site 105
Site 107

Workaround River bed

Site 106

damaged tunnels
Workaround damaged
bridges with fording

Workaround damaged
Workaround Workaround bridges with fixed bridges
damage damaged bridges
Workaround damaged
bridges with floang bridges

Workaround damaged
bridges with ras
Workaround
damaged roads
Figure 4.35. Sample top-level structuring of the reasoning tree.

13:52:06,
.005
4.12. Modeling Guidelines 149

Table 4.5 The Reduction Tree Modeling Process

1. Identify the hypothesis to be assessed and express it with a clear natural language sentence.
2. Select the instances and constants in the hypothesis.
3. Follow each hypothesis or subhypothesis with a single, concise, question relevant to
decomposing it. Ask small, incremental questions that are likely to have a single category of
answer (but not necessarily a single answer). This usually means asking who, what, where, what
kind of, whether it is this or that, and so on, not complex questions such as “Who and what?” or,
“What and where?”
4. Follow each question with one or more answers to that question. Express answers as complete
sentences, restating key elements of the question in the answer. Even well-formed, simple
questions are likely to generate multiple answers. Select the answer that corresponds to the
example solution being modeled and continue down that branch.
5. Select instances and constants in the question/answer pair.
6. Evaluate the complexity of each question and its answers. When a question leads to apparently
overly complex answers, especially answers that contain an “and” condition, rephrase the
question in a simpler, more incremental manner leading to simpler answers.
7. For each answer, form a new subhypothesis, several subhypotheses, or an assessment
corresponding to that answer by writing a clear, natural language sentence describing the new
subhypotheses or assessment. To the extent that it is practical, incorporate key relevant phrases
and elements of preceding hypothesis names in subhypotheses’ names to portray the expert’s
chain-of-reasoning thought and the accumulation of relevant knowledge. If the answer has led
to several subhypotheses, then model their solutions in a depth-first order.
8. Select instances and constants in each subhypothesis.
9. Utilize the formalization and reuse capabilities of Disciple to minimize the amount of new
modeling required, both for the current hypothesis and for other hypotheses.

Guideline 4.4. Guide the reduction by the possible need

of future changes
Use the reduction pattern in Figure 4.37 when you know all the n factors (i.e., A, B, C)
that lead to the reduction of the top-level hypothesis to n simpler hypotheses. Adding,
changing, or deleting factors after a rule was learned from that reduction is a more difficult
operation that also requires deleting the old rule and learning a new one, so you want to
avoid making such changes.
If, however, you anticipate adding new factors in the future, then you can use the
reduction pattern illustrated in Figure 4.38. You can easily add, change, or delete factors
after the rule is learned by adding, changing, or deleting them in the ontology.

Guideline 4.5. Learn and reuse reduction patterns

Disciple-EBR will learn different rules from reduction steps that have different patterns,
even though their meaning is the same and the only difference is in their wording, such as
“with respect to” as opposed to “wrt” or “from the point of view of.” To avoid the learning

13:52:06,
.005
150 Chapter 4. Modeling the Problem-Solving Process

user instance

instance of
Bob Sharp John Doe

George Mason University Artificial Intelligence

Figure 4.36. Identification of instances and constants in the reasoning tree.

<hypothesis>

Which factors should we consider?

<expression containing factors A, B, and C>

<hypothesis with A> <hypothesis with B> <hypothesis with C>

Figure 4.37. Reduction used when all the relevant factors are known.

<hypothesis>

Which is a factor to consider? Which is a factor to consider?

<expression containing factor A> <expression containing factor D>

<hypothesis with A> <hypothesis with D>

<factor>

instance of Which is a factor to consider? Which is a factor to consider?

<expression containing factor B> <expression containing factor C>
A D
B C
<hypothesis with B> <hypothesis with C>

Figure 4.38. Reduction used when new relevant factors may be added in the future.

13:52:06,
.005
 4.13. Project Assignment 3 151

of semantically redundant rules, you should learn and reuse reduction patterns, as
illustrated in Figure 4.39.

4.13 PROJECT ASSIGNMENT 3

Prototype a preliminary version of the agent that you will develop as part of your project by
working as a team to:

 Define a domain knowledge base as a copy of 00-Reference-KB.

 Think of three hypotheses to analyze.
 Model the reduction of one hypothesis down to the level of hypotheses that will be
addressed by individual team members.

John Doe would be a good PhD advisor

with respect to the
professional reputation criterion.

Which is a professional reputation criterion?

research funding criterion

John Doe would be a good PhD advisor

with respect to the
research funding criterion.

Pattern
learning

?O1 would be a good PhD

advisor with respect to the
?O2.

Which is a ?O2?
?O3

?O1 would be a good PhD

advisor with respect to the
?O3.

Pattern
instantiation

John Doe would be a good PhD advisor John Doe would be a good PhD advisor
with respect to the with respect to the
student learning experience criterion. student learning experience criterion.

Which is a student learning experience criterion? Which is a student learning experience criterion?
research group status criterion student opinion criterion

John Doe would be a good PhD advisor John Doe would be a good PhD advisor
with respect to the with respect to the
research group status criterion. student opinion criterion.

Figure 4.39. Pattern learning and instantiation.

13:52:06,
.005
 152 Chapter 4. Modeling the Problem-Solving Process

 Formalize the reduction tree and define the composition functions.

 Analyze the other two hypotheses by reusing the learned patterns, if applicable.
 Present the prototype agent in class.

4.14 REVIEW QUESTIONS

4.1. Review again Figures 4.3 and 4.4. Then illustrate the application of problem
reduction and solution synthesis with another symbolic integration problem.

4.2. Consider the reductions of Problem1 from Figure 4.40. Indicate the corresponding
solution syntheses.

4.3. Illustrate the reasoning in Figure 4.5 with the problem “Travel from Boston to New
York.” Hint: Consider the question, “Which is a transportation means I can use?”

4.4. How could you use the problem-level synthesis from Question 4.3 to obtain an
optimal solution? What might be some possible optimization criteria?

4.5. Illustrate the reasoning in Figure 4.6 with an example of your own.

4.6. You are considering whether a statement S is true. You search the Internet and find
two items of favoring evidence, E1* and E2*. You estimate that the relevance and the
believability of E1* are “almost certain” and “very likely,” respectively. You also
estimate that the relevance and the believability of E2* are “certain” and “likely,”
respectively. Based on this evidence, what is the probability that S is true? Draw a
reasoning tree that justifies your answer.

4.7. Define the concepts of relevance, believability, and inferential force of evidence.
Then indicate the appropriate synthesis functions and the corresponding solutions
in the reasoning tree from Figure 4.41.

4.8. What are the different types of tangible evidence? Provide an example of each type.

4.9. What items of tangible evidence do you see in Table 4.3?

Problem1

Queson/Answer1 Queson/Answer2

Problem11 Problem12 Problem13 Problem14

Figure 4.40. Problem reductions.

13:52:06,
.005
4.14. Review Questions 153

Hypothesis H1 based
on favoring evidence

function =
Which is a favoring Which is a favoring
item of evidence? E1 item of evidence? E2

function = function =
Relevance Believability Relevance Believability
of E1 to H1 of E1 of E2 to H1 of E2
very likely likely certain almost certain

Figure 4.41. Sample reasoning tree for assessing the inferential force of evidence.

4.10. Provide some examples of tangible evidence for the hypothesis that John Doe
would be a good PhD advisor for Bob Sharp.

4.11. Define the believability credentials of demonstrative tangible evidence.

4.12. What is testimonial evidence? Give an example.

4.13. What are the different types of testimonial evidence? Provide an example of
each type.

4.14. Give some examples from your own experience when you have heard people
providing information about which they hedge or equivocate.

4.15. Provide some examples of testimonial evidence for the hypothesis that John Doe
would be a good PhD advisor for Bob Sharp.

4.16. Define missing evidence. Provide an example of missing evidence.

4.17. Consider our discussion on the cesium-137 canister. Upon further investigation, we
identify the person who rented the truck as Omar al-Massari, alias Omer Riley. We
tell him that we wish to see his laptop computer. We are, of course, interested in
what it might reveal about the terrorists with whom he may be associating. He
refuses to tell us where his laptop is. What inferences might we draw from Omar al-
Massari’s refusal to provide us with his laptop computer?

4.18. What other items of evidence are missing so far in our discussion of the cesium-
137 case?

4.19. Provide some examples of missing evidence for the hypothesis that John Doe
would be a good PhD advisor for Bob Sharp.

4.20. Define the term authoritative record. Provide an example of an authoritative record.

4.21. Define the believability credentials of a source of testimonial evidence.

4.22. What are some types of mixed evidence? Provide an example. Do you see any
example of mixed evidence in Table 4.3?

13:52:06,
.005
154 Chapter 4. Modeling the Problem-Solving Process

4.23. Provide some examples of mixed evidence for the hypothesis that John Doe would
be a good PhD advisor for Bob Sharp.

4.24. Can you provide other examples of mixed evidence from your own experience?

4.25. Which is the general reduction and synthesis logic for assessing a PhD advisor?
Indicate another type of problem that can be modeled in a similar way.

4.26. Use the knowledge engineering guidelines to develop a problem reduction tree for
assessing the following hypothesis based on knowledge from the ontology (not
evidence): “John Doe would be a good PhD advisor with respect to the employers
of graduates criterion.” You do not need to develop the ontology, but the questions
and answers from your reasoning tree should make clear what knowledge would
need to be represented in the ontology. The logic should be clear, all the statements
should be carefully defined, and the question/answer pairs should facilitate learn-
ing. Mark all the instances in the reasoning tree.

4.27. Rapidly prototype an agent that can assess the following hypothesis and others with
a similar pattern: “John Doe would be a good PhD advisor with respect to the
research publication criterion.” Hint: You may consider that a certain number of
publications corresponds to a certain probability for the research publications
criterion. For example, if someone has between 41 and 60 publications, you may
consider that it is very likely that he or she would be a good PhD advisor with
respect to that criterion.

4.28. Rapidly prototype an agent that can assess the following hypothesis and others with
a similar pattern: “John Doe would be a good PhD advisor with respect to the
research funding criterion.” Hint: You may consider that a certain average amount
of annual funding corresponds to a certain probability for the research funding
criterion. For example, if someone has between $100,000 dollars and $200,000, you
may consider that it is very likely that he or she would be a good PhD advisor with
respect to that criterion.

4.29. Rapidly prototype an agent that can assess the following hypothesis and others with
a similar pattern: “John Doe would be a good PhD advisor with respect to the
publications with advisor criterion.” Hint: You may consider that a certain number
of publications of PhD students with the advisor corresponds to a certain probabil-
ity for the publications with advisor criterion.

4.30. Rapidly prototype an agent that can assess the following hypothesis and others with
a similar pattern: “John Doe would be a good PhD advisor with respect to the
research group status criterion.”

13:52:06,
.005
 5 Ontologies

5.1 WHAT IS AN ONTOLOGY?

An ontology is an explicit formal specification of the terms that are used to represent
an agent’s world (Gruber, 1993).
In an ontology, definitions associate names of entities in the agent’s world (e.g.,
classes of objects, individual objects, relations, hypotheses, problems) with human-
readable text and formal axioms. The text describes what a name means. The axioms
constrain the interpretation and use of a term. Examples of terms from the ontology of
the PhD advisor assessment agent include student, PhD student, professor, course, and
publication. The PhD advisor assessment agent is a Disciple agent that helps a PhD
student in selecting a PhD advisor based on a detailed analysis of several factors,
including professional reputation, learning experience of an advisor’s students, respon-
siveness to students, support offered to students, and quality of the results of previous
students (see Section 3.3). This agent will be used to illustrate the various ontology issues
discussed in this chapter.
The ontology is a hierarchical representation of the objects from the application
domain. It includes both descriptions of the different types of objects (called concepts or
classes, such as professor or course) and descriptions of individual objects (called instances
or individuals, such as CS580), together with the properties of each object and the
relationships between objects.
The underlying idea of the ontological representation is to represent knowledge in the
form of a graph (similar to a concept map) in which the nodes represent objects, situations,
or events, and the arcs represent the relationships between them, as illustrated in Figure 5.1.
The ontology plays a crucial role in cognitive assistants, being at the basis of knowledge
representation, user–agent communication, problem solving, knowledge acquisition, and
learning.
First, the ontology provides the basic representational constituents for all the elements
of the knowledge base, such as the hypotheses, the hypothesis reduction rules, and the
solution synthesis rules. It also allows the representation of partially learned knowledge,
based on the plausible version space concept (Tecuci, 1998), as discussed in Section 7.6.
Second, the agent’s ontology enables the agent to communicate with the user and with
other agents by declaring the terms that the agent understands. Consequently, the ontol-
ogy enables knowledge sharing and reuse among agents that share a common vocabulary
that they understand. An agreement among several agents to use a shared vocabulary in a
coherent and consistent manner is called ontological commitment.

155

13:52:21,
.006
 156 Chapter 5. Ontologies

Third, the problem-solving methods or rules of the agent are applied by matching them
against the current state of the agent’s world, which is represented in the ontology. The
use of partially learned knowledge (with plausible version spaces) in reasoning allows
assessing hypotheses (or solving problems) with different degrees of confidence.
Fourth, a main focus of knowledge acquisition is the elicitation of the domain concepts
and of their hierarchical organization, as will be discussed in Section 6.3.
And fifth, the ontology represents the generalization hierarchy for learning, in which
specific problem-solving episodes are generalized into rules by replacing instances with
concepts from the ontology.

5.2 CONCEPTS AND INSTANCES

A concept (or class) is a general representation of what is common to a set of instances

(or individuals). Therefore, a concept may be regarded as a representation of that set of
instances. For example, professor in Figure 5.2 represents the set of all professors, which
includes Amanda Rice and Dan Smith.

object
subconcept of
...

publication
course
subconcept of
professor
subconcept of article
subconcept of
university course subconcept of
associate professor
journal paper
instance of instance of
instance of
has as reading has as author
Mason-CS480 Doe 2000 John Doe
has as reading
U Montreal-CS780

Figure 5.1. Fragment of an ontology.

professor

instance of instance of

Amanda Rice Dan Smith

Figure 5.2. A concept and two of its instances.

13:52:21,
.006
 5.3. Generalization Hierarchies 157

An instance (individual) is a representation of a particular entity in the application

domain, such as Amanda Rice. We indicate that an instance belongs to a concept by using
the relation instance of:

Amanda Rice instance of professor

5.3 GENERALIZATION HIERARCHIES

Generalization is a fundamental relation between concepts. A concept P is said to be more

general than (or a generalization of) another concept Q if and only if the set of instances
represented by P includes the set of instances represented by Q.
Figure 5.3 shows several concepts with different degrees of generality. For example,
person is more general than student because any student must be a person or, in other
words, the set of all persons includes the set of all students.
Let us notice that the preceding definition of generalization is extensional, based upon the
instance sets of concepts. In order to show that P is more general than Q, this definition would
require the computation of the (possibly infinite) sets of the instances of P and Q. Therefore, it
is useful in practice only for showing that P is not more general than Q. Indeed, according to
this definition, it is enough to find an instance of Q that is not an instance of P because this
shows that the set represented by Q is not a subset of the set represented by P. Section 8.3
discusses generalization rules that allow the agent to compare the generality of the concepts
by working with their descriptions rather than their sets of instances.
One may express the generality relation between two concepts by using the relation
subconcept of:

student subconcept of person

Other names used for expressing this type of relation are subclass of, type, and isa.
A concept Q is a direct subconcept of a concept P if an only if Q is a subconcept of P and
there is no other concept R such that Q is a subconcept of R and R is a subconcept of P.
One may represent the generality relations between the concepts in the form of a
partially ordered graph that is usually called a generalization hierarchy (see Figure 5.4).
The leaves of the hierarchy in Figure 5.4 are instances of the concepts that are represented
by the upper-level nodes. Notice that (the instance) John Doe is both an associate professor
and a PhD advisor. Similarly, a concept may be a direct subconcept of several concepts.

person
P
university employee
Q
faculty staff student
member member

Figure 5.3. Concepts of different generality.

13:52:21,
.006
 158 Chapter 5. Ontologies

5.4 OBJECT FEATURES

The objects in an application domain may be described in terms of their properties and
their relationships with each other. For example, Figure 5.5 represents Mark White as an
associate professor employed by George Mason University. In general, the value of a feature
may be a number, a string, an instance, a symbolic probability, an interval, or a concept
(see Section 5.9).

5.5 DEFINING FEATURES

A feature is itself characterized by several features that have to be specified when defining
a new feature. They include its domain, range, superfeatures, subfeatures, and
documentation.
The domain of a feature is the concept that represents the set of objects that could have
that feature. The range is the set of possible values of the feature.

person

subconcept of

university employee student

subconcept of
subconcept of
faculty member staff member
graduate undergraduate
subconcept of student student

subconcept of
instructor professor PhD advisor
subconcept of
BS student
subconcept of

assistant associate full

MS student PhD student
professor professor professor

instance of instance of instance of instance of instance of

John Smith John Doe Jane Austin Joan Dean Bob Sharp

Figure 5.4. A generalization hierarchy.

associate professor university

instance of instance of
has as employer
Mark White George Mason University

Figure 5.5. Sample object description.

13:52:21,
.006
5.5. Defining Features 159

For example, Figure 5.6 shows the representation of the has as employer feature. Its
domain is person, which means that only entities who are persons may have an employer.
Its range is organization, meaning that any value of such a feature should be an
organization.
There are several types of ranges that could be defined with Disciple-EBR: Concept,
Number, Symbolic interval, Text, and Any element.
We have already illustrated a range of type “Concept” (see Figure 5.6). A range of type
“Number” could be either a set or an interval of numbers, and the numbers could be
either integer or real. A range of type “Symbolic interval” is an ordered set of symbolic
intervals. A range of type “Text” could be any string, a set of strings, or a natural language
text. Finally, a range of type “Any element” could be any of the aforementioned entities.
As will be discussed in more detail in Chapter 7, the knowledge elements from the
agent’s knowledge base, including features, may be partially learned. Figure 5.7 shows an
example of the partially learned feature has as employer. The exact domain is not yet
known, but its upper and lower bounds have been learned as person and professor,
respectively. This means that the domain is a concept that is less general than or as
general as person. Similarly, the domain is more general than or as general as professor.

feature

subconcept of documentation
indicates the employer of a person
domain
has as employer person

range organization
Figure 5.6. The representation of a feature.

has as employer
Mark White George Mason University

feature

indicates the employer of an individual

subconcept of
documentation
plausible upper bound: person
domain
has as employer plausible lower bound: professor

range plausible upper bound: employer

plausible lower bound: university

Figure 5.7. Partially learned feature.

13:52:21,
.006
 160 Chapter 5. Ontologies

Through further learning, the agent will learn that the actual domain is person, as
indicated in Figure 5.6.
Features are also organized in a generalization hierarchy, as illustrated in Figure 5.8.

5.6 REPRESENTATION OF N-ARY FEATURES

Let us suppose that we want to represent the following information in the ontology: “John
Doe has written Windows of Opportunities.” This can be easily represented by using a
binary feature:

John Doe has as writing Windows of Opportunities.

But let us now suppose that we want to represent “John Doe has written Windows of
Opportunities from 2005 until 2007.” This information can no longer be represented with a
binary feature, such as has as writing, that can link only two entities in the ontology. We
need to represent this information as an instance (e.g., Writing 1) of a concept (e.g.,
writing), because an instance may have any number of features, as illustrated in Figure 5.9.
Disciple-EBR can, in fact, represent n-ary features, and it generates them during
learning, but the ontology tools can display only binary features, and the user can define
only binary features.

domain force
has as economic factor
range
economic factor

domain force
has as information network or system
range
information network or system

domain force
has as commerce authority
range
commerce authority

domain force
has as transportation factor
range
transportation factor

domain force
has as strategic raw material
range
strategic raw material

domain force
has as transportation center
range
domain force transportation center
has as industrial factor
range
industrial factor

domain force
has as transportation network or system
range
transportation network or system

domain force
has as industrial authority
range
industrial authority

domain force
has as industrial center
range
industrial center

Figure 5.8. A generalization hierarchy of features from the COG domain (see Section 12.4).

13:52:21,
.006
 5.7. Transitivity 161

writing
Windows of Opportunities
instance of
has as title
John Doe
has as author
Writing1
has as start time
2005
has as end time

2007
Figure 5.9. Representation of an n-ary relation as binary features.

person person person

subconcept of

student

subconcept of

graduate
student

subconcept of subconcept of

MS student MS student

instance of instance of

Joan Dean Joan Dean

Figure 5.10. Use of the properties of instance of and subconcept of.

5.7 TRANSITIVITY

The instance of and subconcept of relations have the following properties:

8x,8y,8z, (x subconcept of y) ∧ (y subconcept of z) ➔ (x subconcept of z)

8x,8y,8z, (x instance of y) ∧ (y subconcept of z) ➔ (x instance of z)

As one can see, subconcept of is a transitive relation and, in combination with instance of,
allows inferring new instance of relations. Let us consider, for example, the hierarchy
fragment from the middle of Figure 5.10. By applying the aforementioned properties of
instance of and subconcept of, one may infer that:

Joan Dean instance of person

MS student subconcept of person

13:52:21,
.006
 162 Chapter 5. Ontologies

5.8 INHERITANCE

A theorem that is implicitly represented in an ontology is the inheritance of features from a

more general concept to a less general concept or an instance.
An instance inherits the features of the concepts to which it belongs:

8x,8y,8z, (x instance of y) ∧ (y feature z) ➔ (x feature z)

Similarly, a concept inherits the properties of its superconcepts:

8x,8y,8z, (x subconcept of y) ∧ (y feature z) ➔ (x feature z)

For example, in the case of the ontology in Figure 5.11, one can infer:

professor retirement age 66

assistant professor retirement age 66
John Smith retirement age 66

by inheriting the retirement age property from

faculty member retirement age 66

The inheritance of properties is one of the most important strengths of an ontology, allowing a
compact and economical representation of knowledge. Indeed, if all the instances of a concept
C have the property P with the same value V, then it is enough to associate the property P with
the concept C because it will be inherited by each of the concept’s instances. There are,
however, two special cases of inheritance to which one should pay special attention: default
inheritance and multiple inheritance. They are discussed in the following subsections.

5.8.1 Default Inheritance

In many domains, exceptions to general rules exist. For example, it is generally useful to
assume that all birds can fly. Certain birds, however, such as the ostrich and the kiwi,
cannot fly. In such a case, it is reasonable to use a representation scheme in which
properties associated with concepts in a hierarchy are assumed to be true for all sub-
concepts and instances, unless specifically overridden by a denial or modification associ-
ated with the subconcepts or the instances.
Let us consider again the hypothetical example in Figure 5.11. The fact that the retire-
ment age of assistant professor is 66 is inherited from faculty member. On the other hand, full
professor does not inherit this property from faculty member because it is explicitly repre-
sented that the retirement age of full professor is 70. This overrides the default inherited from
faculty member. Therefore, to find a feature of some object, the agent will first check whether
the feature is explicitly associated with the object and take the corresponding value. Only if
the feature is not explicitly associated with the object will the agent try to inherit it from the
superconcepts of the object, by climbing the generalization hierarchy.

5.8.2 Multiple Inheritance

It is possible for a concept or an instance to have more than one direct superconcept.
For example, in the ontology from Figure 5.11, John Doe is both an associate professor and

13:52:21,
.006
 5.9. Concepts as Feature Values 163

retirement
faculty member age 66

subconcept of

instructor professor PhD advisor

subconcept of

retirement
assistant professor associate professor full professor age 70

instance of instance of instance of

John Smith John Doe Jane Austin

Figure 5.11. An illustration of inheritance.

concept-1 feature-1 concept-2

instance-of instance-of

instance-11 ... instance-1m instance-21 ... instance-2n

feature-1
instance-11 intance-21
... ...
feature-1
instance-11 intance-2n
... ...
feature-1
instance-1m intance-21
... ...
feature-1
instance-1m intance-2n

Figure 5.12. Relationship between two concepts with instances.

a PhD advisor. Therefore, John Doe will inherit features from both of them and there is a
potential for inheriting conflicting values. In such a case, the agent should use some
strategy in selecting one of the values. A better solution, however, is to detect such
conflicts when the ontology is built or updated, and to associate the correct feature value
directly with each element that would otherwise inherit conflicting values.

5.9 CONCEPTS AS FEATURE VALUES

The previous section has discussed how the features of the concepts are inherited.
However, in all the examples given, the value of the feature was a number. The
same procedure will work if the value is an instance or a string. But what happens if
the value is a concept, which has itself a set of instances, as shown in the top part of
Figure 5.12?

13:52:21,
.006
 164 Chapter 5. Ontologies

In this case, each instance of concept-1 inherits feature-1, the value of which is concept-
2, which is the set of all the instances of concept-2, as shown in the bottom part of
Figure 5.12.
One has to exercise care when defining features between concepts. For example, the
correct way to express the fact that a parent has a child is to define the following feature:
has as child
domain parent
range child

On the contrary, the expression, “parent has as child child” means that each parent is the
parent of each child.

5.10 ONTOLOGY MATCHING

Ontology matching allows one to ask questions about the objects in the ontology, such as:
“Is there a course that has as reading a publication by John Doe?”
We first need to express the question as a network fragment with variables, as illus-
trated in the top part of Figure 5.13. The variables represent the entities we are looking for.
We then need to match the network fragment with the ontology to find the values of the
variables, which represent the answer to our question.
For example, John Doe in the pattern is matched with John Doe in the ontology, as
shown in the right-hand side of Figure 5.13. Then, following the has as author feature (in
reverse), ?O2 is successfully matched with Doe 2000 because each of them is a publication.
Finally, following the has as reading feature (also in reverse), ?O1 is successfully matched
with Mason-CS480 and with U Montreal-CS780, because each of them is an instance of a
course. Therefore, one obtains two answers of the asked question:
Yes, Mason-CS480 has as reading Doe 2000, and Doe 2000 has as author John Doe.
Yes, U Montreal-CS780 has as reading Doe 2000, and Doe 2000 has as author John Doe.

course publication
is is
has as reading has as author
?O1 ?O2 John Doe

publication
course professor
subconcept of
subconcept-of article subconcept-of

subconcept of
university course associate professor
journal paper
instance-of instance-of
instance-of
has as reading has as author
Mason-CS480 Doe 2000 John Doe
has as reading
U Montreal-CS780

Figure 5.13. Ontology matching.

13:52:21,
.006
 5.11. Hands On: Browsing an Ontology 165

One important aspect to notice is that the structure of the ontology is also a guide in
searching it. This significantly speeds up the matching process as compared, for example,
to a representation of the same information as a set of predicates.

5.11 HANDS ON: BROWSING AN ONTOLOGY

The objective of this case study is to learn how to use the various ontology browsers of
Disciple-EBR: the Hierarchical Browser, the Feature Browser and the Feature Viewer, the
Association Browser, the Object Browser, and the Object Viewer. These tools are very
similar to the ontology browsers from many other knowledge engineering tools, such as
Protégé (2015) and TopBraid Composer (2012).
Figure 5.14 shows the interface and the main functions of the Hierarchical Browser of
Disciple-EBR, an ontology tool that may be used to browse a hierarchy of concepts and
instances. The hierarchy in Figure 5.14 is rotated, with the most general concept (object)
on the left-hand side and its subconcepts on its right-hand side. The hierarchy can be
rotated by clicking on the Rotate View button. Clicking on the Expand View button leads
to showing additional levels of the hierarchy, while clicking on the Reduce View button
leads to showing fewer levels.
Figure 5.15 shows the interface and the main functions of the Association Browser of
Disciple-EBR, which may be used to browse the objects and their features. This browser is
centered on a given object (e.g., John Doe), showing its features (e.g., John Doe has as
employer George Mason University), the features for which it is a value (e.g., Adam Pearce
has as PhD advisor John Doe), its direct concepts (e.g., PhD advisor and associate professor)
and, in the case of a concept, its direct subconcepts or its direct instances. Double-clicking
on any entity in the interface will center the Association Browser on that entity.
One may also browse the objects and their features by using the Object Browser and the
Object Viewer, as illustrated in Figure 5.16 and described in Operation 5.1.

Expand/reduce tree to
show more/fewer
hierarchy levels

Rotate the tree for

better viewing

Invoke the Finder to

locate an object

Invoke the Association

Browser to view the
selected object

Figure 5.14. The interface and the main functions of the Hierarchical Browser.

13:52:21,
.006
166 Chapter 5. Ontologies

Invoke the Finder

to locate an object

Invoke the Hierarchical Browser

to view the selected object
Click on an object or feature to view
it with the Association Browser

Figure 5.15. The Association Browser and its functions.

Figure 5.16. The interface of the Object Browser (left) and the Object Viewer (right).

13:52:21,
.006
5.11. Hands On: Browsing an Ontology 167

Operation 5.1. View an object

 Select the Scenario workspace.
 Click on the Ontology menu and select Object Browser.
 Click on the object in the Object Browser (e.g., John Doe on the left-hand side of
Figure 5.16).
 Click on the View button to view the object’s features with the Object Viewer (see the
right-hand side of Figure 5.16).
 Click on the Association button to view the object’s features with the Association Browser.

You can browse the feature generalization hierarchy by using the Feature Browser, which
is illustrated in the left-hand side of Figure 5.17. To view the definition of a specific feature,
you may follow the steps in Operation 5.2.

Operation 5.2. View a feature definition

 Select the Scenario workspace.
 Click on the Ontology menu and select Feature Browser.
 Click on the feature in the Feature Browser (e.g., is expert in on the left-hand side of
Figure 5.17).
 Click on the View button to view the feature with the Feature Viewer (see the right-
hand side of Figure 5.17).
 Click on the Association button to view the feature with the Association Browser.

Figure 5.17. The Feature Browser (left) and the Feature Viewer (right).

13:52:21,
.006
 168 Chapter 5. Ontologies

Start Disciple-EBR, select the case study knowledge base “07-Ontology-Browsing/Scen,”

and proceed as indicated in the instructions from the bottom of the opened window.

5.12 PROJECT ASSIGNMENT 4

Extend the preliminary version of the agent that you will develop as part of your project by
analyzing one leaf hypothesis based on several items of evidence, as discussed in Section
4.4 and practiced in the case study from Section 4.5.

5.13 REVIEW QUESTIONS

5.1. What is an ontology?

5.2. Why are ontologies important?

5.3. What is a concept?

5.4. What is an instance?

5.5. What does it mean for a concept P to be more general than a concept Q?

5.6. What are the possible relationships between two concepts A and B, from a gener-
alization point of view? Provide examples of concepts A and B in each of the
possible relationships.

5.7. How could one prove that a concept A is more general than a concept B? Is the
proposed procedure likely to be practical?

5.8. How can one prove that a concept A is not more general than a concept B? Is the
proposed procedure likely to be practical?

5.9. Consider the feature hierarchy from Figure 5.18. Indicate the necessary relationship
between: (a) Domain B and Domain 1; (b) Range B and Range 1; (c) Domain A2
and Domain 1; (d) Domain A and Domain B; (e) Domain 1 and Range 1.

5.10. Consider the knowledge represented in Figure 5.11 (p. 163). What is the retirement
age of John Smith? What is the retirement age of Jane Austin?

5.11. Insert the additional knowledge that platypus lays eggs into the object ontology
from Figure 5.19. Explain the result.

5.12. Explain in English what information is represented in Figure 5.20.

5.13. Explain in English what information is represented in Figure 5.21. How could
one encode the additional information that Clyde owned nest1 from spring
2013 to fall 2013? Hint: You need to restructure the ontology and represent an
n-ary relationship.

5.14. Represent the following information as an ontology fragment: “Bob Sharp enrolled
at George Mason University in fall 2014.”

5.15. Consider the generalization hierarchy from Figure 5.4 (p. 158). Consider the
following information: In general, the retirement age of a faculty member is 66,

13:52:21,
.006
5.13. Review Questions 169

domain Domain 1
feature 1
range
Range 1

domain Domain B
feature B
range
Range B

domain Domain A
feature A
range
Range A

domain Domain A1 domain Domain A2

feature A1 feature A2
range range
Range A1 Range A2

Figure 5.18. Sample feature hierarchy.

has as birth mode

mammal live

subconcept of

cow platypus

Figure 5.19. Ontology fragment.

yellow

has as color

has as mass has as temp

mass Sun Sun temp Sun

attracts
greater than greater than
revolves around

has as value has as mass has as temp

3500 mass Earth Earth temp Earth

Figure 5.20. Knowledge representation.

13:52:21,
.006
170 Chapter 5. Ontologies

but a full professor may retire at 70, although Jane Austin opted to retire at 66. How
could you represent this information?

5.16. How can we deal with the inheritance of contradictory properties? Provide an
example.

5.17. Define the main features of a feature and illustrate each of them with an example.

5.18. Provide an example of a partially learned feature.

5.19. Why it is very important to carefully define the features of a feature?

5.20. What is the meaning of the ontology fragment from Figure 5.22?

5.21. How could one represent the fact that a bird has a nest?

5.22. Explain how the following questions are answered based on the ontology fragment
from Figure 5.23, specifying the types of inference used in each case, and providing
the corresponding answers:
What is the color of membrane?
What does contact adhesive1 glue?
Which are the loudspeaker components made of metal?

5.23. Consider the background knowledge consisting of the object hierarchy from
Figure 5.23.
(a) Which are all the answers to the following question: “Is there a part of a
loudspeaker that is made of metal?”
(b) Which are the reasoning operations that need to be performed in order to
answer this question?
(c) Consider one of the answers that require all these operations and show how
the answer is found.

bird

subconcept of

robin nest

instance of instance of

owns
Clyde nest1

Figure 5.21. Representation with binary predicates.

has as child
mother child

instance of instance of

Mary ... Jane Bob ... Julie

Figure 5.22. Example of a wrong concept feature.

13:52:21,
.006
 object

loudspeaker
part of
inflammable object adhesive toxic substance material fragile object loudspeaker component

paper
made of
membrane
glues color
chassis membrane assembly
glues black
part of
contact adhesive mechanical chassis contains
mowicoll state fluid
.006

glues glues
provider
glues chassis membrane assembly1
glues made of mechanical chassis1
13:52:21,

mowicoll1 made of bolt

metal
bolt1
made of
chassis assembly
part of
state fluid
Glue Inc. caoutchouc contact adhesive1
chassis assembly1
Figure 5.23. Ontology fragment from the loudspeaker domain (Tecuci, 1998). Dotted links indicate instance of relationships while continuous unnamed links
indicate subconcept of relationships.
 object

cleaner loudspeaker component

air mover

hard cleaner damages membrane

soft cleaner

waste material may have

emery paper
removes membrane1
emery paper1

air jet device solvent removes

.006

removes removes surplus adhesive

13:52:21,

surplus adhesive1

may have
air press air sucker acetone alcohol dust entrefer may have surplus paint

air press1 air sucker1 acetone1 alcohol1 dust1 entrefer1 surplus paint1
Figure 5.24. Ontology fragment from the loudspeaker domain (Tecuci, 1998). Dotted links indicate instance of relationships while unnamed continuous links
indicate subconcept of relationships.
5.13. Review Questions 173

5.24. Consider the ontology fragment from Figure 5.24. Notice that each of the most
specific concepts, such as dust or air press, has an instance, such as dust1 and air
press1, respectively.
(a) Represent the question “Is there a cleaner X that removes dust?” as a network
fragment.
(b) Find all the possible answers to this question based on the information in the
ontology fragment.
(c) In order to answer this question, the agent would need to use several
reasoning operations. Which are these operations?

5.25. Consider the following description in the context of the ontology fragment from
Figure 5.24:

?z is cleaner
removes surplus-paint1

Determine all the possible values of ?z.

5.26. Consider the following action description in the context of the ontology fragment
from Figure 5.24:

clean object ?x
of ?y
with ?z
condition
?x is entrefer
may have ?y
?y is object
?z is cleaner
removes ?y

Find all the possible values for the variables ?x, ?y, and ?z. Indicate some of the
corresponding actions.

13:52:21,
.006
 6 Ontology Design and Development

6.1 DESIGN AND DEVELOPMENT METHODOLOGY

Ontology design is a creative process whose first step is determining the scope of the
ontology by specifying its main concepts, features, and instances. One approach is to elicit
them from a subject matter expert or some other sources, as will be discussed in Section 6.3.
Another approach is to extract a specification of the ontology from the reasoning trees
developed as part of the rapid prototyping of the agent. During this phase, the subject
matter expert and the knowledge engineer define a set of typical hypotheses (or problems)
that the envisioned agent should be able to assess (or solve). Then they actually assess
these hypotheses the way they would like Disciple-EBR to assess them. This process
identifies very clearly what concepts and features should be present in the ontology to
enable the agent to assess those types of hypotheses. This modeling-based ontology
specification strategy will be discussed in Section 6.4. Once a specification of the ontology
has been developed, one has to complete its design.
Because ontology design and development is a complex process, it makes sense to
import relevant concepts and features from previously developed ontologies (including
those from the Semantic Web) rather than defining them from scratch. In particular, one
may wish to look for general-purpose ontologies, such as an ontology of time, space, or
units of measures, if they are necessary to the agent under development. Significant
foundational and utility ontologies have been developed and can be reused (Obrst et al.,
2012), as discussed in Section 3.2.2.
The actual development of the ontology is performed by using ontology tools such as
Protégé (Noy and McGuinness, 2001) or those that will be presented in this section. As will
be discussed next, ontology development is an iterative process during which additional
concepts, features, and instances are added while teaching the agent to assess hypotheses
(or solve problems).
An important aspect to emphasize is that the ontology will always be incomplete.
Moreover, one should not attempt to represent all of the agent’s knowledge in the
ontology. On the contrary, the ontology is intended to represent only the terms of the
representation language that are used in the definitions of hypotheses and rules. The more
complex knowledge will be represented as rules.

6.2 STEPS IN ONTOLOGY DEVELOPMENT

Table 6.1 presents the ontology development steps, which are also illustrated in Figure 6.1.

174

13:52:56,
.007
6.2. Steps in Ontology Development 175

Table 6.1 Ontology Development Steps

1. Define basic concepts (types of objects) and their organization into a hierarchical structure (the
generalization hierarchy).
2. Define object features by using the previously defined concepts to specify their domains and
ranges.
3. Define instances (specific objects) by using the previously defined concepts and features.
4. Extend the ontology with new concepts, features, and instances.
5. Repeat the preceding steps until the ontology is judged to be complete enough.

1. Define concepts actor

subconcept of
person organization

subconcept of subconcept of
university employee student educational organization
subconcept of subconcept of subconcept of
staff member faculty member graduate student university

undergraduate student
associate professor

feature
documentation the employer
2. Use concepts to subfeature of of a person
define features has as employer domain person

range
organization

associate professor university

3. Use concepts and
features to define instance of instance of
instances and
has as employer
their features Mark White George Mason University

Figure 6.1. Steps in ontology development.

First one needs to define basic concepts and to organize them into a hierarchical
structure. This may be performed by using the Object Browser of Disciple-EBR, as will
be discussed in Section 6.5.
Once a set of basic concepts have been defined, one can define features that use these
concepts as their domains and ranges. For example, one may define the feature has as
employer with the domain person and range organization, which are previously defined
concepts. The features are defined and organized in a hierarchy by using the Feature
Browser and the Feature Editor, as discussed in Section 6.7.

13:52:56,
.007
 176 Chapter 6. Ontology Design and Development

With some of the concepts and features defined, one can define instances of these
concepts and associate features with them, as discussed in Section 6.8. For example, one
can define Mark White as an instance of associate professor and specify its feature has as
employer with the value George Mason University.
Ontology development is an iterative process, as indicated by the last step in Table 6.1.
In the case of Disciple-EBR, one does not develop an ontology from scratch. Rather,
one extends the shared ontology for evidence-based reasoning. Moreover, as part of the
rapid prototyping phase, the user has defined the specific instances and the generic
instances used in the sample reasoning trees. All these entities are represented as
instances of the “user instance” concept. Thus, as part of ontology development, one
needs to move all these instances under their proper concepts, as illustrated in
Figure 6.2.

6.3 DOMAIN UNDERSTANDING AND CONCEPT ELICITATION

Concept elicitation consists of determining which concepts apply in the domain, what do
they mean, what is their relative place in the domain, what are the differentiating criteria
distinguishing similar concepts, and what is the organizational structure giving these
concepts a coherence for the expert (Gammack, 1987).
What are some natural ways of eliciting the basic concepts of a domain? Table 6.2 lists
the most common concept elicitation methods. The methods are briefly described in the
following subsections.

PhD student university

instance of instance of
user instance
George Mason
Bob Sharp
University
instance of

George Mason
Bob Sharp University

Artificial
John Doe Intelligence

associate Computer Science

PhD advisor
professor
instance of
instance of
Artificial Software
John Doe Intelligence Engineering

Figure 6.2. Moving the instances of “user instance” under their corresponding concepts.

13:52:56,
.007
 6.3. Domain Understanding and Concept Elicitation 177

Table 6.2 Basic Concept Elicitation Methods

 Preliminary methods
 Tutorial session delivered by expert
 Ad-hoc list created by expert
 Book index
 Interviews with expert
 Unstructured interview
 Structured interview
 Multiple-choice questions
 Dichotomous questions
 Ranking scale questions
 Protocol analysis
 Concept hierarchy elicitation

6.3.1 Tutorial Session Delivered by the Expert

The knowledge engineer asks the subject matter expert to prepare an introductory talk
outlining the whole domain and to deliver it as a tutorial session. Then the knowledge
engineer extracts concepts from the transcript of the talk.

6.3.2 Ad-hoc List Created by the Expert

The knowledge engineer asks the subject matter expert to generate a list of typical
concepts and then systematically probe for more relevant information (e.g., using free
association).

6.3.3 Book Index

The knowledge engineer identifies a representative book on the expertise domain and
extracts concepts from the index of the book.

6.3.4 Unstructured Interviews with the Expert

These are goal-oriented methods used when the knowledge engineer wants to explore an
issue, where the questions and the responses are open-ended. Examples of unstructured
interview fragments have been presented in Tables 3.4 (p. 85), 3.6 (p. 87), and 3.7 (p. 87) in
Section 3.1.

6.3.5 Structured Interviews with the Expert

These are interviews where the questions are fixed in advance. The types of structured
questions include multiple-choice questions (see Table 6.3); dichotomous (yes/no)
questions (see Table 6.4); and ranking scale questions, where the expert is asked to
arrange some items in a list in the order of their importance or preference (see Table 6.5).

13:52:56,
.007
 178 Chapter 6. Ontology Design and Development

Table 6.3 Multiple-Choice Question for the Diabetic Foot Advisor (Awad, 1996)

If a diabetic patient complains of foot problems, who should he or she see first (check one):
□ Podiatrist
□ General practitioner
□ Orthopedic surgeon
□ Physical therapist

Table 6.4 Dichotomous Question for the Diabetic Foot Advisor (Awad, 1996)

Do patients with neuropathy come for regular checkups?

□ Yes
□ No

Table 6.5 Example of a Ranking Scale Question

Please rank the following professional reputation criteria for a PhD advisor in the order of their
importance from your point of view. Give a rank of 1 to the most important criterion, a rank of 2 to
the second most important one, and so on:
___ research funding criterion
___ publications criterion
___ citations criterion
___ peer opinion criterion

What are the main characteristics of the structured interview method? It offers specific
choices, enables faster tabulation, and has less bias due to the way the questions are
formulated. However, this method is restricted by the requirement to specify choices.

6.3.6 Protocol Analysis (Think-Aloud Technique)

This is the systematic collection and analysis of the thought processes or problem-solving
methods of an expert where the expert is asked to solve problems and to verbalize what
goes through his or her mind, stating directly what the expert thinks. The solving process is
carried out in an automatic fashion while the expert talks. The knowledge engineer does
not interrupt or ask questions. The structuring of the information elicited occurs later
when the knowledge engineer analyzes the protocol. Table 6.6 shows an example of a
protocol where a doctor verbalizes the diagnosis of a diabetic foot patient.
Which are the main strengths and weaknesses of all of the preceding concept elicitation
methods? These methods give the knowledge engineer an orientation to the domain,

13:52:56,
.007
 6.4. Modeling-based Ontology Specification 179

Table 6.6. Sample Protocol (Adapted from Awad, 1996)

1. This woman is in her mid- to late forties.

2. The patient being quite overweight and a diabetic, blisters are common occurrences.
3. Pain is symptomatic of the blister.
4. Patient is experiencing this blister for the first time. She’s probably more worried than being in
pain.
5. Being diabetic, blisters take a long time to heal. It is not likely to get worse.
...
40. I don’t see broken skin or pus accumulating, which is a good sign.
41. I’m going to recommend NSD and soaking the foot in warm water before going to bed and
after getting up.
42. Her husband will have to help.
43. I’m going to recommend that patient wear wide-toed shoes.
...
64. So, for the moment, I am going to tell the patient to see me in two weeks.
65. Right now, I wouldn’t recommend any medical treatment. Surgery is the last thing on my
mind.
66. I’ll relay this diagnosis and decision to the patient.

generate much knowledge cheaply and naturally, and do not require a significant effort on
the part of the expert. However, they have an incomplete and arbitrary coverage, and the
knowledge engineer needs appropriate training and/or social skills.

6.3.7 The Card-Sort Method

This is a simple method to develop a hierarchy of concepts. Its main steps are shown in
Table 6.7.
An example of a concept hierarchy elicited through the card-sort method for the
development of a domestic gas-fired hot water and central heating system is shown in
Figure 6.3.
Which are the main strengths and weaknesses of the card-sort method? This method
produces clusters of concepts and hierarchical organization, splits large domains into
manageable subareas, is easy to use, and is widely applicable. However, it is incomplete
and unguided, produces strict hierarchies that are usually too restrictive, and does not
comply with the knowledge engineering guidelines for ontology structuring.
How could we modify the card-sort method to build a tangled hierarchy? We simply need
to write the same concept on more cards, so that it can be included into different groups.

6.4 MODELING-BASED ONTOLOGY SPECIFICATION

Modeling-based ontology specification has already been introduced in Section 3.3. The
knowledge engineer and the subject matter expert analyze each step of the reasoning trees

13:52:56,
.007
 180 Chapter 6. Ontology Design and Development

Table 6.7 The Card-Sort Method (Gammack, 1987)

1. Type the concepts on small individual index cards.

2. Ask the expert to group together the related concepts into as many small groups as possible.
3. Ask the expert to label each of the groups.
4. Ask the expert to combine the groups into slightly larger groups and to label them.
5. Ask the expert to repeat step 4 until a single group is obtained.
The result will be a hierarchical organization of the concepts.

Satchwell
Time Switch Electric Time Controls
Programmer
Thermostat
Set Point Thermostat
Rotary Control Knob
Gas Control Valve
Gas Control Control Electricity
Solenoid
Electrical System
Electrical Supply Electrical Supply
Electrical Contact
Fuse Electrical Components
Pump
Mechanical Components
Motorized Valve
Figure 6.3. Concept hierarchy elicited through the card-sort method (Gammack, 1987).

developed as part of the rapid prototyping phase of agent development, in order to identify
the concepts and the features that should be in the ontology to enable the agent to perform
that reasoning.
Let us consider the reasoning step from the left-hand side of Figure 6.4. To enable the
agent to answer the question from this step, we may define the ontology fragment from the
right-hand side of Figure 6.4.
However, this reasoning step is just an example. We want the agent to be able to answer
similar questions, corresponding to similar hypotheses. Therefore, the ontology fragment
from the right-hand side of Figure 6.4 should be interpreted only as a specification of the
ontological knowledge needed by the agent. This specification suggests that we should
define various university positions, as well as various employers in the ontology, as shown
in Figure 6.5. It also suggests defining two features, has as position and has as employer, as
illustrated in Figure 6.6.

6.5 HANDS ON: DEVELOPING A HIERARCHY OF CONCEPTS

AND INSTANCES

The hierarchy of concepts and instances can be developed by using the Object Browser,
which was introduced in Section 5.11. Its interface and main functions are shown in
Figure 6.7. This tool shows the hierarchy in a tree structure, which the user can expand
or collapse by selecting a node (e.g., educational organization) and then clicking on Expand

13:52:56,
.007
6.5. Hands On: Developing a Hierarchy 181

John Doe will stay on the faculty of

George Mason University for the
duration of dissertation of Bob Sharp.
long-term faculty position university

Is John Doe likely to stay on the faculty of instance of instance of

George Mason University for the duration of
the dissertation of Bob Sharp? tenured position George Mason University
Yes, because John Doe has a tenured position
which is a long-term position.
has as position has as employer

It is almost certain that John Doe will stay on John Doe

the faculty of George Mason University for the
duration of dissertation of Bob Sharp.

Figure 6.4. Modeling-based ontology specification.

position actor

subconcept of subconcept of

university position person organization

subconcept of subconcept of

staff position faculty position

educational organization company
subconcept of
subconcept of
long-term short-term
college university
faculty position faculty position

instance of instance of

tenured position George Mason University

has as position has as employer
John Doe

Figure 6.5. Concept hierarchy design.

feature

subfeature of

domain person
has as employer
range
organization

domain person
has as position
range position

Figure 6.6. Feature hierarchy design.

13:52:56,
.007
182 Chapter 6. Ontology Design and Development

Expand/collapse the subtree of

the selected element

Display the selected element in

Object Viewer / Association
Browser / Hierarchical Browser

Invoke the Finder

to locate an object

Rename the selected element

Invoke the Object Editor to

modify the selected element

Delete the selected element

Define a subconcept / instance

of the selected concept

Figure 6.7. The Object Browser and its main functions.

All and Collapse All, respectively. Similar effects may be obtained by clicking on the – and
+ nodes. Selecting a node and then clicking on the Hierarchical button will open the
Hierarchical Browser with that node as the top of the displayed hierarchy.
The Object Browser can be used to develop a generalization hierarchy by defining
concepts and instances, as described in Operation 6.1 and illustrated in Figure 6.8.

Operation 6.1. Define a subconcept or an instance of a concept

 Open the Object Browser.
 Click on the concept name to select the concept.
 Click on the Concept button (to define a subconcept) or on the Instance button (to
define an instance).
 Write the name of the subconcept/instance.
 Press the Enter key.

If an instance is to be used only in the current scenario, then it should be defined in the
Scenario part of the knowledge base, as described in Operation 6.2. The system will create
it as a specific instance.

Operation 6.2. Define a specific instance of a concept

 Select the Scenario workspace.
 Open the Object Browser and notice that it displays both generic instances and specific
instances.

13:52:56,
.007
6.5. Hands On: Developing a Hierarchy 183

1. Click on the concept

name to select it

3. Write the name of the

subconcept and press Enter

2. Click on “Concept”

Figure 6.8. Defining a subconcept of a concept.

 Define the instance (see Operation 6.1).

 Notice that the instance is displayed in italics, indicating that it is a specific instance.

If an instance needs to be used in more than one scenario, then you have to define it as a
generic instance in the Domain part of the knowledge base, as described in Operation 6.3
and illustrated in Figure 6.9. Such an instance is visible both in the Domain KB and in all
its Scenario KBs and is displayed in regular font, as shown in Figure 6.10.

Operation 6.3. Define a generic instance of a concept

 Select the Domain workspace.
 Open the Object Browser and notice that it displays only generic instances.
 Define the instance (see Operation 6.1).
 Notice that the instance is displayed in regular font, indicating that it is a generic instance.

In the case of the PhD advisor assessment agent, the following instances have been
defined in the Domain part of the knowledge base, and are therefore generic instances:

 The criteria to assess a PhD advisor (e.g., professional reputation criterion)

 Faculty positions (e.g., tenured position)
 PhD research areas (e.g., Artificial Intelligence)

The Scenario part still contains specific instances, such as John Doe, Bob Sharp, and George
Mason University. A specific instance is visible only in the corresponding Scenario KB and is
displayed in italics, as shown in Figure 6.10.

13:52:56,
.007
 2. Notice the domain KB workspace

. Select a Domain KB

3. Define instances
.007
13:52:56,

4. The browser shows only

the generic instances

Figure 6.9. Defining a generic instance.

6.5. Hands On: Developing a Hierarchy 185

1. Notice the scenario

KB (Scen) workspace.

2. The browser will show both specific

instances and generic instances.

Concepts
(light blue straight)

Specific instances
(dark blue italic)

Elements from the

Shared KB (shaded)

Generic instances
(dark blue straight)

Figure 6.10. Specific instances and generic instances.

Notice that the specific instances are displayed with italic font only in the ontology
interfaces. In all the other interfaces, they are displayed in the regular font.
In addition to defining concepts and instances, one should also be able to rename or
delete them. These operations are performed as explained in Operations 6.4 and 6.5,
respectively. Deletion is a particularly complex operation. Disciple-EBR prevents the
deletion of an entity if this would lead to an inconsistent knowledge base where some of
the knowledge base elements refer to the element to be deleted.

13:52:56,
.007
 186 Chapter 6. Ontology Design and Development

Operation 6.4. Rename a concept or an instance

 Open the Object Browser.
 Click on the concept or instance to select it.
 Click on the Rename button.
 Write the new name.
 Press the Enter key.

Operation 6.5. Delete a concept or an instance

 Open the Object Browser.
 Click on the concept or instance to select it.
 Click on the Delete button.

Start Disciple-EBR, select the knowledge base “08-Ontology-Development-Objects/Scen,”

and use the Object Browser to extend the ontology with the following information:

 course subconcept of object

 operating systems course subconcept of course
 artificial intelligence course subconcept of course
 CS571 instance of operating systems course
 CS580 instance of artificial intelligence course

6.6 GUIDELINES FOR DEVELOPING GENERALIZATION

HIERARCHIES

6.6.1 Well-structured Hierarchies

Siblings in a generalization hierarchy are the concepts (or instances) that are direct
subconcepts (instances) of the same concept. For example, assistant professor, associate
professor, and full professor in Figure 5.4 (p. 158) are siblings because they are all direct
subconcepts of professor.

Guideline 6.1. Define similar siblings

In a well-structured generalization hierarchy, all the siblings should have a comparable
level of generality. In particular, they should be either all concepts or all instances. If some
of them are instances, insert one or several concepts that include them, as illustrated in
Figure 6.11.

M M

subconcept of instance of subconcept of

A B C D1 A B C D

instance of

D1

Figure 6.11. Define similar siblings.

13:52:56,
.007
 6.6. Guidelines for Developing Generalization Hierarchies 187

Guideline 6.2. Group similar siblings under natural concepts

The siblings should reflect concepts from the real world. However, if there are too many
siblings, consider whether some of them may be grouped under another (natural) con-
cept, as illustrated in Figure 6.12.

Guideline 6.3. Recognize that a single subconcept may indicate

ontology incompleteness or error
A case of a single subconcept, such as ABC in Figure 6.13, may be an indication of either
an incomplete ontology (where the siblings of ABC are missing) or a modeling error
(where ABC should not defined and A, B, and C are linked directly to M).

6.6.2 Instance or Concept?

In general, a set of individuals is represented as a concept, while a single individual is
represented as an instance. But sometimes a set can be regarded as an individual and
represented as an instance. Consider the hierarchy from Figure 6.14. In that hierarchy,

subconcept of

A B C D E F … X Y Z

subconcept of

ABC DEF … XYZ

subconcept of subconcept of subconcept of

A B C D E F … X Y Z

Figure 6.12. Grouping of similar siblings under natural concepts.

subconcept of

ABC

subconcept of

A B C

Figure 6.13. Single subconcept of a concept.

13:52:56,
.007
 188 Chapter 6. Ontology Design and Development

Figure 6.14. Representation of entities as concepts or as instances.

Artificial Intelligence Magazine is represented as an instance of a journal. But one could have
also represented it as a concept, the instances of which would have been specific issues of
the Artificial Intelligence Magazine. Whether something is represented as an instance or as a
concept influences how Disciple-EBR learns. In particular, when learning a reduction rule
from a reduction example, instances are generalized while the concepts are preserved as
such. Therefore, in many cases, learning considerations determine whether an entity is
represented as an instance or as a concept.

6.6.3 Specific Instance or Generic Instance?

If an instance needs to be used in more than one scenario (such as Artificial Intelligence,
which is needed to evaluate advisors at various universities), then you have to create it in
the Domain knowledge base, and it will be automatically defined as a generic instance.
Similarly, if it would make sense to have general rules that explicitly mention that
instance, then again it should be defined as a generic instance. For example, we may
envision a general rule related to Artificial Intelligence, but not a general rule related to John
Doe. Thus Artificial Intelligence may be defined as a generic instance, but John Doe should be
defined as a specific instance.
The instances created in the Scenario KB will automatically be defined as specific instances.
Notice that, in the ontology interfaces, the names of the generic instances are displayed
in dark blue straight characters (such as Artificial Intelligence), while the names of the
specific instances are displayed in dark blue italics characters (such as John Doe).

6.6.4 Naming Conventions

The following guidelines suggest conventions for naming concepts.

Guideline 6.4. Adopt and follow a naming convention

It is recommended to adopt a naming convention and to strictly adhere to it. This will
both facilitate the understanding of the ontology and will help avoid modeling mistakes.

13:52:56,
.007
 6.7. Hands On: Developing a Hierarchy 189

In Disciple-EBR, it is recommended to use lowercase letters for concept names. It is

also recommended to use either singular or plural consistently, but not both, in naming
the concepts.

Guideline 6.5. Name subconcepts based on superconcepts

Often the names of the subconcepts of a concept include the name of the concept, as
shown in several places in Figure 5.4 (p. 158) and illustrated in Figure 6.15.

6.6.5 Automatic Support

Disciple-EBR will not accept ontology operations that will make the ontology inconsistent
or introduce a circularity, where a concept is both a superconcept and a subconcept of
another concept. Disciple-EBR will not accept definitions of concepts or instances that
have the same names as previously defined ones.

6.7 HANDS ON: DEVELOPING A HIERARCHY OF FEATURES

The hierarchy of features can be developed by using the Feature Browser. Its interface and
main functions are shown in Figure 6.16. Like the Object Browser, this tool shows the
feature hierarchy in a tree structure that the user can expand or collapse by selecting a
node (e.g., has as part) and then by clicking on Expand All and Collapse All, respectively.
Similar effects may be obtained by clicking on the – and + nodes. Selecting a node and
then clicking on the Hierarchical button opens the Hierarchical Browser with that node at
the top of the displayed hierarchy.
The features are defined and organized in a hierarchy by using the Feature Browser,
similarly to how the Object Browser is used to develop a concept hierarchy. The steps
needed to define a new feature (e.g., works for) are those described in Operation 6.6.

Operation 6.6. Define a feature

 Open the Feature Browser.
 Select the superfeature (e.g., feature) of the feature to be defined (e.g., works for).
 Click on the Feature button.
 Specify the name of the feature to be defined (i.e., works for).
 Press the Enter key.

When a user defines a subfeature of a given feature, the domain and the range of the
subfeature are set to be the ones of the superfeature. The user can change them by clicking
on the Modify button of the Feature Browser. This invokes the Feature Editor, which is

student
subconcept of

graduate student undergraduate student

Figure 6.15. Naming concepts.

13:52:56,
.007
190 Chapter 6. Ontology Design and Development

Expand/collapse the subtree of

the selected feature

Display the selected feature in

Feature Viewer or Association
Browser or Hierarchical
Browser

Invoke the Finder to locate a

feature

Rename the selected feature

Invoke the Feature Editor to

modify the selected feature

Delete the selected feature

Define a fact or subfeature of

the selected feature

Figure 6.16. The Feature Browser and its main functions.

illustrated in Figure 6.17. Using the Feature Editor, the user can add or delete super-
features or subfeatures of the selected feature, in addition to modifying its domain
and range.
Figure 6.18 illustrates the process of changing the domain of the works for feature from
object to actor. This process consists of the steps described in Operation 6.7.

Operation 6.7. Change the domain of a feature

 Open the Feature Browser.
 Select the feature with the domain to be changed (e.g., works for).
 Click on the Modify button to invoke the Feature Editor.
 Select the <P1>[domain] tab in the Feature Editor.
 Click on the Add button to open an Object Browser pane.
 Browse and select a concept for the new domain (e.g., actor).
 Click on the Add to domain button in the Object Browser pane.
 Click on the Apply Domain button in the <P1>[domain] tab to commit the addition in
the ontology.

A range of type “Concept” is modified in the same way as a domain (see Operation 6.8).

Operation 6.8. Change the range of a feature

 Open the Feature Browser.
 Select the feature with the range to be changed (e.g., works for).
 Click on the Modify button to invoke the Feature Editor.
 Select the <P2>[range] tab in the Feature Editor.
 Click on the Add button to open an Object Browser pane.

13:52:56,
.007
6.7. Hands On: Developing a Hierarchy 191

Feature to be modified

Invoke the Feature Browser to add

superfeatures of the feature

Delete the selected superfeature

Add/delete subfeatures

Invoke the Object Browser to define

the domain/range of the feature

Delete concepts from domain/range

Commit the modifications

in the ontology

Figure 6.17. The Feature Editor and its main functions.

 Browse and select a concept for the new range (e.g., actor).
 Click on the Add to range button in the Object Browser pane.
 Click on the Apply button in the <P2>[range] tab to commit the addition in the ontology.

For range types other than “Concept,” a type-specific editor is invoked after clicking on the
Add button. For example, Figure 6.19 illustrates the definition of a range, which is the
integer interval [0, 10].
Start Disciple-EBR, select the knowledge base “09-Ontology-Development-Features/
Scen,” and use the Feature Browser and the Feature Editor to represent the following
related features in a hierarchy:

 works for, a subfeature of feature, with domain actor and range actor
 is employed by, a subfeature of works for, with domain employee and range organization

13:52:56,
.007
 192 Chapter 6. Ontology Design and Development

Figure 6.18. Changing the domain of a feature.

Optionally, you may also add the following feature:

 contracts to, a subfeature of works for, with domain actor (inherited from works for) and
range organization

6.8 HANDS ON: DEFINING INSTANCES AND THEIR FEATURES

With some concepts and features defined, one may use the Object Editor to define
instances of these concepts (as discussed in Section 6.5) and associate features with them.
Figure 6.20 illustrates the process of defining the is interested in feature of John Doe. The
steps of this process are those described in Operation 6.9.

13:52:56,
.007
6.8. Hands On: Defining Instances 193

Figure 6.19. Defining a feature range as an integer interval.

Operation 6.9. Define a feature of an object

 Open the Object Browser and select the object (e.g., John Doe).
 Click on the Modify button to open the Object Editor.
 Click on the Add feature button in the Object Editor to open a Feature Browser pane.
 Browse and select the feature to be added to the object (e.g., is interested in).
 Click on the Define for object button in the Feature Browser pane.
 The Finder is automatically opened to locate its value in the ontology (see Figure 6.21).
 Write a part of the name of the value in the Finder text field.
 Click on the Find button.
 Click on the correct value from the list returned by the Finder.
 Click on the Select button.
 If the range of the feature is a symbolic interval or a set of values, a selection pane
is opened in which you can choose the right value of the feature. For a number, simply
type it.

The Object Editor can also be used to update the list of the direct superconcepts of
an object (instance or concept) and the list of the direct subconcepts or instances of a
concept. The actual steps to perform are presented in Operations 6.10 and 6.11. As with all
the ontology operations, Disciple-EBR will not perform them if they would lead to an
inconsistent ontology or a cycle along the subconcept of relation.

13:52:56,
.007
194 Chapter 6. Ontology Design and Development

Figure 6.20. Defining a feature of an instance.

Operation 6.10. Add a direct superconcept to an object

 Locate the object (e.g., John Doe) with the Object Browser and the Finder and click on
the Modify button to open the Object Editor (see Figure 6.20).
 Click on the Add button in the “Super-concepts” pane to open an Object
Browser pane.
 Browse and select the superconcept.
 Click on the Add as parent button in the Object Browser pane.
 Click on the Apply links button in the Object Editor to commit the addition in the
ontology.

Operation 6.11. Add a direct subconcept or an instance to a concept

 Locate the concept with the Object Browser and the Finder and click on the Modify
button to open the Object Editor.
 Click on the Add button in the “Sub-concepts and instances” pane to open an Object
Browser pane.
 Browse and select the subconcept or the instance to be added.
 Click on the Add as child button in the Object Browser pane.
 Click on the Apply links button in the Object Editor to commit the addition in the
ontology.

Start Disciple-EBR, select the knowledge base “10-Ontology-Development-Facts/Scen,” and

use the Ontology tools to extend the representation of John Doe with the following fact:

John Doe is interested in Software Engineering

13:52:56,
.007
 6.9. Guidelines for Defining Features and Values 195

Figure 6.21. Using the Finder to define the value of a feature.

6.9 GUIDELINES FOR DEFINING FEATURES AND VALUES

6.9.1 Concept or Feature?

Guideline 6.6. Represent well-established categories from the real world

as concepts
Almost any distinction from the real world may be represented either as a concept, or as a
feature. While there are no absolute rules for this difficult modeling decision, several
guidelines are useful to follow. For example, two alternative ways of representing the fact
that Bob Evens is a PhD student are:

13:52:56,
.007
 196 Chapter 6. Ontology Design and Development

Bob Evens instance of PhD student

Bob Evens has as student level PhD

The first one is preferable because PhD student is a well-established category.

Guideline 6.7. Define concepts and instances to represent knowledge

corresponding to n-ary relations
A feature in an ontology can naturally represent a binary relation. However, if the relation
is not binary, you need to define a concept or an instance with which you can associate
any number of features. Let us consider the fact that John Doe has written “Windows of
Opportunities.” This can be represented as:

John Doe has as writing Windows of Opportunities

But what if we want to represent the additional knowledge that he has written it from
2005 until 2007? We can no longer associate this information with the has as writing
feature. A solution is to define the concept writing and its instance Writing1, instead of
the feature has as writing, as was illustrated in Figure 5.9 (p. 161).

6.9.2 Concept, Instance, or Constant?

Another modeling issue is to decide whether to represent the value of a feature as a
concept, as an instance, or as a constant (i.e., a string).
If one needs to associate additional properties with a value, then it cannot be represented
as a constant.
Let us consider again the representation from Figure 5.9 (p. 161). If we want to
represent any additional features of Windows of Opportunities (e.g., that it is 258 pages
long), then we have to represent it as an instance (which will have number of pages as a
property with value 258).
If one anticipates learning concepts that will be characterized by various subsets of values,
then one may elect to define the colors as strings, such as white, yellow, and orange. A learned
concept might then be {white, orange}. However, if concepts such as warm color or cold color
are important in the application domain, then one should define a hierarchy of colors where
white, yellow, and orange are defined as generic instances or concepts.
The same considerations apply to numbers. While generally they are represented as
values, one may wish to define an ordered set of intervals (see Section 8.3.6). In this case,
each interval will have a specific name (such as toddler, youth, or mature), which can be
used as value.

6.9.3 Naming of Features

Guideline 6.8. Define feature names that distinguish them from

concept names
It is useful to define names that allow for easy distinction between a concept and a feature.
For example, “author” would not be a good name for a feature because it would not be
clear that “author” denotes a feature. Indeed, you may have a concept named “author.”

13:52:56,
.007
 6.10. Ontology Maintenance 197

Two common practices are to add the “has as” prefix or the “of” suffix to the feature name,
resulting in has as author or author of, respectively.

6.9.4 Automatic Support

The ontology tools of Disciple-EBR have several features that facilitate their use. For
example, when modifying the domain or the range of a feature (see the right-hand
side of Figure 6.18 on p. 192), Disciple-EBR will display only the concepts that are
acceptable values.
When defining a new feature for a given concept or an instance (see the right-hand side
of Figure 6.20, p. 194), Disciple-EBR will only display the features the domain of which
include the given concept or instance.
When defining the value of a given feature (see Figure 6.21 on p. 195), the Finder will
display only those values that are in the range of the feature.

6.10 ONTOLOGY MAINTENANCE

Maintaining the consistency of the ontology is a very complex knowledge engineering

activity because the definitions of the objects and features interact in complex ways. For
example, deleting a concept requires the updating of all the knowledge base elements that
refer to it, such as any feature that contains it in its range or domain, or any concept that
inherits its features.
Consider the ontology fragment from the left part of Figure 6.22. Let us assume that in
the initial state of this ontology, the domain of the feature f is the concept A. Let us further
assume that the instance C has the feature f with value 7. If we now delete the relation “B
subconcept of A” (i.e., B is no longer a subconcept of A), the modified ontology is
inconsistent. Indeed, in the initial state C can have the feature f because the domain of f
is A, and C is included in A (C is an instance of B, which is a subconcept of A). In the
modified ontology however, C can no longer have the feature f because it is no longer
included in A. The important thing to remember from this example is that a modification
done to B generated an error in another part of the knowledge base (at C).

domain domain
f A f A

A A

B B

C f 7 C f 7

Initial State Inconsistent State

Figure 6.22. Complexity of ontology maintenance.

13:52:56,
.007
 198 Chapter 6. Ontology Design and Development

6.11 PROJECT ASSIGNMENT 5

Use the reasoning trees developed as part of the rapid prototyping of your agent and
employ the modeling-based ontology specification method to extend the ontology of
your agent.

6.12 REVIEW QUESTIONS

6.1. What are the basic concept elicitation methods? What are their main strengths?
What are their main weaknesses?

6.2. Briefly describe the card-sort method. What are its main strengths? What are its
main weaknesses? How could one modify this method to build a tangled hierarchy?

6.3. Describe and illustrate the “Well-structured Hierarchies” knowledge engineering

guidelines presented in Section 6.6.1.

6.4. Describe and illustrate the decision of representing an entity as instance or concept.

6.5. Describe and illustrate the “Concept or Feature?” knowledge engineering guide-
lines presented in Section 6.9.1.

6.6. Describe and illustrate the decision of representing an entity as concept, instance,
or constant.

6.7. Describe and illustrate the “Naming Conventions” guidelines, presented in Section
6.6.4 for concepts, and in Section 6.9.3 for features.

6.8. Consider the question/answer pair from Figure 6.23. Specify the ontology frag-
ments that are suggested by this question/answer pair, including instances, con-
cepts, and features definitions (with appropriate domains and ranges that will
facilitate learning). Do not limit yourself to the concepts that are explicitly referred
to, but define additional ones as well, to enable the agent to assess similar hypoth-
eses in a similar way.

Hint 1: Notice that the answer represents an n-ary relation while in an ontology you
may only represent binary relations.
Hint 2: You need to define a hierarchy of concepts that will include those used in
the domains and ranges of the defined features.
Hint 3: Your solution should reflect the use of knowledge engineering guidelines.

Has Mary graduated?

Yes, Mary graduated from George Mason University in 2012 with an MSCS degree.

… … …

Figure 6.23. Question/answer pair corresponding to a reduction step.

13:52:56,
.007
6.12. Review Questions 199

6.9. Consider the reduction step from Figure 6.24. Specify the ontology fragments that
are suggested by this reasoning step. Do not limit yourself to the concepts and
features that are explicitly mentioned, but define additional ones as well, to enable
the agent to assess similar hypotheses in a similar way.

6.10. What instances, concepts, and relationships should be defined in the agent’s
ontology, based on the analysis of the reduction step from Figure 6.25?

6.11. What instances, concepts, and relationships should be defined in the agent’s
ontology, based on the analysis of the reduction step from Figure 6.26?

6.12. What instances, concepts, and relationships should be defined in the agent’s
ontology, based on the analysis of the reduction step from Figure 6.27?

6.13. Consider the design of an agent for assessing whether some actor (e.g.,
Aum Shinrikyo) is developing weapons of mass destruction (e.g., chemical
weapons). We would like our agent to be able to perform the sample reasoning
step from Figure 6.28, where the entities in blue are represented as specific
instances.

John Doe would be a good PhD advisor

with respect to the publication with advisor
criterion, considering the publications with
John Doe of Adam Pearce.

Is there any joint publication of

John Doe and Adam Pearce?
Yes, Pearce and Doe 2006.

It is certain that John Doe would be a

good PhD advisor with respect to the
publication with advisor criterion.

Figure 6.24. Reduction step for modeling-based ontology specification.

Jill Knox will stay on the faculty of

George Mason University for the
duration of dissertation of Bob Sharp.

Is Jill Knox likely to stay on the faculty of

George Mason University for the duration
of the dissertation of Bob Sharp?
Yes, because Jill Knox has a tenure-track
position and is likely to get tenure.

It is likely that Jill Knox will stay on the

faculty of George Mason University for the
duration of dissertation of Bob Sharp.

Figure 6.25. Reduction step for modeling-based ontology specification.

13:52:56,
.007
200 Chapter 6. Ontology Design and Development

Dan Smith will stay on the faculty of

George Mason University for the duration
of the dissertation of Bob Sharp.

Is Dan Smith likely to stay on the faculty of

George Mason University for the duration of
the dissertation of Bob Sharp?
No, because Dan Smith plans to retire from
George Mason University.

There is no support for the hypothesis

that Dan Smith will stay on the faculty of
George Mason University for the duration
of the dissertation of Bob Sharp.

Figure 6.26. Reduction step for modeling-based ontology specification.

Amanda Rice will stay on the faculty of

George Mason University for the duration
of the dissertation of Bob Sharp.

Is Amanda Rice likely to stay on the faculty of

George Mason University for the duration of
the dissertation of Bob Sharp?
No, because Amanda Rice has a visiting
position which is a short-term position.

There is no support for the hypothesis that

Amanda Rice will stay on the faculty of
George Mason University for the duration
of the dissertation of Bob Sharp.

Figure 6.27. Reduction step for modeling-based ontology specification.

Aum Shinrikyo has members

trained in chemistry.

Is there any member of Aum Shinrikyo

who is trained in chemistry?
Yes, Masami Tsuchiya, who has a
master’s degree in chemistry.

It is certain that Aum Shinrikyo has

members trained in chemistry.

Figure 6.28. Reduction step for modeling-based ontology specification.

(a) What relationships should you define in the agent’s ontology in order to
represent the meaning of the question/answer pair?
(b) Represent this meaning as a network fragment showing the relationships and
the related instances.

13:52:56,
.007
6.12. Review Questions 201

(c) Network fragments such as the one at (b) represent a specification of the
needed ontology, guiding you in defining a hierarchy of concepts to which
the identified instances belong, as well as the siblings of these concepts.
Indicate eight such concepts. Also indicate which might be the domain and
the range of each of the identified feature.

6.14. Develop an ontology that represents the following information:

Apple1 is an apple.
The color of Apple1 is red.
Apple2 is an apple.
The color of Apple2 is green.
Apples are fruits.
Hint: You should define concepts, features, and instances.

6.15. Develop an ontology that represents the following information:

Puss is a calico.
Herb is a tuna.
Charlie is a tuna.
All tunas are fishes.
All calicos are cats.
All cats like to eat all kinds of fish.
Cats and fishes are animals.
Hint: You should define concepts, features, and instances.

6.16. Develop an ontology that represents the following information:

Basketball players are tall.

Muresan is a basketball player.
Muresan is tall.
Hint: Define concepts, features, and instances.

6.17. Develop an ontology that represents the following information:

Birds are animals.

Birds have feathers, fly, and lay eggs.
Albatross is a bird.
Donald is a bird.
Tracy is an albatross.
Hint: Define concepts, features, and instances.

6.18. Explain why maintaining the consistency of the ontology is a complex knowledge
engineering activity.

6.19. One of the principles in the development of a knowledge base with a tool such as
Disciple-EBR is to maintain its consistency because correcting an inconsistent
knowledge base is a very complex problem. Therefore, the tool will not allow the
deletion of a knowledge base element (e.g., an instance, a fact, a concept, or a
feature definition) if that operation will make the knowledge base inconsistent. List
and explain five possible ways in which the deletion of a concept may render the
knowledge base inconsistent.

13:52:56,
.007
 7 Reasoning with Ontologies and Rules

7.1 PRODUCTION SYSTEM ARCHITECTURE

In Chapter 4, we presented the problem reduction and solution synthesis paradigm. In this
chapter, we will present how a knowledge-based agent can employ this paradigm to solve
problems and assess hypotheses.
Figure 7.1 shows the architecture of the agent, which is similar to that of a production
system (Waterman and Hayes-Roth, 1978). The knowledge base is the long-term memory,
which contains an ontology of concepts and a set of rules expressed with these concepts.
When the user formulates an input problem, the problem reduction and solution synthesis
inference engine applies the learned rules from the knowledge base to develop a problem
reduction and solution synthesis tree, as was discussed in Section 4. This tree is developed
in the reasoning area, which plays the role of the short-term memory.
The ontology from the knowledge base describes the types of objects (or concepts) in
the application domain, as well as the relationships between them. Also included are the
instances of these concepts, together with their properties and relationships.
The rules are IF-THEN structures that indicate the conditions under which a general
problem (or hypothesis) can be reduced to simpler problems (hypotheses), or the solu-
tions of the simpler problems (hypotheses) can be combined into the solution of the more
complex problem (hypothesis).
The applicability conditions of these rules are complex concepts that are expressed by
using the basic concepts and relationships from the ontology, as will be discussed in
Section 7.2. The reduction and synthesis rules will be presented in Section 7.3. This section

Problem Problem input

Reduction and
Solution Synthesis Formulation

Knowledge Base P1 S1 Reasoning

employee
person
subconcept of

student
IF the problem to solve is P1g
IF the problem to solve is P1g
IF the problem to solve is P1g
Condition
P11 S11 … P S
1n
Area
1n
subconcept of
IF the problem to solve is P1g
Condition
university employee Except-When
… ConditionIF the problem to solve is P1g
Condition
IF the problem to solve is P1g
subconcept of
Except-When Condition
staff member
subconcept of
faculty member graduate
undergraduate
student
…
Except-When
Condition
Except-When
… Condition
Except-When Except-When
Condition
Condition
Condition
Condition P1n1 S1n1 … P1nm S1nm
student … Condition
subconcept of Except-When
… Condition Condition
THEN solve itsExcept-When
sub-problems
subconcept of
Except-When
… Condition Condition
instructor professor PhD advisor subconcept of
BS student THEN1 …
P1g solve
1 itsExcept-When
Png sub-problems
subconcept of THEN1 …
P1g solve1 itsExcept-When
Png sub-problems Condition
assistant associate full
THEN
P1g1 … solve1 itsExcept-When
Png sub-problems Condition
solve1 its sub-problems
professor

instance of
professor

instance of
professor

instance of
MS student

instance of
PhD student

instance of
THEN
P1g1 …
THEN
P1g
Png
1 … solve1 its sub-problems
Png
1 … P1
P1n11 S1n11 … P1n1q S1n1q
P1g ng
John Smith John Doe Jane Austin Joan Dean Bob Sharp

Ontology output
Solution
Development and
Rule Learning Browsing

Figure 7.1. Production system architecture of the knowledge-based agent.

202

13:52:57,
.008
 7.2. Complex Ontology-based Concepts 203

will also present the overall reduction and synthesis algorithm. Section 7.4 will present the
simplified reduction and synthesis rules used for evidence-based hypotheses analysis, and
Section 7.5 will present the rule and ontology matching process. Finally, Section 7.6 will
present the representation of partially learned knowledge, and Section 7.7 will present the
reasoning with this type of knowledge.

7.2 COMPLEX ONTOLOGY-BASED CONCEPTS

Using the concepts and the features from the ontology, one can define more complex
concepts as logical expressions involving these basic concepts and features. For example,
the concept “PhD student interested in an area of expertise” may be expressed as
shown in [7.1].

?O1 instance of PhD student [7.1]

is interested in ?O2
?O2 instance of area of expertise

Because a concept represents a set of instances, the user can interpret the preceding
concept as representing the set of instances of the tuple (?O1, ?O2), which satisfy the
expression [7.1], that is, the set of tuples where the first element is a student and the
second one is the research area in which the student is interested. For example, Bob
Sharp, a PhD student interested in artificial intelligence, is an instance of the concept [7.1].
Indeed, the following expression is true:

Bob Sharp instance of PhD student

is interested in Artificial Intelligence
Artificial Intelligence instance of area of expertise

In general, the basic representation unit (BRU) for a more complex concept has the
form of a tuple (?O1, ?O2, . . . , ?On), where each ?Oi has the structure indicated by [7.2],
called a clause.

?Oi instance of concepti [7.2]

featurei1 ?Oi1
...
featureiq ?Oim

Concepti is either an object concept from the object ontology (such as PhD student), a
numeric interval (such as [50 , 60]), a set of numbers (such as {1, 3, 5}), a set of strings
(such as {white, red, blue}), or an ordered set of intervals (such as (youth, mature)). ?Oi1 . . .
?Oim are distinct variables from the sequence (?O1, ?O2, . . . , ?On).
A concept may be a conjunctive expression of form [7.3], meaning that any instance of
the concept satisfies BRU and does not satisfy BRU1 and . . . and does not satisfy BRUp.

BRU ∧ not BRU1 ∧ . . . ∧ not BRUp [7.3]

However, instead of “not,” we write “Except-When,” as shown in [7.4].

BRU ∧ Except-When BRU1 ∧ . . . ∧ Except-When BRUp [7.4]

13:52:57,
.008
 204 Chapter 7. Reasoning with Ontologies and Rules

For example, expression [7.5] represents the concept “PhD student interested in an area of
expertise that does not require programming.”

?O1 instance of PhD student [7.5]

is interested in ?O2
?O2 instance of area of expertise
Except-When
?O2 instance of area of expertise
requires programming

7.3 REDUCTION AND SYNTHESIS RULES AND

THE INFERENCE ENGINE

An agent can solve problems through reduction and synthesis by using problem reduction
rules and solution synthesis rules. The rules are IF-THEN structures that indicate the
conditions under which a general problem can be reduced to simpler problems, or the
solutions of the simpler problems can be combined into the solution of the more complex
problem.
The general structure of a problem reduction rule Ri is shown in Figure 7.2. This rule
indicates that solving the problem P can be reduced to solving the simpler problems
Pi1, . . ., Pini, if certain conditions are satisfied. These conditions are expressed in two
equivalent forms: one as a question/answer pair in natural language that is easily under-
stood by the user of the agent, and the other as a formal applicability condition expressed
as a complex concept having the form [7.4], as discussed in the previous section.
Consequently, there are two interpretations of the rule in Figure 7.2:

(1) If the problem to solve is P, and the answer to the question QRi is ARi, then one
can solve P by solving the subproblems Pi1, . . . , Pini.
(2) If the problem to solve is P, and the condition CRi is satisfied, and the conditions
ERi1, . . ., ERiki are not satisfied, then one can solve P by solving the subproblems
Pi1, . . . , Pini.

IF the problem to solve is P

Question QRi
Answer ARi

Main Condition CRi

Except-When Condition ERi1
…
Except-When Condition ERiki

THEN solve
Subproblem Pi1
…
Subproblem Pini

Figure 7.2. Structure of a problem reduction rule Ri.

13:52:57,
.008
7.3. Reduction and Synthesis Rules 205

An example of a problem reduction rule is shown in Figure 7.3. It reduces the IF problem
to two simpler problems. This rule has a main condition and no Except-When conditions.
Notice also that the main condition is expressed as the concept (?O1, ?O2, ?O3, ?O4).
As discussed in Section 4.2 and illustrated in Figure 4.6 (p. 116), there are two synthesis
operations associated with a reduction operation: a reduction-level synthesis and a
problem-level synthesis. These operations are performed by employing two solution
synthesis rules that are tightly coupled with the problem reduction rule, as illustrated in
Figure 7.4 and explained in the following paragraphs.
For each problem reduction rule Ri that reduces a problem P to the subproblems
P 1, . . . , Pini (see the left-hand side of Figure 7.4), there is a reduction-level solution
i

synthesis rule (see the upper-right-hand side of Figure 7.4). This reduction-level solution
synthesis rule is an IF-THEN structure that expresses the condition under which the
solutions Si1, . . . , Sini of the subproblems Pi1, . . . , Pini of the problem P can be combined
into the solution Si of P corresponding to the rule Ri.
Let us now consider all the reduction rules Ri, . . . , Rm that reduce problem P to simpler
problems, and the corresponding synthesis rules SRi, . . . , SRm. In a given situation, some
of these rules will produce solutions of P, such as the following ones: Based on Ri, the

IF the problem to solve is

Analyze a strategic COG candidate corresponding to
the ?O1 which is an industrial economy.

Q: Who or what is a strategically critical element

with respect to the ?O1 ?

A: ?O2 because it is an essential generator of war

material for ?O3 from the strategic perspective.

Main Condition
?O1 is industrial economy
?O2 is industrial capacity
generates essential war material from
the strategic perspective of ?O3
?O3 is multistate force
has as member ?O4
?O4 is force
has as economy ?O1
has as industrial factor ?O2

THEN solve the subproblems

Identify ?O2 as a strategic COG candidate with
respect to the ?O1.
Test ?O2 which is a strategic COG candidate with
respect to the ?O1.

Figure 7.3. An example of a problem reduction rule.

13:52:57,
.008
 206 Chapter 7. Reasoning with Ontologies and Rules

Reduction rule Ri Synthesis rule SRi corresponding to Ri

IF the problem to solve is P IF

The solution of Pi1 is Si1
…
Question QRi The solution of Pini is Sini
Answer ARi
Question QSRi
Condition CRi
Answer ASRi
THEN solve Condition CSRi
Subproblem Pi1
… THEN
Subproblem Pini Based on Ri the solution of P is Si

Synthesis rule SP corresponding to P

IF
Based on Ri the solution of P is Si
…
Based on Rm the solution of P is Sm
Question QP
Answer AP
Condition CP
THEN
The solution of P is S

Figure 7.4. A reduction rule and the corresponding synthesis rules.

solution of P is Si, . . . , and based on Rm, the solution of P is Sm. The synthesis rule SP
corresponding to the problem P combines all these rule-specific solutions of the problem
P (named Si, . . . , Sm) into the solution S of P (see the bottom-right side of Figure 7.4).
The problem-solving algorithm that builds the reduction and synthesis tree is pre-
sented in Table 7.1.

7.4 REDUCTION AND SYNTHESIS RULES FOR

EVIDENCE-BASED HYPOTHESES ANALYSIS

In Section 4.3, we discussed the specialization of inquiry-driven analysis and synthesis for
evidence-based reasoning, where one assesses the probability of hypotheses based on
evidence, such as the following one:

John Doe would be a good PhD advisor for Bob Sharp.

Moreover, the assessment of any such hypothesis has the form, “It is <probability> that
H,” which can be abstracted to the actual probability, as in the following example, which
can be abstracted to “very likely:”

It is very likely that John Doe would be a good PhD advisor for Bob Sharp.

13:52:57,
.008
 7.5. Rule and Ontology Matching 207

Table 7.1 The Basic Operation of the Inference Engine

Procedure SOLVE( PN(n0, P, x0) )

Input
P – a problem class;
x0 – parameters instantiation
P(x0) – a specific problem
n0 – the index of the problem node in the reasoning tree
PN(n0, P, x0) – a problem node for problem P(x0) with index n0
SOL={ }
FOR ALL Ri 2 GET-APPLICABLE-RULES-OF( P )
FOR ALL yj 2 GENERATE-RULE-INSTATIATIONS (Ri, x0 )
RN(nij, Ri, yj) = GENERATE-REDUCTION-NODE ( PN(n0,P,x0), Ri (yj) )
FOR ALL TNnk(nk, Tk, zk) 2 GET-THEN-STATEMENT-NODE (RN(nij, Ri, yj))
IF TNnk IS PROBLEM-NODE THEN
SNnk SOLVE(TNnk)
IF ALL TNnk HAVE SOLUTIONS THEN
SN(nij, S, uk) COMPUTE-FIRST-REDUCTION-LEVEL-SOLUTION-NODE
(RN(nij, Ri, yj), (SNnk)k)
IF SN(nij, S, uk) THEN
SOL SOL [ {S(uk)}
IF SOL 6¼ ∅ THEN
SN(n0, S, u) COMPUTE-FIRST-PROBLEM-LEVEL-SOLUTION-NODE (PN(n0, P, x0), SOL)
RETURN SNn0

As a result, the solution synthesis rules have a simplified form, as shown in Figure 7.5. In
particular, notice that the condition of a synthesis rule is reduced to computing the
probability of the THEN solution based on the probabilities for the IF solutions.
In the current implementation of Disciple-EBR, the function for a reduction-level
synthesis rule SR could be one of the following (as discussed in Section 4.3): min, max,
likely indicator, very likely indicator, almost certain indicator, or on balance. The
function for a problem-level synthesis can be only min or max.
Figure 7.6 shows an example of a reduction rule and the corresponding synthesis rules.
The next section will discuss how such rules are actually applied in problem solving.

7.5 RULE AND ONTOLOGY MATCHING

Let us consider the following hypothesis to assess:

Bob Sharp is interested in an area of expertise of John Doe.

The agent (i.e., its inference engine) will look for all the reduction rules from the know-
ledge base with an IF hypothesis that matches the preceding hypothesis. Such a rule is the
one from the right-hand side of Figure 7.7. As one can see, the IF hypothesis becomes
identical with the hypothesis to be solved if ?O1 is replaced with Bob Sharp and ?O2 is
replaced with John Doe. The rule is applicable if the condition of the rule is satisfied for
these values of ?O1 and ?O2.

13:52:57,
.008
208 Chapter 7. Reasoning with Ontologies and Rules

Reduction rule Ri Synthesis rule SRi corresponding to Ri

IF the hypothesis to assess is H

IF
Question QRi It is <probability pi1> that Hi1
Answer ARi …
It is < probability pini > that Hini
Condition CRi
pi = function (pi1, … , pini )
THEN assess
Subhypothesis Hi1
… THEN
Subhypothesis Hini Based on Ri it is < probability pi> that H

Synthesis rule SH corresponding to H

IF
Based on Ri it is < probability pi> that H
…
Based on Rm it is < probability pm> that H

p = function (pi, …, pm)

THEN
It is < probability p> that H

Figure 7.5. Reduction and synthesis rules for evidence-based hypotheses analysis.

The partially instantiated rule is shown in the right-hand side of Figure 7.8. The agent
has to check that the partially instantiated condition of the rule can be satisfied. This
condition is satisfied if there is any instance of ?O3 in the object ontology that satisfies all
the relationships specified in the rule’s condition, which is shown also in the left-hand side
of Figure 7.8. ?Sl1 is an output variable that is given the value certain, without being
constrained by the other variables.
The partially instantiated condition of the rule, shown in the left-hand side of Fig-
ures 7.8 and 7.9, is matched successfully with the ontology fragment shown in the right-
hand side of Figure 7.9. The questions are: How is this matching performed, and is it
efficient?
John Doe from the rule’s condition (see the left-hand side of Figure 7.9) is matched with
John Doe from the ontology (see the right-hand side of Figure 7.9).
Following the feature is expert in, ?O3 has to match Artificial Intelligence:

John Doe is expert in ?O3

John Doe is expert in Artificial Intelligence

This matching is successful because both ?O3 and Artificial Intelligence are areas of
expertise, and both are the values of the feature is interested in of Bob Sharp. Indeed,
Artificial Intelligence is an instance of Computer Science, which is a subconcept of area
of expertise.

13:52:57,
.008
7.5. Rule and Ontology Matching 209

Reduction Rule Ri Synthesis Rule SRi corresponding to Ri

IF the hypothesis to assess is IF

?O1 would be a good PhD advisor for ?O2. It is <probability pi1> that ?O2 is interested in an area
of expertise of ?O1.
Q: Which are the necessary conditions? It is <probability pi2> that ?O1 will stay on the faculty of
?O3 for the duration of dissertation of ?O2.
A: ?O2 should be interested in the area
It is <probability pi3> that ?O1 would be a good PhD
of expertise of ?O1 who should stay
advisor with respect to the ?O4.
on the faculty of ?O3 for the
duration of the dissertation of ?O2,
and should have the qualities of a pi = min(pi1, pi2, pi3)
good PhD advisor. THEN
Based on Ri it is <probability pi> that ?O1 would be a
Main Condition good PhD advisor for ?O2.
?O1 is PhD advisor
has as employer ?O3 Synthesis Rule SH
?O2 is PhD student
IF
?O3 is university
Based on Ri it is <probability pi> that ?O1 would be a
?O4 is PhD advisor quality criterion
good PhD advisor for ?O2.
…
THEN assess the subhypotheses
Based on R m it is <probability pm> that ?O1 would be a
?O2 is interested in an area of expertise of good PhD advisor for ?O2.
?O1.
?O1 will stay on the faculty of ?O3 for the l = max (pi,…, pm)
duration of dissertation of ?O2. THEN
?O1 would be a good PhD advisor with It is <probability l> that ?O1 would be a good
respect to the ?O4. PhD advisor for ?O2.

Figure 7.6. Examples of reduction and synthesis rules for evidence-based hypotheses analysis.

Hypothesis
Bob Sharp is interested in an area of
expertise of John Doe.

?O1  Bob Sharp

?O2  John Doe
Rule
IF the hypothesis to assess is
Rule’s condition
?O1 is interested in an area of
expertise of ?O2.
PhD student
Q: Is ?O1 interested in an area of
instance of expertise of ?O2 ?
A: Yes,?O3.
?O1
area of expertise
Condition
is interested in ?O1 is PhD student
instance of
is interested in?O3
?O3 ?O2 is PhD advisor
PhD advisor
is expert in ?O3
?O3 is area of expertise
instance of is expert in
?SI1 is-in [certain - certain]

?Sl1 = certain THEN conclude

?O2
It is ?SI1 that ?O1 is interested in an
area of expertise of ?O2.

Figure 7.7. Rule selection based on hypotheses matching.

13:52:57,
.008
210 Chapter 7. Reasoning with Ontologies and Rules

Hypothesis
Bob Sharp is interested in an area of
expertise of John Doe.

?O1  Bob Sharp

?O2  John Doe
Rule
IF the hypothesis to assess is
Rule’s condition
Bob Sharp is interested in an area of
expertise of John Doe.
PhD student
Q: Is Bob Sharp interested in an area of
instance of expertise of John Doe?
A: Yes,?O3.
Bob Sharp
area of expertise
Condition
is interested in Bob Sharp is PhD student
instance of
is interested in ?O3
?O3 John Doe is PhD advisor
PhD advisor
is expert in ?O3
?O3 is area of expertise
instance of is expert in ?SI1 is-in [certain - certain]

THEN conclude
John Doe ?Sl1 = certain
It is ?SI1 that Bob Sharp is interested in
an area of expertise of John Doe.

Figure 7.8. Partial instantiation of the rule based on hypotheses matching.

Rule’s condition Agent’s ontology

PhD student PhD student

area of expertise
instance of instance of subconcept of

Bob Sharp area of expertise Bob Sharp Computer Science

is interested in instance of instance of

is interested in

PhD advisor ?O3 PhD advisor Artificial Intelligence

instance of is expert in instance of is expert in

Probability =
John Doe ?Sl1 = certain John Doe { …, almost certain, certain }

?O1  Bob Sharp ?O2  John Doe ?O3  Artificial Intelligence ?Sl1  certain

Figure 7.9. Ontology matching.

As the result of this matching, the rule’s ?O3 variable is instantiated to Artificial
Intelligence:
?O3 ⟵ Artificial Intelligence

Also, ?Sl1 will take the value certain, which is one of the values of probability, as
constrained by the rule’s condition.

13:52:57,
.008
7.5. Rule and Ontology Matching 211

The matching is very efficient because the structure used to represent knowledge (i.e.,
the ontology) is also a guide for the matching process, as was illustrated previously and
discussed in Section 5.10.
Thus the rule’s condition is satisfied for the following instantiations of the variables:

?O1 ⟵ Bob Sharp

?O2 ⟵ John Doe
?O3 ⟵ Artificial Intelligence
?SI1 ⟵ certain

Therefore, the rule can be applied to reduce the IF hypothesis to an assessment. This
entire process is summarized in Figure 7.10, as follows:

(1) The hypothesis to assess is matched with the IF hypothesis of the rule, leading to
the instantiations of ?O1 and ?O2.
(2) The corresponding instantiation of the rule’s condition is matched with the
ontology, leading to instances for all the variables of the rule.
(3) The question/answer pair and the THEN part of the rule are instantiated, gener-
ating the following reasoning step shown also in the left-hand side of Figure 7.10:

Hypothesis to assess:

Bob Sharp is interested in an area of expertise of John Doe.

Q: Is Bob Sharp interested in an area of expertise of John Doe?

A: Yes, Artificial Intelligence.

Assessment:

It is certain that Bob Sharp is interested in an area of expertise of John Doe.

1 Match hypothesis with IF part

Rule
Hypothesis to assess ?O1  Bob Sharp
?O2  John Doe IF the hypothesis to assess is
Bob Sharp is interested in an area of
expertise of John Doe. ?O1 is interested in an area of
expertise of ?O2.

Q: Is Bob Sharp interested in an

area of expertise of John Doe?
3 Instantiate Q/A Q: Is ?O1 interested in an area of
expertise of ?O2 ?
A: Yes, Artificial Intelligence. A: Yes,?O3.

Condition
2 Match rule’s condition with the ontology
?O1 is PhD student
?O1  Bob Sharp is interested in?O3
?O2  John Doe ?O2 is PhD advisor
?O3  Artificial Intelligence is expert in ?O3
?SI1  certain ?O3 is area of expertise
?SI1 is-in [certain - certain]
Assessment Instantiate
It is certain that Bob Sharp is interested 3 assessment
THEN conclude

in an area of expertise of John Doe. It is ?SI1 that ?O1 is interested in an

area of expertise of ?O2.

Figure 7.10. Application of a rule in problem solving.

13:52:57,
.008
 212 Chapter 7. Reasoning with Ontologies and Rules

In the preceding example, the agent has found one instance for each rule variable, which
has led to a solution. What happens if it cannot find instances for all the rule’s variables?
In such a case, the rule is not applicable.
But what happens if it finds more than one set of instances? In that case, the agent will
generate an assessment for each distinct set of instances.
What happens if there is more than one applicable reduction rule? In such a case, the
agent will apply each of them to find all the possible reductions.
All the obtained assessments will be combined into the final assessment of the hypoth-
esis, as was discussed in the previous section.
Figure 7.11 shows the successive applications of two reduction rules to assess an initial
hypothesis. Rule 1 reduces it to three subhypotheses. Then Rule 2 finds the assessment of
the first subhypothesis.

7.6 PARTIALLY LEARNED KNOWLEDGE

7.6.1 Partially Learned Concepts

An incremental learning system will maintain representations of the partially learned
concepts. Disciple-EBR is such an incremental learner that learns general concepts from
examples and explanations (Tecuci, 1998; Tecuci et al., 2005b), as will be discussed in
the following chapters. During the learning process, the agent maintains a set of
possible versions of the concept to be learned, called a version space (Mitchell, 1978,
1997; Tecuci, 1988). The concepts in this space are partially ordered, based on the
generalization relationship. For that reason, the version space can be represented by an
upper bound and a lower bound. The upper bound of the version space contains the most

John Doe would be a good PhD advisor for Bob Sharp.

Rule 1
Which are the necessary conditions?
Bob Sharp should be interested in an area of expertise of John Doe, who should
stay on the faculty of George Mason University for the duration of the dissertation
of Bob Sharp and should have the qualities of a good PhD advisor.

Bob Sharp is John Doe will stay on the John Doe would be a
interested in an faculty of George Mason good PhD advisor with
area of expertise University for the duration of respect to the PhD
of John Doe. the dissertation of Bob Sharp. advisor quality criterion.
Rule 2

Q: Is Bob Sharp interested in an

area of expertise of John Doe?
A: Yes, Artificial Intelligence.

It is certain that Bob Sharp

is interested in an area of
expertise of John Doe.

Figure 7.11. Successive rules applications.

13:52:57,
.008
 7.6. Partially Learned Knowledge 213

general concepts from the version space, and the lower bound contains the least general
concepts from the version space. Any concept that is more general than (or as general as) a
concept from the lower bound and less general than (or as general as) a concept from the
upper bound is part of the version space and may be the actual concept to be learned.
Therefore, a version space may be regarded as a partially learned concept.
The version spaces built by Disciple-EBR during the learning process are called plaus-
ible version spaces because their upper and lower bounds are generalizations based on an
incomplete ontology. Therefore, a plausible version space is only a plausible approxima-
tion of the concept to be learned, as illustrated in Figure 7.12.
The plausible upper bound of the version space from the right-hand side of Figure 7.12
contains two concepts: “a faculty member interested in an area of expertise” (see expres-
sion [7.6]) and “a student interested in an area of expertise” (see expression [7.7]).

Concept 1: ?O1 instance of faculty member [7.6]

is interested in ?O2
?O2 instance of area of expertise
Concept 2: ?O1 instance of student [7.7]
is interested in ?O2
?O2 instance of area of expertise

The plausible lower bound of this version space also contains two concepts, “an
associate professor interested in Computer Science,” and “a graduate student interested
in Computer Science.”
The concept to be learned (see the left side of Figure 7.12) is, as an approximation, less
general than one of the concepts from the plausible upper bound, and more general than
one the concepts from the plausible lower bound.
The notion of plausible version space is fundamental to the knowledge representation,
problem-solving, and learning methods of Disciple-EBR because all the partially learned
concepts are represented using this construct, as discussed in the following.

7.6.2 Partially Learned Features

As discussed in Section 5.5, a feature is characterized by a domain and a range. Most often, the
domains and the ranges of the features are basic concepts from the object ontology. However,

Universe of instances Plausible

Upper Bound Plausible Upper Bound
Concept to ?O1 instance of {faculty member,
be learned student}
is interested in ?O2
Plausible ?O2 instance of area of expertise
Lower Bound
Plausible Lower Bound
?O1 instance of {associate professor,
graduate student}
is interested in ?O2
?O2 instance of Computer Science

Figure 7.12. The plausible version space of a concept to be learned.

13:52:57,
.008
 214 Chapter 7. Reasoning with Ontologies and Rules

they could also be complex concepts of the form shown in Section 7.2. Moreover, in the case of
partially learned features, they are plausible version spaces, as illustrated in Figure 5.7 (p. 159).

7.6.3 Partially Learned Hypotheses

A specific hypothesis consists of a natural language sentence that includes at least one
instance and, optionally, additional instances, concepts, numbers, and strings. An example
of a specific hypothesis is the following one:

John Doe is a potential PhD advisor for Bob Sharp. [7.8]

As will be discussed in Section 9.10, the agent learns general hypotheses with applic-
ability conditions from specific hypotheses. An example of a general hypothesis learned
from [7.8] in shown in [7.9].

Name [7.9]
?O1 is a potential PhD advisor for ?O2.
Condition
?O1 instance of faculty member
?O2 instance of person

The condition is a concept that, in general, may have the form [7.4] (p. 203). The purpose
of the condition is to ensure that the hypothesis makes sense for each hypothesis instanti-
ation that satisfies it. For example, the hypothesis from [7.8] satisfies the condition in [7.9]
because John Doe is a faculty member and Bob Sharp is a person. However, the hypothesis
instance in [7.10] does not satisfy the condition in [7.9] because 45 is not a faculty member.
The condition in [7.9] will prevent the agent from generating the instance in [7.10].

45 is a potential PhD advisor for Bob Sharp. [7.10]

A partially learned hypothesis will have a plausible version space condition, as illus-
trated in [7.11].

Name [7.11]
?O1 is a potential PhD advisor for ?O2.
Plausible Upper Bound Condition
?O1 instance of person
?O2 instance of person
Plausible Lower Bound Condition
?O1 instance of {associate professor, PhD advisor}
?O2 instance of PhD student

7.6.4 Partially Learned Rules

As discussed in Section 7.3, the applicability condition of a rule is the concept representing
the set of instances for which the rule is correct. In the case of a partially learned rule, the
applicability condition is a plausible version space. An example of such a rule is presented
in Figure 7.13. Its lower bound condition requires that ?O1 be a PhD student interested in
artificial intelligence, while the upper bound allows ?O1 to be any person interested in any
area of expertise. Similarly, the lower bound condition requires ?O2 to be either a PhD

13:52:57,
.008
 7.7. Reasoning with Partially Learned Knowledge 215

Figure 7.13. Partially learned hypothesis reduction rule.

advisor or an associate professor who is an expert in artificial intelligence, while the upper
bound allows ?O2 to be any person who is an expert in any area of expertise ?O3 in which
?O1 is interested.

7.7 REASONING WITH PARTIALLY LEARNED KNOWLEDGE

A learning agent should be able to reason with partially learned knowledge. Figure 7.14
shows an abstract rule with a partially learned applicability condition. It includes a Main
plausible version space condition (in light and dark green) and an Except-When plausible
version space condition (in light and dark red). The reductions generated by a partially
learned rule will have different degrees of plausibility, as indicated in Figure 7.14.
For example, a reduction r1 corresponding to a situation where the plausible lower bound
condition is satisfied and none of the Except-When conditions is satisfied is most likely to
be correct. Similarly, r2 (which is covered by the plausible upper bound, and is not
covered by any bound of the Except-When condition) is plausible but less likely than r1.
The way a partially learned rule is used depends on the current goal of the agent. If the
current goal is to support its user in problem solving, then the agent will generate the
solutions that are more likely to be correct. For example, a reduction covered by the
plausible lower bound of the Main condition and not covered by any of the Except-When
conditions (such as the reduction r1 in Figure 7.14) will be preferable to a reduction
covered by the plausible upper bound of the Main condition and not covered by any of the
Except-When conditions (such as the reduction r2), because it is more likely to be correct.
However, if the current goal of the agent is to improve its reasoning rules, then it is
more useful to generate the reduction r2 than the reduction r1. Indeed, no matter how the
user characterizes r2 (either as correct or as incorrect), the agent will be able to use it to

13:52:57,
.008
 216 Chapter 7. Reasoning with Ontologies and Rules

Main Plausible Except-When Plausible

IF <Problem> Version Space Condition Version Space Condition
Main
Plausible Version
Space Condition
Except-When
Plausible Version
Space Condition r1
THEN <Subproblem 1>
<Subproblem 2>
…
<Subproblem m> r2

Reduction is Reduction Reduction is Reduction is most

most likely correct is plausible not plausible likely incorrect
Figure 7.14. Plausible reasoning based on a partially learned rule.

refine the rule, either by generalizing the plausible lower bound of the Main condition to
cover r2 (if r2 is a correct reduction), or by specializing the plausible upper bound of the
Main condition to uncover r2 (if r2 is an incorrect reduction), or by learning an additional
Except-When condition based on r2 (again, if r2 is an incorrect reduction).

7.8 REVIEW QUESTIONS

7.1. What does the following concept represent? What would be an instance of it?

?O1 instance of course

has as reading ?O2
?O2 instance of publication
has as author ?O3
?O3 instance of professor

7.2. Illustrate the problem-solving process with the hypothesis, the rule, and the ontol-
ogy from Figure 7.15.
7.3. Consider the reduction rule and the ontology fragment from Figure 7.16. Indicate
whether this agent can assess the hypothesis, “Bill Bones is expert in an area of
interest of Dan Moore,” and if the answer is yes, indicate the result.
7.4. Consider the following problem:

Analyze a strategic COG candidate corresponding to the economy of US 1943, which

is an industrial economy.
Explain how this problem is reduced by applying the reduction rule from
Figure 7.17, in the context of the ontology from Figure 7.18. Show the generated
reduction step.
7.5. Illustrate the problem-solving process with the hypothesis, rule, and ontology
fragment from Figure 7.19.

13:52:57,
.008
7.8. Review Questions 217

object Jane Austin will stay on the faculty of George Mason

University for the duration of dissertation of Bob Sharp.
actor
IF the hypothesis to assess is
?O1 will stay on the faculty of ?O2 for the
person organization duration of dissertation of ?O3.
Q: Will ?O1 stay on the faculty of ?O2 for the
employee student educational duration of the dissertation of O3?
organization
A: Yes, because ?O1 has a ?O4, which is a
graduate long-term position, and ?O1 does not plan
university employee
student subconcept of to retire from ?O2.
Condition
subconcept of
university ?O1 is PhD advisor
faculty member
has as position ?O4
PhD student
has as employer ?O2
subconcept of instance of instance of
?O2 is university
Bob Sharp ?O3 is PhD student
professor George Mason
?O4 is long-term position
University
PhD advisor ?SI1 is-in [almost certain – almost certain]
subconcept of
long-term Except When Condition
full ?O1 is PhD advisor
position
professor has as employer plans to retire from ?O2
instance of
instance of THEN conclude
has as position tenure
Jane Austin It is ?SI1 that ?O1 will stay on the faculty of
position ?O2 for the duration of dissertation of ?O3.

Figure 7.15. Ontology fragment, hypothesis, and reduction rule.

IF the hypothesis to assess is

?O1 is expert in an area of interest
of ?O2.

Q: Is ?O1 expert in an area of

interest of ?O2 ? area of expertise
A: Yes, because ?O2 is interested
in ?O3, which is an area of subconcept-of
expertise of ?O1.
Computer Science
Condition
?O1 is PhD advisor instance-of
is expert in ?O3
?O2 is PhD student Art if icial Sof tware Computer
is interested in ?O3 Intelligence Engineering Networks

?O3 is area of expertise

PhD student PhD advisor
?SI1 is-in [certain - certain]

THEN conclude instance-of instance-of

It is ?SI1 that ?O1 is expert in ?O3, is interested in is expert in
which is an area of interest of ?O2. Dan Moore Bill Bones

Figure 7.16. Reduction rule and ontology fragment.

7.6. Consider the partially learned concept and the nine instances from Figure 7.20.
Order the instances by the plausibility of being instances of this concept and justify
the ordering.

13:52:57,
.008
218 Chapter 7. Reasoning with Ontologies and Rules

IF the problem to solve is

Analyze a strategic COG candidate corresponding
to the ?O1, which is an industrial economy.

Question
Who or what is a strategically critical element
with respect to the ?O1 ?
Answer
?O2 because it is an essential generator of war
material for ?O3 from the strategic perspective.

Condition
?O1 is industrial economy
?O2 is industrial capacity
generates essential war material from
the strategic perspect ive of ?O3
?O3 is multistate force
has as member ?O4
?O4 is force
has as economy ?O1
has as industrial factor ?O2

THEN solve the subproblems

Ident ify ?O2 as a strategic COG candidate with
respect to the ?O1.
Test ?O2, which is a strategic COG candidate with
respect to the ?O1.

Figure 7.17. Reduction rule from the center of gravity analysis domain.

7.7. Consider the ontology fragment from the loudspeaker manufacturing domain
shown in Figure 5.24 (p. 172). Notice that each most specific concept, such as dust
or air press, has an instance, such as dust1 or air press1.
Consider also the following rule:

IF the task to perform is

Clean ?x of ?y
Condition
?x is object
may have ?y
?y is object
?z is cleaner
removes ?y
THEN perform the task
Clean ?x of ?y with ?z

13:52:57,
.008
7.8. Review Questions 219

object

force economy capacity

opposing force mult imember force single-member force group

industrial economy

multigroup multistate single-state force

force force
economy of US 1943 economy of UK 1943

multistate multistate
alliance coalition

US 1943
dominant partner equal partners industrial capacity
multistate alliance multistate alliance UK 1943

European Axis 1943 Allied Forces 1943

generates essential
war material from the
strategic perspect ive of industrial capacity of UK 1943

Finland1943 Germany 1943 industrial capacity of US 1943

has as member has as industrial factor has as economy generates essential war material
domain multimember force domain force domain force from the strategic perspect ive of
range force range industrial capacity range economy domain capacity
range force

Figure 7.18. Feature definitions and ontology fragment from the center of gravity analysis domain.
Dotted links indicate “instance of” relationships while unnamed continuous links indicate “sub-
concept of” relationships.

Describe how this rule is applied to solve the problem:

Clean entrefer1 of dust1

What will be the result?

Describe how this rule is applied to solve the problem:

Clean membrane1 of surplus-adhesive1

What will be the result?

13:52:57,
.008
 object A PhD advisor will stay on the faculty of George Mason
University for the duration of the dissertation of Joe Dill.
actor position
IF the hypothesis to assess is
person organization A PhD advisor will stay on the faculty of ?O2 for
the duration of the dissertation of ?O3.
employee student Q: Will a PhD advisor stay on the faculty of ?O2
educational organization for the duration of the dissertation of O3?

university employee subconcept of A: Yes,?O1 who has a ?O4, a long-term position

graduate student
subconcept of and does not plan to retire from ?O2.

subconcept of Condition
university long-term position ?O1 is PhD advisor
faculty member
PhD student has as position ?O4
has as employer ?O2
.008

subconcept of instance of instance of ?O2 is university

Joe Dill ?O3 is PhD student
professor instance of
13:52:57,

George Mason University ?O4 is long-term position

subconcept of PhD advisor ?SI1 is-in [almost certain – almost certain]
instance of has as employer Except-When Condition
full professor tenure position ?O1 is PhD advisor
plans plans to retire from ?O2
to retire has as position
instance of from THEN conclude
It is ?SI1 that ?O1 will stay on the faculty of ?O2
Jim Barber Ann Smith for the duration of the dissertation of ?O3.

Figure 7.19. Ontology, hypothesis, and potential rule for assessing it.
7.8. Review Questions 221

Main Plausible Version Except-When Plausible

Space Condition Version Space Condition

I6
I9
I2
I3 I7 I8
I1

I4 I5

Figure 7.20. A partially learned concept and several instances.

13:52:57,
.008
 8 Learning for Knowledge-based Agents

The previous chapters introduced the main knowledge elements from the knowledge base
of an agent, which are all based on the notion of concept. This chapter presents the basic
operations involved in learning, including comparing the generality of concepts, general-
izing concepts, and specializing concepts. We start with a brief overview of several
machine learning strategies that are particularly useful for knowledge-based agents.

8.1 INTRODUCTION TO MACHINE LEARNING

8.1.1 What Is Learning?

The following are definitions of learning given by some of the most prominent researchers
in this field:

“Learning denotes changes in the system that are adaptive in the sense that they
enable the system to do the same task or tasks drawn from the same population more
efficiently and more effectively the next time” (Simon, 1983, p. 28).

“‘Learning’ is making useful changes in the workings of our minds” (Minsky, 1986,
p. 120).

“Learning is constructing or modifying representations of what is being experienced”
(Michalski, 1986, p. 10).

“A computer program is said to learn from experience E with respect to some class
of tasks T and performance measure P, if its performance at tasks in T, as measured by
P, improves with experience E.” (Mitchell, 1997, p. 2).

Given the preceding definitions, we may characterize learning as denoting the way in
which people and computers:

Acquire, discover, and organize knowledge by building, modifying, and organizing

internal representations of some external reality.

Acquire skills by gradually improving their motor or cognitive abilities through repeated
practice, sometimes involving little or no conscious thought.

There are two complementary dimensions of learning: competence and efficiency. A system
is improving its competence if it learns to solve a broader class of problems and to make
fewer mistakes in problem solving. The system is improving its efficiency if it learns to solve
the problems from its area of competence faster or by using fewer resources.

222

13:53:43,
.009
 8.1. Introduction to Machine Learning 223

Machine learning is the domain of artificial intelligence that is concerned with building
adaptive computer systems that are able to improve their performance (competence and/or
efficiency) through learning from input data, from a user, or from their own problem-solving
experience.
Research in machine learning has led to the development of many basic learning
strategies, each characterized by the employment of a certain type of:

Inference (e.g., deduction, induction, abduction, analogy)

Computational or representational mechanism (e.g., rules, trees, neural networks)

Learning goal (e.g., to learn a concept, discover a formula, acquire new facts, acquire
new knowledge about an entity, refine an entity)

The following are some of the most representative learning strategies:

Rote learning

Version space learning

Decision trees induction

Clustering

Rule induction (e.g., Learning rule sets, Inductive logic programming)

Instance-based strategies (e.g., K-nearest neighbors, Locally weighted regression, Col-
laborative filtering, Case-based reasoning and learning, Learning by analogy)

Bayesian learning (e.g., Naïve Bayes learning, Bayesian network learning)

Neural networks and Deep learning

Model ensembles (e.g., Bagging, Boosting, ECOC, Staking)

Support vector machines

Explanation-based learning

Abductive learning

Reinforcement learning

Genetic algorithms and evolutionary computation

Apprenticeship learning

Multistrategy learning

In the next sections, we will briefly introduce four learning strategies that are particularly
useful for agent teaching and learning.

8.1.2 Inductive Learning from Examples

The goal of inductive learning from examples is to learn a general description of a concept,
for instance, the concept of “cup,” by analyzing positive examples of cups (i.e., objects that
are cups) and negative examples of cups (i.e., objects that are not cups).
The learning agent will attempt to find out what is common to the cups and what
distinguishes them from non-cups. For instance, in the illustration from Figure 8.1, the
agent may learn that a cup should have a handle because all the positive examples of cups
have handles, and the negative examples of cups do not have handles. However, the color
does not seem to be important for a cup because a particular color is encountered for both
cups and non-cups.
Learning a good concept description through this learning strategy requires a very large
number of positive and negative examples. On the other hand, this is the only information

13:53:43,
.009
 224 Chapter 8. Learning for Knowledge-based Agents

Positive examples of cups: P1 P2 ...

Negative examples of cups: N1 N2 …

Description of the cup concept: has-handle(x), ...

Figure 8.1. Illustration of inductive concept learning from examples.

An example of a cup
cup(o1): color(o1, white), made-of(o1, plastic), light-mat(plastic),
has-handle(o1), has-flat-bottom(o1), up-concave(o1), ...
Proof that o1 is a cup Generalized proof
cup(o1) cup(x)

lif table(o1) stable(o1) open-vessel(o1) lif table(x) stable(x) open-vessel(x)

... ... ... ...
light(o1) graspable(o1) light(x) graspable(x)

light-mat(plastic) light-mat(y)
made-of(o1,plastic) has-handle(o1) made-of(x, y) has-handle(x)

Learned rule
∀x ∀y made-of(x, y), light-mat(y), has-handle(x) ...  cup(x)

Figure 8.2. Illustration of explanation-based learning.

the agent needs. For instance, the agent does not require any prior knowledge to perform
this type of learning.
The result of this learning strategy is the increase of the problem-solving competence of
the agent. Indeed, the agent will learn to perform tasks it was not able to perform before,
such as recognizing the cups from a set of objects.

8.1.3 Explanation-based Learning

The goal of explanation-based learning is to improve the agent’s efficiency in problem
solving. The agent is able to perform some task, for example, to recognize cups, but in
an inefficient manner. Through explanation-based learning, the agent learns to perform its
tasks faster, as illustrated in the following.
Let us assume that the agent receives the description of the cup o1 from the top part of
Figure 8.2. Using its prior knowledge, it will recognize that this object is a cup by
performing a complex reasoning process demonstrating that o1 is indeed a cup. This
process is illustrated by the proof tree from the left-hand side of Figure 8.2.
The agent’s reasoning process proceeds as follows:

o1 is made of plastic, which is a light material. Therefore, o1 is light.

o1 has a handle, and it is therefore graspable.

13:53:43,
.009
 8.1. Introduction to Machine Learning 225

Being light and graspable, o1 is liftable.

...

Being liftable, stable, and an open vessel, o1 is a cup.

Notice that the agent used the fact that o1 has a handle in order to prove that o1 is a cup.
This means that having a handle is an important feature. On the other hand, the agent did
not use the color of o1 to prove that o1 is a cup. This means that color is not important.
Notice how the agent reaches the same conclusions as in inductive learning from
examples, but through a different line of reasoning, and based on a different type of
information (i.e., prior knowledge instead of multiple examples).
The next step in the learning process is to generalize the proof tree from the left-hand
side of Figure 8.2 into the general tree from the right-hand side. This is done by using the
agent’s prior knowledge of how to generalize the individual inferences from the specific tree.
While the tree from the left-hand side proves that the specific object o1 is a cup, the tree
from the right-hand side proves that any object x that satisfies the leaves of the general tree
is a cup. Thus the agent has learned the general cup recognition rule from the bottom of
Figure 8.2.
To recognize that another object, o2, is a cup, the agent needs only to check that it
satisfies the rule, that is, to check for the presence of these features discovered as
important (i.e., light-mat, has-handle, etc.). The agent no longer needs to build a complex
proof tree. Therefore, cup recognition is done much faster.
Finally, notice that the agent needs only one example from which to learn. However, it
needs a lot of prior knowledge to prove that this example is a cup. Providing such prior
knowledge to the agent is a very complex task.

8.1.4 Learning by Analogy

Learning by analogy is the process of learning new knowledge about some entity by
transferring it from a known similar entity. The central intuition supporting the learning
by analogy paradigm is that if two entities are similar in some respects, then they could be
similar in other respects as well.
An important result of the learning by analogy research (Winston, 1980; Carbonell, 1983,
1986; Gentner, 1983; Davies and Russell, 1990; Forbus et al., 1994; Veloso, 1994) is that the
analogy involves mapping some underlying causal network of relations between analogous
situations. A causal network of relations generally means a set of relations related by higher-
order causal-like relations, such as “physical-cause(ri, rj),” “logically-implies(ri, rj),” “enables
(ri, rj),” “explains(ri, rj),” “justifies(ri, rj),” “determines(ri, rj),” and so on. The idea is that such
similar “causes” are expected to have similar effects. This idea is illustrated in Figure 8.3.
Because A’ is similar to A, which causes B, A’ is expected to cause something (B’) that is
similar to B. Thus analogy involves mapping some underlying “causal network of rela-
tions” between a (better-known) source entity A and a (less-known) target entity B, with
the goal of transferring knowledge from A to B.
For instance, we can teach students about the structure of the hydrogen atom by using the
analogy with the solar system (see Figure 8.4): The hydrogen atom is like our Solar System.
We are telling the students that the hydrogen atom has a similar structure with that of
the Solar System, where the electrons revolve around the nucleus as the planets revolve
around the sun.

13:53:43,
.009
 226 Chapter 8. Learning for Knowledge-based Agents

Source similar Target

A A’

causes causes?

similar
B B’

A causes B and A is similar to A’.

Therefore, it is possible that A’ causes something (B’), which is similar to B.
Figure 8.3. Analogical reasoning.

yellow

has as color

Sun nucleus
has as mass has as temp has as mass

mass Sun at tracts temp Sun mass nucleus at tracts

greater than greater than greater than

causes
revolves causes?
around revolves
mass Earth temp Earth mass electron around
has as mass has as temp has as mass
Earth electron

The Solar System The hydrogen atom

Figure 8.4. Illustration of analogical learning.

The students may then infer that other features of the Solar System are also features of
the hydrogen atom. For instance, in the Solar System, the greater mass of the sun and its
attraction of the planets cause the planets to revolve around it. Therefore, the students may
hypothesize that this causal relationship is also true in the case of the hydrogen atom: The
greater mass of the nucleus and its attraction of the electrons cause the electrons to revolve
around the nucleus. This is indeed true and represents a very interesting discovery.
The main problem with analogical reasoning is that not all the features of the Solar
System are true for the hydrogen atom. For instance, the sun is yellow, but the nucleus is
not. Therefore, the information derived by analogy has to be verified.

8.1.5 Multistrategy Learning

As illustrated in Table 8.1, the individual learning strategies have complementary strengths
and weaknesses. For instance, inductive learning from example requires a lot of examples
while explanation-based learning requires only one example. On the other hand, inductive
learning from examples does not require any prior knowledge while explanation-based
learning requires complete prior knowledge.
Multistrategy learning attempts to integrate synergistically such complementary learn-
ing strategies in order to take advantage of their relative strengths to compensate for

13:53:43,
.009
 8.2. Concepts 227

Table 8.1 Illustration of the Complementariness of the Learning Strategies

Inductive learning Explanation-based Multistrategy

from examples learning learning

Examples needed many one several

Prior knowledge very little complete incomplete
Type of inference induction deduction induction, deduction, . . .
Effect on agent’s improves competence improves efficiency improves competence
behavior and/or efficiently

positive example concept descript ion

negative exception + ?O1 instance of PhD student
negat ive example - is interested in ?O2
- ?O2 instance of area of expertise
positive exception +
Figure 8.5. Examples and exceptions of a concept.

their relative weaknesses. Multistrategy learning is concerned with developing learning

agents that synergistically integrate two or more learning strategies in order to solve
learning problems that are beyond the capabilities of the individual learning
strategies that are integrated (Tecuci and Michalski, 1991; Tecuci 1993; Michalski and
Tecuci, 1994).
An example of multistrategy learning system is Disciple which integrates learning from
examples, learning from explanations, and learning by analogy and experimentation, as
will be discussed in Chapters 9 and 10.

8.2 CONCEPTS

8.2.1 Concepts, Examples, and Exceptions

As discussed in Section 5.2, a concept represents a set of instances.
With respect to the description of a concept, an instance can be a positive example of
the concept, a negative example, a positive exception, or a negative exception, as illus-
trated in Figure 8.5.
A positive example is an instance of the concept that is covered by the description of the
concept.
A negative example is an instance that does not belong to the concept and is not covered
by the description of the concept.
A positive exception is a positive example that is not covered by the description of the
concept, and any generalization of the description that would cover it would also cover some
negative examples.

13:53:43,
.009
 228 Chapter 8. Learning for Knowledge-based Agents

A negative exception is a negative example that is covered by the description of the

concept, and any specialization of the description that would uncover it would also uncover
some positive examples.
The generalization and the specialization of concepts will be discussed in the following
sections.

8.2.2 Examples and Exceptions of a Partially Learned Concept

Figure 8.6 shows the representation of a partially learned concept consisting of the
Plausible Version Space of the Main Condition (with the plausible upper bound in light
green and the plausible lower bound in dark green) and the Plausible Version Space of an
Except-When Condition (with the plausible upper bound in light red and the plausible
lower bound in dark red). In general, a partially learned concept has a plausible version
space of a Main condition, and none, one, or several plausible version spaces of Except-
When conditions.
A partially learned concept may have known positive and negative examples. For the
other instances of the representation space, one may estimate their nature based on their
actual position with respect to the plausible bounds of the concept, as illustrated in
Figure 8.6 and defined in the following.
An instance covered by the plausible lower bound of the Main condition of the concept
and not covered by the plausible version space of any Except-When condition is most
likely a positive example of the concept.
An instance covered by the plausible upper bound of the Main condition of the
concept, but not covered by the plausible lower bound of the Main condition and not
covered by the plausible version space of any Except-When condition, is likely to be
a positive example of the concept.
An instance covered by the plausible lower bound of one of the Except-When
plausible version space conditions of the concept is most likely a negative example of
the concept.
Finally, an instance covered by the plausible upper bound of an Except-When plaus-
ible version space condition of the concept is likely to be a negative example of the
concept.

Plausible Version Space Plausible Version Space of

of the Main Condition the Except-When Condition

Positive
exception +

Negative
exception –
+
Positive
example
Negative –
example
Most likely a More likely a More likely a Most likely a
positive example positive example negative example negative example

Figure 8.6. Examples and exceptions of a partially learned concept.

13:53:43,
.009
 8.3. Generalization and Specialization Rules 229

In the next sections, we will describe in detail the basic learning operations dealing with
concepts: the generalization of concepts, the specialization concepts, and the comparison of
the generality of the concepts.

8.3 GENERALIZATION AND SPECIALIZATION RULES

A concept was defined as representing a set of instances. In order to show that a concept
P is more general than a concept Q, this definition would require the computation and
comparison of the (possibly infinite) sets of the instances of P and Q. In this section, we
will introduce generalization and specialization rules that will allow one to prove that a
concept P is more general than another concept Q by manipulating the descriptions of
P and Q, without computing the sets of instances that they represent.
A generalization rule is a rule that transforms (the description of) a concept into (the
description of) a more general concept. The generalization rules are usually inductive
transformations. The inductive transformations are not truth preserving but falsity pre-
serving. That is, if P is true and is inductively generalized to Q, then the truth of Q is not
guaranteed. However, if P is false, then Q is also false.
A specialization rule is a rule that transforms a concept into a less general concept. The
reverse of any generalization rule is a specialization rule. Specialization rules are deduct-
ive, truth-preserving transformations.
A reformulation rule transforms a concept into another, logically equivalent concept.
Reformulation rules are also deductive, truth-preserving transformations.
If one can transform concept P into concept Q by applying a sequence of generalization
rules, then Q is more general than P.
Consider the phrase, “Students who have majored in computer science at George
Mason University between 2007 and 2008.” The following are some of the phrases that
are obvious generalizations of this phrase:

“Students who have majored in computer science between 2007 and 2008”

“Students who have majored in computer science between 2000 and 2012”

“Students who have majored in computer science at George Mason University”

“Students who have majored in computer science”

Some of the phrases that are specializations of the preceding phrase follow:

“Graduate students who have majored in computer science at George Mason Univer-
sity between 2007 and 2008”

“Students who have majored in computer science at George Mason University in 2007”

“Undergraduate students who have majored in both computer science and mathemat-
ics at George Mason University in 2008”

It will be easy to demonstrate the generalization relationships between the preceding

phrases by using some of the following generalization rules, which will be described in the
following sections:

Turning constants into variables

Turning occurrences of a variable into different variables

Climbing the generalization hierarchies

13:53:43,
.009
 230 Chapter 8. Learning for Knowledge-based Agents

Dropping conditions

Extending intervals

Extending ordered sets of intervals

Extending discrete sets

Using feature definitions

Using inference rules

8.3.1 Turning Constants into Variables

The turning constants into variables generalization rule consists in generalizing an expres-
sion by replacing a constant with a variable. For example, expression [8.1] represents the
following concept: “The set of professors with 55 publications.”

E1 = ?O1 instance of professor [8.1]

number of publications 55

By replacing 55 with the variable ?N1, which can take any value, we generalize this
concept to the one shown in [8.2]: “The set of professors with any number of publications.”
In particular, ?N1 could be 55. Therefore the second concept includes the first one.

E2 = ?O1 instance of professor [8.2]

number of publications ?N1

Conversely, by replacing ?N1 with 55, we specialize the concept [8.2] to the concept
[8.1]. The important thing to notice here is that by a simple syntactic operation (turning a
number into a variable), we can generalize a concept. This is one way in which an agent
generalizes concepts.

8.3.2 Turning Occurrences of a Variable into Different Variables

According to this rule, the expression [8.3] may be generalized to the expression [8.4]
by turning the two occurrences of the variable ?O3 in E1 into two variables, ?O31 and ?
O32:

E1 = ?O1 instance of paper [8.3]

is authored by ?O3
?O2 instance of paper
is authored by ?O3
?O3 instance of professor
E2 = ?O1 instance of paper [8.4]
is authored by ?O31
?O2 instance of paper
is authored by ?O32
?O31 instance of professor
?O32 instance of professor

E1 may be interpreted as representing the concept: “the papers ?O1 and ?O2 authored
by the professor ?O3.” E2 may be interpreted as representing the concept: “the papers ?O1

13:53:43,
.009
 8.3. Generalization and Specialization Rules 231

and ?O2 authored by the professors ?O31 and ?O32, respectively.” In particular, ?O31 and
?O32 may represent the same professor. Therefore, the second set includes the first one,
and the second expression is more general than the first one.

8.3.3 Climbing the Generalization Hierarchies

One can generalize an expression by replacing a concept from its description with a more general
concept, according to some generalization hierarchy. For instance, the expression [8.5] may be
generalized to [8.6] by replacing the concept assistant professor with the more general concept
professor (see the generalization hierarchy in Figure 5.4, p. 158). The reverse operation of
replacing a concept with a less general one leads to the specialization of an expression.

E1 = ?O1 instance of assistant professor [8.5]

has as employer ?O2
?O2 instance of state university
E2 = ?O1 instance of professor [8.6]
has as employer ?O2
?O2 instance of state university

8.3.4 Dropping Conditions

The agent can generalize a concept by dropping a condition, that is, by dropping a constraint
that its instances must satisfy. For example, the expression [8.7] may be generalized to [8.8]
by removing the constraint on the professor to be employed by a state university.

E1 = ?O1 instance of assistant professor [8.7]

has as employer ?O2
?O2 instance of state university

E2 = ?O1 instance of assistant professor [8.8]

8.3.5 Extending Intervals

A number may be generalized to an interval containing it. For example, the expression [8.9]
(the set of professors with 55 publications) may be generalized to [8.10] (the set of
professors with 50 to 60 publications), which in turn may be generalized to [8.11] (the
set of professors with 25 to 75 publications).

E1 = ?O1 instance of professor [8.9]

number of publications 55
E2 = ?O1 instance of professor [8.10]

number of publications ?N1

?N1 is in [50, 60]
E3 = ?O1 instance of professor [8.11]
number of publications ?N1
?N1 is in [25, 75]

13:53:43,
.009
 232 Chapter 8. Learning for Knowledge-based Agents

8.3.6 Extending Ordered Sets of Intervals

An ordered set of intervals is an ordered set of symbols where each symbol corresponds to
an interval. For example, “infant” corresponds to the interval “(0.0, 1.0),” and “toddler”
corresponds to the interval “[1.0, 4.5).” Obviously infant < toddler. Such an ordered set of
intervals may be regarded as an ordered generalization hierarchy, where the nodes are
ordered from left to right, as illustrated in Figure 8.7.
Using such an ordered set of intervals, one may generalize the expression [8.12]
(persons from youth to teen) to the expression [8.13] (persons from youth to mature)
by replacing the symbolic interval “[youth – teen]” with the larger interval “[youth –
mature].”

E1 = ?O1 instance of person [8.12]

has as age [youth – teen]
E1 = ?O1 instance of person [8.13]
has as age [youth – mature]

8.3.7 Extending Symbolic Probabilities

A symbolic probability may be generalized by replacing it with a symbolic probability
interval containing it. For example, “very likely” may be generalized to the interval “[likely –
very likely]” (which includes the values “likely” and “very likely”). This interval can be
further generalized by replacing it with a larger interval, such as “[likely – certain].”

8.3.8 Extending Discrete Sets

An expression may be generalized by replacing a discrete set with a larger discrete set that
includes the first set. For example, the expression [8.14] (the set of flags with white color, or
red color, or both) may be generalized to the expression [8.15] (the set of flags with white
color, or red color, or blue color, or any combination of these colors).

E1 = ?O1 instance of flag [8.14]

includes colors from {white, red}
E2 = ?O1 instance of flag [8.15]
includes colors from {white, red, blue}

human age

infant toddler youth teen mature elder

(0.0, 1.0) [1.0, 4.5) [4.5, 12.5) [12.5, 19.5) [19.5, 65.5) [65.5, 150.0]
Figure 8.7. Ordered set of intervals as an ordered generalization hierarchy.

13:53:43,
.009
 8.3. Generalization and Specialization Rules 233

8.3.9 Using Feature Definitions

This rule generalizes an expression containing a feature, such as, “A feature B,” by replacing
A and B with the domain and the range of feature, respectively.
This rule is illustrated by the generalization of the expression [8.16] (professors who are
experts in computer science) to the expression [8.17] (persons who are experts in some
area of expertise).

E1 = ?O1 instance of professor [8.16]

is expert in ?O2
?O2 instance of Computer Science

E2 = ?O1 instance of person [8.17]

is expert in ?O2
?O2 instance of area of expertise
In the preceding example, professor was generalized to person (the domain of is expert
in), and Computer Science was generalized to area of expertise (the range of is expert in).
In general, a generalization rule indicates how one can generalize a concept through a
simple syntactic transformation without suggesting the actual generalization to perform,
but only a set of possible generalizations. The using feature definition rule actually suggests
a generalization to perform, which will be useful during learning. In fact, it indicates the
most general generalization that can be performed. Indeed, ?O1 in [8.16] has the feature is
expert in. Therefore, it has to be in the domain of this feature, which is person. Similarly,
?O2 is the value of the feature is expert in. Therefore, it has to be in the range of this
feature, which is area of expertise.

8.3.10 Using Inference Rules

Given an inference rule of the form “A ➔ B,” one may generalize an expression by replacing A with
B. For example, using the theorem [8.18] one may generalize the expression [8.19] (students
and their PhD advisors) to the expression [8.20] (students and the professors they know).

8X, 8Y ((X has as PhD advisor Y) ➔ (X knows Y)) [8.18]

E1 = ?O1 instance of student [8.19]
has as PhD advisor ?O2
?O2 instance of professor
E2 = ?O1 instance of student [8.20]
knows ?O2
?O2 instance of professor
Indeed, by applying the preceding inference rule, one may transform E1 into the
equivalent expression E’1 shown in [8.21]. Then, by dropping the relation has as PhD
advisor, one generalizes E’1 into E2.

E’1 = ?O1 instance of student [8.21]

has as PhD advisor ?O2
knows ?O2
?O2 instance of professor

13:53:43,
.009
 234 Chapter 8. Learning for Knowledge-based Agents

8.4 TYPES OF GENERALIZATIONS AND SPECIALIZATIONS

Up to this point we have only defined when a concept is more general than another
concept. Learning agents, however, would need to generalize sets of examples and
concepts. In the following we define some of these generalizations.

8.4.1 Definition of Generalization

As defined in Section 5.3, a concept P is said to be more general than another concept
Q if and only if the set of instances represented by P includes the set of instances
represented by Q.
If a concept Q can be transformed into another concept P by applying a sequence of
generalization rules, then P is more general than Q. Indeed, the application of each
successive generalization rule generalizes the concept Q, as shown in the previous
sections. As an illustration, consider the concept C1 in [8.22] and the concept C2 in [8.23].

C1 = ?O1 instance of assistant professor [8.22]

number of publications 10
is employed by George Mason University
C2 = ?O1 instance of professor [8.23]
number of publications ?N1
?N1 is in [10, 35]

To show that [8.23] is more general than [8.22] it is enough to show that [8.22] can be
transformed into [8.23] by applying a sequence of generalization rules. The sequence is the
following one:

Generalize assistant professor to professor (climbing the generalization hierarchy).

Generalize 10 to [10, 35] (extending a number to an interval).

Drop “?O1 is employed by George Mason University” (dropping condition).

8.4.2 Minimal Generalization

The concept P is a minimal generalization of a different concept Q if and only if P is a
generalization of Q, and P is not more general than any other generalization of Q.
Consider the generalization hierarchy from Figure 8.8 and the concepts from Figure 8.9.
G3 is a minimal generalization of G2 because there is no concept in Figure 8.8 which is
more general than student and less general than person.
However, G3 is not a minimal generalization of G1. Indeed, the concept G4 in [8.24] is
more general than G1 and less general than G3.

G4 = ?O1 instance of employee [8.24]

is interested in ?O2
?O2 instance of area of expertise

If the minimal generalization of a concept is unique, it is called the least general

generalization.

13:53:43,
.009
 8.4. Types of Generalizations and Specializations 235

person

employee student

university employee graduate student

graduate graduate
research teaching
assistant assistant

Figure 8.8. Fragment of a generalization hierarchy.

G3 ?O1 instance of person

is interested in ?O2
?O2 instance of area of exper tise

G1 ?O1 instance of university employee G2 ?O1 instance of student

is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise

C1 ?O1 instance of graduate research assistant

is interested in ?O2
?O2 instance of area of expertise

Figure 8.9. Illustration of a minimal generalization.

G1 ?O1 instance of university employee

is interested in ?O2
?O2 instance of area of expertise

S1 ?O1 instance of graduate research assistant S2 ?O1 instance of graduate teaching assistant
is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise

C1 ?O1 instance of graduate research assistant C2 ?O1 instance of graduate teaching assistant
is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise
requires programming requires fieldwork

Figure 8.10. Illustration of a minimal specialization.

8.4.3 Minimal Specialization

The concept Q is a minimal specialization of a different concept P if and only if Q is a
specialization of P, and no other specialization of P is more general than Q. For example,
consider the generalization hierarchy from Figure 8.8 and the concepts from Figure 8.10.
Clearly C1 and C2 are not minimal specializations of G1. However, both S1 and S2 are
minimal specializations of G1.

13:53:43,
.009
 236 Chapter 8. Learning for Knowledge-based Agents

8.4.4 Generalization of Two Concepts

The concept Cg is a generalization of the concepts C1 and C2 if and only if Cg is more general
than C1 and Cg is more general than C2. One may show that if both the concept C1 and the
concept C2 can be transformed into the concept Cg by applying generalization rules, then
Cg is a generalization of C1 and C2.
As an illustration, consider the concept C1 described by the expression [8.25], and the
concept C2 described by the expression [8.26]. One may easily show that the concept
C described by the expression [8.27] is a generalization of C1 and C2. Indeed, by general-
izing assistant professor to professor (through climbing the generalization hierarchy), by
generalizing 10 to [10, 35] (through replacing a number with an interval containing it), and
by dropping the condition “?O1 is employed by George Mason University,” one generalizes
C1 to C. Similarly, by generalizing associate professor to professor, and 35 to [10, 35], one
generalizes C2 into C. Therefore, C is a generalization of C1 and C2.

C1 = ?O1 instance of assistant professor [8.25]

number of publications 10
is employed by George Mason University
C2 = ?O1 instance of associate professor [8.26]
number of publications 35
C = ?O1 instance of professor [8.27]
number of publications ?N1
?N1 is in [10, 35]

In general, to determine a generalization of two concepts, each represented as a clause,

such as [8.25] and [8.26], one first matches the features of the two concepts. Then one
applies the dropping condition rule to remove the unmatched features (and possibly even
matched features). Finally, one applies other generalization rules to determine the gener-
alizations of the matched feature values. Notice that usually there are several generaliza-
tions of two concepts.
In a similar way, one can determine a generalization G of two more complex expres-
sions E1 and E2, each consisting of a conjunction of clauses. G will consist of the
conjunction of the generalizations of some of the corresponding clauses in the two
expressions E1 and E2.

8.4.5 Minimal Generalization of Two Concepts

The concept G is a minimal generalization of the concepts C1 and C2 if and only if G is a
generalization of C1 and C2, and G is not more general than any other generalization of
C1 and C2.
To determine a minimal generalization of two clauses, one has to keep all the common
features of the clauses and determine a minimal generalization of each of the matched
feature values. In a similar way, one determines the minimal generalization of two
conjunctions of clauses by matching the clauses and determining the minimal generaliza-
tions of the matched clauses. These procedures are correct if we assume that there are no
other common features due to theorems. Otherwise, all the common features will have to
first be made explicit by applying the theorems.

13:53:43,
.009
 8.4. Types of Generalizations and Specializations 237

Notice, however, that there may be more than one minimal generalization of two
expressions. For instance, according to the generalization hierarchy from the middle of
Figure 8.8, there are two minimal generalizations of graduate research assistant and
graduate teaching assistant. They are university employee and graduate student. Conse-
quently, there are two minimal generalizations of S1 and S2 in Figure 8.11: mG1 and
mG2. The generalization mG1 was obtained by generalizing graduate research assistant and
graduate teaching assistant to university employee. mG2 was obtained in a similar fashion,
except that graduate research assistant and graduate teaching assistant were generalized to
graduate student. Neither mG1 nor mG2 is more general than the other. However, G3 is
more general than each of them.
Disciple agents employ minimal generalizations, also called maximally specific gener-
alizations (Plotkin, 1970; Kodratoff and Ganascia, 1986). They also employ maximal
generalizations, also called maximally general generalizations (Tecuci and Kodratoff,
1990; Tecuci, 1992; Tecuci 1998).

8.4.6 Specialization of Two Concepts

The concept Cs is a specialization of the concepts C1 and C2 if and only if Cs is less general
than C1 and Cs is less general than C2. For example, research professor is a specialization of
researcher and professor.
One can easily show that if both the concept C1 and the concept C2 can be transformed
into the concept Cs by applying specialization rules (or Cs can be transformed into C1, and
it can also be transformed into C2 by applying generalization rules), then Cs is a special-
ization of C1 and C2.

8.4.7 Minimal Specialization of Two Concepts

The concept C is a minimal specialization of two concepts C1 and C2 if and only if C is
a specialization of C1 and C2 and no other specialization of C1 and C2 is more general
than C.

G3 ?O1 instance of person

is interested in ?O2
?O2 instance of area of expertise

mG1 ?O1 instance of university employee mG2 ?O1 instance of graduate student
is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise

S1 ?O1 instance of graduate research assistant S2 ?O1 instance of graduate teaching assistant
is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise

Figure 8.11. Several generalizations of the concepts S1 and S2.

13:53:43,
.009
 238 Chapter 8. Learning for Knowledge-based Agents

The minimal specialization of two clauses consists of the minimal specialization of the
matched feature-value pairs, and of all the unmatched feature-value pairs. This procedure
assumes that no new clause feature can be made explicit by applying theorems. Otherwise,
one has first to make all the features explicit.
The minimal specialization of two conjunctions of clauses C1 and C2 consists of the
conjunction of the minimal specializations of each of the matched clauses of C1 and C2,
and of all the unmatched clauses from C1 and C2.
Figure 8.12 shows several specializations of the concepts G1 and G2. mS1 and mS2 are
two minimal specializations of G1 and G2 because graduate research assistant and graduate
teaching assistant are two minimal specializations of university employee and graduate
student.
Notice that in all the preceding definitions and illustrations, we have assumed that the
clauses to be generalized correspond to the same variables. If this assumption is not
satisfied, then one would need first to match the variables and then compute the general-
izations. In general, this process is computationally expensive because one will need to try
different matchings.

8.5 INDUCTIVE CONCEPT LEARNING FROM EXAMPLES

Inductive concept learning from examples has already been introduced in Section 8.1.2. In
this section, we will discuss various aspects of this learning strategy that are relevant to
agent teaching and learning. The problem of inductive concept learning from examples
can be more precisely defined as indicated in Table 8.2.
The bias of the learning agent is any basis for choosing one generalization over another,
other than strict consistency with the observed training examples (Mitchell, 1997). In the
following, we will consider two agents that employ two different preference biases: a
cautious learner that always prefers minimal generalizations, and an aggressive learner
that always prefers maximal generalizations.
Let us consider the positive examples [8.28] and [8.29], and the negative example [8.30]
of a concept to be learned by these two agents in the context of the generalization
hierarchies from Figure 8.13.

G1 ?O1 instance of university employee G2 ?O1 instance of graduate student

is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise

mS1 ?O1 instance of graduate research assistant mS2 ?O1 instance of graduate teaching assistant
is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise

C1 ?O1 instance of graduate research assistant C2 ?O1 instance of graduate teaching assistant
is interested in ?O2 is interested in ?O2
?O2 instance of area of expertise ?O2 instance of area of expertise
requires programming requires fieldwork

Figure 8.12. Several specializations of the concepts G1 and G2.

13:53:43,
.009
8.5. Inductive Concept Learning from Examples 239

Table 8.2 The Problem of Inductive Concept Learning from Examples

Given

A language of instances.

A language of generalizations.

A set of positive examples (E1, . . ., En) of a concept.

A set of negative (or counter) examples (C1, . . ., Cm) of the same concept.

A learning bias.

Other background knowledge.
Determine
A concept description that is a generalization of the positive
examples and that does not cover any of the negative examples.
Purpose of concept learning
Predict if an instance is a positive example of the learned concept.

person
university

staf f member professor

state private
university university
computer full associate assistant
technician professor professor professor

George Janet Mark George Mason University Stanford

Dean Rice White University of Virginia University

Figure 8.13. Two generalization hierarchies.

Positive examples:
Mark White instance of assistant professor [8.28]
is employed by George Mason University

Janet Rice instance of assistant professor [8.29]

is employed by University of Virginia

Negative example:
George Dean instance of computer technician [8.30]
is employed by Stanford University

What concept might be learned by the cautious learner from the positive examples
[8.28] and [8.29], and the negative example [8.30]? The cautious learner would learn a
minimal generalization of the positive examples, which does not cover the negative
example. Such a minimal generalization might be the expression [8.31], “an assistant
professor employed by a state university,” obtained by minimally generalizing George
Mason University and University of Virginia to state university.

13:53:43,
.009
240 Chapter 8. Learning for Knowledge-based Agents

?O1 instance of assistant professor [8.31]

is employed by ?O2
?O2 instance of state university

The concept learned by the cautious learner is represented in Figure 8.14 as the
minimal ellipse that covers the positive examples without covering the negative example.
Assuming a complete ontology, the learned concept is included into the actual concept.
How will the cautious learner classify each of the instances represented in Figure 8.14
as black dots? It will classify the dot covered by the learned concept as positive example,
and the two dots that are not covered by the learned concept as negative examples.
How confident are you in the classification, when the learner predicts that an instance
is a positive example? When a cautious learner classifies an instance as a positive example
of a concept, this classification is correct because an instance covered by the learned
concept is also covered by the actual concept.
But how confident are you in the classification, when the learner predicts that an
instance is a negative example? The learner may make mistakes when classifying an
instance as a negative example, such as the black dot that is covered by the actual concept
but not by the learned concept. This type of error is called “error of omission” because
some positive examples are omitted – that is, they are classified as negative examples.
Let us now consider the concept that might be learned by the aggressive learner from
the positive examples [8.28] and [8.29], and the negative example [8.30]. The aggressive
learner will learn a maximal generalization of the positive examples that does not cover the
negative example. Such a maximal generalization might be the expression [8.32], “a
professor employed by a university.” This is obtained by generalizing assistant professor
to professor (the most general generalization that does not cover computer technician from
the negative example) and by maximally generalizing George Mason University and Univer-
sity of Virginia to university. Although university covers Stanford University, this is fine
because the obtained concept [8.32] still does not cover the negative example [8.30].

?O1 instance of professor [8.32]

is employed by ?O2
?O2 instance of university

An alternative maximally general generalization of the positive examples [8.28] and

[8.29], which does not cover the negative example [8.30], is [8.33].

?O1 instance of person [8.33]

is employed by ?O2
?O2 instance of state university

Concept learned by
the cautious learner
-
Actual concept

+
+
Figure 8.14. Learning and classifications by a cautious learner.

13:53:43,
.009
8.5. Inductive Concept Learning from Examples 241

The concept learned by the aggressive learner is represented in Figure 8.15 as the
maximal ellipse that covers the positive examples without covering the negative example.
Assuming a complete ontology, the learned concept includes the actual concept.
How will the aggressive learner classify each of the instances represented in Figure 8.15
as black dots? It will classify the dot that is outside the learned concept as a negative
example, and the other dots as positive examples.
How confident are you in the classification when the learner predicts that an instance is
negative example? When the learner predicts that an instance is a negative example, this
classification is correct because that instance is not covered by the actual concept, which is
itself covered by the learned concept.
But, how confident are you in the classification when the learner predicts that an
instance is a positive example? The learner may make mistakes when predicting that an
instance is a positive example, as is the case with the dot covered by the learned concept,
but not by the actual concept. This type of error is called “error of commission” because
some negative examples are committed – that is, they are classified as positive examples.
Notice the interesting fact that the aggressive learner is correct when it classifies
instances as negative examples (they are indeed outside the actual concept because they
are outside the concept learned by the aggressive learner) while the cautious learner is
correct when it classifies instances as positive examples (they are inside the actual concept
because they are inside the concept learned by the cautious learner). How could one
synergistically integrate these two learning strategies to take advantage of their comple-
mentariness? An obvious solution is to use both strategies, learning both a minimal and a
maximal generalization from the examples, as illustrated in Figure 8.16.
What class will be predicted by a dual-strategy learner for the instances represented as
black dots in Figure 8.16? The dot covered by the concept learned by the cautious learner

Concept learned by the aggressive learner

-
Actual concept

+
+

Figure 8.15. Learning and classifications by an aggressive learner.

Concept learned by the aggressive learner

-
Actual concept

+
+
Concept learned by the cautious learner
Figure 8.16. Learning and classifications by a dual-strategy learner.

13:53:43,
.009
 242 Chapter 8. Learning for Knowledge-based Agents

will be classified, with high confidence, as a positive example. The dot that is not covered
by the concept learned by the aggressive learner will be classified, again with high
confidence, as a negative example. The dual-strategy learner will indicate that it cannot
classify the other two dots.

8.6 LEARNING WITH AN INCOMPLETE REPRESENTATION

LANGUAGE

Let us consider the ontology from Figure 8.17. What is the maximal generalization of the
positive examples John Doe and Jane Austin that does not cover the given negative example
Bob Sharp, in the case where graduate research assistant is included into the ontology? The
maximal generalization is faculty member.
But what is the maximal generalization in the case where graduate research assistant is
missing from the ontology? In this case, the maximal generalization is employee, which is,
in fact, an overgeneralization.
What is the minimal specialization of person that does not cover Bob Sharp in the case
where graduate research assistant is included into the ontology? It is faculty member.
But what is the minimal specialization in the case where graduate research assistant is
missing from the ontology? In this case, the minimal specialization is employee, which
is an underspecialization.

person
subconcept of

employee student
subconcept of
subconcept of
university employee
graduate
subconcept of student

undergraduate
staff member faculty member student
subconcept of
subconcept of
subconcept of
instructor professor PhD advisor graduate
research teaching
assistant assistant
subconcept of BS student

assistant associate full PhD student MS student

professor professor professor

instance of instance of instance of instance of instance of

John Smith John Doe Jane Ausn Bob Sharp Joan Dean
_
+ +
Figure 8.17. Plausible generalizations and specializations due to ontology incompleteness.

13:53:43,
.009
 8.7. Formal Definition of Generalization 243

Notice that the incompleteness of the ontology causes the learner both to overgeneral-
ize and underspecialize. In view of the preceding observations, what can be said about the
relationships between the concepts learned using minimal and maximal generalizations
and the actual concept when the ontology and the representation language are incom-
plete? The minimal and maximal generalizations are only approximations of the actual
concept, as shown in Figure 8.18.
Why is the concept learned with an aggressive strategy more general than the one
learned with a cautious strategy? Because they are based on the same ontology and
generalization rules.

8.7 FORMAL DEFINITION OF GENERALIZATION

8.7.1 Formal Representation Language for Concepts

A knowledge representation language defines the syntax and the semantics for expressing
knowledge in a form that an agent can use. We define a formal representation language for
concepts as follows:

Let be a set of variables. For convenience in identifying variables, their names start
with “?,” as in, for instance, ?O1. Variables are used to denote unspecified instances of
concepts.

Let be a set of constants. Examples of constants are the numbers (such as “5”), strings
(such as “programming”), symbolic probability values (such as “very likely”), and
instances (such as “John Doe”). We define a term to be either a variable or a constant.

Let be a set of features. The set includes the domain independent features
“instance of,” “subconcept of,” and “direct subconcept of,” as well as other domain-
specific features, such as “is interested in.”

Let be an object ontology consisting of a set of concepts and instances defined using the
clause representation [7.2] presented in Section 7.2, where the feature values (vi1 . . . vim)
are constants, concepts, instances, or intervals (numeric or symbolic). That is, there are no
variables in the definition of a concept or an instance from , such as the following one:

instance-k instance of concept-i

feature-i1 vi1
...
feature-in vim

With complete ontology With incomplete ontology

Learned by aggressive learner Learned by aggressive learner

Actual concept
Actual concept
Learned by Learned by
cauous learner cauous learner

Figure 8.18. The influence of the ontology on concept learning.

13:53:43,
.009
244 Chapter 8. Learning for Knowledge-based Agents

The concepts and the instances from are related by the generalization relations
“instance of” and “subconcept of.” includes the concept object, which represents all
the instances from the application domain and is therefore more general than any
other object concept.

Let be the set of the theorems and the properties of the features, variables, and
constants.

Two properties of any feature are its domain and its range. Other features may have
special properties. For instance, the relation subconcept of is transitive (see Section
5.7). Also, a concept or an instance inherits the features of the concepts that are more
general than it (see Section 5.8).

Let be a set of connectors. includes the logical connectors AND (∧), OR (∨), and
NOT (Except–When), the connectors “{” and “}” for defining alternative values of a
feature, the connectors “[” and “]” as well as “(” and “)” for defining a numeric or a
symbolic interval, the delimiter “,” (a comma), and the symbols “Plausible Upper
Bound” and “Plausible Lower Bound.”

We call the tuple a representation language for concepts.

The basic representation unit (BRU) for a concept in has the form of a tuple (?O1,
?O2,. . ., ?On), where each ?Oi is a clause:

?O1 instance of concept-i

feature-11 ?Oj
...
feature-1i ?Ol
...
?On instance of concept-n
feature-n1 ?Op
...
feature-nk ?Ot

In the preceding expression, each of concept-i . . . concept-n is either an object concept from
the object ontology (such as PhD student), a numeric interval (such as [50, 60]), a set of
numbers (such as {1, 3, 5}), a set of strings (such as {white, red, blue}), a symbolic probability
interval (such as [likely - very likely]), or an ordered set of intervals (such as [youth - mature]).
?Oj, . . . , ?Ol,. . ., ?Op, . . . , ?Ot are distinct variables from the sequence (?O1, ?O2, . . . , ?On).
When concept-n is a set or interval such as “[50, 60],” we use “is in” instead of “instance of.”
A more complex concept is defined as a conjunctive expression “BRU ∧ not BRU1 ∧ . . . ∧
not BRUp,” where “BRU” and each “BRUk (k = 1, . . . , p)” is a conjunction of clauses. This is
illustrated by the following example, which represents the set of instances of the tuple
(?O1, ?O2, ?O3), where ?O1 is a professor employed by a university ?O2 in a long-term
position ?O3, such that it is not true that ?O1 plans to retire from ?O2 or to move to some
other organization:

?O1 instance of professor

has as employer ?O2
has as position ?O3
?O2 instance of university
?O3 instance of long-term faculty position

13:53:43,
.009
 8.7. Formal Definition of Generalization 245

Not
?O1 instance of professor
plans to retire from ?O2
?O2 instance of university
Not
?O1 instance of professor
plans to move to ?O4
?O4 instance of organization

We generally use “Except-When” instead of “Not.”

In the following sections, we will provide a formal definition of generalization in the
representation language , based on substitutions.
A substitution is a function σ = (x1 t1, . . . , xn tn), where each xi (i = 1,. . .,n) is a
variable and each ti (i = 1,. . .,n) is a term.
If li is an expression in the representation language , then σli is the expression
obtained by substituting each variable xi from li with the term ti.

8.7.2 Term Generalization

In the representation language , a term is a constant (e.g., number, string, symbolic
interval, or instance) or a variable. An unrestricted variable ?X is more general than any
constant and is as general as any other unrestricted variable (such as ?Y).

8.7.3 Clause Generalization

Let us consider two concepts described by the following two clauses, C1 and C2, respect-
ively, where v1, v2, v11, . . ., v2n are variables, b1 and b2 are concepts, and f11, . . ., f2n are
features from the ontology.

C1 = v1 instance of b1
f11 v11
...
f1m v1m

C2 = v2 instance of b2

f21 v21
...
f2n v2n

We say that the clause C1 is more general than the clause C2 if there exists a substitution
σ such that:

σv1 = v2
b1 = b2
8i2{1,. . .,m}, 9 j2{1,. . .,n} such that f1i = f2j and σv1i = v2j.

For example, the concept

C1 = ?X instance of student
enrolled at George Mason University
13:53:43,
.009
 246 Chapter 8. Learning for Knowledge-based Agents

is more general than the concept

C2 = ?Y instance of student
enrolled at George Mason University
has as sex female

Indeed, let σ = (?X ?Y). As one can see, σC1 is a part of C2, that is, each feature of
σC1 is also a feature of C2. The first concept represents the set of all students enrolled at
George Mason University, while the second one represents the set of all female students
enrolled at George Mason University. Obviously the first set includes the second one, and
therefore the first concept is more general than the second one.
Let us notice, however, that this definition of generalization does not take into account
the theorems and properties of the representation language . In general, one needs to
use these theorems and properties to transform the clauses C1 and C2 into equivalent
clauses C’1 and C’2, respectively, by making explicit all the properties of these clauses.
Then one shows that C’1 is more general than C’2. Therefore, the definition of the more
general than relation in is the following one:
A clause C1 is more general than another clause C2 if and only if there exist C’1, C’2,
and a substitution σ, such that:

C0 1 ¼L C1

C0 2 ¼L C2

σv1 ¼L v2

b1 is more general than b2 in

8 i2f1; . . . ; mg, 9 j2f1; . . . ; ng such that f 0 1i ¼L f 0 2j and σv0 1i ¼L v0 2j :

In the following sections, we will always assume that the equality is in and we will no
longer indicate this.

8.7.4 BRU Generalization

As discussed in Section 8.7.1, a basic representation unit (BRU) is a conjunction of clauses.
An example of BRU is the following one where, for notational convenience, we have
dropped the AND connector between the clauses:

?O1 instance of course

has as reading ?O2
?O2 instance of publication
has as author ?O3
?O3 instance of professor

Therefore, anytime there is a sequence of clauses, they are to be considered as being

connected by AND (∧).
Let us consider two concepts, A and B, defined by the following expressions

A ¼ A1 ∧A2 ∧ . . . ∧An

B ¼ B1 ∧B2 ∧ . . . ∧Bm

where each Ai (i = 1, ...,n) and each Bj (j = 1, ...,m) is a clause.

13:53:43,
.009
 8.8. Review Questions 247

A is more general than B if and only if there exist A’, B’, and σ such that:

A0 ¼ A, A0 ¼ A0 1 ∧A0 2 ∧ . . . ∧A0 p

B0 ¼ B, B0 ¼ B0 1 ∧B0 2 ∧ . . . ∧B0 q

8 i2f1; . . . ; pg, 9 j2f1; . . . ; qg such that σA0 i ¼ B0 j :

Otherwise stated, one transforms the concepts A and B, using the theorems and the
properties of the representation language, so as to make each clause from A’ more general
than a corresponding clause from B’. Notice that some clauses from B’ may be “left over,”
that is, they are not matched by any clause of A’.

8.7.5 Generalization of Concepts with Negations

By concept with negations, we mean an expression of the following form:

BRU∧not BRU1 ∧ . . . ∧not BRUp

where each BRU is a conjunction of clauses.

Let us consider two concepts with negations, A and B, defined by the following
expressions:

A ¼ BRUa ∧not BRUa1 ∧ . . . ∧not BRUap

B ¼ BRUb ∧not BRUb1 ∧ . . . ∧not BRUbq

A is more general than B if and only if there exist A0 , B0 , and σ, such that:

A0 ¼ A, A0 ¼ BRU0 a ∧not BRU0 a1 ∧ . . . ∧not BRU0 ap

B0 ¼ B, B0 ¼ BRU0 b ∧not BRU0 b1 ∧ . . . ∧not BRU0 bq

σBRU0 a is more general than BRU0 b

8 i2f1; . . . ; pg, 9 j2f1; . . . ; qg such that BRU0 bj is more general than σBRU0 ai :

8.7.6 Substitutions and the Generalization Rules

One can use the definition of generalization based on substitution to prove that the
generalization rules transform concepts into more general concepts.
As an illustration, let us consider the turning constants into variables generalization
rule (see Section 8.3.1) that transformed the expression E1 from [8.1] on p. 230, into the
expression E2 from [8.2] on p. 230. E2 is indeed a generalization of E1 because E1 = σE2,
where σ = (?N1 55).

8.8 REVIEW QUESTIONS

8.1. What is a positive example of a concept? What is a negative example of a concept?

8.2. What is a generalization rule? What is a specialization rule? What is a

reformulation rule?
13:53:43,
.009
248 Chapter 8. Learning for Knowledge-based Agents

8.3. Name all the generalization rules you know.

8.4. Define and illustrate the following generalization rule: climbing generalization
hierarchies.
8.5. Briefly describe and illustrate the “turning constants into variables” generalization rule.
8.6. Define and illustrate the “dropping conditions” generalization rule.
8.7. Define the minimal generalization G of two concepts A and B. Then define their
least general generalization. Do these generalizations always exist?
8.8. What is a negative exception? What is a positive exception?
8.9. Indicate three different generalizations of the phrase “History books in the Fenwick
Library,” and demonstrate that each of them is more general than the given phrase.
8.10. Give an example of a natural language phrase C that has a concept interpretation.
(a) Formulate a phrase G that is a generalization of C, and use the generalization
rules to demonstrate that G is a generalization of C.
(b) Formulate a phrase S that is a specialization of C, and use the generalization
rules to demonstrate that S is a specialization of C.
(c) Formulate a phrase D that is neither a generalization of C nor a specialization of C.
8.11. P3 in Figure 8.19 is a natural language sentence that represents a concept. Fill in each
of the other boxes with sentences such that the arrows from the figure represent all the
“subconcept of” relationships between these concepts. This means, for instance, that
P3 is less general than P5, P3 is more general than P1, and P3 is more general than P2.
However, P3 is not more general than P4, and P4 is not more general than P3. Justify
each shown “subconcept of” relationship by labeling it with the name of the general-
ization rule(s) that transform the less general concept into the more general concept.
You should use at least three different generalization rules while solving this problem.
8.12. Draw a picture representing a plausible version space, as well as a positive
example, a negative example, a positive exception, and a negative exception. Then
briefly define each of these notions.
8.13. Consider the cells consisting of two bodies, each body having two attributes,
number of nuclei (1 or 2), and color (yellow or green). The relative position of

P5

P3 The U.S. athletes who have competed in the P4

2012 Olympic games in individual competitions.

P1 P2

Figure 8.19. Determining concepts that satisfy given “subconcept of” relationships.

13:53:43,
.009
8.8. Review Questions 249

the bodies is not relevant because they can move inside the cell. For example,
((1 green) (2 yellow)) is the same with ((2 yellow) (1 green)) and represents a cell
where one body has one nucleus and is green, while the other body has two nuclei
and is yellow. You should also assume that any generalization of a cell is also
described as a pair of pairs ((s t) (u v)).
(a) Indicate all the possible generalizations of the cell from Figure 8.20 and the
generalization relations between them.
(b) Determine the number of the distinct sets of instances and the number of the
concept descriptions from this problem.
(c) Consider the cell descriptions from Figure 8.21 and determine the following
minimal generalizations: g(E1, E2), g(E2, E3), g(E3, E1), g(E1, E2, E3).

8.14. Consider the ontology fragment from the loudspeaker manufacturing domain,
shown in Figure 5.23 (p. 171), and the following expressions:

E1: ?X instance of membrane E2: ?X instance of mechanical chassis

made of ?M made of ?M
?M instance of paper ?M instance of metal
?Z instance of contact adhesive ?Z instance of mowicoll
glues ?M glues ?M
state fluid

(a) Find the minimal generalizations of E1 and E2.

(b) Find two generalizations of E1 and E2 that are not minimal generalizations.
(c) Consider one of the generalizations found at (b) and demonstrate that it is a
generalization of E1 and E2 but it is not a minimal generalization.
(d) What would be the least general generalization of E1 and E2? Does it exist?
(e) Indicate a specialization of E1.

8.15. Consider the ontology fragment from the loudspeaker manufacturing domain,
shown in Figure 5.24 (p. 172). Notice that each most specific concept, such as dust
or air press, has an instance, such as dust1 or air press1.
Consider also the following two expressions:

+ ((1 green) (2 yellow))

Figure 8.20. An example of a cell.

E1: ((1 green) (1 green)) E2: ((1 yellow) (2 green)) E3: ((1 green) (2 green))
Figure 8.21. The descriptions of three cells.

13:53:43,
.009
250 Chapter 8. Learning for Knowledge-based Agents

E1: ?X instance of soft cleaner

removes ?Z
?Y instance of membrane
?Z instance of waste material

E2: ?X instance of air sucker

removes ?Z
?Y instance of membrane
?Z instance of dust

Use the generalization rules to show that E1 is more general than E2.
8.16. Determine a generalization of the following two expressions in the context of the
ontology fragment from Figure 5.24 (p. 172):

E1: ?X instance of entrefer

may have ?Y
?Y instance of dust
?Z instance of air sucker
removes ?Y
E2: ?X instance of membrane
may have ?Y
?Y instance of surplus adhesive
?Z instance of alcohol
removes ?Y
8.17. Consider the background knowledge represented by the generalization hier-
archies shown in Figure 8.22.
Consider also the following concept:

E: ?O instance of object
color yellow
shape circle
radius 5
Indicate five different generalization rules. For each such rule, determine an
expression Eg that is more general than E according to that rule.

any color any shape

polygon round
warm color cold color

triangle rectangle circle ellipse

red orange yellow blue green black

square

Figure 8.22. Generalization hierarchies for color and shape.

13:53:43,
.009
8.8. Review Questions 251

adhesive toxic substance inflammable object

scotch tape super glue mowicoll contact adhesive

loudspeaker component material

membrane chassis assembly bolt caoutchouc paper metal

Figure 8.23. Ontology fragments.

8.18. Consider the following two concepts:

C1: ?X instance of screw C2: ?X instance of nut

head ?M cost 6
cost 5

Indicate two different generalizations of them.

8.19. Consider the following two concepts G1 and G2, and the ontology fragments in
Figure 8.23. Indicate four specializations of G1 and G2 (including a minimal
specialization).

G1:?X instance of loudspeaker component G2: ?X instance of loudspeaker component

made of ?M made of ?M
?M instance of material ?M instance of material
?Z instance of adhesive ?Z instance of inflammable object
glues ?M glues ?M

8.20. Illustrate the clause generalization defined in Section 8.7.3 with an example from
the PhD Advisor Assessment domain.

8.21. Illustrate the BRU generalization defined in Section 8.7.4 with an example from the
PhD Advisor Assessment domain.

8.22. Illustrate the generalization of concepts with negations defined in Section 8.7.5 by
using an example from the PhD Advisor Assessment domain.

8.23. Use the definition of generalization based on substitution to prove that each of the
generalization rules discussed in Section 8.3 transforms a concept into a more
general concept.

13:53:43,
.009
 9 Rule Learning

In this and the next chapter on rule refinement, we will refer to both problems and
hypotheses, interchangeably, to emphasize the fact that the learning methods presented
are equally applicable in the context of hypotheses analysis and problem solving.

9.1 MODELING, LEARNING, AND PROBLEM SOLVING

Figure 9.1 summarizes the interactions between the subject matter expert and the learning
agent that involve modeling, learning, and problem solving.
The expert formulates the problem to be solved (or the hypothesis to be analyzed), and
the agent uses its knowledge to generate a (problem-solving or argumentation) tree to be
verified by the expert.
Several cases are possible. If the problem is not completely solved, the expert will extend
the tree with additional reductions and provide solutions for the leaf problems/hypotheses.

P1

Problem Question
Answer

1 1 1
P1 S1 Pn

Mixed-Initiative
Question
Problem Solving Question
Answer
Answer

2 2 2 2
P1 S1 Pm Sm

Question
Reasoning Tree Question
Answer
Answer Ontology + Rules
3 3 3
Extend Tree Accept step P1 S1 Sp
3
Sp

1
Pn

Modeling Reject step Rule Refinement Refined

Question
Answer with Positive Rules
2
Pn1
2
Pnk
Example
Rule Refinement
Rule with Negative Refined Rules
Learning Example

Explanation-based Refined Ontology

Ontology Extension

Learned Rules

Figure 9.1. Mixed-initiative modeling, learning, and problem solving.

252

13:59:50,
.010
 9.2. Rule Learning and Refinement 253

From each new reduction provided by the expert, the agent will learn a new rule, as will be
presented in the following sections.
If the expert rejects any of the reasoning steps generated by the agent, then an explan-
ation of why that reduction is wrong needs to be determined, and the rule that generated it
will be refined to no longer generate the wrong reasoning step.
If the expert accepts a reasoning step as correct, then the rule that generated it may be
generalized. The following section illustrates these interactions.

9.2 AN ILLUSTRATION OF RULE LEARNING AND REFINEMENT

As will be discussed in the following, the subject matter expert helps the agent to learn
by providing examples and explanations, and the agent helps the expert to teach it by
presenting attempted solutions.
First, as illustrated in Figure 9.2, the expert formulates the problem to solve or the
hypothesis to analyze which, in this illustration, is the following hypothesis:

John Doe would be a good PhD advisor for Bob Sharp.

In this case, we will assume that the agent does not know how to assess this hypothesis.
Therefore, the expert has to teach the agent how to assess it. The expert will start by
developing a reduction tree, as discussed in Chapter 4 and illustrated in the middle of
Figure 9.2. The initial hypothesis is first reduced to three simpler hypotheses, guided
by a question/answer pair. Then each of the subhypotheses is further reduced, either
to a solution/assessment or to an elementary hypothesis to be assessed based on
evidence. For example, the bottom part of Figure 9.2 shows the reduction of the first
subhypothesis to an assessment.
After the reasoning tree has been developed, the subject matter expert interacts with
the agent, helping it “understand” why each reduction step is correct, as will be discussed
in Section 9.5. As a result, from each reduction step the agent learns a plausible version

1. Modeling 2. Learning
Expert explains Agent learns
how to solve a general
specific problem reducon rules
Rule1

Rule2

Figure 9.2. Modeling and rule learning.

13:59:50,
.010
254 Chapter 9. Rule Learning

space rule, as a justified generalization of it. This is illustrated in the right-hand side of
Figure 9.2 and discussed in Section 9.7. These rules are not shown to the expert, but they
may be viewed with the Rule Browser.
The agent can now use the learned rules to assess by itself similar hypotheses formu-
lated by the expert, as illustrated in Figure 9.3, where the expert formulated the following
hypothesis:

Dan Smith would be a good PhD advisor for Bob Sharp.

The reduction tree shown in Figure 9.3 was generated by the agent. Notice how the agent
concluded that Bob Sharp is interested in an area of expertise of Dan Smith, which is
Information Security, by applying the rule learned from John Doe and Bob Sharp, who share
a common interest in Artificial Intelligence.
The expert has to inspect each reduction generated by the agent and indicate whether it
is correct or not. Because the reductions from Figure 9.3 are correct, the agent generalizes
the lower bound conditions of the applied rules, if the reductions were generated based on
the upper bound conditions of these rules.
The bottom part of Figure 9.4 shows a reduction generated by the agent that is rejected
by the expert. While Dan Smith has indeed a tenured position, which is a long-term faculty
position, he plans to retire. It is therefore wrong to conclude that it is almost certain that he
will stay on the faculty of George Mason University for the duration of the dissertation of
Bob Sharp.
Such failure explanations are either proposed by the agent and accepted by the expert,
or are provided by the expert, as discussed in Section 9.5.2.
Based on this failure explanation, the agent specializes the rule that generated this
reduction by adding an Except-When plausible version space condition, as illustrated in
the right-hand side of Figure 9.4. From now on, the agent will check not only that the
faculty member has a long-term position (the main condition of the rule), but also that he
or she does not plan to retire (the Except-When condition). The refined rule is not shown
to the expert, but it may be viewed with the Rule Browser.

3. Solving 5. Refinement

Agent applies Agent refines

learned rules rules with
to solve a new new positive
problem examples
4. Critiquing

Expert accepts
reasoning

Figure 9.3. Problem solving, critiquing, and rule refinement.

13:59:50,
.010
9.2. Rule Learning and Refinement 255

1. Solving 3. Refinement

Agent refines
rule with
negative
example
2. Critiquing

Incorrect
because
Dan Smith
plans to retire
- +

Figure 9.4. Problem solving, critiquing, and refinement.

1. Solving 3. Refinement

Agent refines rule with

negative example

2. Critiquing

- +

Incorrect
because
Jane Austin
+ -
plans to move -

Figure 9.5. Problem solving, critiquing, and refinement.

Figure 9.5 shows another reasoning tree generated by the agent for an expert-
formulated hypothesis. Again the expert rejects one of the reasoning steps: Although
Jane Austin has a tenured position and does not plan to retire, she plans to move from
George Mason University and will not stay on the faculty for the duration of the dissertation
of Bob Sharp.
Based on this failure explanation, the agent specializes the rule that generated the
reduction by adding an additional Except-When plausible version space condition, as
shown in the right-hand side of Figure 9.5. From now on, the agent will check not only that
the faculty member has a long-term position, but also that he or she does not plan to retire
or move from the university.

13:59:50,
.010
256 Chapter 9. Rule Learning

The refined rule is shown in Figure 9.6. Notice that this is a quite complex rule that was
learned based only on one positive example, two negative examples, and their explan-
ations. The rule may be further refined based on additional examples.
The following sections describe in more detail the rule-learning and refinement pro-
cesses. Before that, however, let us notice a significant difference between the develop-
ment of a knowledge-based learning agent and the development of a (nonlearning)
knowledge-based agent. As discussed in Sections 1.6.3.1 and 3.1, after the knowledge base
of the (nonlearning) agent is developed by the knowledge engineer, the agent is tested
with various problems. The expert has to analyze the solutions generated by the agent, and
the knowledge engineer has to modify the rules manually to eliminate any identified
problems, testing the modified rules again.

+
-
-

Figure 9.6. Partially learned rule.

13:59:50,
.010
 9.3. The Rule-Learning Problem 257

In the case of a learning agent, both rule learning and rule refinement take place as part
of agent teaching. Testing of the agent is included into this process. This process will also
continue as part of knowledge base maintenance. If we would like to extend the agent
to solve new problems, we simply need to teach it more. Thus, in the case of a learning
agent, such as Disciple-EBR, there is no longer a distinction between knowledge base
development and knowledge base maintenance. This is very important because it is well
known that knowledge base maintenance (and system maintenance, in general) is much
more challenging and time consuming than knowledge base (system) development.
Thus knowledge base development and maintenance are less complex and much faster
in the case of a learning agent.

9.3 THE RULE-LEARNING PROBLEM

The rule-learning problem is defined in Table 9.1 and is illustrated in Figures 9.7 and 9.8.
The agent receives an example of a problem or hypothesis reduction and learns a plausible
version space rule that is an analogy-based generalization of the example. There is no
restriction with respect to what the example actually represents. However, it has to be
described as a problem or hypothesis that is reduced to one or several subproblems,
elementary hypotheses, or solutions. Therefore, this example may also be referred to as a
problem-solving episode. For instance, the example shown in the top part of Figure 9.8
reduces a specific hypothesis to its assessment or solution, guided by a question and
its answer.
The expert who is training the agent will interact with it to help it understand why the
example is a correct reduction. The understanding is done in the context of the agent’s
ontology, a fragment of which is shown in Figure 9.7.
The result of the rule-learning process is a general plausible version space rule that will
allow the agent to solve problems by analogy with the example from which the rule was
learned. The plausible version space rule learned from the example at the top of Figure 9.8
is shown at the bottom part of the figure. It is an IF-THEN structure that specifies the
conditions under which the problem from the IF part has the solution from the THEN part.
The rule is only partially learned because, instead of a single applicability condition, it has
two conditions:

Table 9.1 The Rule-Learning Problem

GIVEN

A knowledge base that includes an ontology and a set of (previously learned) rules

An example of a problem reduction expressed with the concepts and instances from the agent’s
knowledge base

An expert who will interact with the agent to help it understand why the example is correct

DETERMINE

A plausible version space rule, where the upper bound is a maximal generalization of the
example, and the lower bound is a minimal generalization that does not contain any specific
instance

An extended ontology, if any extension is needed for the understanding of the example

13:59:50,
.010
 258 Chapter 9. Rule Learning

object

actor position area of expertise

person organization university position subconcept of

employee educational Computer

student faculty position
organization Science

university employee subconcept of long-term instance of

graduate faculty position
student university
faculty member
Artificial Information
subconcept of instance of tenured position
Intelligence Security
professor
George Mason Indiana
PhD advisor
University University has as position
subconcept of domain person
PhD student range position
associate plans to retire from
is expert in
professor instance of domain person
domain person
range area of expertise range organization
instance of
is interested in plans to move to
is expert in is interested in domain person domain person
John Doe Bob Sharp range area of expertise range organization

Figure 9.7. Fragment of the agent’s ontology.

A plausible upper bound condition that is a maximal generalization of the instances and
constants from the example (e.g., Bob Sharp, certain), in the context of the agent’s ontology

A plausible lower bound condition that is a minimal generalization that does not
contain any specific instance

The relationships among the variables ?O1, ?O2, and ?O3 are the same for both conditions
and are therefore shown only once in Figure 9.8, under the conditions.
Completely learning the rule means learning an exact condition, where the plausible
upper bound is identical with the plausible lower bound.
During rule learning, the agent might also extend the ontology with new features or
concepts, if they are needed for the understanding the example.

9.4 OVERVIEW OF THE RULE-LEARNING METHOD

An overview of the rule-learning method is presented in Figure 9.9 and in Table 9.2. As in
explanation-based learning (DeJong and Mooney, 1986; Mitchell et al., 1986), it consists of
two phases: an explanation phase and a generalization phase. However, in the explanation
phase the agent does not automatically build a deductive proof tree but an explanation
structure through mixed-initiative understanding. Also, the generalization is not a deduct-
ive one, but an analogy-based one.
In the following, we will describe this learning method in more detail and illustrate it.
First we will present the mixed-initiative process of explanation generation and example

13:59:50,
.010
9.4. Overview of the Rule-Learning Method 259

Rule learning

Figure 9.8. Rule learned from an example.

Mixed-Initiative Analogy-based
Understanding Generalization

Plausible version
space rule
Example of a
reduction step

Explanation
NLP, Analogy

Knowledge Base

Figure 9.9. The rule-learning method.

13:59:50,
.010
 260 Chapter 9. Rule Learning

Table 9.2 Basic Steps of the Rule-Learning Method

(1) Mixed-Initiative Understanding (Explanation Generation)

Through a mixed-initiative interaction with the subject matter expert, determine the set of
relationships EX from the agent’s ontology that collectively explain why the example is correct.
EX represents the explanation of the example E. In general, these relationships express the
meaning of the question/answer pair from the example E, as well as other conditions that need
to be satisfied by the instances and constants from the example E. During this process, new
objects and features may be elicited from the expert and added to the ontology, if the ontology
is incomplete.
(2) Example Reformulation
Generate a variable for each instance and each constant (i.e., number, string, or symbolic
probability) that appears in the example E and its explanation EX. Then use these variables to
create an instance I of the concept C representing the applicability condition of the rule R to be
learned. C is the concept to be learned as part of rule learning and refinement. Finally, reformulate
the example as a very specific IF-THEN rule with I as its applicability condition. The IF problem, the
question, the answer, and the THEN subproblems (or solutions) are obtained by replacing each
instance or constant from the example E with the corresponding variable.
(3) Analogy-based Generalizations
Generate the plausible upper bound condition of the rule R as the maximal generalization of I in
the context of the agent’s ontology. Generate the plausible lower bound condition of the rule R as
the minimal generalization of I, which does not contain any specific instance.
(4) Rule Analysis
If there is any variable from the THEN part of a rule that is not linked to some variable from the
other parts of the rule, or if the rule has too many instances in the knowledge base, then interact
with the expert to extend the explanation of the example and update the rule if new explanation
pieces are found.
(5) Determining the Generalized Example
Generate a minimal generalization of the example E and its explanation EX that does not contain
any specific instance, and associate it with the learned rule.

understanding, which is part of the first phase. Then we will present and justify the
generalization method, which is based on analogical reasoning.

9.5 MIXED-INITIATIVE EXAMPLE UNDERSTANDING

9.5.1 What Is an Explanation of an Example?

The mixed-initiative example understanding process is illustrated in Figure 9.10. The
expert has defined the example during the modeling process. Now the expert and the
agent have to collaborate to learn a general rule from this specific example.
The first step consists in finding the set of relationships from the agent’s ontology
that collectively explain why the example is correct. In general, these relation-
ships express the meaning of the question/answer pair from the example, as well
as other conditions that need to be satisfied by the instances and constants from
the example.

13:59:50,
.010
9.5. Mixed-Initiative Example Understanding 261

Consider the following question/answer pair:

Is Bob Sharp interested in an area of expertise of John Doe?

Yes, Artificial Intelligence.

The following is an approximate representation of the meaning of the question/answer

pair using the terms from the agent’s ontology:

Bob Sharp is interested in Artificial Intelligence

John Doe is expert in Artificial Intelligence

Adding “Probability of solution is always certain” to the preceding relationships results in

the explanation of the example, which is shown on the left-hand side of Figure 9.10. This is
a formal explanation of the example and consists of a set of relationships (or explanation
pieces) involving the instances and the constants from the example.
One can distinguish between different types of explanation pieces, as shown in
Table 9.3. Formally, the explanation of the example consists of several explanation pieces.
Each explanation piece corresponds to a path in the ontology between an object in the
example and another object or constant, as illustrated in Figure 9.10. In principle, the path
could have any length. In practice, however, one has to limit the maximum length of the
path for a certain type of explanation, in order to reduce the combinatorial explosion in
the generation of plausible explanations by the agent.

Table 9.3 Types of Explanation Pieces

Association – a relation between two objects from the example

Correlation – a common feature of two objects from the example

Property – a property of an object from the example

Relation – a relation between an object from the example and one from the knowledge base

Generalization – a generalization of an object from the example

Specific value – a specific value for an entity from the example (constant or generic instance)

Bob Sharp John Doe

is interested in is expert in

Artif icial Intelligence

Probability of solution
is always certain

Figure 9.10. Example understanding.

13:59:50,
.010
 262 Chapter 9. Rule Learning

Because the agent’s ontology is incomplete, sometimes the explanation includes only
an approximate representation of the meaning of the question/answer (natural language)
sentences.
The next section presents the explanation generation method.

9.5.2 Explanation Generation

Fully automatic explanation generation by the agent, while highly desirable, is problematic
because it would require human-like natural language understanding capabilities from an
agent that has only an incomplete ontology.
Manual explanation definition by an expert who is not a knowledge engineer is also
problematic. For one thing, he or she would need to use the formal language of the agent.
But this would not be enough. The expert would also need to know the names of the
potentially many thousands of concepts and features from the agent’s ontology (such as “is
interested in”).
While defining the formal explanation of a reduction step is beyond the individual
capabilities of both the expert and the agent, it is not beyond their joint complementary
capabilities. Therefore, finding such explanation pieces is modeled as a mixed-initiative
process (Tecuci et al., 2007a), based on the following observations:

It is easier for an expert to understand sentences in the formal language of the agent
than it is to produce such formal sentences

It is easier for the agent to generate formal sentences than it is to understand sentences
in the natural language of the expert

In essence, the agent will use basic natural language processing, various heuristics,
analogical reasoning, and help from the expert in order to identify and propose a set of
plausible explanation pieces, ordered by their plausibility of being correct explanations.
Then the expert will select the correct ones from the generated list.
The left-hand side of Figure 9.11 shows an example to be understood, and the upper-
right part of Figure 9.11 shows all the instances and constants from the example. The agent
will look for plausible explanation pieces of the types from Table 9.3, involving those
instances and constants. The most plausible explanation pieces identified, in plausibility

Figure 9.11. Explanation generation.

13:59:50,
.010
9.5. Mixed-Initiative Example Understanding 263

order, are shown in the bottom-right of Figure 9.11. Notice that the two most plausible
explanation pieces from Figure 9.11 are the correct explanation pieces shown in Figure 9.10.
The expert will have to select each of them and click on the Accept button. As a result, the
agent will move them in the Explanations pane from the left side of Figure 9.11.
Notice in the upper-right of Figure 9.11 that all the objects and constants from the example
are selected. Consequently, the agent generates the most plausible explanations pieces
related to all these objects and displays those with the highest plausibility. The expert may
click on the See More button, asking the agent to display the next set of plausible explanations.
The expert may also deselect some of the objects and constants, asking the agent to
generate only plausible explanations involving the selected elements. For example,
Figure 9.12 illustrates a situation where only the constant “certain” is selected. As a result,
the agent generated only the explanation “The value is specifically certain,” which means
that this value should be kept as such (i.e., not generalized) in the learned rule.
The expert may also provide a new explanation, even using new instances, concepts, or
features. In such a case, the expert should first define the new elements in the ontology.
After that, the expert may guide the agent to generate the desired explanations.
If the example contains any generic instance, such as “Artificial Intelligence,” the agent
will automatically select the explanation piece “Artificial Intelligence is Artificial Intelligence”
(see Explanation pane on the left side of Figure 9.11), meaning that this instance will
appear as such in the learned rule. If the expert wants Artificial Intelligence to be general-
ized, he or she should simply remove that explanation by clicking on it and on the Remove
button at its right.
The expert may also define explanations involving functions and comparisons, as will
be discussed in Sections 9.12.3 and 9.12.4.
Notice, however, that the explanation of the example may still be incomplete for at least
three reasons:

The ontology of the agent may be incomplete, and therefore the agent may not be able
to propose all the explanation pieces of the example simply because they are not
present in the ontology

1. Select an object

2. Click on “Search”

3. The agent generates

plausible explanations

Figure 9.12. Guiding explanation generation.

13:59:50,
.010
 264 Chapter 9. Rule Learning

The agent shows the plausible explanation pieces incrementally, as guided by the
expert, and if one of the actual explanation pieces is not among the first ones shown, it
may not be seen and selected by the expert

It is often the case that the human expert forgets to provide explanations that corres-
pond to common-sense knowledge that also is not represented in the question/
answer pair

The incompleteness of the explanation is not, however, a significant problem because the
explanation may be further extended during the rule refinement process, as discussed
in Chapter 10.
To conclude, Table 9.4 summarizes the mixed-initiative explanation generation method.
Once the expert is satisfied with the identified explanation pieces, the agent will generate the
rule, as discussed in the following sections.

9.6 EXAMPLE REFORMULATION

As indicated in Table 9.2 (p. 260), once the explanation of the example is found, the agent
generates a very specific IF-THEN rule with an applicability condition that covers only that
example. The top part of Figure 9.13 shows an example, and the bottom part shows the
generated specific rule that covers only that example. Notice that each instance (e.g., Bob
Sharp) and each constant (e.g., certain) is replaced with a variable (i.e., ?O1, ?Sl1).
However, the applicability condition restricts the possible values of these variables to
those from the example (e.g., “?O1 is Bob Sharp”). The applicability condition also includes
the properties and the relationships from the explanation. Therefore, the rule from the
bottom of Figure 9.13 will cover only the example from the top of Figure 9.13. This rule will
be further generalized to the rule from Figure 9.8, which has a plausible upper bound
condition and a plausible lower bound condition, as discussed in the next section. In
particular, the plausible upper bound condition will be obtained as the maximal general-
ization of the specific condition in the context of the agent’s ontology. Similarly, the
plausible lower bound condition will be obtained as the minimal generalization of the
specific condition that does not contain any specific instance.

Table 9.4 Mixed-Initiative Explanation Generation

Let E be an example.
Repeat

The expert focuses the agent’s attention by selecting some of the instances and constants from
the example.

The agent proposes what it determines to be the most plausible explanation pieces related to
the selected entities, ordered by their plausibility.

The expert chooses the relevant explanation pieces.

The expert may ask for the generation of additional explanation pieces related to the selected
instances and constants, may select different ones, or may directly specify explanation pieces.

until the expert is satisfied with the explanation of the example.

13:59:50,
.010
 9.7. Analogy-based Generalization 265

Example

Specific rule

IF ?O1 is interested in an area of expertise of ?O2.

Q: Is ?O1 interested in an area of expertise of ?O2 ?

A: Yes,?O3.
Condition
?O1 is Bob Sharp
is interested in ?O3
?O2 is John Doe
is expert in ?O3
?O3 is Artificial Intelligence
?SI1 is-exactly certain

THEN It is ?SI1 that ?O1 is interested in an area of expertise of ?O2.

Figure 9.13. Specific rule covering only the example.

9.7 ANALOGY-BASED GENERALIZATION

9.7.1 Analogical Problem Solving Based on Explanation Similarity

The goal of the learning agent is to learn a general rule that will enable it to solve new
problems by analogy with the example provided by the expert. As discussed in Section
8.1.4, analogy involves mapping some underlying “causal network of relations” between a
source and a target, with the goal of transferring knowledge from the source to the target.
This applies to analogical problem solving as illustrated in Figure 9.14.
The explanation in the upper-left part of Figure 9.14 explains why the input example
in the lower-left part is correct. The expression from the upper-right part of Figure 9.14 is
similar with this explanation because both have the same structure. Therefore, one may
infer, by analogy, that this expression explains an example that is similar to the initial
example. This example is shown in the bottom-right part of Figure 9.14.
The more general question is: When should we consider that two expressions or explan-
ations are similar? In general, one considers that two expressions, explanations, or situ-
ations are similar if they match within a certain predefined threshold (Winston, 1980). As
shown in Kodratoff and Ganascia (1986), generalization may be reduced to structural
matching. Therefore, we may consider that two expressions (explanations) are similar if

13:59:50,
.010
 266 Chapter 9. Rule Learning

explanation similar explanation

Bob Sharp John Doe Peter Jones Dan Smith

is interested in is expert in similar is interested in is expert in

Artificial Intelligence Information Security

Probability of solution is always certain Probability of solution is always certain

explains explains?
initial example similar example
I need to I need to
Bob Sharp is interested in an area of Peter Jones is interested in an area of
expertise of John Doe. expertise of Dan Smith.
similar
Therefore I conclude that Therefore I conclude that
It is certain that Bob Sharp is interested It is certain that Peter Jones is interested
in an area of expertise of John Doe. in an area of expertise of Dan Smith.

Figure 9.14. Analogical problem solving based on explanation similarity.

they are both less general then a given expression that represents the analogy criterion.
Consequently, the preceding question may be rephrased as:
Given the explanation EX of an example E, which generalization of EX should be
considered an analogy criterion, enabling the agent to generate reductions that are analo-
gous to E?
There are two interesting answers of this question, one given by a cautious learner, and
the other given by an aggressive learner, as discussed in the next sections.

9.7.2 Upper Bound Condition as a Maximally General

Analogy Criterion
An aggressive learner always prefers maximal generalizations, as discussed in Section 8.5.
Consequently, it will determine the analogy criterion as a maximal generalization of the
explanation, as illustrated in Figure 9.15 and explained in the following.
The explanation of the initial example corresponds to the applicability condition of the
specific rule representing the example (see Figure 9.13). Thus the aggressive learner will
learn a rule by maximally generalizing the specific applicability condition, as illustrated in
the upper right side of Figure 9.16.
The specific instance Bob Sharp is generalized to the most general concept from the
ontology that covers it and satisfies all the constraints imposed by the domains and the
ranges of the features related to it in the specific condition (explanation). As shown in the
left-hand side of Figure 9.16, the most general generalization of Bob Sharp in the ontology
is object. On the other hand, in the specific condition (explanation), Bob Sharp has the
feature “is interested in,” the domain of which is person. Thus the maximal generalization
of Bob Sharp is: object \ domain(is interested in) = object \ person = person.
In general, if an instance in the explanation has the features f11, . . . , f1m, and appears as
value of the features f21, . . . , f2n, then its maximal general generalization is:

object \ domain(f11) \ . . . \ domain(f1m) \ range(f21) \ . . . \ range(f2n)

13:59:50,
.010
9.7. Analogy-based Generalization 267

Analogy person person

criterion
area of
instance of expertise instance of
?O1 ?O2
instance of
is interested in is expert in
?O3
Probability of solution is always certain

less general than less general than

explanation similar explanation
Bob Sharp John Doe Peter Jones Dan Smith

is interested in is expert in similar is interested in is expert in

Artificial Intelligence Information Security

Probability of solution is always certain Probability of solution is always certain

Figure 9.15. Upper bound condition as a maximally general analogy criterion.

object
Most general generalization
?O1 is person
actor area of UB is interested in ?O3
expertise ?O2 is person
person UB
is expert in ?O3
subconcept of
?O3 is area of expertise
employee student ?SI1 is in [certain – certain]
Computer
Science Maximal
university employee generalization
instance of Specif ic condit ion
faculty member graduate ?O1 is Bob Sharp
student Artificial is interested in ?O3
subconcept of
Intelligence LB ?O2 is John Doe
professor PhD advisor ?O3 is expert in ?O3
LB ?O3 is Artificial Intelligence
subconcept of subconcept of
is expert in ?SI1 is exactly certain
associate LB PhD student LB domain person
professor range area of expert ise
instance of instance of is interested in
John Doe Bob Sharp domain person
range area of expertise
?O2 ?O1
Figure 9.16. Maximal generalization of the specific applicability condition.

Now consider John Doe. Its most general generalization is object \ domain(is expert in) =
object \ person = person.
Consider now Artificial Intelligence. It appears as a value of the features “is interested in”
and “is expert in.” Therefore, its maximal generalization is: object \ range(is interested in) \
range(is expert in) = object \ area of expertise \ area of expertise = area of expertise.

13:59:50,
.010
 268 Chapter 9. Rule Learning

On the other hand, the maximal generalization of “certain” is the interval with a single
value “[certain – certain]” because ?Sl1 is restricted to this value by the feature “is exactly.”
Let us consider again the example from the top part of Figure 9.13, but now let us
assume that “The value is specifically certain” was not identified as an explanation piece.
That is, the explanation of the example consists only of the following pieces:

Bob Sharp is interested in Artificial Intelligence

John Doe is expert in Artificial Intelligence

In this case, the generated specific condition is the one from the bottom of Figure 9.17, and
its maximal generalization is the one from the top of Figure 9.17. Notice that the maximal
generalization of “certain” is the entire interval [no support – certain] because there is no
restriction on the possible values of ?Sl1.

9.7.3 Lower Bound Condition as a Minimally General

Analogy Criterion
A cautious learner prefers minimal generalizations, as discussed in Section 8.5. Conse-
quently, it will determine the analogy criterion as a minimal generalization of the explan-
ation, as illustrated in Figure 9.18.
The explanation of the initial example corresponds to the applicability condition of the
specific rule representing the example (see Figure 9.13). Thus the cautious learner will
learn a rule by minimally generalizing the specific applicability condition, as illustrated in
Figure 9.19.
The specific instance Bob Sharp is minimally generalized to the most specific concept
from the ontology that covers it, that is, to PhD student. This generalization is allowed
because the domain of the feature “is interested in” is person, which includes PhD student.

Most general generalization

?O1 is person
is interested in ?O3
?O2 is person
is expert in ?O3
?O3 is area of expertise
?SI1 is in [no support – certain]
Maximal
generalization
Specific condition
?O1 is Bob Sharp
is interested in ?O3
?O2 is John Doe
is expert in ?O3
?O3 is Artificial Intelligence
?SI1 is certain

Figure 9.17. Maximal generalization of a symbolic probability value when no explanation is identified.

13:59:50,
.010
9.7. Analogy-based Generalization 269

Analogy associate
criterion PhD student PhD advisor
professor
instance of instance of
Artificial Intelligence
?O1 ?O2
is interested in is is expert in
?O3
Probability of solution is always certain

less general than less general than

explanation similar explanation
Bob Sharp John Doe Peter Jones Dan Smith

is interested in is expert in similar is interested in is expert in

Artificial Intelligence Information Security

Probability of solution is always certain Probability of solution is always certain

Figure 9.18. Lower bound condition as a minimally general analogy criterion.

Most specific generalization

?O1 is PhD student
is interested in ?O3
?O2 is (PhD advisor, associate professor)
is expert in ?O3
?O3 is Artificial Intelligence
?SI1 is in [certain – certain]
Minimal
generalization
Specific condition
?O1 is Bob Sharp
is interested in ?O3
?O2 is John Doe
is expert in ?O3
?O3 is Artificial Intelligence
?SI1 is exactly certain

Figure 9.19. Minimal generalization of the specific applicability condition.

Similarly, the specific instance John Doe is minimally generalized to PhD advisor or
associate professor, because these are the minimal generalizations of John Doe and neither
is more specific than the other. Additionally, both these concepts are subconcepts of
person, the domain of is expert in.
Because Artificial Intelligence is a generic instance, it can appear in the learned rule (as
opposed to the specific instances Bob Sharp and John Doe). Therefore, its minimal general-
ization is Artificial Intelligence itself. Similarly, the constants (such as certain) can also appear
in the learned rule, and they are kept as such in the minimal generalization.

13:59:50,
.010
 270 Chapter 9. Rule Learning

Notice that if you want an instance to appear in the condition of a learned rule, it needs
to be defined as a generic instance. Specific instances are always generalized to concepts
and will never appear in the condition.

9.8 RULE GENERATION AND ANALYSIS

The partially learned rule is shown in the bottom part of Figure 9.8 (p. 259). Notice that the
features are listed only once under the bounds because they are the same for both bounds.
The generated rule is analyzed to determine whether there are any variables in the
THEN part that are not linked to some variable from the IF part. If such an unlinked
variable exists, then it can be instantiated to any value, leading to solutions that make no
sense. Therefore, the agent will interact with the expert to find an additional explanation
that will create the missing link and update the rule accordingly.
The generated rule is also analyzed to determine whether it has too many instances in
the knowledge base, which is also an indication that its explanation is incomplete and
needs to be extended.

9.9 GENERALIZED EXAMPLES

The rule learned from an example and its explanation depends on the ontology of the
agent at the time the rule was generated. If the ontology changes, the rule may need to be
updated, as will be discussed in Chapter 10. For example, the minimal generalization of a
specific instance will change if a new concept is inserted between that instance and the
concept above it. To enable the agent to update its rules automatically when relevant
changes occur in the ontology, minimal generalizations of the examples and their explan-
ations are associated with the learned rules.
Why is the agent maintaining minimal generalizations of examples instead of the
examples themselves? Because the examples exist only in Scenario KBs, where the specific
instances are defined, while the rules are maintained in the Domain KB. If a scenario is no
longer available, the corresponding examples are no longer defined. However, generalized
examples (which do not contain specific instances) will always be defined in the Domain
KB. Thus the generalized examples represent a way to maintain a history of how a rule was
learned, independent of the scenarios. They are also a compact way of preserving this
history because one generalized example may correspond to many actual examples.
Figure 9.20 shows the minimal generalization of the example and its explanation from
which the rule in Figure 9.8 (p. 259) was learned.
One should notice that the minimal generalization of the example shown at the top part
of Figure 9.20 is not the same with the plausible lower bound condition of the learned rule
from Figure 9.8. Consider the specific instance John Doe from the example.
In the ontology, John Doe is both a direct instance of PhD advisor and of associate
professor (see Figure 9.16). In the lower bound condition of the rule, John Doe is general-
ized to PhD advisor or associate professor, indicated as (PhD advisor, associate professor),
because each of these two concepts is a minimal generalization of John Doe in the
ontology. Thus the agent maintains the two concepts as part of the lower bound of the
rule’s version space: one corresponding to PhD advisor, and the other corresponding to

13:59:50,
.010
 9.10. Hypothesis Learning 271

Generalized Example
?O1 is PhD student
is interested in ?O3
?O2 is PhD advisor
is associate professor
is expert in ?O3
?O3 is Artificial Intelligence
?SI1 is in [certain - certain]
Covered positive examples: 1
Covered negative examples: 0

Minimal example
generalization
Example and its explanation
?O1 is Bob Sharp
is interested in ?O3
?O2 is John Doe
is expert in ?O3
?O3 is Artificial Intelligence
?SI1 is exactly certain

Figure 9.20. Determining the generalized example.

associate professor. During further learning, the agent will choose one of these generaliza-
tions or a more general generalization that covers both of them.
In the minimal generalization of the example, John Doe is generalized to PhD advisor
and associate professor because this is the best representation of the minimal generaliza-
tion of the example that can be used to regenerate the rule, when changes are made to the
ontology. This minimal generalization is expressed as follows:
?O2 is PhD advisor
is associate professor

Initially, the generalized example shown at the top of Figure 9.20 covers only one
specific example. However, when a new (positive or negative) example is used to refine
the rule, the agent checks whether it is already covered by an existing generalized example
and records this information. Because a generalized example may cover any number of
specific positive and negative examples, its description also includes the number of
specific examples covered, as shown in the top part of Figure 9.20.
Cases of using generalized examples to regenerate previously learned rules are pre-
sented in Section 10.2.

9.10 HYPOTHESIS LEARNING

In addition to learning a general reduction rule from a specific reduction example, Disciple-
EBR also learns general hypotheses (or problems). The left-hand side of Figure 9.21 shows

13:59:50,
.010
 Hypothesis Generalizaon

1. Modeling 2. Learning

Agent learns general

rule and hypotheses
Reducon Rule

Expert explains how to

solve a specific problem

Reducon Example
.010
13:59:50,

Hypothesis Generalizaon Hypothesis Generalizaon Hypothesis Generalizaon

Figure 9.21. A reduction rule and four hypotheses learned from a specific hypothesis reduction.
9.10. Hypothesis Learning 273

an example of a reduction where a hypothesis is reduced to three subhypotheses.

From this example, Disciple-EBR learns five general knowledge pieces that share the
same variables:

A general hypothesis reduction rule that is a generalization of the reduction example

where each instance was replaced with a variable

A generalization of the top hypothesis with a plausible version space condition

Generalizations of the three subhypotheses, each with its own plausible version space
condition

The hypothesis learning method, shown in Table 9.5, is very similar with the rule-learning
method.
Figure 9.22 illustrates the automatic learning of a general hypothesis from the specific
hypothesis, “John Doe would be a good PhD advisor for Bob Sharp,” when no explanation is
provided.
The specific instances, John Doe and Bob Sharp, are replaced with the variables ?O1 and
?O2, respectively, as in the reduction rule.
The lower bounds of these variables are obtained as the minimal generalizations of
John Doe and Bob Sharp, according to the agent’s ontology from the left-hand side of
Figure 9.22, because both of them are specific instances. Notice that there are two minimal
generalizations of John Doe: PhD advisor and associate professor. The minimal generaliza-
tion of Bob Sharp is PhD student.
The upper bounds are obtained as the maximum generalizations of John Doe and
Bob Sharp, according to the agent’s ontology from the left-hand side of Figure 9.22. They
are both object.
During the explanation generation process, the user may wish to restrict the general-
ization of the hypothesis by providing the following explanations:

Table 9.5 Basic Steps of the Hypothesis Learning Method

(1) Mixed-Initiative Understanding (Explanation Generation)

Through a mixed-initiative interaction with the subject matter expert, determine the set of
constraints EX that need to be satisfied by the instances and constants from the example
hypothesis EH in order for the hypothesis statement to make sense. These include the types of the
instances from the hypothesis.
(2) Hypothesis Reformulation
Use the constraints EX and the variables generated during the learning of the corresponding
reduction rule to create an instance I of the concept C representing the applicability condition of
the hypothesis H to be learned. Then reformulate the specific hypothesis as a hypothesis pattern
with I as its applicability condition.
(3) Analogy-based Generalizations
Generate the plausible upper bound condition of the general hypothesis H as the maximal
generalization of I in the context of the agent’s ontology. Generate the plausible lower bound
condition of the hypothesis H as the minimal generalization of I that does not contain any specific
instance.
(4) Determining the Generalized Example
Generate a minimal generalization of the example hypothesis EH and its explanation EX that does
not contain any specific instance and associate it with the learned hypothesis.

13:59:50,
.010
274 Chapter 9. Rule Learning

object UB
Most specific generalization Most general generalization
?O1 is (PhD advisor, associate professor) ?O1 is object
actor
?O2 is PhD student ?O2 is object

person Min Max

generalization generalization
employee student Specific condition
?O1 is John Doe
university employee ?O2 is Bob Sharp

faculty member graduate Hypothesis Example

student
subconcept of

professor PhD advisor LB Hypothesis Generalization

subconcept of
subconcept of

associate
LB PhD student LB
professor

instance of instance of

John Doe Bob Sharp

?O1 ?O2

Figure 9.22. Generation of the applicability conditions of a hypothesis.

John Doe is faculty member

Bob Sharp is student

In this case, the lower bound condition of the learned hypothesis remains the same, but
the upper bound condition becomes:

?O1 is faculty member

?O2 is student

In general, hypothesis learning is done automatically, as a byproduct of rule learning, but

the explanations found for learning the reduction rule will not be used for learning the
hypothesis. Indeed, these explanations indicate only why this particular reduction is
correct. There may be other reductions of the same hypothesis, which will have different
explanations. If desired, the user may provide specific explanations for the hypothesis that
limit the possible instantiations of the learned hypothesis to those that make sense, as
illustrated previously.
Notice that for each specific hypothesis, the agent may learn a single general hypothesis
but several reduction rules. Then, when trying to reduce a hypothesis to simpler hypoth-
eses, the agent checks that both the applicability condition of the corresponding general
hypothesis and that of the considered rule are satisfied. It is therefore important that the
applicability condition of the general hypothesis is general enough to enable the use of all
the learned reduction rules associated with that general hypothesis.
A general hypothesis may be learned in several circumstances. As illustrated in
Figure 9.21, a hypothesis may be learned as a byproduct of rule learning. It may also be
learned during the modeling and formalization of a reasoning tree by using the Learn

13:59:50,
.010
 9.11. Hands On: Rule and Hypotheses Learning 275

Hypothesis Pattern or the Learn Tree Patterns commands that were introduced in
Section 4.10. Finally, it may be learned by specifically invoking hypothesis (problem)
learning when working with the Mixed-Initiative Reasoner. But it is only this last situation
that also allows the definitions of explanations, as will be discussed later in this section. In
all the other situations, a hypothesis is automatically learned, with no explanations, as was
illustrated in Figure 9.22.
The overall user–agent interactions during the hypothesis explanation process are
illustrated in Figure 9.23 and described in Operation 9.1. Here it is assumed that the
reasoning tree was already formalized and thus a hypothesis pattern was already learned.
If it was not learned, the pattern will be automatically learned before the explanations are
identified.

Operation 9.1. Define explanations for a hypothesis

Select the Scenario workspace.

Click on the Reasoning menu and select Mixed-Initiative Reasoner.

Select the hypothesis to analyze, and click on the Select button on the right side of
the window.

In the Reasoning Hierarchy viewer of the Mixed-Initiative Reasoner, select a hypoth-
esis by clicking on it.

In the right panel, click on the Modify Explanations button.

Explanation generation is automatically invoked, the left pane showing the hypothesis
in the Reasoning Step viewer, and the right pane showing the selected entities and a
corresponding ordered list of plausible explanations (see the upper-right part of
Figure 9.23).

Inspect the explanation pieces proposed by the agent in the right panel, click on a
correct explanation piece (if any), and then click on the Accept button.

Repeat explanation selection and acceptance to select all the needed explanation pieces.

If necessary, click on the See More button, asking the agent to generate additional
possible explanation pieces related to the selected entities.

If necessary, select a subset of instances and constants and click on the Search button,
asking the agent to generate explanations related to the selected entities.

All the accepted explanations are added to the hypothesis (see the bottom-left part of
Figure 9.23). You may delete any of them by clicking on it and then clicking on the
Remove button.

Click on the Finish button to end the learning of the current hypothesis.

9.11 HANDS ON: RULE AND HYPOTHESES LEARNING

This case study will guide you to use Disciple-EBR to learn rules and hypotheses from
examples. More specifically, you will learn how to:

Invoke the rule learning module

Interact with the agent to find the explanation of the example

Learn the rule (and the associated hypotheses)

The overall user–agent interactions during the rule- (and hypotheses-) learning process
are illustrated in Figure 9.24 and described in Operation 9.2. It is assumed that the

13:59:50,
.010
 1. In “Reasoning Hierarchy” (a) Select Hypothesis 4. In “Reasoning Step” (b) Accept Explanation Pieces

6. Click on “Accept”
5. Select a relevant
3. Click on explanation piece
“Modify Explanations”
2. Click on a
hypothesis to select it

(d) Finish Hypothesis Learning (c) Accept Explanation Pieces

.010
13:59:50,

7. Repeat steps 5 and 6 to select

8. The accepted explanation pieces all the needed explanation pieces.
were added to the example 9. Click on “Finish” to end the
hypothesis learning process

Figure 9.23. Overview of the user–agent interactions during the hypothesis explanation process.
 1. In “Reasoning Hierarchy” 4. In “Reasoning Step” (b) Accept Explanation Pieces
(a) Select Reasoning Step

6. Select a relevant
explanation piece 7. Click on “Accept”
3. Click on
2. Click on question/answer “Learn Condition” 5. If necessary, select and remove any
to select the reasoning step automatically selected explanation piece

(d) Finish Rules Learning (c) Accept Explanation Pieces

.010

8. Repeat steps 6 and 7 to select

13:59:50,

all the needed explanation pieces.

10. Click on “Finish” to end

the rule learning process
9. The accepted explanation pieces
were added to the example

Figure 9.24. Overview of the user–agent interactions during rule (and hypotheses) learning.
278 Chapter 9. Rule Learning

reasoning tree is formalized. If not, one can easily formalize it in the Evidence workspace
before invoking rule learning by simply right-clicking on the top node and selecting
Learn Tree.

Operation 9.2. Learn rule

Select the Scenario workspace.

Click on the Reasoning menu and select Mixed-Initiative Reasoner.

Select the hypothesis to analyze, and click on the Select button on the right side of
the window.

In the Reasoning Hierarchy viewer, select a reasoning step by clicking on the corres-
ponding Q/A node (see the upper-left part of Figure 9.24).

In the right panel, click on the Learn Condition button.

Explanation generation is automatically invoked, the left pane showing the reduction
in the Reasoning Step viewer, and the right pane showing the selected entities and a
corresponding ordered list of plausible explanations (see the upper-right part of
Figure 9.24).

If the example contains generic instances, check whether you want to remove any of
the corresponding explanations that were automatically accepted by the agent. To
remove such an explanation piece, click on it in the left panel and then click on the
Remove button.

Inspect the explanation pieces proposed by the agent in the right panel, click on a
correct explanation piece (if any), and then click on the Accept button.

Repeat explanation selection and acceptance to select all the needed explanation
pieces.

If necessary, click on the See More button, asking the agent to generate additional
possible explanation pieces related to the selected entities.

If necessary, select a subset of instances and constants and click on the Search button,
asking the agent to generate explanations related to the selected entities.

All the accepted explanations are added to the reduction (see the bottom-left part of
Figure 9.24). You may delete any of them by clicking on them and then clicking on the
Remove button.

Click on the Finish button to end the learning of the current rule.

Continue with selecting another reasoning step and learning the corresponding
rule.

You are now ready to perform a rule-learning case study. There are two of them, a shorter
one and a longer one. In the shorter case study, you will guide the agent to learn the rule
from Figure 9.8, as discussed in the previous sections. In the longer case study, you will
guide the agent to learn several rules, including the rule from Figure 9.8.
Start Disciple-EBR, select one of the case study knowledge bases (either “11-Rule-Learning-
short/Scen” or “11-Rule-Learning/Scen”), and proceed as indicated in the instructions at the
bottom of the opened window.
A learned rule can be displayed as indicated in the following operation.

Operation 9.3. Display a learned rule with the Rule Viewer

In the Scenario workspace, in the Reasoning Hierarchy viewer select a reasoning step
for which a rule has already been learned by clicking on the corresponding Q/A node.

At the bottom of the right panel, click on the Reduction Rule button to see the rule that
generated the reduction step.

13:59:50,
.010
 9.12. Explanation Generation Operations 279

At the top of the Rule Viewer, notice the name of the rule (e.g., DDR.00018). You may
also display or delete this rule with the Rule Browser, as described in Operations 10.4
and 10.5.

Click on the X button of the Rule Viewer to close it.

Operation 9.4. Display a learned hypothesis with the Problem Viewer

In the Scenario workspace, in the Reasoning Hierarchy viewer click on a hypothesis to
select one for which a general hypothesis has already been learned.

In the right panel, click on the View Problem button to see the learned hypothesis.

Click on the X button of the Problem Viewer to close it.

9.12 EXPLANATION GENERATION OPERATIONS

An important part of rule learning is finding an explanation of the example that is as

complete as possible. The following sections discuss various operations that support
this process.

9.12.1 Guiding Explanation Generation

You can guide the agent in explanation generation by selecting the instances and con-
stants from the example for which explanation pieces will be proposed. As a result, the
agent generates a list of potential explanations, ordered by their plausibility, from which
you can select the correct ones. The corresponding user–agent interaction is described in
Operation 9.5 and illustrated in Figure 9.25.

Operation 9.5. Guide the generation of explanations

In the Scenario workspace, during rule learning, in the Reasoning Hierarchy viewer,
select a reasoning step by clicking on the corresponding Q/A node.

In the right panel, click on the Learn Condition button.

1. Select an object

2. Click on “Search”

3. The agent generates

plausible explanations

Figure 9.25. Guiding explanation generation.

13:59:50,
.010
 280 Chapter 9. Rule Learning

Select one or several entities from the “Elements to search for” pane, such as certain in
Figure 9.25. You may also need to deselect some entities by clicking on them.

Click on the Search button, asking the agent to generate explanation pieces related to
the selected entities.

Select an explanation piece and click on the Accept button.

Click on the See More button to see more of the generated explanation pieces.

Repeat the preceding steps until all the desired explanations are generated and selected.

9.12.2 Fixing Values

You may block the generalizations of numbers (e.g., “5”), symbolic probability values (e.g.,
“certain”), or generic instances (e.g., “Artificial Intelligence”) by selecting explanations of the
following forms, as described in Operation 9.6:

The value is specifically 5

The value is specifically certain
Artificial Intelligence is Artificial Intelligence

Operation 9.6. Generate explanations with fixed values

In the Scenario workspace, during rule learning, in the Reasoning Hierarchy viewer,
select a reasoning step by clicking on the corresponding Q/A node.

In the right panel, click on the Learn Condition button.

Select one or several generic instances or constants (e.g., certain in Figure 9.25) in the
“Elements to search for” pane. No specific instances should be selected.

If a selected element is a constant, then the agent will generate a possible explanation
piece of the form “The value is specifically <selected constant>.”

If a selected element is a generic instance, then the agent will generate a possible
explanation piece of the form “<selected instance> is <selected instance>,” as well as
other explanations containing the selected instance.

Accept the explanation “The value is specifically <selected constant>” (or “<selected
instance> is <selected instance>”) to fix that value in the learned rule.

Notice that these types of explanation pieces are generated when only constants or generic
instances are selected in the “Elements to search for” pane. For example, only “certain”
was selected in the “Elements to search for” pane in the upper-right part of Figure 9.25,
and therefore the potential explanation piece “The value is specifically certain” was
generated.
Notice also that explanations such as “Artificial Intelligence is Artificial Intelligence” can be
generated only for generic instances. They cannot be generated for specific instances
because these instances are always generalized in the learned rules.

9.12.3 Explanations with Functions

We will illustrate these explanations based on the reasoning tree in Figure 9.26. Notice that
all the numbers are correctly recognized as numbers and appear in green.
The Q/A pair from the first reduction explains how the price of the Apple iPad 16GB is
computed. The corresponding explanation pieces express the meaning of the Q/A pair by

13:59:50,
.010
9.12. Explanation Generation Operations 281

Figure 9.26. Reasoning tree involving numbers.

The function is defined with

the Expression Editor, using
numbers from the example

Figure 9.27. Explanations with functions.

using relationships from the agent’s ontology. They are the following ones, and are shown
also in the left-hand side of Figure 9.27:

Apple iPad 16GB has as price 495.0

MicroCenter Fairfax is located in Virginia
Virginia has as state sale tax 0.05

13:59:50,
.010
282 Chapter 9. Rule Learning

Additionally, you have to teach the agent how the price is actually computed. You invoke
the Expression Editor by clicking on the Edit Expression button, which displays a pane to
define the expression (see the bottom right of Figure 9.27). Then you fill in the left side of
the equality with the price, and the right side with the expression that leads to this price, by
using the numbers from the example:

519:75 ¼ 495:0  0:05 þ 495:0

Each of these numbers is generalized to a variable, and a general relationship between

these variables is learned (i.e., “?N1 = ?N2 * ?N3 + ?N2”). The learned rule, containing this
expression in its applicability condition, is shown in Figure 9.28.
In general, to define an explanation containing a function, you have to follow the steps
from Operation 9.7.

Figure 9.28. Learned rule with a learned function in the applicability condition.

13:59:50,
.010
 9.12. Explanation Generation Operations 283

Operation 9.7. Define explanations containing functions

In the Scenario workspace, during rule learning, in the Reasoning Hierarchy viewer,
select a reasoning step by clicking on the corresponding Q/A node.

In the right panel, click on the Learn Condition button.

In the “Possible explanations” pane, click on the Edit Expression button.

The agent displays the Expression Editor.

In the left part of the Expression Editor, start typing the result of the expression (e.g., the
actual price of “519.75” shown in Figure 9.27) and select it from the completion pop-up.

In the right part of the Expression Editor, type the expression used to compute the
result (i.e., “495.0 * 0.05 + 495.0”), making sure you select the values from the completion
pop-up.

Click on the OK button under the Expression Editor to define the expression.

9.12.4 Explanations with Comparisons

We will illustrate these explanations based on the reasoning tree from Figure 9.26.
Notice that the Q/A pair corresponding to the bottom reduction explains why Mike has
enough money to buy the Apple iPad 16GB. The following is the corresponding explanation
piece:

Mike has as available funds 650.35

Additionally, you have to indicate that 650.35 is greater than 519.75. You click on the
Create New… button, which opens a window allowing you to define a new explanation as
an object-feature-value triplet (see the bottom of Figure 9.29). In the left editor, you
start typing the amount of money Mike has (i.e., 650.35) and select it from the completion

The comparison is defined with

the Create Explanation Editor,
using numbers from the example

Figure 9.29. Defining explanations with comparisons.

13:59:50,
.010
284 Chapter 9. Rule Learning

pop-up. In the center editor, you type >=. Then, in the right editor, you start typing the
actual cost (519.75) and select it from the completion pop-up. Finally, you click on the OK
button in the Create explanation window to select this explanation:

650:35 >¼ 519:75

Each of these numbers is generalized to a variable, and a general relationship between

these variables is learned (i.e., “?N2 >= ?N1”). The learned rule, containing this relation-
ship in its applicability condition, is shown in Figure 9.30.
In general, to define an explanation containing a comparison you have to follow the
steps from Operation 9.8.

Operation 9.8. Define explanations containing comparisons

In the Scenario workspace, during rule learning, in the Reasoning Hierarchy viewer,
select a reasoning step by clicking on the corresponding Q/A node.

In the right panel, click on the Learn Condition button.

In the “Possible explanations” pane, click on the Create New… button.

The agent opens a window for defining an explanation as an object-feature-value
triplet (see the bottom of Figure 9.29).

In the left editor, type the number corresponding to the left side of the comparison
(e.g., “650.35” in Figure 9.29) and select it from the completion pop-up.

Figure 9.30. Learned rule with a condition containing a comparison.

13:59:50,
.010
 9.13. Guidelines for Rule and Hypothesis Learning 285

In the center editor, type the comparison operator (<, <=, =, !=, >=, or >).

In the right editor, type the number corresponding to the right side of the comparison.

Click on the OK button in the Create explanation window to accept the explanation.

9.12.5 Hands On: Explanations with Functions and Comparisons

The objective of this case study is to learn how to define explanations involving functions
(as discussed in Section 9.12.3) and comparisons (as discussed in Section 9.12.4).
Start Disciple-EBR, select the case study knowledge base “12-Explanations/Scen,” and
proceed as indicated in the instructions at the bottom of the opened window.

9.13 GUIDELINES FOR RULE AND HYPOTHESIS LEARNING

Guideline 9.1. Properly identify all the entities in the example before
starting rule learning
Before starting rule learning, make sure that all the elements are properly recognized as
instances, numbers, symbolic intervals, or strings. This is important because only the
entities with one of these types will be replaced with variables as part of rule learning, as
shown in the top part of Figure 9.31. Recognizing concepts is also recommended, but it is
optional, since concepts are not generalized. However, recognizing them helps the agent
in explanation generation.
Notice the case from the middle part of Figure 9.31. Because “the United States”
appears as text (in black) and not as instance (in blue), it will not be replaced with a
variable in the learned rule. A similar case is shown at the bottom of Figure 9.31. Because
600.0 is not recognized as number (in green), it will appear as such in the learned rule,
instead of being generalized to a variable.

concept instance Learning concept variable

The price of Apple iPad is under The price of ?O1 is under

600.0 dollars in the United States. ?N1 dollars in ?O2.

number instance variable variable

The price of Apple iPad is under The price of ?O1 is under

600.0 dollars in the United States. ?N1 dollars in the United States.

text (not recognized as an instance) same text (not generalized)

The price of Apple iPad is under The price of ?O1 is under

600.0 dollars in the United States. 600.0 dollars in the United States.

text (not recognized as a number) same text (not generalized)

Figure 9.31. Variables generation during rule learning.

13:59:50,
.010
286 Chapter 9. Rule Learning

Guideline 9.2. Avoid learning from examples that are too specific
It is important to teach the agent with good examples from which it can learn general
rules. A poor example is illustrated in the upper-left part of Figure 9.32. In this case, the
amount of money that Mike has is the same as the price of the Apple iPad 16GB. As a result,
both occurrences of 519.75 are generalized to the same variable ?N1, and the agent will
learn a rule that will apply only to cases where the amount of money of the buyer is exactly
the same as the price of the product (see the upper-right part of Figure 9.32).
You need instead to teach the agent with an example where the numbers are different,
such as the one from the bottom-left part of Figure 9.32. In this case, the agent will
generalize the two numbers to two different variables. Notice that the learned rule will
also apply to cases where ?N1 = ?N2.

Guideline 9.3. Use modeling-based ontology extension before

starting rule learning
Let us consider the reduction step from the left-hand side of Figure 9.33.
If you have used modeling-based ontology specification, then the ontology contains the
basic explanation piece of this reduction step:
John Doe has as number of publications 53

Therefore, before starting rule learning, review the modeling and check that you have
defined the features suggested by the Q/A pair in the ontology.
What are the other explanation pieces for this reduction step? You will need to define
two additional explanation pieces involving comparisons, as well as explanation pieces
fixing the values 41, 53, and very likely, as shown in the left-hand side of Figure 9.34. The
learned rule is shown in the right-hand side of Figure 9.34.

Guideline 9.4. Carefully define the domains and the ranges

of the features
Thoughtfully defined features significantly simplify and speed up the rule-learning process
because their domains and ranges are used in determining the minimal and maximal
generalizations of the examples, as discussed in Sections 9.7.2 and 9.7.3.

Does Mike have enough money Learning Does ?O1 have enough money
to buy Apple iPad 16GB? to buy ?O2?
Yes, because Mike has 519.75 Yes, because ?O1 has ?N1
dollars and Apple iPad 16GB dollars and ?O2
costs 519.75 dollars. costs ?N1 dollars.

Does Bob have enough money Does ?O1 have enough money
to buy Apple iPad 16GB? to buy ?O2?
Yes, because Bob has 620.25 Yes, because ?O1 has ?N1
dollars and Apple iPad 16GB dollars and ?O2
costs 519.75 dollars. costs ?N2 dollars.

Figure 9.32. Variable generation for identical entities.

13:59:50,
.010
9.13. Guidelines for Rule and Hypothesis Learning 287

Figure 9.33. Specification of the ontology knowledge for learning.

Select or define the necessary explanations

Figure 9.34. Another example of rule learning.

When you define a new feature, make sure you define both the domain and the range.
Do not leave the ones generated automatically. The automatically generated range (Any
Element) is too general, and features with that range are not even used in explanations.
Therefore, make sure that you select a more specific domain and range, such as a concept,
a number interval, or a symbolic interval.
If you have already defined facts involving that feature, you need to remove them first
before you can change the domain and the range of the feature.

13:59:50,
.010
288 Chapter 9. Rule Learning

Guideline 9.5. Provide hints to guide explanation generation

You may guide explanation generation by selecting one or several entities from the
example, asking the agent to propose only plausible explanation pieces involving those
entities, as was discussed in Section 9.12.1. This is a very effective way of reducing the
combinatorial explosion of explanation generation.

Guideline 9.6. Avoid learning rules without explanations

When you invoke Learn Condition and then Finish without selecting any explanation, the
agent automatically generates a plausible version space rule with a very general upper
bound condition. The rule is then automatically applied to reduce problems. There is a
danger that the rule will generate too many reductions, which may even block the
system. To alleviate this situation, the current implementation of Disciple-EBR uses the
following heuristic:
Generalization heuristic: When no explanation is provided for a reduction step, the
upper bound of the learned rule is obtained by generalizing each instance not to “object,” but
to a direct child of it.
Even with this heuristic, the plausible upper bound condition is still very general. To
alleviate this situation further, consider selecting explanations of the form “<specific
instance> is <concept>,” such as “John Doe is PhD advisor.” You can always provide
such explanations that limit the possible values in similar reductions to those that
make sense.
An additional explanation type to consider is “<generic instance> is <generic
instance>,” such as “Artificial Intelligence is Artificial Intelligence.” However, make sure that
such an explanation makes sense. Do you want “Artificial Intelligence” in all the similar
reductions? If yes, then this explanation is necessary. Otherwise, use “Artificial Intelligence
is <some concept>.”
Yet another explanation type to consider is, “The value is specifically <some value>,”
such as, “The value is specifically 5.0,” or, “The value is specifically certain.” However,
make sure that these explanations make sense. Do you want 5.0 in all the similar reduc-
tions? Do you want “certain” in all the similar reductions?

Guideline 9.7. Recognize concepts in the reasoning tree

When editing a node in the reasoning tree, Disciple-EBR proposes completions of the
current word based on the names of the instances, concepts, and symbolic probabilities
from the ontology. Selecting one of the proposed completions introduces it into the edited
statement with a color characteristic to its type (e.g., dark blue for concepts).
Even if the concepts were not recognized at the time of the development of the
reasoning tree, they can later be recognized automatically by right-clicking on the top
hypothesis and selecting Recognize Concepts in Tree. However, only the concepts that
appear exactly as in the ontology will be recognized, and only in the nodes that have not
yet been formalized.
Having the concepts recognized in the reasoning tree helps the agent in identifying the
relevant explanation pieces during the explanation generation process.

13:59:50,
.010
 9.15. Review Questions 289

Operation 9.9. Recognize concepts in the reasoning tree

In the Evidence workspace, click on the Reasoner menu at the top of the window.

In the right panel, right-click on the top hypothesis and select Recognize Concepts
in Tree.

9.14 PROJECT ASSIGNMENT 6

Learn rules from the reasoning trees developed in the previous project assignments.

9.15 REVIEW QUESTIONS

9.1. Define the rule-learning problem.

9.2. What are the basic steps of rule learning?

9.3. Consider the following expression, where both Jane Austin and Bob Sharp are
specific instances:

?O1 is Jane Austin

?O2 is Bob Sharp
has as advisor ?O1

Find its minimal generalization that does not contain any instance, in the context
of the ontological knowledge from Figure 9.35. Find also its maximal generalization.
9.4. Consider the following explanation of a reduction:

Dan Smith plans to retire from George Mason University.

(a) Reformulate this explanation as a concept with variables.

(b) Determine the minimal generalization of the concept, in the context of the
ontology from Figure 9.36, where all the instances are specific instances.
(c) Determine the maximal generalization of the concept.

9.5. Consider the following explanation of a reduction:

Jane Austin plans to move to Indiana University.

(a) Reformulate this explanation as a concept with variables.

(b) Determine the minimal generalization of the concept, in the context of the
ontology from Figure 9.36, where all the instances are specific instances.
(c) Determine the maximal generalization of the concept.

9.6. Consider the ontological knowledge from Figure 9.37, where Dana Jones, Rutgers
University, and Indiana University are specific instances.
(a) What are the minimal generalization and the maximal generalization of the
following expression?

?O1 is Dana Jones

plans to move to ?O2
?O2 is Indiana University

13:59:50,
.010
290 Chapter 9. Rule Learning

object

subconcept of
has as advisor
actor research area domain student
range faculty member
subconcept of

organization person

subconcept of
employee student

subconcept of

university employee subconcept of

subconcept of
graduate undergraduate
faculty member staff member student student

subconcept of subconcept of
subconcept of
instructor professor PhD advisor graduate graduate subconcept of
research teaching
subconcept of assistant assistant BS student

assistant associate full

professor professor professor PhD student MS student

instance of instance of instance of instance of instance of

John Smith John Doe Jane Austin Bob Sharp Joan Dean

Figure 9.35. Ontology fragment.

object plans to rere from

domain person
range organizaon
actor
plans to move to
person organizaon domain person
range organizaon
employee
educaonal organizaon

university employee subconcept of

university
faculty member
instance of
subconcept of
George Mason Indiana
professor PhD advisor University University
subconcept of

full professor plans to

rere from plans to move to
ce of
instance instance of
Dan Smith Jane Ausn

Figure 9.36. Ontology fragment.

13:59:50,
.010
9.15. Review Questions 291

is faculty member at object

domain faculty member
range university
actor
place
plans to move to
domain person person organizaon
range organizaon
employee educaonal subconcept of
organizaon
university employee subconcept of fun place

university
faculty member instance of
subconcept of instance of
Rutgers Indiana
professor PhD advisor University University

subconcept of plans to move to

full professor is faculty
member at
instance of

Dana Jones

Figure 9.37. Ontology fragment.

Figure 9.38. A reduction example and its explanation.

(b) What are the minimal generalization and the maximal generalization of the
following expression?

?O1 is Dana Jones

?O2 is Indiana University

9.7. Consider the example problem reduction and its explanation from Figure 9.38.
Which is the specific rule condition covering only this example? What rule will be
learned from this example and its explanation, assuming the ontology fragment

13:59:50,
.010
292 Chapter 9. Rule Learning

from Figure 9.39? What general problem will be learned from the specific IF
problem of this reduction?
Notice that some of the instances are specific (e.g., Aum Shinrikyo and Masami
Tsuchiya), while others are generic (e.g., chemistry).

9.8. Consider the problem reduction example and its explanation from Figure 9.40.
Which is the specific rule covering only this example? What rule will be learned
from this example and its explanation, assuming the ontology fragment from
Figure 9.41? What general problems will be learned from this example? Assume
that all the instances are specific instances.

has as member
domain organization
range person

has master degree in

domain person
range expertise area
Figure 9.39. Ontology fragment.

Figure 9.40. A problem reduction example and its explanation.

13:59:50,
.010
9.15. Review Questions 293

object
is tesmony by is tesmony about
subconcept of domain evidence domain evidence
range source range evidence
actor source evidence

subconcept of subconcept of

tesmonial evidence
person

subconcept of subconcept of

direct tesmonial
author terrorist
item of evidence evidence
instance of instance of subconcept of
subconcept of
Hamid Mir Osama bin Laden tesmonial tesmonial evidence
evidence obtained non-elementary elementary piece based on direct
interview at second hand piece of evidence of evidence observaon
instance of instance of

EVD-Dawn-Mir-01-01 EVD-Dawn-Mir-01-01c

Figure 9.41. Ontology fragment.

9.9. Compare the rule-learning process with the traditional knowledge acquisition
approach, where a knowledge engineer defines such a rule by interacting with
a subject matter expert. Identify as many similarities and differences as possible,
and justify the relative strengths and weaknesses of the two approaches, but be as
concise as possible.

13:59:50,
.010
 10 Rule Refinement

10.1 INCREMENTAL RULE REFINEMENT

10.1.1 The Rule Refinement Problem

The triggering event for rule refinement is the discovery of a new positive or negative
example of the rule. There are several possible origins for such an example:

It is generated by the agent during its regular problem solving

It is generated by the agent through active experimentation for rule testing

It is provided by the expert

It is obtained from an external source (e.g., a repository of examples)

Regardless of the origin of the example, the goal of the agent is to refine the rule to be
consistent with the example. A possible effect of rule refinement is the extension of the
ontology.
The rule refinement problem is defined in Table 10.1 and an overview of the rule
refinement method is presented in the next section.

Table 10.1 The Rule Refinement Problem

GIVEN

A plausible version space reduction rule

A positive or a negative example of the rule (i.e., a correct or an incorrect reduction)

A knowledge base that includes an ontology and a set of (previously learned) reduction rules

An expert who will interact with the agent, helping it understand why the example is positive
(correct) or negative (incorrect)

DETERMINE

An improved rule that covers the example if it is positive, and does not cover the example if it is
negative

An extended ontology, if this is needed for rule refinement

294

13:54:13,
.011
 10.1. Incremental Rule Refinement 295

10.1.2 Overview of the Rule Refinement Method

Figure 10.1 shows an overview of rule refinement during problem solving. The agent
applies existing partially learned rules to solve the current problem or assess the current
hypothesis. As discussed in Section 9.7, the reductions generated by a rule R are analogous
with the reduction from which the rule was initially learned. Thus this process can be
understood as learning by analogy and experimentation, as indicated in the top part
of Figure 10.1.
The subject matter expert has to verify the generated reduction and accept it if it is
correct, or reject it otherwise. In both cases, the rule will be improved to become
consistent with this new example. We therefore refer to this learning phase as learning
from examples, as indicated in the bottom part of Figure 10.1.
If the generated reduction is correct, then it is a new positive example of the rule, which
may need to be generalized to cover it, by minimally generalizing its applicability condi-
tion. There are various ways to generalize the applicability condition, depending on the
position of the positive example with respect to it. For instance, if the example is covered
by the plausible upper bound of the main condition (the light green ellipse), but it is not
covered by the plausible lower bound (the dark green ellipse), then the plausible lower
bound is generalized as little as possible to cover this new positive example, while still
remaining less general than (or included into) the plausible upper bound.
If the generated example is incorrect, then the expert has to interact with the agent to
help it understand why the reduction is wrong. This interaction is similar to that taking
place during the explanation generation phase of rule learning (see Section 9.5). The
identified failure explanation will be used to refine the rule. We therefore refer to this
phase as learning from explanations, as indicated in the left part of Figure 10.1.
There are various strategies to refine the rule based on a negative example and its
failure explanation (if identified), as illustrated in the top part of Figure 10.1. For instance,
if the negative example is covered only by the upper bound of the main condition, then
this bound is minimally specialized to no longer cover the negative example, while still
remaining more general than the plausible lower bound. Or, a new Except-When plausible

Learning by Analogy
Knowledge Base and Experimentation Rule
Rule’s condition IF we have to solve
– <Problem>

– +
+ Main
PVS Condition

Except-When
Failure PVS Condition
explanation
Examples of problem reductions
THEN solve
generated by the agent <Subproblem 1>
…
<Subproblem m>
Incorrect Correct
example example
Learning from
Explanations Learning from Examples
Figure 10.1. Multistrategy rule refinement.

13:54:13,
.011
 296 Chapter 10. Rule Refinement

version space condition may be learned, starting from that negative example and its failure
explanation. This plausible version space Except-When condition is represented by the red
ellipses at the top of Figure 10.1.
The refined rule is shown in the right-hand side of Figure 10.1. The applicability
condition of a partially learned rule consists of a main applicability condition and zero,
one, or more Except-When conditions. The way the rule is refined based on a new
example depends on the type of the example (i.e., positive or negative), on its position
with respect to the current conditions of the rule, and on the type of the explanation of
the example (if identified). The refinement strategies will be discussed in more detail in
the next sections by considering a rule with a main condition and an Except-When
condition, as shown in Figure 10.2. We will consider all the possible nine positions of the
example with respect to the bounds of these conditions. Notice that the presented
methods will similarly apply when there is no Except-When condition or more than
one Except-When condition.
We will first illustrate rule refinement with a positive example and then we will present
the general method.

10.1.3 Rule Refinement with Positive Examples

10.1.3.1 Illustration of Rule Refinement with a Positive Example
The upper-left part of Figure 10.3 shows an example generated by the agent based on the
partially learned rule shown in the right-hand side of the figure. This example is generated
because it satisfies the plausible upper bound condition of the rule. Indeed, the condition
corresponding to this example, shown in the bottom-left part of Figure 10.3, is covered the
plausible upper bound condition: Bob Sharp is a person, Dan Smith is a person, Information
Security is an area of expertise, and certain is in the interval [certain – certain]. Additionally,
the entities from the example are in the relationships from the rule’s condition, that is,
“Bob Sharp is interested in Information Security,” and “Dan Smith is expert in Information
Security.”
Notice also that the example is not covered by the plausible lower bound condition of
the rule because Dan Smith (corresponding to ?O2) is a full professor, and Information

MU: Main Condition

. XU: Except-When Condition

. 4 Plausible Upper Bound

Plausible Upper Bound

.
5
XL: Except-When Condition
ML: Main Condition
6 Plausible Lower Bound

.
Plausible Lower Bound

. 1
..
7

. 8
Universe of

.
2 9
Instances
3

Figure 10.2. Partially learned condition and various positions of a new example.

13:54:13,
.011
 Positive example that satisfies the upper bound but not the lower bound

Rule to be refined through the generalization of the lower bound

.011

Condition corresponding to the example

13:54:13,

?O1 is Bob Sharp

is interested in ?O3

?O2 is Dan Smith

is expert in ?O3

?O3 is Informaon Security

?SI1 is-exactly certain

Figure 10.3. Positive example generated by the agent.

298 Chapter 10. Rule Refinement

Security (corresponding to ?O3) is different from Artificial Intelligence. Therefore, this

example corresponds to Case 2 in Figure 10.2.
Because the example is accepted as correct by the expert, the agent will minimally
generalize the plausible lower bound condition of the rule to cover it, while still keeping
this condition less general than the plausible upper bound condition, as shown in
Figure 10.4.
Notice that, with respect to the variable ?O3, Artificial Intelligence and Information
Security are generalized to Computer Science, which is the least general concept covering
both of them, and is less general than area of expertise, the corresponding concept from
the plausible upper bound. With respect to the variable ?O2, PhD advisor does not need to
be generalized because it already covers Dan Smith. However, associate professor is
minimally generalized to professor to cover Dan Smith, which is a full professor. This is
possible because professor is less general than person, the concept from the plausible
upper bound. The refined rule is shown in Figure 10.5.

10.1.3.2 The Method of Rule Refinement with a Positive Example

The general rule refinement method illustrated by the preceding example (i.e., Case 2 in
Figure 10.2) is presented in Table 10.2. It is similar with the part of the Candidate Elimination
Algorithm (Mitchell, 1997) corresponding to the treatment of a positive example.
The method described in Table 10.2 corresponds to the more complex case where the
entity EX from the example is an instance, such as Bob Sharp or Information Security, and
the “climbing the generalization hierarchies” rule is used (see Section 8.3.3). Similar
methods apply when the entity EX corresponds to a number, string, or symbolic probabil-
ity, except that other generalization rules are used, as discussed in Section 8.3.
One of the most difficult problems in computing a good generalization of some
expressions is to establish the objects to be matched (Kodratoff and Ganascia, 1986). This

area of expertise person

Generalized plausible lower bound condition
?O1 is PhD student
is interested in ?O3 professor
Computer Science PhD advisor
?O2 is (PhD advisor, professor) subconcept of
is expert in ?O3
instance of full professor
?O3 is Computer Science
associate professor
Artificial Intelligence ?SI1 is-exactly certain instance of
Information Security Dan Smith
minimal generalization

Condition corresponding to the example Rule’s plausible lower bound condition

?O1 is Bob Sharp ?O1 is PhD student
is interested in ?O3 is interested in ?O3
?O2 is Dan Smith ?O2 is (PhD advisor, associate professor)
is expert in ?O3 is expert in ?O3
?O3 is Information Security ?O3 is Artificial Intelligence
?SI1 is-exactly certain ?SI1 is-exactly certain

Figure 10.4. Minimal generalization of the rule’s plausible lower bound condition.

13:54:13,
.011
10.1. Incremental Rule Refinement 299

Table 10.2 The Example-based Incremental Inductive Generalization Method

Let R be a plausible version space rule, U its main plausible upper bound condition, L its main plausible
lower bound condition, and P a positive example of R covered by U and not covered by L.

Repeat for each variable ?X from U

Let the concept corresponding to ?X from U be UX = {u1 . . . um}.

Let the concept corresponding to ?X from L be LX = {l1 . . . ln}.
Each concept ui from UX is a maximal generalization of all the known positive examples of ?X
that does not cover any of the known negative examples of ?X and is more general than (or
as general as) at least one concept lk from LX. Also, ui is not covered by any other element uk
of UX.
Each concept li is a minimal generalization of the known positive examples of ?X that does
not cover any of the known negative examples of ?X and is less general than (or as general
as) at least one concept uk from UX. Also li does not cover any other element lk of LX.

Let EX be the new positive example of ?X from the rule’s example P. EX is covered by at least
one element ui of UX.

Remove from UX any element that does not cover EX.
Repeat for each li from LX that does not cover EX

Remove li from LX.

Add to LX all minimal generalizations of li and EX that are less general or at most as general
as an element of UX.

Remove from LX all the elements that are more general than or as general as other
elements from LX.
end

end

Determine Pg, the minimal generalization of the example P (see Section 9.9).

Return the generalized rule R with the updated conditions U and L, and Pg in the list of
generalized examples of R.

Refined rule

Condition corresponding to the example

?O1 is Bob Sharp
is interested in ?O3

?O2 is Dan Smith

is expert in ?O3

?O3 is Information Security

?SI1 is-exactly certain

Rule’s condition

Figure 10.5. Rule generalization based on a positive example.

13:54:13,
.011
 300 Chapter 10. Rule Refinement

problem, however, is trivial in Disciple-EBR because both the plausible lower bound
condition and the condition corresponding to the example have exactly the same struc-
ture, and the corresponding variables have the same names, as shown in Figure 10.5. This
is a direct consequence of the fact that the example is generated from the plausible upper
bound condition of the rule.

10.1.3.3 Summary of Rule Refinement with a Positive Example

The general method for rule refinement with a positive example, covering all the nine
cases from Figure 10.2, is presented in Table 10.3. Notice that a positive example covered
by an Except-When condition leads to the specialization of that condition to no longer
cover it. This is similar with a negative example covered by the main applicability condi-
tion, which is discussed in the next section.
Next we will illustrate rule refinement with negative examples, and then we will present
the general refinement method.

10.1.4 Rule Refinement with Negative Examples

10.1.4.1 Illustration of Rule Refinement with Except-When Conditions
The left side of Figure 10.6 shows an example generated by the agent based on the partially
learned rule shown in the right-hand side of Figure 10.6. This example is generated
because it satisfies the plausible upper bound condition of the rule. However, the example
is rejected as incorrect by the expert, with the following explanation:

Dan Smith plans to retire from George Mason University

Based on this failure explanation, the agent generates an Except-When plausible version
space condition by applying the method described in Sections 9.6 and 9.7. First it reformu-
lates the explanation as a specific condition by using the corresponding variables from the
rule, or by generating new variables (see also the bottom-left part of Figure 10.7):

?O1 is Dan Smith

plans to retire from ?O2
?O2 is George Mason University

Then the agent generates a plausible version space by determining maximal and minimal
generalizations of the preceding condition. Finally, the agent adds it to the rule as an
Except-When plausible version space condition, as shown in the bottom-right part of
Figure 10.7. The Except-When condition should not be satisfied to apply the rule. Thus, in
order to conclude that a professor will stay on the faculty for the duration of the disserta-
tion of a student, the professor should have a long-term position (the main condition) and
it should not be the case that the professor plans to retire from the university (the Except-
When condition).
Figure 10.8 shows the further refinement of the rule with an additional negative
example. This example satisfies the rule in Figure 10.7. Indeed, Jane Austin has a long-term
position and she does not plan to retire from George Mason University. Nevertheless, the
expert rejects the reasoning represented by this example because Jane Austin plans to
move to Indiana University. Therefore, she will not stay on the faculty of George Mason
University for the duration of the dissertation of Bob Sharp.

13:54:13,
.011
10.1. Incremental Rule Refinement 301

Table 10.3 Rule Refinement with a Positive Example

MU XU
1. If the positive example E is covered by ML and is not
covered by XU (case 1), then the rule does not need to ML
XL

be refined because the example is correctly classified

1
as positive by the current rule.

MU XU XU
MU
2. If E is covered by MU, is not covered by ML, and is not
covered by XU (case 2), then minimally generalize ML ML
XL
ML
XL

to cover E and remain less general than MU. Remove

from MU the elements that do not cover E.
2 +

MU XU MU XU
3. If E is not covered by MU (cases 3 and 5), or if E is 5 +
covered by XL (cases 5, 6, and 7), then keep E as a ML 6 XL
ML + XL

positive exception of the rule. Alternatively, learn a 7 +

new rule starting from this example.
3 +
MU + XU
4. If E is not covered by MU but is covered by XU (case
XL
4), then E is kept as a positive exception (or a new rule ML

is learned). Additionally, interact with the expert to MU 4 XU

find an explanation of the form, “The problem
XL Ii is Ci
reduction step is correct because Ii is Ci,” where Ii is an ML
+
instance from the example E and Ci is a concept from MU XU

the ontology. If such an explanation is found, then XU XL

ML
is minimally specialized to no longer cover E.

5. If E is covered by ML and XU, but it is not covered by

MU XU
XL (case 8), then interact with the expert to find an
explanation of the form, “The problem reduction step ML
XL

is correct because Ii is Ci,” where Ii is an instance from XU

MU
the example E and Ci is a concept from the ontology. If +
XL Ii is Ci
such an explanation is found, then XU is minimally ML
XU
specialized to no longer cover E. Otherwise, E is kept MU

as a positive exception. 8
ML
XL

6. If E is covered by MU and XU, it is not covered by ML,

and it is not covered by XL (case 9), then minimally XU
MU
generalize ML to cover E and remain less general than
MU. Also remove from MU the elements that do not ML
XL

cover E. Then interact with the expert to find an XU

MU
explanation of the form, “The problem reduction step
XL Ii is Ci .+
is correct because Ii is Ci,” where Ii is an instance from ML
XU
the example E and Ci is a concept from the ontology. If MU

such an explanation is found, then XU is minimally 9 ML

XL

specialized to no longer cover E. Otherwise, E is kept as

a positive exception.
+

As in the case of the previous negative example, the failure explanation

Jane Austin plans to move to Indiana University

is rewritten as a specific condition

?O1 is Jane Austin

plans to move to ?O5
?O5 is Indiana University

13:54:13,
.011
 Negative example generated by the rule Rule that generated the negative example
.011
13:54:13,

Failure explanation
Dan Smith plans to retire from George Mason University

Figure 10.6. Negative example generated by the agent.

 Negative example Specialized rule
.011

+
-
13:54:13,

Failure explanation
Dan Smith plans to rere from George Mason University

Rewrite as
Specific Except-When condition
?O1 is Dan Smith
plans to retire from ?O2
?O2 is George Mason University

Figure 10.7. Rule specialization with a negative example.

 Specialized rule

Negative example
.011
13:54:13,

+
-

Failure explanation -

Jane Austin plans to move to Indiana University

Rewrite as
Specific Except-When condition
?O1 is Jane Austin
Max generalization
plans to move to ?O5 Min generalization
?O5 is Indiana University

Figure 10.8. Rule refinement with an additional negative example.

10.1. Incremental Rule Refinement 305

Notice that the agent has introduced a new variable ?O5 because Indiana University does
not correspond to any entity from the previous form of the rule (as opposed to Jane Austin
who corresponds to ?O1).
Then the agent generates a plausible version space by determining maximal and
minimal generalizations of the preceding condition. Finally, the agent adds it to the rule
as an additional Except-When plausible version space condition, as shown at the bottom-
right part of Figure 10.8.

10.1.4.2 The Method of Rule Refinement with Except-When Conditions

The basic steps for learning an Except-When condition based on the explanation of a
negative example are summarized in Table 10.4.

10.1.4.3 Illustration of Rule Refinement through Condition

Specialization
The preceding two sections presented the refinement of a rule through the addition of
Except-When plausible version space conditions. This section will illustrate rule refinement
through the specialization of the upper bound of the main applicability condition of the rule.
Figure 10.9 illustrates the mixed-initiative development of a reasoning tree. The expert
formulated the hypothesis:

Jill Knox would be a good PhD advisor for Peter Jones.

Then the agent developed a partial reasoning tree, but it was unable to assess one of the
subhypotheses:

Jill Knox will stay on the faculty of George Mason University for the duration of the
dissertation of Peter Jones.

Table 10.4 Basic Steps for Learning an Except-When Condition

Let R be a plausible version space rule, N an instance of R rejected by the expert as an incorrect
reasoning step (a negative example of R), and EX an explanation of why N is incorrect (a failure
explanation).
(1) Reformulation of the Failure Explanation
Generate a new variable for each instance and each constant (i.e., number, string, or symbolic
probability) that appears in the failure explanation EX but does not appears in the negative
example N. Use the new variables and the rule’s variables to reformulate the failure explanation EX
as an instance I of the concept EC representing an Except-When condition of the rule R.
(2) Analogy-based Generalizations of the Failure Explanation
Generate the plausible upper bound XU of the concept EC as the maximal generalization of I in the
context of the agent’s ontology.
Generate the plausible lower bound LU of the concept EC as the minimal generalization of I that
does not contain any specific instance.
(3) Rule Refinement with an Except-When Plausible Version Space Condition
Add an Except-When plausible version space condition (XU, LU) to the existing conditions of the
rule R. This condition should not be satisfied for the rule to be applicable in a given situation.

13:54:13,
.011
306 Chapter 10. Rule Refinement

Therefore, the expert defined the reduction of this hypothesis, which includes its assess-
ment, as shown at the bottom of Figure 10.9.
Based on this example, the agent learned a general rule, as illustrated in the right-hand
side of Figure 10.9 and as discussed in Chapter 9. The rule is shown in Figure 10.10.

1. Solving

Agent applies
learned rules to
solve new
problems

3. Learning

2. Modeling
Agent learns
a new rule

Expert extends the

reasoning tree

Figure 10.9. Problem solving, modeling, and learning.

Figure 10.10. Partially learned rule.

13:54:13,
.011
10.1. Incremental Rule Refinement 307

This and the other learned rules enabled the agent to develop the reasoning tree from
Figure 10.11 for assessing the following hypothesis:

Bill Bones would be a good PhD advisor for June Allison.

However, the expert rejected the bottom reasoning step as incorrect. Indeed, the correct
answer to the question, “Is Bill Bones likely stay on the faculty of George Mason University
for the duration of the PhD dissertation of June Allison?” is, “No,” not, “Yes,” because there
is no support for Bill Bones getting tenure.
The user–agent interaction during example understanding is illustrated in Figure 10.12.
The agent identified an entity in the example (the symbolic probability “no support”) that
would enable it to specialize the upper bound of the main condition of the rule to no
longer cover the negative example. Therefore, it proposed the failure explanation shown at
the right-hand side of Figure 10.12:

The example is incorrect because of the value no support.

The expert accepted this explanation by clicking on OK, and the rule was automatically
specialized as indicated in Figure 10.13. More precisely, the upper bound of the main
condition for the variable ?Sl1 was minimally specialized from the interval [no support –
certain] to the interval [likely – certain], in order to no longer cover the value no support,
while continuing to cover the interval representing the lower bound, which is [almost
certain – almost certain].

10.1.4.4 The Method of Rule Refinement through Condition

Specialization
The previous example illustrated the specialization of an upper bound of a condition with
a negative example. The corresponding general method is presented in Table 10.5.

3. Solving

Agent applies
learned rules
5. Refinement
to solve new
problems

Agent refines
rule with
negative example
4. Critiquing

Incorrect because + -
of the “no support”
probability

Figure 10.11. Solving, critiquing, and rule refinement.

13:54:13,
.011
308 Chapter 10. Rule Refinement

Figure 10.12. Explaining why the problem reduction is incorrect.

Specialized rule

Failure explanation

Min specialization

[no support – certain]

Figure 10.13. Specialization of the upper bound of a plausible version space condition.

10.1.4.5 Summary of Rule Refinement with a Negative Example

We have illustrated only some of the nine cases of negative examples from Figure 10.2
(p. 296). The general method for rule refinement with a negative example, covering all the
cases, is presented in Table 10.6.

13:54:13,
.011
 10.2. Learning with an Evolving Ontology 309

Table 10.5 The Example-based Incremental Inductive Specialization Method

Let R be a plausible version space rule, U the plausible upper bound of the main condition, L the
plausible lower bound of the main condition, N a negative example covered by U and not covered
by L, and C an entity from N that is blamed for the failure.
1. Let ?X be the variable from the rule’s conditions that corresponds to the blamed
entity C.
Let UX and LX be the classes of ?X in the two bounds.
If each concept from LX covers C
then Continue with step 2.
else Continue with step 3.
2. The rule cannot be specialized to uncover the current negative example.
The negative example N is associated with the rule as a negative exception.
Return the rule R.
3. There are concepts in LX that do not cover C. The rule can be specialized to uncover
N by specializing UX, which is known to be more general than C.
3.1. Remove from LX any element that covers C.
3.2. Repeat for each element ui of UX that covers C

Remove ui from UX.

Add to UX all minimal specializations of ui that do not cover C and are more general than
or at least as general as a concept from LX.

Remove from UX all the concepts that are less general than or as general as other
concepts from UX.

end
4. Return the specialized rule R.

10.2 LEARNING WITH AN EVOLVING ONTOLOGY

10.2.1 The Rule Regeneration Problem

The applicability conditions of the learned rules depend of the agent’s ontology, which
is used as a generalization hierarchy. When the ontology changes in significant ways,
the rules need to be updated accordingly. Let us consider that the ontology fragment
shown in Figure 9.7 (p. 258) has evolved into the one from Figure 10.14. The changed
elements are underlined. In particular, PhD advisor was moved from under professor to
above professor. A new concept, expert, was introduced between PhD advisor and
person. Additionally, the domains of the features is expert in, has as position, and plans
to retire from have been changed from person to expert, employee, and employee,
respectively. These are all changes that affect the minimal and the maximal generaliza-
tions of the examples from which the rules have been learned. Therefore, the rules
need to be updated.
In Disciple-EBR, the adaptation of the rules to a changed ontology is done by simply
automatically regenerating the rules based on the minimal generalizations of the examples

13:54:13,
.011
 310 Chapter 10. Rule Refinement

Table 10.6 Rule Refinement with a Negative Example

MU XU

1. If the negative example N is covered by ML and it is XL

not covered by XU (case 1), then interact with the ML
XU
subject matter expert to find an explanation of why MU _

N is an incorrect problem reduction step. If an XL

ML EX
explanation EX is found, then generate a new
1 MU XU
Except-When plausible version space condition and
add it to the rule. Otherwise, keep N as a negative ML
XL

exception. _

2. If N is covered by MU but it is not covered by ML MU XU

and by XU (case 2), then interact with the expert to
XL
find an explanation of why N is an incorrect ML

problem reduction step. If an explanation EX is MU XU

found and it has the form “because of v,” then XL “because of v” _

ML
minimally specialize MU to no longer cover N while
XU
still remaining more general than ML. Otherwise, if MU

another type of explanation EX is found, then learn a 2 EX XL

ML
new Except-When plausible version space condition
based on it, and add it to the rule.
_

3. If N is not covered by MU (cases 3, 5), or it is

covered by XL (cases 5, 6, 7), then the rule does not MU 4 XU MU _ XU

5 _
need to be refined because the example is correctly XL _ XL
ML 6 ML
classified as negative by the current rule. If N is not
7 _
covered by MU, is not covered by XL, and is covered
by XU (case 4), then minimally generalize XL to
_
cover N and remain less general than XU. 3

MU XU MU XU
4. If N is covered by ML and by XU, but it is not covered
by XL (case 8), or N is covered by MU and by XU, but ML
XL
ML
XL

it is not covered by ML and XL (case 9), then

minimally generalize XL to cover N while still 8 _
9 _
remaining less general than XU.

and their explanations in the context of the updated ontology. This is, in fact, the reason
why the generalized examples are maintained with each rule, as discussed in Section 9.9.
The rule regeneration problem is presented in Table 10.7. Notice that not all the
changes of an ontology lead to changes in the previously learned rules. For example,
adding a new concept that has no instance, or adding a new instance, will not affect the
previously learned rules. Also, renaming a concept or a feature in the ontology automatic-
ally renames it in the learned rules, and no additional adaptation is necessary.

10.2.2 On-Demand Rule Regeneration

Rules adaptation in response to changes in the ontology is an expensive operation if it
means regenerating all the rules after each significant change of the ontology. Therefore,
the question is: When should the rules be regenerated? The solution adopted in Disciple-
EBR is on-demand regeneration. That is, a rule is regenerated only when the agent needs
to apply it in problem solving, and only if the current version of the ontology is different
from the one when the rule was last refined or regenerated. Thus, the agent associates with

13:54:13,
.011
10.2. Learning with an Evolving Ontology 311

the ontology a version, and each time a rule is learned or refined, it associates the version
of the ontology with the rule. The version of the ontology is incremented each time a
significant change – that is, a change that may affect the conditions of the previously
learned rules – is made. Then, before using a rule in problem solving, the agent checks
the rule’s ontology version with the current version of the ontology. If the versions are
the same, the rule is up to date and can be used. Otherwise, the agent regenerates the
rule based on the current ontology and also updates the rule’s ontology version to the
current version of the ontology. The on-demand rule regeneration method is presented
in Table 10.8.

Table 10.7 The Rule Regeneration Problem

GIVEN

A plausible version space reduction rule R corresponding to a version v of the ontology

Minimal generalizations of the examples and explanations from which the rule R was learned, in
the context of the version v of the ontology

An updated ontology with a new version v’

DETERMINE

An updated rule that corresponds to the same generalized examples, but in the context of the
new version v’ of the ontology

Updated minimal generalizations of the specific examples from the current scenario, if any

object

actor position area of expertise

person organization university position subconcept of

educational
expert employee student organization faculty position Computer Science

university employee subconcept of instance of

long-term
faculty position
faculty member university
tenured position Artificial Information
graduate instance of Intelligence Security
PhD advisor student
George Mason Indiana
subconcept of University University has as position
professor domain employee
PhD student range position
subconcept of
is expert in plans to retire from
associate full domain expert domain employee
professor professor range area of expertise range organization
instance of instance of is is interested in plans to move to
is expert in interested domain person domain person
John Doe Bob Sharp in range area of expertise range organization

Figure 10.14. Evolution of the ontology fragment from Figure 9.7.

13:54:13,
.011
 312 Chapter 10. Rule Refinement

We will first illustrate the regeneration of the rules presented in the previous sections
and then provide the general regeneration method.

10.2.3 Illustration of the Rule Regeneration Method

Let us first consider the rule from the right-hand side of Figure 10.5 (p. 299). This rule was
learned from two positive examples and their explanations, based on the ontology from
Figure 9.7 (p. 258), as discussed in Section 10.1.3.1. This rule can be automatically relearned
from its generalized examples, in the context of the updated ontology, as illustrated in
Figure 10.15. The bottom part of Figure 10.15 shows the generalized examples of the rule,
which were determined in the context of the ontology from Figure 9.7 (p. 258), as discussed

Table 10.8 On-Demand Rule Regeneration

Let R be a plausible version space rule, and O the current ontology with version v.
If R’s ontology version is v, the same as the version of the current ontology O
then Return R (no regeneration is needed).
else Regenerate rule R (see Table 10.9).
Set R’s ontology version to v.
Return R

Plausible Lower Plausible Upper

Bound Condition (LB) Bound Condition (UB)
?O1 is PhD student ?O1 is person
is interested in ?O3 is interested in ?O3
?O2 is professor ?O2 is expert
is expert in ?O3 is expert in ?O3
?O3 is Computer Science ?O3 is area of expertise
?SI1 is in [certain - certain] ?SI1 is in [certain - certain]

Min Max
generalization generalization

Example generalization Example generalization

?O1 is PhD student ?O1 is PhD student
is interested in ?O3 is interested in ?O3
?O2 is PhD advisor ?O2 is PhD advisor
is associate professor is full professor
is expert in ?O3 is expert in ?O3
?O3 is Artificial Intelligence ?O3 is Information Security
?SI1 is exactly certain ?SI1 is exactly certain
Covered positive examples: 1 Covered positive examples: 1
Covered negative examples: 0 Covered negative examples: 0

Figure 10.15. The updated conditions of the rule from Figure 10.5 in the context of the ontology
from Figure 10.14.

13:54:13,
.011
10.2. Learning with an Evolving Ontology 313

in Section 9.9. The top part of Figure 10.15 shows the updated bounds of the rule, in the
context of the updated ontology from Figure 10.14. The new plausible lower bound condi-
tion is the minimal generalization of the generalized examples, in the context of the updated
ontology. Similarly, the new plausible upper bound condition is the maximal generalization
of the generalized examples in the context of the updated ontology.
Notice that in the updated plausible lower bound condition (shown in the upper-left
part of Figure 10.15), ?O2 is now a professor, instead of a professor or PhD advisor. Indeed,
note the following expression from the first generalized example shown in the lower-left
part of Figure 10.15:

?O2 is PhD advisor [10.1]

is associate professor
is expert in ?O3

Based on the updated ontology, where associate professor is a subconcept of PhD advisor,
this expression is now equivalent to the following:

?O2 is associate professor [10.2]

is expert in ?O3

Similarly, note the following expression from the second generalized example shown in
the lower-right part of Figure 10.15:

?O2 is PhD advisor [10.3]

is full professor
is expert in ?O3

Because full professor is now a subconcept of PhD advisor, this expression is now equivalent
to the following:

?O2 is full professor [10.4]

is expert in ?O3

Then the minimal generalization of the expressions [10.2] and [10.4] is the following
expression because professor is the minimal generalization of associate professor and full
professor:

?O2 is professor [10.5]

is expert in ?O3

Also, in the updated plausible upper bound condition (shown in the upper-right part of
Figure 10.15), ?O2 is an expert instead of a person, because this is the maximal generaliza-
tion of associate professor and full professor, which is included into the domain of the is
expert in feature of ?O2, which is expert.
Let us now consider the rule from the right-hand part of Figure 10.8 (p. 304). The
minimal generalizations of the examples from which this rule was learned are shown
under the updated conditions in Figure 10.16. They were determined based on the
ontology from Figure 9.7. You remember that this rule was learned from one positive
example and two negative examples. However, each of the negative examples was used as
a positive example of an Except-When plausible version space condition. That is why each
of the generalized examples in Figure 10.16 has a positive example.

13:54:13,
.011
 Except-When Condition 1
Plausible Lower Plausible Upper
Bound Condition (LB) Bound Condition (UB)
?O1 is full professor ?O1 is employee
plans to retire from ?O2 plans to retire from ?O2
?O2 is university ?O2 is organization

Min Max
Main Condition generalization generalization
Plausible Lower Plausible Upper
Bound Condition (LB) Bound Condition (UB) Example generalization
?O1 is full professor
?O1 is associate professor ?O1 is employee
is PhD advisor
has as position ?O4 has as position ?O4
plans to retire from ?O2
?O2 is university ?O2 is actor
?O2 is university
?O3 is PhD student ?O3 is actor
?O4 is tenured position ?O4 is long-term faculty position Covered positive examples: 1
?SI1 is in [almost certain - almost certain] ?SI1 is in [almost certain - almost certain] Covered negative examples: 0

Min Max Except-When Condition 2

generalization generalization Plausible Lower
.011

Plausible Upper
Example Generalization Bound Condition (LB) Bound Condition (UB)
?O1 is associate professor ?O1 is full professor ?O1 is person
plans to move to ?O5 plans to move to ?O5
13:54:13,

is PhD advisor
has as position ?O4 ?O5 is university ?O5 is organization
?O2 is university
?O3 is PhD student Min Max
?O4 is tenured position generalization generalization
?SI1 is exactly almost certain Example Generalization
Covered positive examples: 1 ?O1 is full professor
Covered negative examples: 0 is PhD advisor
plans to move to ?O5
?O5 is university

Covered positive examples: 1

Covered negative examples: 0
Figure 10.16. The updated conditions of the rule from Figure 10.8, in the context of the ontology from Figure 10.14.
10.2. Learning with an Evolving Ontology 315

The lower and upper bounds of the rule in Figure 10.8 (p. 304) were updated by
computing the minimal and maximal generalizations of these generalized examples in
the context of the updated ontology from Figure 10.14. Let us first consider the updated
version space of the main condition. Notice that in the lower bound, ?O1 is now associate
professor, instead of PhD advisor or associate professor (see Figure 10.8). Also, in the upper
bound, ?O1 is employee instead of person. The version spaces of the Except-When condi-
tions have also been updated. In the first Except-When condition, ?O1 is now full professor
in the lower bound and employee in the upper bound, instead of PhD advisor or full
professor and person, respectively. Similarly, in the second Except-When condition, ?O1
is now full professor in the lower bound, instead of PhD advisor or full professor.
Finally, let us consider the rule from Figure 10.13 (p. 308), which was learned based on
the ontology from Figure 9.7, as discussed in Section 10.1.4. The minimal generalizations
of the positive and negative examples from which this rule was learned are shown at the
bottom of Figure 10.17. They were determined based on the ontology from Figure 9.7.
Notice that these generalized examples include all the explanations from which the rule
was learned. In particular, the explanation that fixed the value of “tenure-track position” is
represented as “?O4 is exactly tenure-track position” and that which excluded the value “no
support” is represented as “?SI1 is-not no support in main condition.”

Plausible Lower Plausible Upper

Bound Condition (LB) Bound Condition (UB)
?O1 is assistant professor ?O1 is professor
has as position ?O4 has as position ?O4
probability of tenure ?SI1 probability of tenure ?SI1
?O2 is university ?O2 is actor
?O3 is PhD student ?O3 is actor
?O4 is tenure-track position ?O4 is tenure-track position
?O5 is tenured position ?O5 is tenured position
?SI1 is in [almost certain - almost certain] ?SI1 is in [likely - almost certain]

Min Max
generalization generalization

Example generalization Example generalization

?O1 is assistant professor ?O1 is assistant professor
has as position ?O4 has as position ?O4
probability of tenure ?SI1 probability of tenure ?SI1
?O2 is university ?O2 is university
?O3 is PhD student ?O3 is PhD student
?O4 is exactly tenure-track position ?O4 is exactly tenure-track position
?O5 is exactly tenured position ?O5 is exactly tenured position
?SI1 is almost certain ?SI1 is no support
is-not no support in main condition
Covered positive examples: 1
Covered negative examples: 0 Covered positive examples: 0
Covered negative examples: 1

Figure 10.17. The updated conditions of the rule from Figure 10.13 in the context of the ontology
from Figure 10.14.

13:54:13,
.011
 316 Chapter 10. Rule Refinement

The new lower and upper bounds of the rule in the context of the updated ontology
from Figure 10.14 are shown at the top of Figure 10.17. Notice that in this case, the
regenerated rule is actually the same with the previous rule. The changes made to the
ontology did not affect this rule. However, the agent did recompute it because it cannot
know, a priori, whether the rule will be changed or not. The only change made to the rule
is to register that it was determined based on the new version of the ontology.

10.2.4 The Rule Regeneration Method

The previous section has illustrated the automatic rule regeneration method. Table 10.9
and the ones following it describe the actual method in a more formal way.

10.3 HYPOTHESIS REFINEMENT

Hypothesis refinement is performed using methods that are very similar to the preceding
methods for rule refinement, as briefly summarized in this section.
Remember that general hypotheses are automatically learned as a byproduct of reduc-
tion rule learning, if they have not been previously learned. When a reduction rule is
refined with a positive example, each included hypothesis is also automatically refined
with its corresponding positive example. Indeed, when you say that a reduction is correct,
you are also implicitly saying that each of the included hypotheses is correct.
However, when a reduction rule is refined with a negative example, the hypotheses
are not affected. Indeed, a negative reduction example means that the corresponding
reduction is not correct, not that the any of the involved hypotheses is incorrect. For this
reason, an explanation of why a specific reduction is incorrect does not automatically
apply to the hypotheses from that reduction. If you want to say that a specific hypothesis
is incorrect, you have to select it and click on the Incorrect problem button. Then the

Table 10.9 Rule Regeneration Method

Let O be the current ontology with version v, and R a plausible version space rule with a different
version.
1. Recompute the formal parameters P for rule R (see Table 10.10).
2. Refresh examples for rule R (see Table 10.11).
3. If R is no longer valid (i.e., the rule has no longer any generalized positive example)
then Return null
4. Recompute plausible version space for rule R (see Table 10.12).
5. Repeat for each specific example EX of R
If the upper bound of the main plausible version space condition (PVS) of R does not cover
EX and EX is a positive example
then make EX a positive exception.
If the upper bound of PVS of R does cover EX and EX is a negative example
then make EX a negative exception.
end
6. Return R

13:54:13,
.011
 10.4. Characterization of Rule Refinement 317

Table 10.10 Method for Recomputing the Formal Parameters of a Rule

Let R be a plausible version space rule having the list of parameters P.

Recompute the parameters P of R as follows:
1. P = IF task parameters [ question parameters [ answer parameters
2. Repeat for each THEN statement of R
P = P [ THEN statement parameters
end
3. Repeat for each explanation of R
P = P [ explanation parameters
end
4. Return P

Table 10.11 Method for Refreshing the Examples of a Rule

Let R be a plausible version space rule.

1. Repeat for each specific example EX of the rule R
Keep only the values corresponding to the current list of parameters in the rule R.
end
2. Repeat for each generalized example GE of the rule R
Keep only the values corresponding to the current list of parameters in the rule R.
end
3. Repeat for each specific example EX of the rule R
If the stored generalized example is no longer a minimal generalization then
Unregister the example from the generalized example.
If the generalized example has no examples any more
then delete the generalized example.
Register the example with its new minimal generalization.
end
4. If the rule R has no generalized or specific positive examples any more
then Remove the rule R.

selected hypothesis will be refined basically using the same methods as those for rule
refinement.
Just as a refined rule, a refined hypothesis may include, in addition to the plausible
version space of the main condition, one or several plausible version spaces of Except-When
conditions, and generalized positive and negative examples. When the ontology is changed,
the hypotheses can be automatically regenerated based on the associated generalized
examples. They are actually regenerated when the corresponding rules are regenerated.

10.4 CHARACTERIZATION OF RULE LEARNING

AND REFINEMENT

The presented rule learning and refinement methods have the following characteristics:

They use multistrategy learning that synergistically integrates mixed-initiative learn-

ing from examples, from explanations, and by analogy, to take advantage of the

13:54:13,
.011
318 Chapter 10. Rule Refinement

Table 10.12 Method for Recomputing the Plausible Version Space of a Rule

Let R be a plausible version space rule having the main condition M and the list of the Except-
When conditions LX.
1. Let MP be the list of the parameters from the main condition M.
Let IP be the list of parameters from the natural language part of the rule R, referred to as
informal parameters.
MP = IP
2. Repeat for each explanation of a positive example EP in the rule R
MP = MP [ the new parameters from EP
end
3. Let LGS be the list of the generalized examples that have at least one specific positive
example.
Let LEP be the list of the explanations EP of the positive examples in the rule R.
Create the multivariable condition MC based on MP, LGS, LEP, IP (see Table 10.13).
4. Let LX be the list of Except-When conditions.
LX = [ ]
5. Repeat for each group EEp of Except-When explanations in R.
Let EP be the list of parameters used in EEp.
Let EX be the list of the generalized negative examples associated with EEp that have
at least one specific negative example.
Create the multivariable condition XC based on MP = EP, LGS = EX, LEP = EEp, and
IP = ∅ (see Table 10.13).
LX = LX [ XC
end
6. Return R

complementary strengths of these individual learning strategies and compensate for

their relative weaknesses.

The methods efficiently capture the expert’s tacit knowledge, significantly reducing the
complexity of developing cognitive assistants.

They use the explanation of the first positive example and analogical reasoning to
generate a much smaller initial version space than Mitchell’s classical version space
method (Mitchell, 1997).

The methods efficiently search the version space, guided by explanations obtained
through mixed-initiative interaction with the user (both the upper bounds and
the lower bounds being both generalized and specialized to converge toward
one another).

They learn from only a few examples, in the context of an incomplete and evolving
ontology.

These methods enable agents to learn even in the presence of exceptions.

They keep minimally generalized examples and explanations to regenerate the rules
automatically when the ontology changes.

The methods have been successfully applied to develop cognitive assistants for many
complex real-world domains (e.g., intelligence analysis, military strategy and planning,
education, collaborative emergency response planning), because they enable agents to
learn within a complex representation language.

13:54:13,
.011
 10.5. Hands On: Rule Refinement 319

Table 10.13 Method for Creating the Multivariable Condition Structure

Let R be a plausible version space rule, MP be the list of the parameters from the main condition,
LGS be the list of generalized examples that have at least one specific positive example, LEP be the
list of the explanations EP of the positive examples in the rule R, and IP be the list of informal
parameters of R.
1. Let A be the list of the generalized explanations fragments (such as “?Oi is interested in
?Oj”) from the generalized examples of R.
Compute A based on MP and LEP.
2. Let D be the domains of the variables from MP, each domain consisting of a lower bound
and an upper bound.
D=[]
3. Repeat for each parameter ?Oi in MP
Determine the list GE = {ge1, . . ., geg} of the concepts from LGS corresponding to ?Oi
(e.g., “assistant professor” from “?Oi is assistant professor”).
Determine the list PC = {PC1, . . ., PCp} of the concepts to which ?Oi must belong,
corresponding to positively constraining explanations
(e.g., “?Oi is exactly tenured position”).
Determine the list NC = {NC1, . . ., NCn} of the concepts to which ?Oi must not belong,
corresponding to negatively constraining explanations (e.g., “Sl1 is-not no support”).
Create the domains Do from the examples GE and the constraints PC and NC
(see Table 10.14).
D = D [ Do
end
4. Create the multivariable condition structure from MP, D, and A.
Return MVC

10.5 HANDS ON: RULE REFINEMENT

There are two rule refinement case studies, a shorter one and a longer one. In the shorter
case study, you will guide the agent to refine the rule from Figure 9.8 (p. 259), as discussed
in the previous sections. In the longer case study, you will guide the agent to refine several
rules, including the rule from Figure 9.8. You may perform the short case study, or the long
one, or both of them.
Start Disciple-EBR, select the case study knowledge base (either “13-Rule-Refinement-
short/Scen” or “13-Rule-Refinement/Scen”), and proceed as indicated in the instructions
at the bottom of the opened window.
The following are the basic operations for rule refinement, as well as additional operations
that are useful for knowledge base refinement, such as changing a generated reasoning step
into a modeling step, visualizing a rule with the Rule Editor, and deleting a rule.

Operation 10.1. Refine rule with positive example

In the Scenario workspace, during rule refinement, in the Reasoning Hierarchy viewer,
select a reasoning step for which a rule has already been learned by clicking on the
corresponding Q/A node.

In the right panel, click on the Correct Reduction button.

13:54:13,
.011
320 Chapter 10. Rule Refinement

Table 10.14 Method for Creating Domains from Examples and Constraints

Let ?Oi be a parameter for which we compute a plausible version space.

GE = {ge1, . . ., geg} be the set of generalized examples corresponding to variable ?Oi.
PC = {PC1, . . ., PCp} be the set of concepts to which ?Oi must belong, representing positive
constraints.
NC = {NC1, . . ., NCn} be the set of concepts to which ?Oi must not belong, representing
negative constraints.
1. Compute UB, the maximal generalization of the generalized examples in GE.
2. Compute LB, the minimal generalization of the generalized examples in GE.
3. Compute UB, the minimal specialization of UB based on all the positive constraints PC.
4. Compute LB by selecting from LB the elements covered by UB.
5. Repeat for each NCi in NC
Compute LB, the minimal specialization of LB that is disjoint with
the negative constraint NCi.
Compute UB, the minimal specialization of UB covering LB that is disjoint with the
negative constraint NCi.
end
6. Create the list (UB, LB) as the domain Do for ?Oi.
7. Return Do

Operation 10.2. Refine rule with negative example

In the Scenario workspace, during rule refinement, in the Reasoning Hierarchy viewer,
select a reasoning step by clicking on the corresponding Q/A node.

In the right panel, click on the Incorrect Reduction button.

Follow the agent’s instructions to select or provide a failure explanation.

Operation 10.3. Replace a generated reasoning step with

a modeling step

In the Scenario workspace, in the Reasoning Hierarchy viewer, right-click on the Q/A
node of the reasoning step that was generated through the application of a learned rule
and select Replace with modeling.

As a result, the reasoning step is replaced with a modeled one that can be changed.
Additionally, the initial reasoning step becomes a negative exception of the rule.

Operation 10.4. View a rule with the Rule Browser

Select the Scenario workspace.

Click on the Rules menu and select Rule Browser.

The left pane displays the names of the rules.

Click on one of these rules (e.g., DDR.00018) and the right pane displays its descrip-
tion in a given format (e.g., “Formal Description”) when you select it in the
corresponding tab.

Generally, you would first find the name of a rule of interest from a reasoning tree (e.g.,
DDR.00018), as described in Operation 9.3, and then you can view or delete it with the
Rule Browser.

Click on the X button of the Rule Browser, to close it.

13:54:13,
.011
 10.6. Guidelines for Rule Refinement 321

Operation 10.5. Delete a rule with the Rule Browser

Generally, you would first find the name of the rule to be deleted (e.g., DDR.00018) by
inspecting the reasoning tree, as described in Operation 9.3.

In the Scenario workspace, click on the Rules menu and select Rule Browser.

The left pane displays the names of the rules.

Click on the name of the rule to be deleted (e.g., DDR.00018), and the right pane
displays its description in a given format (e.g., “Formal Description”) when you select it
in the corresponding tab.

At the bottom of the right pane, click on the Edit Rule button. As a result, new options
are displayed.

At the bottom of the right pane, click on the Delete Rule button.

Click on the X button of the Rule Browser to close it.

10.6 GUIDELINES FOR RULE REFINEMENT

Guideline 10.1. Assess similar hypotheses to refine the rules

Formulate hypotheses (or problems) that are similar to those from which the rules were
learned. Refine the rules with positive examples by selecting (clicking on) the corresponding
Q/A pair, and then clicking on Correct Reduction (see the right-hand side of Figure 10.18).
You can also click on Modify Explanations to interact with the agent and update the
explanations (e.g., by adding new ones and/or deleting some of the existing ones).
Refine the rules with negative examples by selecting (clicking on) the corresponding
Q/A pair and then clicking on Incorrect Reduction (see the right-hand side of Figure 10.18).
Then interact with the agent to identify an explanation of why the reduction is incorrect.

Guideline 10.2. Extend the ontology to define failure explanations

If necessary, extend the ontology to provide explanations for negative examples. However,
do not limit yourself to a minimal extension of the ontology (that consists of only the
failure explanation) if you anticipate that related concepts and features may be needed in
the future for related hypotheses or problems.

Figure 10.18. Rule refinement with a selected example.

13:54:13,
.011
 322 Chapter 10. Rule Refinement

10.7 PROJECT ASSIGNMENT 7

Refine the learned reduction rules by assessing hypotheses that are similar to the ones
considered in the previous assignments.

10.8 REVIEW QUESTIONS

10.1. Define the rule refinement problem.

10.2. What are the basic steps of rule refinement?

10.3. Consider the version space from Figure 10.19. In light of the refinement strategies
studied in this chapter, how will the plausible version space be changed as a result
of a new negative example labeled 1? Draw the new version space(s).

10.4. Consider the version space from Figure 10.20. In light of the refinement strategies
studied in this chapter, what are three alternative ways in which this version space
may be changed as a result of the negative example 2?

Upper Bound of the

Upper Bound of Except-When Condition
the Main Condition

Lower Bound of the

Lower Bound of
Except-When Condition
the Main Condition

. 1

Universe of
Instances

Figure 10.19. Version space and a negative example covered by the lower bound of the main
condition.

Upper Bound of the

Upper Bound of Except-When Condition
the Main Condition

Lower Bound of the

Lower Bound of
Except-When Condition
the Main Condition

. 2
Universe of
Instances

Figure 10.20. Version space and a negative example covered by the upper bound of the main
condition.

13:54:13,
.011
10.8. Review Questions 323

Positive Example 1
We need to
Determine a strategic center of gravity for a
member of Allied Forces 1943.
Explanation
Which is a member of Allied Forces 1943?
Allied Forces 1943 has as member US 1943
US 1943
Therefore we need to
Determine a strategic center of gravity for US 1943.

Figure 10.21. A reduction example and its explanation.

object

… force force goal

opposing force multimember force single-member force group

multigroup multistate single-state single-group illegal legal

force force force force group group
… … … …

multistate multistate
alliance coalition

dominant partner equal partners

multistate alliance multistate alliance
US 1943
UK 1943
Allied Forces 1943
has as member
European Axis 1943 Germany 1943
has as member
domain multimember force
Finland 1943 range force

Figure 10.22. Ontology fragment from the center of gravity analysis domain. Dotted links
indicate instance of relationships while continuous unnamed links indicate subconcept of
relationships.

10.5. (a) Consider the example and its explanation from Figure 10.21. What rule will
be learned from them, assuming the ontology from Figure 10.22, where all
the instances are considered specific instances?
(b) Consider the additional positive example from Figure 10.23. Indicate the
refined rule.
(c) Consider the negative example, its failure explanation, and the additional
ontological knowledge from Figure 10.24. Indicate the refined rule.

13:54:13,
.011
324 Chapter 10. Rule Refinement

Positive Example 2
We need to
Determine a strategic center of gravity for a member of
European Axis 1943.

Which is a member of European Axis 1943?

Germany 1943
Therefore we need to
Determine a strategic center of gravity for Germany 1943.

Figure 10.23. An additional positive example.

Negative Example 2 Ontology Fragment

We need to contribution
Determine a center of gravity for a member of
European Axis 1943. subconcept of

Which is a member of European Axis 1943? military contribution

Finland 1943 subconcept of

Therefore we need to minor military intermediary military major military

Determine a center of gravity for Finland 1943. contribution contribution contribuon

Failure Explanation instance of

Finland 1943 has as military contribution has as military contribution
military contribution of Finland 1943 military contribution domain force
is minor military contribution of Finland 1943 range military contribution

Figure 10.24. Negative example, failure explanation, and additional ontology fragment.

war material and transports of US 1943

is a major generator of

industrial capacity of US 1943

has as industrial factor

US 1943

explains

IF the problem to solve is

Identify each strategic COG candidate with
respect to the industrial civilization of US 1943.
THEN
industrial capacity of US 1943 is a strategic
COG candidate for US 1943.

Figure 10.25. An example and its explanation.

10.6. Consider the example and its explanation shown in Figure 10.25. Find the plaus-
ible version space rule that will be learned based on the ontology fragments from
Figures 10.26, 10.27, and 10.28, where all the instances are defined as generic
instances.

13:54:13,
.011
10.8. Review Questions 325

is a major generator of economic factor

domain economic factor
range product
has as industrial factor
domain force
industrial commerce information
range industrial factor
factor authority network or system

raw material
industrial transportation
industrial farm implement center factor
authority industry
strategic raw
material industrial
farm implement capacity
industry of Italy 1943 transportation transportation
center network or system
oil chromium copper and
bauxite of Germany 1943
is critical to the production of

war material of Germany 1943 industrial capacity of US 1943

is obtained
from has as strategic
raw material industrial capacity of Germany 1943 is a major generator of
has as industrial factor has as industrial factor
Balkans
Germany 1943 US 1943 war material and transports of US 1943

Figure 10.26. Ontology of economic factors. Dotted links indicate instance of relationships while
continuous unnamed links indicate subconcept of relationships.

object
… …
force

group

multigroup force opposing force multistate force single-state force single-group force

US 1943 Germany 1943

component state component state
Allied Forces 1943 European axis 1943
component state component state
Britain 1943 Italy 1943

Figure 10.27. An ontology of forces. Dotted links indicate instance of relationships while continuous
unnamed links indicate subconcept of relationships.

10.7. Minimally generalize the rule from the left side of Figure 10.29 in order to cover
the positive example from the right side of Figure 10.29, considering the back-
ground knowledge from Figures 10.26, 10.27, and 10.28.

13:54:13,
.011
326 Chapter 10. Rule Refinement

object
… …
resource or infrastructure element

resource

strategically essential resource

or infrastructure element raw material

product
strategic
raw material

strategically essential strategically essential non-strategically essential

infrastructure element goods or material goods or services

war material and fuel war material and transports farm implements

main airport main seaport

war material and fuel war material and farm implements
sole airport sole seaport of Germany 1943 transports of US 1943 of Italy 1943

Figure 10.28. An ontology of resources. Dotted links indicate instance of relationships while con-
tinuous unnamed links indicate subconcept of relationships.

Rule

IF Positive example that satisfies the upper bound

Identify each strategic COG candidate with respect
to the industrial civilization of ?O1. IF the task to accomplish is
Identify each strategic COG candidate with respect
Plausible Upper Bound Condition
to the industrial civilization of Germany 1943.
?O1 is force
has as industrial factor ?O2 THEN accomplish the task
?O2 is industrial factor industrial capacity of Germany 1943 is a strategic
is a major generator of ?O3 COG candidate for Germany 1943.
?O3 is product

Plausible Lower Bound Condition

?O1 is US 1943 Explanation
has as industrial factor ?O2 Germany 1943
?O2 is industrial capacity of US 1943 has as industrial factor
is a major generator of ?O3 industrial capacity of Germany 1943
?O3 is war material and transports of US 1943
industrial capacity of Germany 1943
THEN is a major generator of
?O2 is a strategic COG candidate for ?O1. war material and fuel of Germany 1943

Figure 10.29. Rule and positive example.

10.8. Minimally specialize the rule from the left side of Figure 10.30, in order to cover
the positive example from the right side of Figure 10.30, considering the back-
ground knowledge from Figures 10.26, 10.27, and 10.28.

13:54:13,
.011
10.8. Review Questions 327

Table 10.15 Rule Refinement with Learning Agent versus Rule Refinement by
Knowledge Engineer

Rule Refinement with Rule Refinement by

Learning Agent Knowledge Engineer

Description (highlighting
differences and similarities)

Strengths

Weaknesses

Rule

IF
Identify each strategic COG candidate with respect Negative example that satisfies the upper bound
to the industrial civilization of ?O1.
IF the task to accomplish is
Plausible Upper Bound Condition Identify each strategic COG candidate with
?O1 is force respect to the industrial civilization of Italy 1943.
has as industrial factor ?O2
?O2 is industrial factor THEN accomplish the task
is a major generator of ?O3 farm implement industry of Italy 1943 is a
?O3 is product strategic COG candidate for Italy 1943.

Plausible Lower Bound Condition

?O1 is single state force Explanation
has as industrial factor ?O2
Italy 1943
?O2 is industrial capacity has as industrial factor
is a major generator of ?O3 farm implement industry of Italy 1943
?O3 is strategically essential goods or material
farm implement industry of Italy 1943
THEN is a major generator of
?O2 is a strategic COG candidate for ?O1. farm implements of Italy 1943

Figure 10.30. Rule and negative example.

10.9. Consider the problem reduction step and its explanation from Figure 10.25, as
well as the ontology of economic factors from Figure 10.26. Show the correspond-
ing analogy criterion generated by a cautious learner, and an analogous reduction
made by that cautious learner.

10.10. Indicate five qualitatively different possible uses of an ontology in an instructable

agent such as Disciple-EBR.

10.11. Compare the learning-based rule refinement process discussed in this chapter
with the traditional knowledge acquisition approach discussed in Section 3.1.4 by
filling in Table 10.15. Identify similarities and differences and justify the relative
strengths and weaknesses of the two approaches.

13:54:13,
.011
328 Chapter 10. Rule Refinement

Main Plausible Version Except-When Plausible

Space Condition Version Space Condition

I2

I3
I1

I4

Figure 10.31. A partially learned concept and several instances.

10.12. Consider the partially learned concept and the four instances from Figure 10.31.
Order the instances by the plausibility of being positive examples of this concept
and justify the ordering.

13:54:13,
.011
11 Abstraction of Reasoning

Up until this point, the methodology for developing intelligent agents has encouraged
the expert to be very explicit and detailed, to provide clear descriptions of the
hypotheses (or problems), and to formulate detailed questions and answers that guide
the reduction of hypotheses (or problems) to subhypotheses (or subproblems). This is
important because it facilitates a clear and correct logic and the learning of the
reasoning rules.
The developed agents can solve complex problems through the generation of
reasoning trees that can be very large, with hundreds or even thousands of nodes.
In such cases, browsing and understanding these reasoning trees become a
challenge.
In this section, we will discuss an approach to abstract a large reasoning tree that
involves abstracting both hypotheses/problems and subtrees. The goal is to obtain a
simpler representation where the abstract tree has fewer nodes and each node has a
simpler description. At the same time, however, we want to maintain the correspondence
between the abstract tree and the original tree, in order to have access to the full descrip-
tions of the nodes.

11.1 STATEMENT ABSTRACTION

By abstraction of a statement (hypothesis or problem), we simply mean a shorter state-

ment summarizing its meaning. Consider, for example, the following hypothesis:
John Doe will stay on the faculty of George Mason University for the duration of the PhD
dissertation of Bob Sharp.
Any of the following shorter statements is an abstraction of the preceding
hypothesis:
John Doe will stay on the faculty of George Mason University.
John Doe will stay on the faculty.
John Doe will not leave.
The expert needs to define abstractions that are short enough to simplify the display of
the reasoning tree while still conveying the meaning of the original hypotheses. One
abstraction technique is to eliminate some of the words, as illustrated by the first two of
the preceding examples. Additionally, one may abstract phrases by using new words, as
illustrated by the last example.
The specific hypothesis, “John Doe will stay on the faculty of George Mason University for
the duration of the PhD dissertation of Bob Sharp,” is shown also in the upper-left part of

329

13:54:15,
.012
330 Chapter 11. Abstraction of Reasoning

Figure 11.1. As discussed in Section 9.10, from each specific hypothesis Disciple-EBR
automatically learns a general hypothesis with applicability conditions, which can be
further refined. The bottom-left part of Figure 11.1 shows the learned hypothesis whose
upper bound condition was further refined. Let us assume that the expert abstracts the
specific hypothesis to “John Doe will stay on the faculty” (see the upper-right part of
Figure 11.1). As a result, the agent automatically learns the abstraction pattern “?O1 will
stay on the faculty,” which corresponds to the abstraction of the learned hypothesis (see
the lower-right part of Figure 11.1).
The data structures and their relationships illustrated in Figure 11.1 have several
consequences. First, all the instances of the learned hypothesis will have the same
abstraction pattern. For example, the instance, “Dan Barker will stay on the faculty of
University of Virginia for the duration of the PhD dissertation of Sandra Lee,” will automatic-
ally be abstracted to “Dan Barker will stay on the faculty.”
Conversely, if you change the abstraction of a specific hypothesis, the pattern of the
learned abstraction will change accordingly. For example, if you now change the abstrac-
tion of “Dan Barker will stay on the faculty of University of Virginia for the duration of the
PhD dissertation of Sandra Lee,” to “Dan Barker will stay with University of Virginia,” then the
abstraction of the learned hypothesis in Figure 11.1 is automatically changed to “?O1 will
stay with ?O2.” Therefore, the abstraction of the specific hypothesis from Figure 11.1 is
automatically changed to “John Doe will stay with George Mason University.”
Thus, although at the beginning of this section we have provided several examples
of abstracting a hypothesis, each hypothesis may have only one abstraction at a time.
Notice, however, that the same abstract pattern may be an abstraction of different
learned hypotheses. Consequently, different specific hypotheses may also have the
same abstraction.
Finally, notice that the agent automatically abstracts the probabilistic solution of a
hypothesis to the actual probability. Thus, the solution, “It is almost certain that John Doe
will stay on the faculty of George Mason University for the duration of the PhD dissertation of
Bob Sharp,” is automatically abstracted to “almost certain.”

Specific hypothesis
John Doe will stay on the faculty of George Mason Abstract hypothesis
has as
University for the duration of the PhD dissertation of John Doe will stay on the faculty
abstraction
Bob Sharp.

Learned hypothesis Learned abstraction

has as
?O1 will stay on the faculty
?O1 will stay on the faculty of ?O2 for the abstraction
duration of the PhD dissertation of ?O3.
MAIN CONDITION
Var Lower bound Upper Bound
(PhD advisor,
?O1 (person)
associate professor)
?O2 (university) (organization)

?O3 (PhD student) (person)

Figure 11.1. Hypotheses and their abstractions.

13:54:15,
.012
 11.4. Hands On: Abstraction of Reasoning 331

11.2 REASONING TREE ABSTRACTION

One simple way of abstracting a reasoning step is to abstract all the hypotheses and to
eliminate the question/answer pair, as illustrated in Figure 11.2. The right-hand side of
Figure 11.2 shows the complete description of a reasoning step. The left-hand side shows
only the abstraction of the top hypothesis and the abstractions of its subhypotheses.
Notice that the meaning of an abstract subhypothesis is to be understood in the context of its
parent abstract hypothesis, and therefore it can be shorter. For example, “reasons” is under-
stood as “United States has reasons to be a global leader in wind power.” Notice also that the
abstraction of a hypothesis also includes the abstraction of its assessment, or “unknown” (if an
assessment has not yet been made). The assessments (or solutions) may also be made visible in
the detailed reasoning tree by clicking on [SHOW SOLUTIONS] at the top of the window.
One may also abstract an entire subtree, not just a reduction step. The right-hand side
of Figure 11.3 shows the reasoning tree that is abstracted in the left-hand side of the figure.
In particular, the abstract tree consists of the abstraction of the top hypothesis and the
abstractions of the leaf hypotheses from the detailed tree.

11.3 REASONING TREE BROWSING

Once a reasoning tree has been abstracted, it can be browsed as illustrated in Figure 11.4.
The left-hand side of Figure 11.4 shows the entire abstract tree. Each node in this tree is an
abstraction of a hypothesis and of its assessment (or “unknown” if an assessment has not
yet been made). The user can browse this tree by expanding or collapsing its nodes
through clicking on the + or – nodes. Once you click on a node in the abstract tree, such
as “desire: unknown,” the detailed description of that node is shown in the right-hand side.

11.4 HANDS ON: ABSTRACTION OF REASONING

The objective of this case study is to learn how to use Disciple-EBR to abstract a reasoning
tree. More specifically, you will learn how to:

 Introduce a hypothesis/problem into the abstract tree (see Operation 11.1)

 Define the abstraction of a hypothesis/problem (see Operation 11.2)
 Remove a hypothesis/problem from the abstract tree (see Operation 11.3)
 Browse the abstract tree

Figure 11.2. Abstraction of a reasoning step.

13:54:15,
.012
.012
13:54:15,

Figure 11.3. Abstraction of a reasoning tree.

.012
13:54:15,

Figure 11.4. Browsing a reasoning tree.

 334 Chapter 11. Abstraction of Reasoning

This case study will guide you through the process of abstracting the analysis of the
hypothesis, “John Doe would be a good PhD advisor for Bob Sharp,” with which you have
practiced in the previous case studies.
Start Disciple-EBR, select the case study knowledge base “14-Abstractions/Scen” and
proceed as indicated in the instructions at the bottom of the opened window.
This case study illustrates the following operations for the abstraction of reasoning:

Operation 11.1. Introduce a hypothesis into the abstract

reasoning tree
 In the Evidence workspace, in the right panel of the Reasoner module, right-click on
the hypothesis to be abstracted and select Add to TOC.
 Notice that the hypothesis is now displayed in the left panel, while the right panel no
longer displays its reasoning subtree.
 If you click on the hypothesis in the left panel, you will see its subtree in the right panel.

Operation 11.2. Modify the abstraction of a hypothesis

 In the Evidence workspace, in the left panel of the Reasoner module, right-click on the
hypothesis to be abstracted and select Modify.
 In the opened editor, edit the hypothesis to abstract it, making sure that all the
remaining instances and constants are recognized (i.e., appear in blue or green). Then
click outside the editor.

Operation 11.3. Remove a hypothesis from the abstract reasoning tree

 In the Evidence workspace, in the left panel of the Reasoner module, right-click on the
hypothesis to be deleted and select Remove from TOC.
 Notice that the hypothesis is no longer displayed in the left panel, while the right panel
now displays its reasoning subtree.

11.5 ABSTRACTION GUIDELINE

Guideline 11.1. Define short context-dependent hypothesis names for the

abstract tree
Because you always have access to the complete description of an abstracted hypothesis,
define the shortest possible abstraction, taking into account that its meaning is to be
understood in the context of its upper-level hypotheses. Consider, for example, the
abstract hypothesis “US Democratic Party,” selected in the left-hand side of Figure 11.5.
Its upper-level hypotheses indicate that this refers to one of the “major political parties,”

Figure 11.5. Understanding the meaning of an abstracted hypothesis in the context of its upper
hypotheses.

13:54:15,
.012
 11.7. Review Questions 335

which contributes to the “desire” component for the hypothesis “United States global
leader in wind power.”

11.6 PROJECT ASSIGNMENT 8

Abstract the reasoning trees developed in the previous project assignments.

11.7 REVIEW QUESTIONS

11.1. How is the abstraction of a hypothesis defined? Provide an example.

11.2. How is the abstraction of a reasoning tree defined? Provide an example.

11.3. Consider that the top hypothesis is “John Doe would be a good PhD advisor for Bob
Sharp,” and that “John Doe would be a good PhD advisor with respect to the
professional reputation criterion” is one of its subhypotheses. Indicate several pos-
sible abstractions of this subhypothesis, where each such abstraction is to be
understood in the context of the top hypothesis.

11.4. Consider the hypothesis, “John Doe would be a good PhD advisor with respect to
the professional reputation criterion,” and its subhypothesis, “John Doe would be a
good PhD advisor with respect to the peer opinion criterion.” Indicate abstractions of
these two hypotheses, where these abstractions are to be understood in the context
of the top hypothesis “John Doe would be a good PhD advisor for Bob Sharp.”

11.5. Consider the detailed reasoning tree from Figure 11.6. Provide a corresponding
abstract tree, thinking carefully about how to best abstract each of the hypotheses.

11.6. Consider the detailed reasoning tree from Figure 11.7. Provide a corresponding
abstract tree, thinking carefully about how to best abstract each of the hypotheses.

13:54:15,
.012
.012
13:54:15,

Figure 11.6. Detailed reasoning tree to be abstracted.

.012
13:54:15,

Figure 11.7. Another detailed reasoning tree to be abstracted.

 12 Disciple Agents

12.1 INTRODUCTION

The agent building theory, methodology and tool presented in this book evolved over
many years, with developments presented in numerous papers and a series of PhD
theses (Tecuci, 1988; Dybala, 1996; Hieb, 1996; Keeling, 1998; Boicu 2002; Bowman,
2002; Boicu, 2006; Le, 2008; Marcu, 2009). Although this book has emphasized the
development of Disciple agents for evidence-based reasoning applications, the learning
agent theory and technology are applicable and have been applied to a wide range of
knowledge-intensive tasks, such as those discussed in Section 1.6.2.
A previous book (Tecuci, 1998) presented the status of this work at that time and
included descriptions of Disciple agents for designing plans for loudspeaker manufactur-
ing, for assessing students’ higher-order thinking skills in history or in statistics, for
configuring computer systems, and for representing a virtual armored company com-
mander in distributed interactive simulations.
More recent Disciple agents and their applications include Disciple-WA, an agent for
the development of military engineering plans; Disciple-COA, for the critiquing of military
courses of action; Disciple-COG, for military center of gravity determination; Disciple
agents representing virtual experts for collaborative emergency response planning;
Disciple-LTA, for intelligence analysis; Disciple-FS, for regulatory compliance in financial
services industries; Disciple-WB, for assessing the believability of websites; and Disciple
agents for modeling the behavior of violent extremists.
The following sections present four of these agents and their applications. While all
illustrate the general agent development approach discussed in this book, they differ in
some of their capabilities and appearance, each reflecting a different stage or trajectory in
the development of the Disciple approach.

12.2 DISCIPLE-WA: MILITARY ENGINEERING PLANNING

12.2.1 The Workaround Planning Problem

The workaround planning problem consists of assessing how rapidly and by what method a
military unit can reconstitute or bypass damage to a transportation infrastructure, such as
a damaged and/or mined bridge, a blocked tunnel, or a cratered road (Cohen et al. 1998;
Jones, 1998).

338

13:54:21,
.013
12.2. Disciple-WA: Military Engineering Planning 339

The input to the agent includes three elements:

A description of the military unit that needs to work around some damage (e.g., an
armored tank brigade or a supply company)

A description of the damage (e.g., a span of the bridge is dropped and the area is
mined) and of the terrain (e.g., the soil type; the slopes of the riverbanks; the river’s
speed, depth, and width)

A detailed description of the resources in the area that could be used to repair the
damage. This includes a description of the engineering assets of the military unit that
has to work around the damage, as well as the descriptions of other military units in
the area that could provide additional resources

The output of the agent consists of the most likely repair strategies, each described in
terms of three elements:

A reconstitution schedule, giving the transportation capacity of the repaired link

(bridge, road, or tunnel), as a function of time, including both a minimum time and
an expected time

A time line of engineering actions to perform the repair, the minimum as well as the
expected time that these actions require, and the temporal constraints among them

A set of required assets for the entire strategy and for each action

Workaround plan generation requires detailed knowledge about the capabilities of

the engineering equipment and its use. For example, repairing damage to a bridge
typically involves different types of mobile bridging equipment and earth-moving equip-
ment. Each kind of mobile bridge takes a characteristic amount of time to deploy,
requires different kinds of bank preparation, and is owned by different echelons in the
military hierarchy. This information was available from military experts and U.S. Army
field manuals.
Figure 12.1 illustrates a sample input problem for Disciple-WA, a destroyed bridge over
a 25-meter-wide river that needs to be crossed by UNIT 91010.
Disciple-WA automatically generates all the possible plans and selects the one that has
the minimum duration. The best plan for the problem illustrated in Figure 12.1 is shown in

Site 103:cross-section
Damage 200: destroyed bridge

Near approach Bridge/River Far approach

(Right approach) gap = 25 meters (Left approach)
UNIT 91010 Site 108 Site 104

Site 100

Left bank
Right bank
Site 105
Site 107

River bed
Site 106

Figure 12.1. Sample input problem for Disciple-WA.

13:54:21,
.013
340 Chapter 12. Disciple Agents

Figure 12.2. Notice that, in Disciple-WA, we used a structured representation of the

actions, each consisting of a name and a sequence of feature-value pairs.
The top part of Figure 12.2 shows a summary of the plan. It consists of installing an
armored vehicle launched bridge (AVLB) over the river gap. It is estimated that this will
take a minimum of 11h:4m:58s, but it is expected to take 14h:25m:56s. UNIT91010 will
need the help of UNIT202, which has the AVLB equipment, and of UNIT201, which has a
bulldozer. After the bridge is installed, it will allow a traffic rate of 135.13 vehicles per hour.

Initial task: 25 m
WORKAROUND-DAMAGE 17 m
FOR-DAMAGE DAMAGE200 UNIT 91010
BY-INTERDICTED-UNIT UNIT91010 AVLB 70

Engineering action: INSTALL AVLB

MIN-DURATION 11H:4M:58S SITE 100
EXPECTED-DURATION 14H:25M:56S
RESOURCES-REQUIRED AVLB-UNIT202
BULLDOZER-UNIT201 SITE 103
LINK CAPACITY-AFTER-RECONSTRUCTION
135.13 VEHICLES/HR

Detailed plan:
S1 OBTAIN-OPERATIONAL-CONTROL-FROM-CORPS S7 NARROW-GAP-BY-FILLING-WITH-BANK
OF-UNIT UNIT202 FOR-GAP SITE103
BY-UNIT UNIT91010 FOR-BR-DESIGN AVLB70
MIN-DURATION 4H:0M:0S MIN-DURATION 5H:19M:44S
EXPECTED-DURATION 6H:0M:0S EXPECTED-DURATION 6H:7M:42S
TIME-CONSTRAINTS NONE RESOURCES-REQUIRED BULLDOZER-UNIT201
TIME-CONSTRAINTS AFTER S6
S2 MOVE-UNIT
FOR-UNIT UNIT202 S8 EMPLACE-AVLB
FROM-LOCATION SITE0 FOR-BR-DESIGN AVLB70
TO-LOCATION SITE100 MIN-DURATION 5M:0S
MIN-DURATION 1H:8M:14S EXPECTED-DURATION 10M:0S
EXPECTED-DURATION 1H:8M:14S RESOURCES-REQUIRED AVLB-UNIT202
TIME-CONSTRAINTS AFTER S1 TIME-CONSTRAINTS AFTER S3, S7

S3 REPORT-OBTAINED-EQUIPMENT S9 REPORT-EMPLACED-FIXED-BRIDGE
FOR-EQ-SET AVLB-UNIT202 FOR-MIL-BRIDGE AVLB-UNIT202
MIN-DURATION 0S MIN-DURATION 0S
EXPECTED-DURATION 0S EXPECTED-DURATION 0S
TIME-CONSTRAINTS AFTER S2 TIME-CONSTRAINTS AFTER S8

S4 OBTAIN-OPERATIONAL-CONTROL-FROM-CORPS S10 MOVE-EQUIPMENT-OVER-UNSTABILIZED-MIL-BRIDGE

OF-UNIT UNIT201 FOR-EQ-SET BULLDOZER-UNIT201
BY-UNIT UNIT91010 FOR-BR-DESIGN AVLB70
MIN-DURATION 4H:0M:0S MIN-DURATION 2M:0S
EXPECTED-DURATION 6H:0M:0S EXPECTED-DURATION 10M:0S
TIME-CONSTRAINTS NONE RESOURCES-REQUIRED AVLB-UNIT202
TIME-CONSTRAINTS AFTER S9
S5 MOVE UNIT
FOR-UNIT UNIT201 S11 MINOR-BANK-PREPARATION
FROM-LOCATION SITE0 OF-BANK SITE105
TO-LOCATION SITE100 MIN-DURATION 30M:0S
MIN-DURATION 1H:8M:14S EXPECTED-DURATION 50M:0S
EXPECTED-DURATION 1H:8M:14S RESOURCES-REQUIRED BULLDOZER-UNIT201
TIME-CONSTRAINTS AFTER S4 TIME-CONSTRAINTS AFTER S10

S6 REPORT-OBTAINED-EQUIPMENT S12 RESTORE-TRAFFIC-LINK

FOR-EQ-SET BULLDOZER-UNIT201 FOR-UNIT UNIT91010
MIN-DURATION 0S FOR-LINK AVLB70
EXPECTED-DURATION 0S LINK-CAPACITY 135.13 VEHICLES/H
TIME-CONSTRAINTS AFTER S5 MIN-DURATION 0S
EXPECTED-DURATION 0S

S1 S2 S3
S8 S9 S10 S11 S12
S4 S5 S6 S7

Figure 12.2. The best workaround plan.

13:54:21,
.013
 12.2. Disciple-WA: Military Engineering Planning 341

The detailed plan is shown under its summary and consists of twelve elementary
actions. UNIT91010 has to obtain operational control of UNIT202, which has the AVLB.
Then UNIT202 has to come to the site of the destroyed bridge. Also, UNIT91010 has to
obtain operational control of UNIT201, which has a bulldozer. Then UNIT201 will have
to move to the site of the destroyed bridge and to narrow the river gap from 25 meters to
17 meters. These actions can take place in parallel with the actions of bringing UNIT202 to
the bridge site, as shown at the bottom of Figure 12.2. Then the AVLB bridge is emplaced,
and the bulldozer moves over the bridge to clear the other side of the river in order to
restore the flow of traffic. This plan was generated by successively reducing the
WORKAROUND-DAMAGE task to simpler subtasks, until this task was reduced to the
twelve tasks shown in Figure 12.2.
The process of developing this agent has followed the general Disciple methodology, as
briefly discussed in the next sections.

12.2.2 Modeling the Workaround Planning Process

As discussed in Section 4.2, the subject matter expert has to model the planning process of
the agent as problem/task reduction and solution synthesis. The top-level organization
of the reasoning tree was already introduced in Figure 4.35 (p. 148), as part of Guideline
4.1. Figure 12.3 shows the reduction tree corresponding to the workaround problem from
Figure 12.1.
The top-level task is to work around whatever obstacle is encountered by UNIT91010.
This general task is successively reduced to simpler and simpler tasks, guided by ques-
tions whose answers either are obtained from the description of the problem or indicate
ways to perform an action. In this case, for instance, the obstacle is a damaged bridge but
without having any mines, at SITE100. Possible solutions to work around this obstacle are
to ford the river, to use a fixed bridge over the river, or to use a floating bridge. All these
alternatives are considered, but the tree in Figure 12.3 shows only the fixed bridge
solution. There are also alternatives ways to use a fixed bridge. In the case of the
considered problem, the river is too wide to install a fixed bridge directly, but it is narrow
enough to be further narrowed in order to install the bridge. Therefore, the solution is
first to reduce the gap by moving dirt into it with a bulldozer, and then to install an AVLB.
However, because UNIT91010 has neither the bulldozer nor the fixed bridge, it has to
obtain them from other units. As a result, the task, USE-FIXED-BRIDGE-WITH-GAP-
REDUCTION-OVER-GAP, is reduced to three subtasks:

OBTAIN-AVLB
OBTAIN-BULLDOZER-AND-NARROW-GAP
INSTALL-AVLB-OVER-NARROWED-GAP

Each of these subtasks is further reduced to elementary tasks that will constitute the
generated plan shown in Figure 12.2. For example, OBTAIN-AVLB is reduced to the
following elementary tasks:

S1 OBTAIN-OPERATIONAL-CONTROL-FROM-CORPS
S2 MOVE-UNIT
S3 REPORT-OBTAINED-EQUIPMENT

13:54:21,
.013
342 Chapter 12. Disciple Agents

WORKAROUND-OBSTACLE
BY-UNIT UNIT91010

What is the type of obstacle?

tunnel bridge at site100 crater

WORKAROUND-BRIDGE-OBSTACLE
AT-LOCATION SITE100
BY-UNIT UNIT91010

What is the type of bridge obstacle?

only mines damaged bridge but no mines mines and damaged bridge

WORKAROUND-UNMINED-DAMAGED-BRIDGE
AT-LOCATION SITE100
BY-UNIT UNIT91010

What is a possible solution?

fording floating bridge
fixed bridge over site103
Learned Rule
IF the task to accomplish is
WORKAROUND-UNMINED-DESTROYED-BRIDGE-WITH-FIXED-BRIDGE

WORKAROUND-UNMINED-DAMAGED-BRIDGE-WITH-FIXED-BRIDGE AT-LOCATION
FOR-GAP
?O1
?O2
BY-UNIT ?O3
AT-LOCATION SITE100 Plausible upper bound Plausible lower bound
FOR-GAP SITE103 ?O1
?O2
IS
IS
BRIDGE
CROSS-SECTION
?O1
?O2
IS
IS
SITE100
SITE103
HAS-WIDTH ?N4 HAS-WIDTH ?N4
BY-UNIT UNIT91010 ?O3 IS MILITARY-UNIT ?O3 IS UNIT91010
MAX-TRACKED-MLC ?N3 MAX-TRACKED-MLC ?N3
MAX-WHEELED-MLC ?N2 MAX-WHEELED-MLC ?N2
?O4 IS AVLB-EQ ?O4 IS AVLB-EQ
What strategy can be used? CAN-BUILD ?O5
MAX-REDUCIBLE-GAP ?N5
CAN-BUILD ?O5
MAX-REDUCIBLE-GAP ?N5
MAX-GAP ?N6 MAX-GAP ?N6

Fixed bridge with gap reduction because the gap is large but can be reduced ?O5 IS AVLB70
MLC-RATING ?N1
?O5 IS AVLB70
MLC-RATING ?N1
?N1 IS-IN [0.0 150.0] ?N1 IS-IN [70.0 70.0]
?N2 IS-IN [0.0 150.0] ?N2 IS-IN [20.0 20.0]
≤ ?N1 ≤ ?N1

USE-FIXED-BRIDGE-WITH-GAP-REDUCTION-OVER-GAP ?N3 IS-IN

≤
[0.0 150.0]
?N1
?N3 IS-IN
≤
[40.0 40.0]
?N1

AT-LOCATION SITE100 ?N4

?N5
IS-IN
IS-IN
[0.0 100.0]
[0.0 100.0]
?N4
?N5
IS-IN
IS-IN
[25.0 25.0]
[26.0 26.0]
≥ ?N4 ≥ ?N4
FOR-GAP SITE103 ?N6 IS-IN
<
[0.0 100.0)
?N4
?N6 IS-IN
<
[17.0 17.0]
?N4
BY-UNIT UNIT91010 THEN accomplish the task
WITH-BR-EQ AVLB-EQ USE-FIXED-BRIGE-WITH-GAP-REDUCTION-OVER-GAP
AT-LOCATION ?O1
FOR-GAP ?O2
BY-UNIT ?O3
WITH-BR-EQ ?O4

Does the unit have bulldozer and AVLB?

The unit has neither bulldozer nor AVLB

OBTAIN-BULLDOZER-AND-NARROW-GAP
AT-LOCATION SITE100
FOR-GAP SITE103
BY-UNIT UNIT91010

OBTAIN-AVLB INSTALL-AVLB-OVER-NARROWED-GAP
AT-LOCATION SITE100 AT-LOCATION SITE100
BY-UNIT UNIT91010 BY-UNIT UNIT91010

Which close-by unit has a bulldozer?

UNIT202
S3 REPORT-OBTAINED-EQUIPMENT
S1 OBTAIN-OPERATIONAL-CONTROL-FROM-CORPS FOR-EQ-SET AVLB-UNIT202
OF-UNIT UNIT202 MIN-DURATION 0S
BY-UNIT UNIT91010 EXPECTED-DURATION 0S
MIN-DURATION 4H:0M:0S TIME-CONSTRAINTS AFTER S2
EXPECTED-DURATION 6H:0M:0S
TIME-CONSTRAINTS NONE S2 MOVE-UNIT
FOR-UNIT UNIT202
FROM-LOCATION SITE0
TO-LOCATION SITE100
MIN-DURATION 1H:8M:14S
EXPECTED-DURATION 1H:8M:14S
TIME-CONSTRAINTS AFTER S1

Figure 12.3. Modeling the planning process.

Since these are elementary tasks that can be directly performed by the actual units, they
have known minimum and expected durations that are specified in military manuals.
Notice that these durations are specified in the descriptions of these tasks, as shown at the
bottom of Figure 12.3.
Similarly, OBTAIN-BULLDOZER-AND-NARROW-GAP from Figure 12.3 is reduced to
the tasks S4, S5, S6, and S7, as shown in Figure 12.2. Also, INSTALL-AVLB-OVER-
NARROWED-GAP is reduced to the tasks S8, S9, S10, S11, and S12.

13:54:21,
.013
 12.2. Disciple-WA: Military Engineering Planning 343

Notice that the generated plan is a partially ordered one, where some of the actions are
actually performed in parallel. Disciple-WA uses a very simple strategy to generate such
plans that does not require maintaining complex state descriptions and operators with
preconditions and effects, as other planning systems do. Instead, it generates REPORT
actions with duration 0 that mark the achievement of conditions used in ordering elemen-
tary actions generated in different parts of the planning tree. For example, the action S8
EMPLACE-AVLB from Figure 12.2 can be performed only after S3 (a REPORT action) and S7:

S3 REPORT-OBTAINED-EQUIPMENT
S7 NARROW-GAP-BY-FILLING-WITH-BANK

12.2.3 Ontology Design and Development

As discussed in Section 6.4, modeling trees like the one in Figure 12.3 provide a specification
of the concepts and features to be represented in the ontology of the agent. We have defined
these entities by using information from several military engineering manuals. We have also
imported elements of the military unit ontology, as well as various characteristics of military
equipment (such as their tracked and wheeled vehicle military load classes) from the Loom
server (MacGregor, 1991). Figure 12.4, for instance, presents a fragment of the hierarchy of
concepts from the ontology of the workaround agent. Included are several types of military
bridges that can be used to cross a river. Notice that, in Disciple-WA, we used “CLASS” to
refer to a concept, and “SUBCLASS-OF” to denote a “subconcept of” relation.

BRIDGE-CONCEPT TEMPORARY-TRAFFIC-LINK

MILTARY-BRIDGE

FLOATING-MILITARY-BRIDGE FIXED-MILITARY-BRIDGE

M4T6-BRIDGE RIBBON-BRIDGE AVLB MGB BAILEY-BRIDGE

M4T6-30 RB-HEAVY-45 AVLB70 MGB-16 BB-24 BB-30

MGB-SS16 BB-TT24

MGB-SS16-11 MGB-SS16-12

Figure 12.4. A fragment of the concept hierarchy.

13:54:21,
.013
344 Chapter 12. Disciple Agents

Figure 12.5 contains the descriptions of two concepts from the hierarchy in
Figure 12.4, AVLB and AVLB70. An AVLB is a subclass of fixed military bridge that
has additional features. AVLB70 is a subclass of AVLB bridge. Each such concept
inherits all of the features of its superconcepts. Therefore, all the features of AVLB
are also features of AVLB70.
The features are defined in the same way as the concepts, in terms of more general
features. Figure 12.6, for instance, presents a fragment of the feature hierarchy. Two
important characteristics of any feature are its domain (the set of objects that could have
this feature) and its range (the set of possible values of the feature). The features may also
specify functions for computing their values.

AVLB SUBCLASS-OF FIXED-MILITARY-BRIDGE

MIN-CROSSING-TIME-FOR-UNSTABILIZED-END 2 MIN
EXPECTED-CROSSING-TIME-FOR-UNSTABILIZED-END 10 MIN
MIN-EMPLACEMENT-TIME 5 MIN
EXPECTED-EMPLACEMENT-TIME 10 MIN
MAX-DOWNHILL-SLOPE-FOR-EQ 19 %
MAX-TRANSVERSE-SLOPE 11 %
MAX-UPHILL-SLOPE-FOR-EQ 28 %

AVLB70 SUBCLASS-OF AVLB

HAS-WIDTH 19.2 METERS
MAX-GAP 17 METERS
MAX-REDUCIBLE-GAP 26 METERS
MLC-RATING 70 TONS
WEIGHT 15 TONS

Figure 12.5. Descriptions of two concepts from the hierarchy in Figure 12.4.

DIMENSION

SUBCLASS-OF SUBCLASS-OF SUBCLASS-OF

LENGTH WIDTH HEIGHT

SUBCLASS-OF SUBCLASS-OF

GAP-WIDTH HAS-WIDTH

SUBCLASS-OF SUBCLASS-OF

MAX-GAP MAX-REDUCIBLE-GAP

Figure 12.6. A fragment of the feature hierarchy.

13:54:21,
.013
 12.2. Disciple-WA: Military Engineering Planning 345

12.2.4 Rule Learning

As presented in Chapters 9 and 10, a Disciple agent learns rules from individual reasoning
steps and their explanations, as illustrated in the middle of Figure 12.3. That task reduction
step is shown again in the top part of Figure 12.7. It states that in order to work around the
damaged bridge at SITE100, one can use a bridge equipment of type AVLB-EQ and to
reduce the size of the gap.
The bottom part of Figure 12.7 shows the explanation of this task reduction step,
obtained through an interaction with the expert. The first two explanation pieces justify
why one can use gap reduction.
The width of the SITE103 gap is 25 meters, and the AVLB-EQ allows building a bridge of
type AVLB70 that can span gaps only as long as 17 meters. Therefore, the gap is too wide to
install an AVLB70 directly. However, any gap that is smaller than 26 meters can be reduced
to a 17 meter gap on which one can install an AVLB70 bridge.
The next two explanation pieces show that an AVLB70 bridge is strong enough to
sustain the vehicles of UNIT91010.
The maximum load class of the wheeled vehicles of UNIT91010 is 20 tons, and AVLB70
can sustain vehicles with a load of up to 70 tons. Similarly, the AVLB70 bridge can sustain
the tracked vehicles of UNIT91010.

TASK
WORKAROUND-UNMINED-DESTROYED-BRIDGE-WITH-FIXED-BRIDGE
AT-LOCATION SITE100
FOR-GAP SITE103
BY-UNIT UNIT91010

Question: What strategy can be used?

Answer: Fixed bridge with gap reduction because the gap is
large but can be reduced

SUBTASK
USE-FIXED-BRIDGE-WITH-GAP-REDUCTION-OVER-GAP
AT-LOCATION SITE100
FOR-GAP SITE103
BY-UNIT UNIT91010
WITH-BR-EQ AVLB-EQ

EXPLANATIONS

SITE103 HAS-WIDTH 25M,

AVLB-EQ CAN-BUILD AVLB70 MAX-GAP 17M < 25M

SITE103 HAS-WIDTH 25M,

AVLB-EQ CAN-BUILD AVLB70 MAX-REDUCIBLE-GAP 26M ≥ 25M

UNIT91010 MAX-WHEELED-MLC 20T,

AVLB-EQ CAN-BUILD AVLB70 MLC-RATING 70T ≥ 20T

UNIT91010 MAX-TRACKED-MLC 40T,

AVLB-EQ CAN-BUILD AVLB70 MLC-RATING 70T ≥ 40T

Figure 12.7. An example of task reduction and its explanation.

13:54:21,
.013
 346 Chapter 12. Disciple Agents

Figure 12.8 shows the plausible version space rule learned from the task reduction
example and its explanation in Figure 12.7.

12.2.5 Experimental Results

Disciple-WA (Tecuci et al., 1999; 2000) was tested together with three other systems in a
two-week intensive study, in June 1998, as part of the Defense Advance Research Projects
Agency’s (DARPA) annual High Performance Knowledge Bases (HPKB) program evalu-
ation (Cohen et al. 1998). The evaluation, performed by Alphatech, consisted of two
phases, each comprising a test and a retest. In the first phase, the systems were tested
on twenty problems that were similar to those used for systems development. Then the
solutions were provided, and the developers had one week to improve their systems,
which were tested again on the same problems. In the second phase, the systems were
tested on five new problems, partially or completely out of the scope of the systems. For
instance, they specified a new type of damage (cratered roads) or required the use of new
types of engineering equipment (TMM bridges, ribbon rafts, and M4T6 rafts). Then again
the correct solutions were provided, and the developers had one week to improve and
develop their systems, which were tested again on the same five problems and five new
ones. Solutions were scored along five equally weighted dimensions: (1) generation of
the best workaround solutions for all the viable options; (2) correctness of the overall

IF the task to accomplish is

WORKAROUND-UNMINED-DESTROYED-BRIDGE-WITH-FIXED-BRIDGE
AT-LOCATION ?O1
FOR-GAP ?O2
BY-UNIT ?O3

Plausible upper bound Plausible lower bound

?O1 IS BRIDGE ?O1 IS SITE100
?O2 IS CROSS-SECTION ?O2 IS SITE103
HAS-WIDTH ?N4 HAS-WIDTH ?N4
?O3 IS MILITARY-UNIT ?O3 IS UNIT91010
MAX-TRACKED-MLC ?N3 MAX-TRACKED-MLC ?N3
MAX-WHEELED-MLC ?N2 MAX-WHEELED-MLC ?N2
?O4 IS AVLB-EQ ?O4 IS AVLB-EQ
CAN-BUILD ?O5 CAN-BUILD ?O5
MAX-REDUCIBLE-GAP ?N5 MAX-REDUCIBLE-GAP ?N5
MAX-GAP ?N6 MAX-GAP ?N6
?O5 IS AVLB70 ?O5 IS AVLB70
MLC-RATING ?N1 MLC-RATING ?N1
?N1 IS-IN [0.0 150.0] ?N1 IS-IN [70.0 70.0]
?N2 IS-IN [0.0 150.0] ?N2 IS-IN [20.0 20.0]
≤ ?N1 ≤ ?N1
?N3 IS-IN [0.0 150.0] ?N3 IS-IN [40.0 40.0]
≤ ?N1 ≤ ?N1
?N4 IS-IN [0.0 100.0] ?N4 IS-IN [25.0 25.0]
?N5 IS-IN [0.0 100.0] ?N5 IS-IN [26.0 26.0]
≥ ?N4 ≥ ?N4
?N6 IS-IN [0.0 100.0) ?N6 IS-IN [17.0 17.0]
< ?N4 < ?N4

THEN accomplish the task

USE-FIXED-BRIGE-WITH-GAP-REDUCTION-OVER-GAP
AT-LOCATION ?O1
FOR-GAP ?O2
BY-UNIT ?O3
WITH-BR-EQ ?O4

Figure 12.8. Rule learned by Disciple-WA from the example and the explanation in Figure 12.7.

13:54:21,
.013
12.2. Disciple-WA: Military Engineering Planning 347

time estimate for each workaround solution; (3) correctness of each solution step;
(4) correctness of temporal constraints among these steps; and (5) appropriateness of
engineering resources used. Scores were assigned by comparing the systems’ answers with
those of Alphatech’s human expert. Bonus points were awarded when systems gave better
answers than the expert, and these answers were used as standard for the next phase of
the evaluation.
The participating teams were not uniform in terms of prior system development and
human resources. Consequently, only one of them succeeded to enter the evaluation with
a system that had a fully developed knowledge base. The other three teams (including the
Disciple team) entered the evaluation with systems that had incompletely developed
knowledge bases. Figure 12.9 shows a plot of the overall coverage of each system against
the overall correctness of that system for each of the two phases of the evaluation.
The Disciple team entered the evaluation with a workaround agent the knowledge base
of which was covering only about 40 percent of the workaround domain (equivalent to
11,841 binary predicates). The coverage of our agent was declared prior to each release of
the testing problems, and all the problems falling within its scope were attempted and
scored. During the evaluation period, we continued to extend the knowledge base to cover
more of the initially specified domain, in addition to the developments required by the
modification phase. At the end of the two weeks of evaluation, the knowledge base of our
agent grew to cover about 80 percent of the domain (equivalent to 20,324 binary predi-
cates). This corresponds to a rate of knowledge acquisition of approximately 787 binary
predicates per day, as indicated in Figure 12.10. This result supports the claim that the
Disciple approach enables rapid acquisition of relevant problem-solving knowledge from
subject matter experts.
With respect to the quality of the generated solutions, within its scope, the Disciple-WA
agent performed at the level of the human expert. There were several cases during the
evaluation period where the Disciple-WA agent generated more correct or more complete

Correctness
Disciple Disciple
1.0 initial final

AIAI-initial ISI-final
.75 ISI-initial
TFS-final

.5

.25
TFS-initial

.25 .5 .75 1.0

Coverage

Figure 12.9. Evaluation results for the coverage of the problem space and the correctness of the
solutions (reprinted with permission from Eric Jones).

13:54:21,
.013
 348 Chapter 12. Disciple Agents

25000

KB Size – Predicates
20000 y = 787.63x + 5521.6

15000

10000

5000

16

18

20

22

24

26

28
June 14

June 30
Days
Figure 12.10. Knowledge base development time.

solutions than those of the human expert. There were also cases where the agent gener-
ated new solutions that the human expert did not initially consider. For instance, it
generated solutions to work around a cratered road by emplacing a fixed bridge over
the crater in a way similar to emplacing a fixed bridge over a river gap. Or, in the case of
several craters, it generated solutions where some of the craters were filled while for others
fixed bridges were emplaced. These solutions were adopted by the expert and used as
standard for improving all the systems. For this reason, although the agent also made
some mistakes, the overall correctness of its solutions was practically as high as that of the
expert’s solutions. This result supports the second claim that the acquired problem-
solving knowledge is of good enough quality to ensure a high degree of correctness of
the solutions generated by the agent.
Finally, our workaround generator had also a very good performance, being able to
generate a solution in about 0.3 seconds, on a medium-power PC. This supports the third
claim, that the acquired problem-solving knowledge ensures high performance of the
problem solver.
Based on the evaluation results, the Disciple-WA agent was selected by DARPA and
Alphatech to be further extended and was integrated by Alphatech into a larger system that
supports air campaign planning by the Joint Force Air Component Commander (JFACC)
and his or her staff. The integrated system was one of the systems selected to be
demonstrated at EFX’98, the Air Force’s annual showcase of promising new technologies.

12.3 DISCIPLE-COA: COURSE OF ACTION CRITIQUING

12.3.1 The Course of Action Critiquing Problem

A military course of action (COA) is a preliminary outline of a plan for how a military unit
might attempt to accomplish a mission. A COA is not a complete plan in that it leaves out
many details of the operation, such as exact initial locations of friendly and enemy forces.
After receiving orders to plan for a mission, a Commander and staff complete a detailed
and practiced process of analyzing the mission. They conceive and evaluate potential
COAs, select a COA, and prepare detailed plans to accomplish the mission based on the
selected COA. The general practice is for the staff to generate several COAs for a mission

13:54:21,
.013
12.3. Disciple-COA: Course of Action Critiquing 349

and then to compare those COAs based on many factors, including the situation, the
commander’s guidance, the principles of war, and the tenets of Army operations. Then the
Commander makes the final decision on which COA will be used to generate his or her
plan based on the recommendations of the staff and his or her own experience with the
same factors considered by the staff (Jones, 1998).
The COA critiquing problem consisted of developing a knowledge-based agent that can
automatically critique COAs for ground force operations, can systematically assess
selected aspects of a COA, and can suggest repairs to it. The role of this agent is to act
as an assistant to the military Commander, helping the Commander in choosing between
several COAs under consideration for a certain mission. The agent could also help military
students learn how to develop courses of action.
The input to the COA critiquing agent consists of the description of a COA that includes
the following aspects:

1. The COA sketch, such as the one in the top part of Figure 12.11, which is a
graphical depiction of the preliminary plan being considered. It includes enough
of the high-level structure and maneuver aspects of the plan to show how the
actions of each unit fit together to accomplish the overall purpose, while omitting
much of the execution detail that will be included in the eventual operational plan.
The three primary elements included in a COA sketch are (a) control measures that
limit and control interactions between units; (b) unit graphics that depict known,
initial locations and makeup of friendly and enemy units; and (c) mission graphics
that depict actions and tasks assigned to friendly units. The COA sketch is drawn
using a palette-based sketching utility.
2. The COA statement, such as the partial one shown in the bottom part of
Figure 12.11, which clearly explains what the units in a course of action will do to
accomplish the assigned mission. This text includes a description of the mission
and the desired end state, as well as standard elements that describe purposes,
operations, tasks, forms of maneuver, units, and resources to be used in the COA.
The COA statement is expressed in a restricted but expressive subset of English.
3. Selected products of mission analysis, such as the areas of operations of the units,
avenues of approach, key terrain, unit combat power, and enemy COAs.

Based on this input, the critiquing agent has to assess various aspects of the COA, such as
its viability (i.e., its suitability, feasibility, acceptability, and completeness), its correctness
(which considers the array of forces, the scheme of maneuver, and the command and
control), and its strengths and weaknesses with respect to the principles of war and the
tenets of Army operations. The critiquing agent should also be able to clearly justify the
assessments made and to propose improvements to the COA.
Disciple-COA was developed in the DARPA’s HPKB program to solve part of the COA
critiquing problem (Cohen et al., 1998). In particular, Disciple-COA identifies the strengths
and the weaknesses of a course of action with respect to the principles of war and the tenets
of Army operations (FM 100–5, 1993). There are nine principles of war: objective, offensive,
mass, economy of force, maneuver, unity of command, security, surprise, and simplicity.
They provide general guidance for the conduct of war at the strategic, operational, and
tactical levels. For example, Table 12.1 provides the definition of the principle of mass.
The tenets of Army operations describe the characteristics of successful operations.
They are initiative, agility, depth, synchronization, and versatility.

13:54:21,
.013
350 Chapter 12. Disciple Agents

Mission: BLUE-BRIGADE2 attacks to penetrate RED-MECH-REGIMENT2 at 130600 Aug in order to enable the completion
of seize OBJ-SLAM by BLUE-ARMOR-BRIGADE1.
Close: BLUE-TASK-FORCE1, a balanced task force (MAIN-EFFORT) attacks to penetrate RED-MECH-COMPANY4, then
clears RED-TANK-COMPANY2 in order to enable the completion of seize OBJ-SLAM by BLUE-ARMOR-
BRIGADE1.
BLUE-TASK-FORCE2, a balanced task force (SUPPORTING-EFFORT1) attacks to fix RED-MECH-COMPANY1
and RED-MECH-COMPANY2 and RED-MECH-COMPANY3 in order to prevent RED-MECH-COMPANY1 and
RED-MECH-COMPANY2 and RED-MECH-COMPANY3 from interfering with conducts of the MAIN-EFFORT1,
then clears RED-MECH-COMPANY1 and RED-MECH-COMPANY2 and RED-MECH-COMPANY3 and RED-TANK-
COMPANY1.
BLUE-MECH-BATTALION1, a mechanized infantry battalion (SUPPORTING-EFFORT2) attacks to fix RED-MECH-
COMPANY5 and RED-MECH-COMPANY6 in order to prevent RED-MECH-COMPANY5 and RED-MECH-
COMPANY6 from interfering with conducts of the MAIN-EFFORT1, then clears RED-MECH-COMPANY5 and RED-
MECH-COMPANY6 and RED-TANK-COMPANY3
Reserve: The reserve, BLUE-MECH-COMPANY8, a mechanized infantry company, follows MAIN-EFFORT, and is prepared
to reinforce MAIN-EFFORT.
Security: SUPPORTING-EFFORT1 destroys RED-CSOP1 prior to begin moving across PL-AMBER by MAIN-EFFORT in
order to prevent RED-MECH-REGIMENT2 from observing MAIN-EFFORT.
SUPPORTING-EFFORT2 destroys RED-CSOP2 prior to begin moving across PL-AMBER by MAIN-EFFORT in
order to prevent RED-MECH-REGIMENT2 from observing MAIN-EFFORT.
Deep: Deep operations will destroy RED-TANK-COMPANY1 and RED-TANK-COMPANY2 and RED-TANK-COMPANY3.
Rear: BLUE-MECH-PLT1, a mechanized infantry platoon secures the brigade support area.
Fire: Fires will suppress RED-MECH-COMPANY1 and RED-MECH-COMPANY2 and RED-MECH-COMPANY3 and
RED-MECH-COMPANY4 and RED-MECH-COMPANY5 and RED-MECH-COMPANY6.
End State: At the conclusion of this operation, BLUE-BRIGADE2 will enable accomplishing conducts forward passage of lines
through BLUE-BRIGADE2 by BLUE-ARMOR-BRIGADE1.
MAIN-EFFORT will complete to clear RED-MECH-COMPANY4 and RED-TANK-COMPANY2.
SUPPORTING-EFFORT1 will complete to clear RED-MECH-COMPANY1 and RED-MECH-COMPANY2 and RED-
MECH-COMPANY3 and RED-TANK-COMPANY1.
SUPPORTING-EFFORT2 will complete to clear RED-MECH-COMPANY5 and RED-MECH-COMPANY6 and RED-
TANK-COMPANY3.

Figure 12.11. COA sketch and a fragment of a COA statement (reprinted with permission from
Eric Jones).

Table 12.2, for instance, shows some of the strengths of the COA from Figure 12.11 with
respect to the principle of mass, identified by Disciple-COA.
In addition to generating answers in natural language, Disciple-COA also provides the
reference material based on which the answers are generated, as shown in the bottom part
of Table 12.2. Also, the Disciple-COA agent can provide justifications for the generated

13:54:21,
.013
 12.3. Disciple-COA: Course of Action Critiquing 351

Table 12.1 The Principle of Mass (from FM 100–5)

Mass the effects of overwhelming combat power at the decisive place and time.
Synchronizing all the elements of combat power where they will have decisive effect on an enemy
force in a short period of time is to achieve mass. To mass is to hit the enemy with a closed fist, not
poke at him with fingers of an open hand. Mass must also be sustained so the effects have staying
power. Thus, mass seeks to smash the enemy, not sting him. This results from the proper
combination of combat power with the proper application of other principles of war. Massing
effects, rather than concentrating forces, can enable numerically inferior forces to achieve decisive
results, while limiting exposure to enemy fire.

Table 12.2 Strengths of the COA from Figure 12.11 with Respect to the Principle of Mass,
Identified by Disciple-COA

Major Strength: There is a major strength in COA411 with respect to mass because
BLUE-TASK-FORCE1 is the MAIN-EFFORT1 and it acts on the decisive point of the COA (RED-
MECH-COMPANY4) with a force ratio of 10.6, which exceeds a recommended force ratio of 3.0.
Additionally, the main effort is assisted by supporting action SUPPRESS-MILITARY-TASK1, which
also acts on the decisive point. This is good evidence of the allocation of significantly more than
minimum combat power required at the decisive point and is indicative of the proper
application of the principle of mass.
Strength: There is a strength in COA411 with respect to mass because BLUE-TASK-FORCE1 is the
main effort of the COA and it has been allocated 33% of available combat power, but this is
considered just a medium-level weighting of the main effort.
Strength: There is a strength in COA411 with respect to mass because BLUE-MECH-COMPANY8 is a
COMPANY-UNIT-DESIGNATION level maneuver unit assigned to be the reserve. This is
considered a strong reserve for a BRIGADE-UNIT-DESIGNATION–level COA and would be
available to continue the operation or exploit success.
Reference: FM 100–5 pg 2–4, KF 113.1, KF 113.2, KF 113.3, KF 113.4, KF 113.5 – To mass is to
synchronize the effects of all elements of combat power at the proper point and time to achieve
decisive results. Observance of the principle of mass may be evidenced by allocation to the main
effort of significantly greater combat power than the minimum required throughout its mission,
accounting for expected losses. Mass is evidenced by the allocation of significantly more than
the minimum combat power required at the decisive point.

answers at three levels of detail, from a very abstract one that shows the general line of
reasoning followed, to a very detailed one that indicates each of the knowledge pieces
used in generating the answer.

12.3.2 Modeling the COA Critiquing Process

To critique a course of action with respect to a specific principle or tenet, one needs a
certain amount of information about that course of action, information related to that
principle or tenet. This information is obtained by asking a series of questions. The answer
to each question allows one to reduce the current critiquing task to a more specific and

13:54:21,
.013
 352 Chapter 12. Disciple Agents

simpler one. This process continues until one has enough information to recognize a
weakness or a strength. Consider, for example, the principle of surprise, whose definition
is provided in Table 12.3. As you can see, this is a very general description. How to apply
this general principle in actual situations is knowledge that is learned by military officers
during their lifetime. Therefore, developing an agent able to identify to what extent a
specific COA conforms to the principle of surprise involves capturing and representing the
knowledge of a military expert into the agent’s knowledge base.
Guided by this general definition and the COA from Figure 12.11, our subject matter
expert (Colonel Michael Bowman) has developed the reduction tree from Figure 12.12.
Notice how each successive question identifies a surprise-related feature of the COA until
a strength is recognized.
Through this kind of modeling of the COA critiquing process, each leaf may lead to
the identification of a strength or weakness. Then, the bottom-up solution synthesis
process consists only in accumulating all the identified strengths and weaknesses, just
as in the case of Disciple-WA, where the solution synthesis process accumulates the
elementary actions.
Notice also that, as in the case of Disciple-WA, the tasks are structured, consisting of a
name and a sequence of feature-value pairs.

12.3.3 Ontology Design and Development

For Disciple-COA, an initial ontology was defined by importing the input ontology built by
Teknowledge and Cycorp for the COA challenge problem (Boicu et al., 1999). The
imported ontology was further developed by using the ontology building tools of
Disciple-COA. Figure 12.13, for instance, presents a fragment of the COA ontology.
The top part represents the upper level of the ontology that identifies the types of
concepts represented in the ontology (called classes in Disciple-COA). They include
GEOGRAPHICAL-REGION, ORGANIZATION, EQUIPMENT, and ACTION. Each of these
concepts is the top of a specialized hierarchy, such as the hierarchy of organizations, a
fragment of which is shown in the left part of Figure 12.13. The leaves of this hierarchy are
specific military units (e.g., BLUE-TASK-FORCE1), corresponding to a specific COA to be
critiqued by Disciple-COA. Each concept and instance of the object hierarchy is described

Table 12.3 The Principle of Surprise (from FM 100–5)

Strike the enemy at a time or place or in a manner for which he is unprepared.

Surprise can decisively shift the balance of combat power. By seeking surprise, forces can achieve
success well out of proportion to the effort expended. Rapid advances in surveillance technology
and mass communication make it increasingly difficult to mask or cloak large-scale marshaling or
movement of personnel and equipment. The enemy need not be taken completely by surprise, but
only become aware too late to react effectively. Factors contributing to surprise include speed,
effective intelligence, deception, application of unexpected combat power, operations security
(OPSEC), and variations in tactics and methods of operation. Surprise can be in tempo, size of
force, direction or location of main effort, and timing. Deception can aid the probability of
achieving surprise.

13:54:21,
.013
12.3. Disciple-COA: Course of Action Critiquing 353

ASSESS-COA-WRT-PRINCIPLE-OF-SURPRISE
FOR-COA COA411

Which is an aspect that characterizes surprise?

The presence of surprise factors The presence of deception actions

ASSESS-SURPRISE-WRT-THE- ASSESS-SURPRISE-WRT-THE-
PRESENCE-OF-SURPRISE-FACTORS PRESENCE-OF-DECEPTION-ACTIONS
FOR-COA COA411 FOR-COA COA411

Enemy reconnaissance The application of surprising levels of combat power

ASSESS-SURPRISE-WRT-COUNTERING- ASSESS-SURPRISE-WRT-THE-APPLICATION-
ENEMY-RECONNAISANCE OF-SURPRISING-LEVELS-OF-COMBAT-POWER
FOR-COA COA411 FOR-COA COA411

Is an enemy reconnaissance unit present?

No Yes, RED-CSOP1, which is performing the reconnaissance action SCREEN1

ASSESS-SURPRISE-WHEN-ENEMY-RECON-IS-PRESENT
FOR-COA COA411
FOR-UNIT RED-CSOP1
FOR-RECON-ACTION SCREEN1

Is the enemy reconnaissance unit destroyed?

Yes, RED-CSOP1 is destroyed by DESTROY1 No

REPORT-STRENGTH-IN-SURPRISE-BECAUSE-OF-COUNTERING-ENEMY-RECON
FOR-COA COA411
FOR-UNIT RED-CSOP1
FOR-RECON-ACTION SCREEN1
FOR-ACTION DESTROY1
WITH-IMPORTANCE “high”
Strength: There is a strength with respect to surprise in COA411 because it contains aggressive
security/counter-reconnaissance plans, destroying enemy intelligence collection units and activities.
Intelligence collection by RED-CSOP1 through SCREEN1 will be disrupted by its destruction by
DESTROY1. This and similar actions prevent the enemy for ascertaining the nature and intent of
friendly operations, thereby increasing the likelihood that the enemy will be surprised. This is a strength
of high importance.
Reference: FM 100-5 pg 2-5, KF 118.1, KF 118.2, KF 118.3 - Surprise is achieved by striking/engaging
the enemy in a time, place or manner for which he is unprepared. The enemy can be surprised by the
tempo of the operation, the size of the force, the direction or location of the main effort, and timing.
Factors contributing to surprise include speed, effective intelligence, deception, application of
unexpected combat power, operations security, and variations in tactics and methods of operation.

Figure 12.12. Sample modeling of the COA critiquing process.

by specific features and values. For instance, the bottom part of Figure 12.13 shows
the description of the specific military unit called BLUE-TASK-FORCE1. BLUE-TASK-
FORCE1 is described as being both an ARMORED-UNIT-MILITARY-SPECIALTY and a
MECHANIZED-INFANTRY-UNIT-MILITARY-SPECIALTY. The other features describe
BLUE-TASK-FORCE1 as being at the battalion level; belonging to the blue side; being
designated as the main effort of the blue side; performing two tasks, PENETRATE1 and
CLEAR1; having a regular strength; and having four other units under its operational control.
The values of the features of BLUE-TASK-FORCE1 are themselves described in the same way.
For instance, one of the tasks performed by BLUE-TASK-FORCE1 is PENETRATE1.

13:54:21,
.013
 OBJECT

GEOGRAPHICAL-REGION
PLAN

ORGANIZATION
ACTION
EQUIPMENT COA-SPECIFICATION-MICROTHEORY
PURPOSE

MILITARY-EVENT
MODERN-MILITARY-ORGANIZATION
MILITARY-EQUIPMENT MILITARY-PURPOSE

MILITARY-OPERATION
MODERN-MILITARY-UNIT-DEPLOYABLE

MILITARY-TASK
MILITARY-MANEUVER

MANEUVER-UNIT-MILITARY-SPECIALTY AVIATION-UNIT-MILITARY-SPECIALTY

COMPLEX-MILITARY-TASK

SUBCLASS-OF
INFANTRY-UNIT-MILITARY-SPECIALTY
ARMORED-UNIT-MILITARY-SPECIALTY SUBCLASS-OF
MILITARY-ATTACK
MECHANIZED-INFANTRY-UNIT-MILITARY-SPECIALTY
.013

SUBCLASS-OF
BLUE-ARMOR-BRIGADE2

INDICATES-MISSION-TYPE military offensive

INSTANCE-OF BLUE-TASK-FORCE2 PENETRATE-MILITARY-TASK operation
13:54:21,

INSTANCE-OF IS-OFFENSIVE-ACTION-FOR
ECHELON-OF-UNIT military offensive
BATALLION-UNIT-DESIGNATION
operation
SOVEREIGN-ALLEGIANCE-OF-ORG BLUE-SIDE RECOMMENDED-FORCE-RATIO
INSTANCE-OF 3
ASSIGNMENT HAS-SURPRISE-FORCE-RATIO
MAIN-EFFORT1 6
TASK
BLUE-TASK-FORCE1 CLEAR1
TASK OBJECT-ACTED-ON
PENETRATE1 RED-MECH-COMPANY4
TROOP-STRENGTH-OF-UNIT
REGULAR-STATUS FORCE-RATIO
10.6
OPERATIONAL-CONTROL-MILITARY-ORG BLUE-MECH-COMPANY1
IS-TASK-OF-OPERATION
OPERATIONAL-CONTROL-MILITARY-ORG ATTACK2
BLUE-MECH-COMPANY2
OPERATIONAL-CONTROL-MILITARY-ORG
BLUE-ARMOR-COMPANY1
OPERATIONAL-CONTROL-MILITARY-ORG
BLUE-ARMOR-COMPANY2

Figure 12.13. Fragment of the COA ontology.

 12.3. Disciple-COA: Course of Action Critiquing 355

PENETRATE1 is defined as being a penetration task, and therefore inherits all the features of
the penetration tasks, in addition to the features that are directly associated with it.
The hierarchy of objects is used as a generalization hierarchy for learning by the
Disciple-COA agent. For instance, one way to generalize an expression is to replace an
object with a more general one from such a hierarchy. In particular, PENETRATE1 from
the bottom-right side of Figure 12.13 can be generalized to PENETRATE-MILITARY-TASK,
COMPLEX-MILITARY-TASK, MILITARY-MANEUVER, etc. The goal of the learning pro-
cess is to select the right generalization.
The features used to describe the objects are themselves represented in the feature
hierarchy.

12.3.4 Training the Disciple-COA Agent

The next step in the development of the Disciple-COA critiquer was to teach it to critique
COAs with respect to the principles of war and the tenets of Army operations, based on the
developed models. The expert loads the description of a specific COA, such as COA411
represented in Figure 12.11, and then invokes the Problem Solver with a task of critiquing
the COA with respect to a certain principle or tenet. Disciple-COA uses its task reduction
rules to reduce the current task to simpler tasks, showing the expert the reductions found.
The expert may accept a reduction proposed by the agent, reject it, or decide to define a
new reduction. From each such interaction, Disciple-COA either refines a previously
learned rule or learns a new task reduction rule. After a new rule is learned or an existing
rule is refined, the Problem Solver resumes the task reduction process until a solution to
the initial problem is found.
Initially Disciple-COA does not contain any rules. Therefore, all the problem-solving
steps (i.e., task reductions) must be provided by the expert, as illustrated in Figure 12.14
and explained in the following.
To assess COA411 with respect to the principle of surprise, the expert and Disciple-COA
need a certain amount of information, which is obtained by asking a series of questions, as
already illustrated in Section 12.3.2. The answer to each question allows the expert and the
agent to reduce the current assessment task to a more detailed one. This process con-
tinues until the expert and Disciple-COA have enough information about COA411 to make
the assessment. As shown in Figure 12.14, the initial task is reduced to that of assessing the
surprise of COA411 with respect to the countering of enemy reconnaissance. Then one
asks whether there is any enemy reconnaissance unit present in COA411. The answer
identifies RED-CSOP1 as being such a unit because it is performing the task SCREEN1.
Therefore, the task of assessing surprise for COA411 with respect to countering enemy
reconnaissance is now reduced to the better defined task of assessing surprise when
enemy reconnaissance is present. The next question to ask is whether the enemy recon-
naissance unit is destroyed or not. In the case of COA411, RED-CSOP1 is destroyed by the
task DESTROY1. Therefore, one can conclude that there is a strength in COA411 with
respect to the principle of surprise because the enemy reconnaissance unit is countered.
Figure 12.15 illustrates the process of teaching Disciple-COA. The left-hand side repre-
sents the reasoning process of the expert, the question and the answer being in free
natural language format. While this line of reasoning is very natural to a human expert,
a learning agent cannot understand it. The explanation that would be understood by the
agent is represented in the upper-right part of Figure 12.15 and consists of various

13:54:21,
.013
356 Chapter 12. Disciple Agents

ASSESS-COA-WRT-PRINCIPLE-OF-SURPRISE
FOR-COA COA411

R$ACWPOS-001
Which is an aspect that characterizes surprise?
Rule
Enemy reconnaissance Learning

ASSESS-SURPRISE-WRT-COUNTERING-ENEMY-
RECONNAISANCE
FOR-COA COA411

R$ASWCER-002
Is an enemy reconnaissance unit present?
Rule
Yes, RED-CSOP1, which is performing
Learning
the reconnaissance action SCREEN1

ASSESS-SURPRISE-WHEN-ENEMY-RECON-IS-PRESENT
FOR-COA COA411
FOR-UNIT RED-CSOP1
FOR-RECON-ACTION SCREEN1

Is the enemy reconnaissance unit destroyed?

R$ASWERIP-003
Yes, RED-CSOP1 is destroyed by DESTROY1 Rule
Learning
REPORT-STRENGTH-IN-SURPRISE-BECAUSE-OF-COUNTERING
ENEMY-RECON
FOR-COA COA411
FOR-UNIT RED-CSOP1
FOR-RECON-ACTION SCREEN1
FOR-ACTION DESTROY1
WITH-IMPORTANCE “high”

Figure 12.14. Task reductions indicated by the expert.

ASSESS-SURPRISE-WRT-COUNTERING-ENEMY-RECONAISSANCE
FOR-COA COA411

Natural Language Logic

Question:
Is an enemy reconnaissance Explanation:
unit present? RED-CSOP1 SOVEREIGN-ALLEGIANCE-OF-ORG RED-SIDE
RED-CSOP1 TASK SCREEN1
SCREEN1 IS INTELLIGENCE-COLLECTION-MILITARY-TASK

Rule: R$ASWCER-001

IF the task to accomplish is

ASSES-SURPRISE-WRT-COUNTERING-ENEMY-RECONAISSANCE
FOR-COA ?O1

Answer: Question: Is an enemy reconnaissance unit present?

Answer: Yes, ?O2 which is performing the

reconnaissance action?O3.

Yes, RED-CSOP1, which is Learn Explanation

?O2 SOVEREIGN-ALLEGIANCE-OF-ORG ?O4 IS RED-SIDE
?O2 TASK ?O3 IS INTELLIGENCE-COLLECTION-MILITARY-TASK

performing the reconnaissance Plausible Upper Bound Condition

the rule ?O1 IS COA-SPECIFICATION-MICROTHEORY

?O2 IS MODERN-MILITARY-UNIT-DEPLOYABLE

action SCREEN1. SOVEREIGN-ALLEGIANCE-OF-ORG ?O4

Main Condition

TASK ?O3
?O3 IS INTELLIGENCE-COLLECTION-MILITARY-TASK
?O4 IS RED-SIDE
Plausible Lower Bound Condition
?O1 IS COA411
?O2 IS RED-CSOP1
SOVEREIGN-ALLEGIANCE-OF-ORG ?O4
TASK ?O3
?O3 IS SCREEN1
?O4 IS RED-SIDE

ASSESS-SURPRISE-WHEN-ENEMY-RECON-IS-PRESENT THEN accomplish the task

ASSES-SURPRISE-WHEN-ENEMY-RECON-IS-PRESENT
FOR-COA
FOR-UNIT
?O1
?O2

FOR-COA COA411 FOR-RECON-ACTION ?O3

FOR-UNIT RED-CSOP1
FOR-RECON-ACTION SCREEN1

Figure 12.15. Teaching Disciple-COA to reduce a task.

13:54:21,
.013
12.3. Disciple-COA: Course of Action Critiquing 357

relations between certain elements from its ontology. The first explanation piece states, in
the formal language of Disciple-COA, that RED-CSOP1 is an enemy unit. The second
explanation piece expresses the fact that RED-CSOP1 is performing the action SCREEN1.
Finally, the last explanation piece expresses the fact that SCREEN1 is a reconnaissance
action. While an expert can understand the meaning of these formal expressions, he
cannot easily define them because he or she is not a knowledge engineer. For one thing,
the expert would need to use the formal language of the agent. But this would not be
enough. The expert would also need to know the names of the potentially many thousands
of concepts and features from the agent’s ontology.
While defining the formal explanations of this task reduction step is beyond the
individual capabilities of the expert or the agent, it is not beyond their joint capabilities.
Finding these explanation pieces is a mixed-initiative process of searching the agent’s
ontology, an explanation piece being a path of objects and relations in this ontology, as
discussed in Section 9.5. In essence, the agent uses analogical reasoning and help from
the expert to identify and propose a set of plausible explanation pieces from which the
expert has to select the correct ones. One explanation generation strategy is based on an
ordered set of heuristics for analogical reasoning. These heuristics exploit the hierarchies
of objects, features, and tasks to identify the rules that are similar to the current
reduction and to use their explanations as a guide to search for similar explanations of
the current example.
From the example reduction and its explanation in Figure 12.15, Disciple-COA auto-
matically generated the plausible version space rule in Figure 12.16. This is an IF-THEN
rule, the components of which are generalizations of the elements of the example in
Figure 12.15.
The rule in Figure 12.16 also contains two conditions for its applicability: a plausible
lower bound condition and a plausible upper bound condition. These conditions approxi-
mate an exact applicability condition that Disciple-COA attempts to learn. Initially, the
plausible lower bound condition covers only the example in Figure 12.15, restricting the
variables from the rule to take only the values from this example. It also includes the
relations between these variables that have been identified as relevant in the explanation
of the example. The plausible upper bound condition is the most general generalization of
the plausible lower bound condition. It is obtained by taking into account the domains and
the ranges of the features from the plausible lower bound condition and the tasks, in order
to determine the possible values of the variables. The domain of a feature is the set of objects
that may have that feature. The range is the set of possible values of that feature. For
instance, ?O2 is the value of the task feature FOR-UNIT, and has as features SOVEREIGN-
ALLEGENCE-OF-ORG and TASK. Therefore, any value of ?O2 has to be in the intersection of
the range of FOR-UNIT, the domain of SOVEREIGN-ALLEGENCE-OF-ORG, and the domain
of TASK. This intersection is MODERN-MILITARY-UNIT-DEPLOYABLE.
The learned rules, such as the one in Figure 12.16, are used in problem solving to
generate task reductions with different degrees of plausibility, depending on which of their
conditions are satisfied. If the plausible lower bound condition is satisfied, then the
reduction is very likely to be correct. If the plausible lower bound condition is not satisfied,
but the plausible upper bound condition is satisfied, then the solution is considered only
plausible. Any application of such a partially learned rule, however, either successful or
not, provides an additional (positive or negative) example, and possibly an additional
explanation, that are used by the agent to improve the rule further through the general-
ization and/or specialization of its conditions.

13:54:21,
.013
358 Chapter 12. Disciple Agents

Rule: R$ASWCER-002

IF the task to accomplish is

ASSESS-SURPRISE-WRT-COUNTERING-ENEMY-RECONAISSANCE
FOR-COA ?O1

Question: Is an enemy reconnaissance unit present?

Answer: Yes, ?O2, which is performing

the reconnaissance action?O3
Explanation
?O2 SOVEREIGN-ALLEGIANCE-OF-ORG ?O4 IS RED-SIDE
?O2 TASK ?O3 IS INTELLIGENCE-COLLECTION-MILITARY-TASK

Plausible Upper Bound Condition

?O1 IS COA-SPECIFICATION-MICROTHEORY
?O2 IS MODERN-MILITARY-UNIT-DEPLOYABLE
SOVEREIGN-ALLEGIANCE-OF-ORG ?O4
Main Condition

TASK ?O3
?O3 IS INTELLIGENCE-COLLECTION-MILITARY-TASK
?O4 IS RED-SIDE

Plausible Lower Bound Condition

?O1 IS COA411
?O2 IS RED-CSOP1
SOVEREIGN-ALLEGIANCE-OF-ORG ?O4
TASK ?O3
?O3 IS SCREEN1
?O4 IS RED-SIDE

THEN accomplish the task

ASSESS-SURPRISE-WHEN-ENEMY-RECON-IS-PRESENT
FOR-COA ?O1
FOR-UNIT ?O2
FOR-RECON-ACTION ?O3

Figure 12.16. Plausible version space rule learned from the example and the explanation in
Figure 12.15.

Let us consider again the specific task reductions from Figure 12.14. At least for the
elementary tasks, such as the one at the bottom of the figure, the expert needs also to
express them in natural language: “There is a strength with respect to surprise in COA411
because it contains aggressive security/counter-reconnaissance plans, destroying enemy
intelligence collection units and activities. Intelligence collection by RED-CSOP1 will be
disrupted by its destruction by DESTROY1.”
Similarly, the expert would need to indicate the reference (source) material for the
concluded assessment. The learned rules contain generalizations of these phrases that
are used to generate answers in natural language, as illustrated in the bottom part
of Figure 12.12. Similarly, the generalizations of the questions and the answers from
the rules applied to generate a solution are used to produce an abstract justification of
the reasoning process.
As Disciple-COA learns plausible version space rules, it can use them to propose
routine or innovative solutions to the current problems. The routine solutions are those
that satisfy the plausible lower bound conditions of the rules and are very likely to be
correct. Those that are not correct are kept as exceptions to the rule. The innovative

13:54:21,
.013
12.3. Disciple-COA: Course of Action Critiquing 359

solutions are those that do not satisfy the plausible lower bound conditions but satisfy the
plausible upper bound conditions. These solutions may or may not be correct, but in each
case they lead to the refinement of the rules that generated them. Let us consider the
situation illustrated in Figure 12.17. After it has been shown how to critique COA411 with
respect to the principle of security, Disciple-COA is asked to critique COA421. COA421
is similar to COA411, except that in this case the enemy recon unit is not destroyed.
Because of this similarity, Disciple-COA is able to propose the two top reductions in
Figure 12.17. Both of them are innovative reductions that are accepted by the expert.
Therefore, Disciple-COA generalizes the plausible lower bound conditions of the corres-
ponding rules, as little as possible, to cover these reductions and to remain less general or
at most as general as the corresponding plausible upper bound conditions.
The last reduction step in Figure 12.17 has to be provided by the expert because no rule
of Disciple-COA is applicable. We call the expert-provided reduction a creative problem-
solving step. From each such reduction, Disciple-COA learns a new task reduction rule, as
was illustrated in the preceding example.
Through refinement, the task reduction rules may become significantly more complex
than the rule in Figure 12.16. For instance, when a reduction proposed by Disciple-COA is
rejected by the expert, the agent attempts to find an explanation of why the reduction is
wrong. Then the rule may be refined with an Except-When plausible version space
condition. The bounds of this version space are generalizations of the explanations that
should not hold in order for the reduction rule to be applicable.

ASSESS-COA-WRT-PRINCIPLE-OF-SURPRISE
FOR-COA COA421

R$ACWPOS-001
Which is an aspect that characterizes surprise?
Rule
Enemy reconnaissance Refinement

ASSESS-SURPRISE-WRT-COUNTERING-ENEMY-
RECONNAISANCE
FOR-COA COA421
R$ASWCER-002

Is an enemy reconnaissance unit present?

Yes, RED-CSOP2, which is performing Rule
the reconnaissance action SCREEN2 Refinement

ASSESS-SURPRISE-WHEN-ENEMY-RECON-IS-PRESENT
FOR-COA COA421
FOR-UNIT RED-CSOP2
FOR-RECON-ACTION SCREEN2
R$ASWERIP-004

Is the enemy reconnaissance unit destroyed?

No Rule
Learning
REPORT-WEAKNESS-IN-SURPRISE-BECAUSE-ENEMY
RECON-IS-NOT-COUNTERED
FOR-COA COA421
FOR-UNIT RED-CSOP2
FOR-RECON-ACTION SCREEN2
WITH-IMPORTANCE “high”

Figure 12.17. Mixed-initiative problem solving and learning.

13:54:21,
.013
 360 Chapter 12. Disciple Agents

In any case, comparing the left-hand side of Figure 12.15 (which is defined by the domain
expert) with the rule from Figure 12.16 (which is learned by Disciple-COA) suggests the
usefulness of a Disciple agent for knowledge acquisition. In the conventional knowledge
engineering approach, a knowledge engineer would need to manually define and debug a
rule such as the one in Figure 12.16. With Disciple, the domain expert needs only to define
an example reduction, because Disciple learns and refines the corresponding rule.

12.3.5 Experimental Results

Like Disciple-WA, Disciple-COA has also been developed and evaluated as part of the
DARPA’s High Performance Knowledge Bases Program, as discussed in the following
(Tecuci et al., 2001).
In addition to George Mason University (GMU), three other research groups have
developed COA critiquers as part of the HPKB program. Teknowledge and Cycorp have
developed a critiquer based on the Cyc system (Lenat, 1995). The other two critiquers
have been developed at the Information Sciences Institute (ISI) at the University of
Southern California (USC), one based on the Expect system (Kim and Gil, 1999), and the
other based on the Loom system (Loom, 1999). All the critiquers were evaluated as part of
the HPKB’s annual evaluation that took place during the period from July 6 to July 16,
1999, and included five evaluation items of increasing difficulty. Each item consisted of
descriptions of various COAs and a set of questions to be answered about each of them.
Item1 consisted of COAs and questions that were previously provided by DARPA to guide
the development of the COA critiquing agents. Item2 included new test questions about
the same COAs. Items 3, 4, and 5 consisted of new COAs that were increasingly more
complex and required further development of the COA agents in order to answer the asked
questions properly. Each of the Items 3, 4, and 5 consisted of two phases. In the first phase,
each team had to provide initial system responses. Then the evaluator issued the model
answers, and each team had a limited amount of time to repair its system, to perform
further knowledge acquisition, and to generate revised system responses.
The responses of each system were scored by a team of domain experts along the
following dimensions and associated weights: (1) Correctness, 50 percent (matches model
answer or is otherwise judged to be correct); (2) Justification, 30 percent (scored on
presence, soundness, and level of detail); (3) Lay Intelligibility, 10 percent (degree to
which a lay observer can understand the answer and the justification); (4) Sources,
10 percent (degree to which appropriate references or sources are noted); and (5) Pro-
activity, 10 percent extra credit (appropriate corrective actions or other information
suggested to address the critique). Based on these scores, several classes of metrics have
been computed, including Recall and Precision. Recall is obtained by dividing the score for
all answers provided by a critiquer to the total number of model answers for the asked
questions. This was over 100 percent in the case of the Disciple-COA critiquer, primarily
because of the extra credit received for generating additional critiques that were not
among the model answers provided by the evaluator. The Precision score is obtained by
dividing the same score by the total number of answers provided by that system (both the
model answers provided by the evaluator and the new answers provided by the critiquer).
The results obtained by the four evaluated critiquers are presented in Figure 12.18. These
graphs show also the averages of these results, which represent the recall and the preci-
sion of the integrated system consisting of the four critiquers.

13:54:21,
.013
 Metric: Recall (Total Score) Recall Breakdown by Criteria

140.00%

120.00% 114.69% 120.00%

100.00% 100.00%
84.20%
80.00% 80.00%
70.20%
63.71%
56.81%
60.00% 60.00%

40.00% 40.00%

20.00% 20.00%

0.00% 0.00%
Correctness Justification Intelligibility Sources Proactivity Total
Tek/Cyc ISI-Expect GMU ISI-Loom ALL
.013

Metric: Precision (Total Score) Precision Breakdown by Criteria

100.00%
100.00%
90.00%
13:54:21,

76.01% 81.99% 75.01% 80.00%

80.00%
70.00%
62.61% 57.48% 60.00%
60.00%
50.00%
40.00%
40.00%
30.00%
20.00%
20.00%
10.00%
0.00%
0.00%
Correctness Justification Intelligibility Sources Proactivity Total
Tek/Cyc ISI-Expect GMU ISI-Loom ALL

Figure 12.18. The performance of the COA critiquers and of the integrated system (reprinted with permission from Eric Jones).
362 Chapter 12. Disciple Agents

Figure 12.19 compares the recall and the coverage of the developed critiquers for the
last three most complex items of the evaluation. For each item, the beginning of each
arrow shows the coverage and recall for the initial testing phase, and the end of the arrow
shows the same data for the modification phase. In this graph, the results that are above
and to the right are superior to the other results. This graph also shows that all the systems
increased their coverage during the evaluation. In particular, the knowledge base of
Disciple-COA increased by 46 percent (from the equivalent of 6,229 simple axioms to
9,092 simple axioms), which represents a very high rate of knowledge acquisition of
286 simple axioms per day.
During August 1999, we conducted a one-week knowledge acquisition experiment with
Disciple-COA, at the U.S. Army Battle Command Battle Lab, in Fort Leavenworth, Kansas,
to test the claim that domain experts who do not have prior knowledge engineering
experience can teach Disciple-COA (Tecuci et al., 2001). The experiment involved four
such military experts and had three phases: (1) a joint training phase during the first three
days, (2) an individual teaching experiment on day four, and (3) a joint discussion of the
experiment on day five. The entire experiment was videotaped. The training for the
experiment included a detailed presentation of Disciple’s knowledge representation,
problem-solving, and learning methods and tools. For the teaching experiment, each
expert received a copy of Disciple-COA with a partial knowledge base. This knowledge
base was obtained by removing the tasks and the rules from the complete knowledge base
of Disciple-COA. That is, the knowledge base contained the complete ontology of objects,

(Evaluaon Items 3, 4, and 5)

160
GMU 3
(Disciple)
140
ISI (Expect) 4
120

100
TEK/CYC 4
GMU 5
Recall

(Disciple)
80 GMU 4
TEK/CYC 5 (Disciple) ISI (Expect) 5
60

40

20
TEK/CYC 3 ISI (Expect) 3
0
25% 50% 75% 100%
Coverage
Figure 12.19. Coverage versus recall, pre-repair and post-repair (reprinted with permission from
Eric Jones).

13:54:21,
.013
12.3. Disciple-COA: Course of Action Critiquing 363

object features, and task features. We also provided the experts with the descriptions of
three COAs (COA411, COA421, and COA51), to be used for training Disciple-COA. These
were the COAs used in the final phases of the DARPA’s evaluation of all the critiquers.
Finally, we provided and discussed with the experts the modeling of critiquing these COAs
with respect to the principles of offensive and security. That is, we provided the experts
with specific task reductions, similar to the one from Figure 12.17, to guide them in
teaching Disciple-COA. After that, each expert taught Disciple-COA independently while
being supervised by a knowledge engineer, whose role was to help the expert if he or she
reached an impasse while using Disciple-COA.
Figure 12.20 shows the evolution of the knowledge base during the teaching process for
one of the experts, being representative for all the four experts. In the morning, the expert
taught Disciple-COA to critique COAs with respect to the principle of offensive, and in the
afternoon he taught it to critique COAs with respect to the principle of security. In both
cases, the expert used first COA411, then COA422, and then COA51. As one can see from
Figure 12.20, Disciple-COA initially learned more rules, and then the emphasis shifted on
rule refinement. Therefore, the increase in the size of the knowledge base is greater toward
the beginning of the training process for each principle. The teaching for the principle of
offensive took 101 minutes. During this time, Disciple-COA learned fourteen tasks and
fourteen rules (147 simple axioms’ equivalent). The teaching for security took place in the
afternoon and consisted of 72 minutes of interactions between the expert and Disciple-
COA. During this time, Disciple-COA learned fourteen tasks and twelve rules (136 simple
axioms’ equivalent). There was no or very limited assistance from the knowledge engineer
with respect to teaching. The knowledge acquisition rate obtained during the experiment
was very high (approximately nine tasks and eight rules per hour, or ninety-eight simple
axioms’ equivalent per hour). At the end of this training process, Disciple-COA was able to
correctly identify seventeen strengths and weaknesses of the three COAs with respect to
the principles of offensive and security.

400

350
Task+Rule axioms
300
Number of Axioms

250

Task axioms
200

150
Rule axioms

100

50

0
Initial 411 421 51 411 421 51 411 421 51 411 421 51

Training for Offensive Training for Security

Figure 12.20. The evolution of the knowledge base during the teaching process.

13:54:21,
.013
 364 Chapter 12. Disciple Agents

After the experiment, each expert was asked to fill in a detailed questionnaire designed
to collect subjective data for usability evaluation. All the answers took into account that
Disciple-COA was a research prototype and not a commercial product and were rated
based on a scale of agreement with the question from 1 to 5, with 1 denoting “not at all”
and 5 denoting “very”. For illustration, Table 12.4 shows three questions and the answers
provided by the four experts.
In conclusion, Disciple-COA demonstrated the generality of its learning methods that
used an object ontology created by another group, namely Teknowledge and Cycorp
(Boicu et al., 1999). It also demonstrated high rule learning rates, as compared with
manual definition of rules, and better performance than the evaluating experts, with many
unanticipated solutions.

12.4 DISCIPLE-COG: CENTER OF GRAVITY ANALYSIS

12.4.1 The Center of Gravity Analysis Problem

Military literature distinguishes among three levels of war – strategic, operational, and
tactical – that help clarify the links between national strategic objectives and tactical
actions. There are no finite limits or boundaries between the levels of war (Joint Chiefs
of Staff, 2008, II–1).
One of the most difficult problems that senior military leaders face at the strategic level
is the determination and analysis of the centers of gravity for friendly and opposing forces.
The concept of the center of gravity of an entity (state, alliance, coalition, or group) was

Table 12.4 Sample Questions Answered by the Four Experts

Questions Answers

Do you think that Disciple is a
Rating 5. Absolutely! The potential use of this tool by
useful tool for knowledge domain experts is only limited by their imagination – not
acquisition? their AI programming skills.

5

4

Yes, it allowed me to be consistent with logical thought.

Do you think that Disciple is a
Rating 5. Yes.

useful tool for problem solving?
5 (absolutely)

4

Yes. As it develops and becomes tailored to the user, it
will simplify the tedious tasks.

Were the procedures/processes
Rating 5. As a minimum yes, as a maximum – better!

used in Disciple compatible with
This again was done very well.
Army doctrine and/or decision-

4
making processes?

4

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 365

introduced by Carl von Clausewitz (1832) as “the foundation of capability, the hub of all
power and movement, upon which everything depends, the point against which all the
energies should be directed.” It is currently defined as comprising the source of power
that provides freedom of action, physical strength, and will to fight (Joint Chiefs of Staff,
2008, IV–X).
It is recognized that, “Should a combatant eliminate or influence the enemy’s strategic
center of gravity, the enemy would lose control of its power and resources and eventually
fall to defeat. Should a combatant fail to adequately protect his own strategic center of
gravity, he invites disaster” (Giles and Galvin, 1996, p. 1). Therefore, the main goal of any
force should be to eliminate or influence the enemy’s strategic center of gravity while
adequately protecting its own.
Correctly identifying the centers of gravity of the opposing forces is of highest import-
ance in any conflict. Therefore, all the U.S. senior military service colleges emphasize
center of gravity analysis in the education of strategic leaders (Warden, 1993; Echevarria,
2003; Strange and Iron, 2004a, 2004b; Eikmeier, 2006).
In spite of the apparently simple definition of the center of gravity, its determination
requires a wide range of background knowledge, not only from the military domain but
also from the economic, geographic, political, demographic, historic, international, and
other domains (Giles and Galvin, 1996). In addition, the adversaries involved, their goals,
and their capabilities can vary in important ways from one situation to another. When
performing this analysis, some experts may rely on their own professional experience and
intuitions without following a rigorous approach.
Recognizing these difficulties, the Center for Strategic Leadership of the U.S. Army War
College started an effort in 1993 to elicit and formalize the knowledge of a number of
experts in center of gravity. This research resulted in a COG monograph (Giles and Galvin,
1996). This monograph made two significant contributions to the theory of center of
gravity analysis. The first was a systematic analysis of the various factors (e.g., politic,
military, economic, etc.) that have to be taken into account for center of gravity determin-
ation. The second significant contribution was the identification of a wide range of center
of gravity candidates.
A significant advancement of the theory of center of gravity analysis was the CG-CC-
CR-CV model introduced by Strange (1996) and summarized by the following definitions:

Centers of gravity Primary sources of moral or physical strength, power, or

(CG): resistance.
Critical capabilities Primary abilities that merit a center of gravity to be identified
(CC): as such, in the context of a given scenario, situation, or
mission.
Critical requirements Essential conditions, resources, and means for a critical
(CR): capability to be fully operative.
Critical vulnerabilities Critical requirements or components thereof that are defi-
(CV): cient, or vulnerable to neutralization, interdiction, or attack
(moral/physical harm), in a manner achieving decisive
results – the smaller the resources and effort applied and
the smaller the risk and cost, the better.

Building primarily on the work of Strange (1996) and Giles and Galvin (1996), we have
developed a computational approach to center of gravity analysis, which is summarized in
Figure 12.21.

13:54:21,
.013
366 Chapter 12. Disciple Agents

Given: A strategic situation (e.g., the invasion of Iraq by the U.S.-led coalition in 2003).
Determine: The strategic centers of gravity of the opposing forces and their critical
vulnerabilities.

Assessment Identification of Testing of

of Situation COG candidates COG candidates
Assemble data and Identify potential primary Test each identified COG
assess the relevant sources of moral or candidate to determine
aspects of the strategic physical strength, power, whether it has all the
situation: and resistance from: critical capabilities:

Opposing forces and Government Which are the required critical

their strategic goals Military capabilities?
Political factors People Are the critical requirements of
Military factors these capabilities satisfied?
Economy
Psychosocial factors Are the critical requirements
Alliances vulnerable?
Economic factors, etc. Etc. Select COG based on analysis.
Figure 12.21. Computational approach to center of gravity analysis.

This approach consists of three main phases: assessment of the strategic situation,
identification of center of gravity candidates, and testing of the identified candidates.
During the assessment of the situation (such as the invasion of Iraq by the U.S.-led
coalition in 2003), one assembles and assesses data and other relevant aspects of the
strategic environment, including the opposing forces (Iraq, on one side, and the U.S.-led
coalition, on the other side), their strategic goals, political factors (e.g., type of government,
governing bodies), military factors (e.g., leaders, will, and capability), psychosocial factors
(e.g., motivation, political activities), economic factors (e.g., type of economy, resources),
and so on. This assessment will be used in the next phases of center of gravity analysis.
During the identification phase, strategic center of gravity candidates are identified
from a belligerent’s elements of power, such as its leadership, government, military,
people, or economy. For example, a strong leader, such as Saddam Hussein or George
W. Bush, could be a center of gravity candidate with respect to the situation at the
beginning of the Iraq War in 2003. The result of this phase is the identification of a wide
range of candidates.
During the testing phase, each candidate is analyzed to determine whether it has all the
critical capabilities that are necessary to be the center of gravity. For example, a leader
needs to be secure; informed; able to maintain support from the government, the military,
and the people; and irreplaceable. For each capability, one needs to determine the
existence of the essential conditions, resources, and means that are required by that
capability to be fully operative. For example, some of the protection means of Saddam
Hussein were the Republican Guard Protection Unit, the Iraqi Military, the Complex of
Iraqi Bunkers, and the System of Saddam doubles. Once these means of protection are
identified, one needs to determine whether any of them, or any of their components, are
vulnerable. For example, the Complex of Iraqi Bunkers is vulnerable because their location
and design are known to the U.S.-led coalition and could be destroyed.
Based on the results of the analysis, one can eliminate any center of gravity candidate
that does not have all the required critical capabilities, and select the centers of gravity

13:54:21,
.013
 12.4. Disciple-COG: Center of Gravity Analysis 367

from the remaining candidates. Moreover, the process also identifies the critical vulner-
abilities of the selected centers of gravity.
An important characteristic of this approach is that it is both natural for a human and
appropriate for automatic processing. By using this approach, we have developed the
Disciple-COG agent, briefly described in the following section.

12.4.2 Overview of the Use of Disciple-COG

Disciple-COG is a computer program that guides a military planner in describing a
strategic situation and performing a center of gravity analysis following the approach
described in the preceding section and summarized in Figure 12.21.
First, Disciple-COG guides the user in identifying, describing, and assessing the aspects
of the strategic situation that are relevant to center of gravity analysis. An example of such
a situation could be World War II in Europe, at the time of the invasion of the island of
Sicily by the Allied Forces, in 1943.
The user–agent interaction is easy and natural for the user, taking place as illustrated in
Figures 12.22 and 12.23. The left part of each window is a table of contents whose elements
indicate various important aspects of the situation. Initially it contains only “situation.”
When the user clicks on one such aspect, Disciple-COG asks specific questions intended to
acquire a description and/or assessment of that aspect or to update a previously specified
description.

Figure 12.22. Initial situation description interface of Disciple-COG.

13:54:21,
.013
368 Chapter 12. Disciple Agents

Figure 12.23. Situation description and assessment interface of Disciple-COG.

Figure 12.22 shows the initial interaction screen after the user has clicked on “situ-
ation.” The right-hand side shows the prompts of Disciple-COG and the information
provided by the user, such as:

Provide a name for the situation to be assessed:

WW II Europe 1943
...
Name the opposing forces in WW II Europe 1943:
Allied Forces 1943
European Axis 1943

Once the user names the opposing forces (i.e., “Allied Forces 1943” and “European Axis
1943”), Disciple-COG includes them into the table of contents, as shown in the left-hand
side of Figure 12.22. Then, when the user clicks on one of the opposing forces (e.g., “Allied
Forces 1943”), Disciple-COG asks for its characteristics, as indicated in the right-hand side
of Figure 12.23 (e.g., “What kind of force is Allied Forces 1943?”). Because the user
characterized “Allied Forces 1943” as a multistate force (by clicking on one of the options
offered by the agent), Disciple-COG further asks for its members and extends the table of
contents with the provided names (i.e., “US 1943,” “Britain 1943,” “USSR 1943,” etc.) and
their relevant aspects (i.e., “Strategic goal,” “Political factors,” “Military factors,” etc.), as
shown in the left-hand side of Figure 12.23. The user can now click on any such aspect and
will be asked specific questions by Disciple-COG.
Thus, the user’s answers lead to the generation of new items in the left-hand side of the
window, and trigger new questions from the agent, which depend on the answers pro-
vided by the user. Through such context-dependent questions, Disciple-COG guides the
user to research, describe, and assess the situation.
As will be discussed in Section 12.4.4, once the user describes various aspects of
the situation, Disciple-COG automatically extends its ontology with the corresponding

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 369

representations. The user is not required to answer all the questions, and Disciple-COG
can be asked, at any time, to identify and test the strategic center of gravity candidates for
the current description of the situation. The COG analysis process uses the problem
reduction and solution synthesis paradigm presented in Section 4.2, following the CG-
CC-CR-CV model (Strange, 1996).
Figure 12.24 shows the interface of the Mixed-Initiative Reasoner of Disciple-COG that
displays the automatically generated analysis. The left-hand side shows an abstract view of
the analysis tree for the problem “Analyze the strategic COG candidates for the WW II
Europe 1943 situation.”
First, the problem of analyzing the strategic COG candidates for this situation is
reduced to analyzing the COG candidates for each of the two opposing forces. Then,
because each of the opposing forces is a multimember force, the problem of analyzing
the COG candidates for an opposing force is reduced to two other problems: (1) the
problem of analyzing the COG candidates for each member of the multimember force
(e.g., US 1943 candidates, Britain 1943 candidates, and USSR 1943 candidates, in the case
of Allied Forces 1943) and (2) the problem of analyzing the multimember COG
candidates.
Continuing, the problem of analyzing the US 1943 candidates is reduced to analyzing
the COG candidates with respect to the main elements of power of US 1943, namely
people of US 1943, government of US 1943, armed forces of US 1943, and economy of
US 1943.
Because the abstract problem “US 1943 candidates” is selected in the left-hand side of
the interface of the Mixed-Initiative Reasoner, the right-hand side shows the detailed
description of the corresponding reduction tree. Notice that the detailed tree shows both
complete problem descriptions and the question/answer pairs that guide their reductions.
The leaves of the detailed tree correspond to the abstract subproblems of “US 1943
candidates,” such as “Candidates wrt people of US 1943.”
The user can browse the entire analysis tree generated by Disciple-COG by clicking on
the nodes and the plus (+) and minus (–) signs. For example, Figure 12.25 shows how the

Figure 12.24. Abstract (left) and detailed (right) COG reduction tree.

13:54:21,
.013
370 Chapter 12. Disciple Agents

Figure 12.25. Reduction tree for testing a national leader as a COG candidate.

problems of analyzing the COG candidates with respect to the main elements of power of
US 1943 (government of US 1943, and armed forces of US 1943) are reduced to identifying
and testing specific COG candidates (i.e., President Roosevelt and military of US 1943).
Testing each of the identified COG candidates is reduced to the problems of testing
whether it has all the necessary critical capabilities. Thus, testing President Roosevelt as a
potential COG candidate is reduced to seven problems, each testing whether President
Roosevelt has a certain critical capability, as shown in the left side of Figure 12.25.
The left-hand side of Figure 12.26 shows how testing of whether a COG candidate has a
certain critical capability is reduced to the testing of whether the corresponding critical
requirements are satisfied. In particular, testing of whether President Roosevelt has the
critical capability to stay informed is reduced to the problem of testing of whether he has
means to receive essential intelligence. These means are identified as US Office of Strategic
Services 1943, US Navy Intelligence 1943, and US Army Intelligence 1943. Consequently, the
user is asked to assess whether each of them has any significant vulnerability. The user
clicks on one of the means (e.g., US Office of Strategic Services 1943 in the left-hand side of
Figure 12.26) and the agent displays two alternative solution patterns for its assessment, in
the right-hand side of Figure 12.26:

The US Office of Strategic Services 1943 that provides essential intelligence to

President Roosevelt has the following significant vulnerability: . . . .
Justification: . . .

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 371

Figure 12.26. Assessing whether a critical requirement has any significant vulnerability.

The US Office of Strategic Services 1943 that provides essential intelligence to

President Roosevelt has no significant vulnerability.
Justification: . . .

The user has to complete the instantiation of one of the two patterns and then click on the
corresponding Save button. In this case, the provided solution is the following one:

The US Office of Strategic Services 1943 that provides essential intelligence to

President Roosevelt has the following significant vulnerability: There is a huge
amount of information that needs to be collected and analyzed by the US Office of
Strategic Services 1943.

Up to this point we have presented the automatic generation of the top-down COG
reduction tree and the evaluation of the elementary problems (i.e., potential vulnerabil-
ities) by the user.
The next stage of the COG analysis process is the bottom-up automatic synthesis of the
elementary solutions. This will be illustrated in the following, starting with Figure 12.27,
which shows how the assessments of the President Roosevelt’s individual means to receive
essential intelligence (i.e., US Office of Strategic Services 1943, US Navy Intelligence 1943, and
US Army Intelligence 1943) are combined to provide an overall assessment of his means to
receive essential intelligence:

President Roosevelt has means to receive essential intelligence (US Office of Strategic
Services 1943, US Navy Intelligence 1943, and US Army Intelligence 1943). The US Office
of Strategic Services 1943 has the following significant vulnerability: There is a huge
amount of information that needs to be collected and analyzed by the US Office of
Strategic Services 1943.

The leaf solutions in Figure 12.27 have a yellow background to indicate that they are
assessments made by the user. The top-level solution obtained through their combination
has a green background to indicate that it was automatically computed by Disciple-COG,
based on a previously learned synthesis rule. This rule indicates the pattern of the solution
and how it is obtained by combining elements of the patterns of the subsolutions. In
particular, notice that the means from individual solutions are gathered into a single list.

13:54:21,
.013
372 Chapter 12. Disciple Agents

Figure 12.27. Obtaining the overall assessment of a critical requirement by combining individual
assessments.

The next bottom-up solution synthesis step is to obtain the assessment of a critical
capability by combining the assessments of its critical requirements. Figure 12.28, for
instance, shows the assessment of President Roosevelt’s critical capability to maintain
support, obtained by combining the assessments of the corresponding means (i.e., means
to secure support from the government, means to secure support from the military, and
means to secure support from the people). This synthesis operation was previously
explained in Section 4.2 and illustrated in Figure 4.8 (p. 117).
Next Disciple-COG obtains the assessment of a COG candidate based on the assess-
ments of its critical capabilities, as illustrated in Figure 12.29:

President Roosevelt is a strategic COG candidate that can be eliminated because

President Roosevelt does not have all the necessary critical capabilities (e.g., be
irreplaceable).

All the identified COG candidates from the analyzed situation are evaluated in a similar
way, and the final solution is a summary of the results of these evaluations, as illustrated
in Figure 12.30. In particular, for the WW II Europe 1943 situation, the solution is
the following one:

For European Axis 1943, choose the strategic center of gravity from the following
candidates: military of Germany 1943 and industrial capacity of Germany 1943. For
Allied Forces 1943, choose the strategic center of gravity from the following
candidates: military of USSR 1943, financial capacity of USSR 1943, industrial capacity
of USSR 1943, will of the people of Britain 1943, military of Britain 1943, financial
capacity of Britain 1943, will of the people of US 1943, military of US 1943, and industrial
capacity of US 1943.

The subsolutions of this top-level solution indicate all the COG candidates considered and
why several of them have been eliminated.

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 373

Figure 12.28. Obtaining the overall assessment of a critical capability based on its critical
requirements.

Figure 12.29. Obtaining the assessment of a COG candidate based on its critical capabilities.

13:54:21,
.013
374 Chapter 12. Disciple Agents

Figure 12.30. Result of the evaluation of the COG candidates corresponding to a situation.

At the end of the analysis, Disciple-COG generates a draft analysis report, a fragment of
which is shown in Figure 12.31. The first part of this report contains a description of the
strategic situation that is generated from the information provided and assessed by the
user, as illustrated in Figures 12.22 and 12.23. The second part of the report includes all the
center of gravity candidates identified by Disciple-COG, together with their analyses, as
previously discussed. The user may now finalize this report by examining the analysis of
each center of gravity candidate and by completing, correcting, or even rejecting it and
providing a different analysis.
Successive versions of Disciple-COG have been used for ten years in courses at the
U.S. Army War College (Tecuci et al., 2008b). It has also been used at the Air War College
and the Joint Forces Staff College. The use of Disciple-COG in such an educational
environment is productive for several reasons. First, the user is guided in performing a
detailed and systematic assessment of the most important aspects of a strategic situation,
which is necessary in order to answer Disciple-COG’s questions. Second, the agent
generates its solutions by employing a systematic analysis, which was learned from a
military expert. Therefore, the user can learn how to perform a similar analysis from
Disciple-COG. Third, the details of the analysis and the actual results reflect the personal
judgment of the user, who has unique military experiences and biases and has a
personal interpretation of certain facts. Thus, the analysis is unique to the user, who
can see how his or her understanding of the situation determines the results yielded by
Disciple-COG.
It is important to note, however, that the solutions generated by Disciple-COG must be
critically analyzed at the end.

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 375

Figure 12.31. Fragment of an automatically generated COG analysis report.

13:54:21,
.013
 376 Chapter 12. Disciple Agents

12.4.3 Ontology Design and Development

The previous section provided a general overview of the reduction and synthesis process
performed by Disciple-COG, which resulted from the modeling of this process by
subject matter experts from the U.S. Army War College, primarily Dr. Jerome Comello
(Colonel, retired). This modeling process also informed the ontology design and devel-
opment process, as discussed in Section 6.4. The top part of the resulting ontology is
shown in Figure 12.32.
Under the leaf concepts from Figure 12.32 are ontologies for those concepts. For
example, Figure 12.33 shows a fragment of the ontology of political factors. Figure 10.22
(p. 323) shows part of the ontology of forces, Figure 10.26 (p. 325) shows part of the
ontology of economic factors, and Figure 10.28 (p. 326) shows part of the ontology of
resources and infrastructure elements.
The features are also organized hierarchically. Several of the feature ontologies have an
almost one-to-one correspondence to a concept ontology. Consider, for example, the
ontology of controlling leaders from the middle of Figure 12.33. For each type of control-
ling leader (e.g., political leader), there is a corresponding feature (i.e., has as political
leader), as shown in Figure 12.34. Similarly, the feature ontology from Figure 5.8 (p. 160)
corresponds to the ontology of economic factors from Figure 10.26 (p. 325).

12.4.4 Script Development for Scenario Elicitation

Disciple-COG has general knowledge about the center of gravity domain, such as problem
reduction and solution syntheses rules (whose applications have been illustrated in
Section 12.4.2), as well as ontologies of concepts and features (as discussed in the previous
section). However, the agent has no specific knowledge about any particular situation
(referred to as scenario in this section).

object

scenario force force goal resource or

infrastructure element
strategic cog relevant factor
operational strategic
goal goal

demographic factor civilization factor

economic factor psychosocial factor

geographic factor political factor

historical factor military factor

international factor

Figure 12.32. The top part of the concept ontology of Disciple-COG.

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 377

political factor

governing body controlling element

state government group governing body controlling leader controlling organization

controlling group
feudal deity
god king monarchy religious
figure ruling
government leader secret political
police party
dictator god
king political
totalitarian theocratic cabinet
government government
or staff
democratic
monarch
government autocratic military
leader religious
leader
body
theocratic chief and
police
democracy tribal council
state
conventional
commander
military religious political military law enforcement
in chief
dictatorship dictatorship leader staff organization

representative democratic council

democracy or board
parliamentary head of head of
democracy government state

Figure 12.33. Fragment of the ontology of political factors.

When the user starts using the agent, Disciple-COG elicits the description of the situation
or scenario to be analyzed, as was illustrated at the beginning of Section 12.4.2 and in
Figures 12.22 (p. 367) and 12.23 (p. 368). Scenario elicitation is guided by elicitation scripts
that are associated with the concepts and features from the ontology of Disciple-COG. For
example, the elicitation script for the feature has as opposing force is shown in Figure 12.35.
The script indicates the question to be asked (“Name the opposing forces in <scenario
name>:”), the variable that will hold the answer received from the user (“<opposing
force>”), the graphical appearance of the interface (“multiple line, height 4”), the way the
ontology will be extended with the elicited opposing force (“<opposing force> instance of
opposing force,” and “<scenario name> has as opposing force <opposing force>”), and the
next script to call (“Elicit properties of the instance <opposing force> in new window”).
An illustration of the execution of this script was provided at the bottom of Figure 12.22
(p. 367). Figure 12.36 shows the effect of this execution on the ontology of Disciple-COG.
Before script execution, the relevant part of the ontology is the one from the top of
Figure 12.36. The execution of the script causes Disciple-COG to prompt the user as
follows: “Name the opposing forces in WW II Europe 1943.” Once the user provides these
names (“Allied Forces 1943” and “European Axis 1943”), Disciple-COG introduces them as
instances of opposing force, and connects the “WW II Europe 1943” scenario to them, as
indicated in the script and illustrated at the bottom part of Figure 12.36.
Disciple-COG contains a Script Editor that allows easy definition of the elicitation
scripts by a knowledge engineer (KE). Figure 12.37 shows how the KE has defined the
script from Figure 12.35. The are a few differences in the naming of some entities in

13:54:21,
.013
378 Chapter 12. Disciple Agents

domain agent
has as controlling leader
range
person

governing body domain governing body

domain
has as god king has as religious leader
range range
person person

domain governing body domain governing body

has as monarch has as military leader
range range
person person

domain governing body

has as political leader
range
person

domain force
has as commander in chief
range
person

domain governing body

has as head of state
range
person

domain governing body

has as head of government
range
person

Figure 12.34. Fragment of the hierarchy of features corresponding to the ontology of controlling
leaders.

feature
scenario
domain

has as opposing force range force

elicitation script

Script type: Elicit the feature has as opposing force for an instance <scenario name>
Controls:
Queson: Name the opposing forces in <scenario name>:
Answer variable: <opposing force>
Control type: multiple line, height 4
Ontology acons:
<opposing force> instance of opposing force
<scenario name> has as opposing force <opposing force>
Script calls:
Elicit properes of the instance <opposing force> in new window

Figure 12.35. Elicitation script for a feature.

Figure 12.37, corresponding to an older version of the Disciple system. In the Feature
Hierarchy Browser, the KE has selected the feature has_as_opposing_force, and then has
clicked on the Script button. As a result, an editor for the elicitation script was opened, as
shown in the right-hand side of Figure 12.37. Then the KE has selected the type of script to

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 379

object
subconcept of
scenario force

subconcept of

opposing force
instance of

WW II Europe 1943

object

subconcept of
scenario force

subconcept of

opposing force
instance of
instance of

Allied Forces 1943

has as opposing force
WW II Europe 1943
has as opposing force
European Axis 1943

Figure 12.36. The effect of the execution of an elicitation script on the ontology.

be specified: “eliciting the values of has_as_opposing_force for an instance.” As a result,

the agent has displayed a panel requesting several items of information from the KE:

Control type: multiple-line

Question: Name the opposing forces in <scenario-name>:
Height: 4
Answer variable: <opposing-force-name>
...

Next, in the “Ontology actions” panel shown in the middle-right of Figure 12.37, the KE has
indicated how the ontology will be extended with the elicited values of the specified variables:

<opposing-force-name> instance-of Opposing_force

<scenario-name> has_as_opposing_force <opposing-force-name>

Finally, as shown at the bottom-right of Figure 12.37, the KE has indicated the script
to be called after the execution of the current script (“elicit properties of an instance”),
how will it be displayed (in a new window), and its parameters (<opposing-force-name>
and Opposing_force).

13:54:21,
.013
 380 Chapter 12. Disciple Agents

Figure 12.37. Interface of the Script Editor of Disciple-COG.

12.4.5 Agent Teaching and Learning

The teaching and learning of Disciple-COG follow the approach described in the previous
sections of this book. Therefore, in this section, we are only going to illustrate them with an
example.
The expert formulates an initial problem, such as “Analyze the strategic COG candi-
dates for WWII Europe 1943” (see Figure 12.38), and shows the agent how to solve this
problem by using the problem reduction paradigm described in Section 4.2. The expert
uses natural language, as if he or she would think aloud. The expert asks a question related
to some piece of information that is relevant to solving the current problem. The answer
identifies that piece of information and leads the expert to reduce the current problem to a
simpler problem (or, in other cases, to several simpler problems): “Analyze the strategic
COG candidates for Allied Forces 1943.” Figure 12.38 shows a sequence of problem
reduction steps.
Each step consists of a problem, a question, its answer, and a subproblem. From each
of these steps, Disciple-COG learns a general problem reduction rule by using its ontology
as a generalization hierarchy and the methods presented in Chapters 9 and 10. They will
be briefly illustrated in the following.
Let us consider the fourth step from the problem reduction tree in Figure 12.38, shown
also in the upper-left part of Figure 12.39. As discussed in Chapter 9, rule learning is a
mixed-initiative process between the expert (who knows why the reduction is correct and
can help the agent to understand this) and the Disciple-COG agent (which is able to
generalize the problem reduction example and its explanation into a general rule by using
the object ontology as a generalization language).

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 381

We need to
1. Modeling Analyze the strategic COG candidates for WWII Europe 1943. 2. Learning
Which is an opposing force in the WWII Europe 1943 scenario?
Rule1
Allied Forces 1943.
Therefore we need to
Learns task
Analyze the strategic COG candidates for Allied Forces 1943. reduction
rules
Provides What type of force is Allied Forces 1943?
examples
Allied Forces 1943 is a multimember force. Rule2
of correct
task Therefore we need to
reduction Analyze the strategic COG candidates for Allied Forces 1943
which is a multimember force.
steps
What type of center of gravity should we
consider for this multimember force?
Rule3
We consider one corresponding to a member of it.
Therefore we need to
Analyze the strategic COG candidates for a
member of Allied Forces 1943.

Which is a member of Allied Forces 1943?

Rule4
US 1943
Therefore we need to
Analyze the strategic COG candidates for US 1943.

3. Solving ••• 5. Refining

Therefore we need to
Applies Rule4 Analyze the strategic COG candidates for a member of
European Axis 1943.
4. Critiquing Which is a member of European Axis 1943? Refines
Rule4
Accepts Germany 1943
?
reduction Therefore we need to
Analyze the strategic COG candidates for Germany 1943.

6. Solving ••• 8. Refining

Therefore we need to
Applies Rule4 Analyze the strategic COG candidates for a member of
European Axis 1943.
7. Critiquing Which is a member of European Axis 1943? Refines
Rule4
Rejects Finland 1943
?
reduction Therefore we need to
Analyze the strategic COG candidates for Finland 1943.

Figure 12.38. An illustration of mixed-initiative modeling, problem solving, and learning.

The question and its answer from the problem reduction step represent the expert’s
reason (or explanation) for performing that reduction. Because they are in natural lan-
guage, the expert has to help Disciple-COG “understand” them in terms of the concepts
and the features from the object ontology. For instance, the meaning of the question/
answer pair from the example in Figure 12.39 (i.e., “Which is a member of Allied Forces
1943? US 1943”) is “Allied Forces 1943 has as member US 1943.”

13:54:21,
.013
382 Chapter 12. Disciple Agents

Posive Example 1 Rule4 learned from Example 1

We need to IF
Analyze the strategic COG candidates for a member Analyze the strategic COG candidates for a
of Allied Forces 1943. member of ?O1.

Which is a member of Allied Forces 1943? Queson: Which is a member of ?O1?

Answer: ?O2
US 1943 Main Condion
Therefore we need to Plausible Upper Bound Condion
Analyze the strategic COG candidates for US 1943. ?O1 is mulmember force
has as member ?O2
?O2 is force
Allied Forces 1943 has as member US 1943
Plausible Lower Bound Condion
Explanaon rewrien as ?O1 is equal partners mulstate alliance
has as member ?O2
?O1 is Allied Forces 1943
has as member ?O2 ?O2 is single-state force
?O2 is US 1943 THEN
Analyze the strategic COG candidates for ?O2.

Figure 12.39. Rule learning from an example and its explanation.

Based on the example and its explanation from the left-hand side of Figure 12.39,
Disciple-COG learns the rule from the right-hand side of Figure 12.39.
The structure of the rule is generated from the structure of the example where each
instance (e.g., Allied Forces 1943) and each constant (if present) is replaced with a variable
(e.g., ?O1). The variables are then used to express the explanation of the example as a very
specific applicability condition of the rule, as shown in the bottom-left part of Figure 12.39.
Finally, the plausible version space condition of the rule is generated by generalizing
the specific applicability condition in two ways.
The plausible lower bound condition is the minimal generalization of the specific
condition, which does not contain any specific instance. This generalization is performed
in the context of the agent’s ontology, in particular the ontology fragment shown in
Figure 10.22 (p. 323). The least general concepts from the ontology in Figure 10.22 that
cover Allied Forces 1943 are opposing force and equal partners multistate alliance. However,
Allied Forces 1943 has the feature has as member, and therefore any of its generalizations
should be in the domain of this feature, which happens to be multimember force. As a
consequence, the minimal generalization of Allied Forces 1943 is given by the following
expression:

fopposing force; equal partners multistate allianceg \ fmultimember forceg

¼ fequal partners multistate allianceg

Similarly (but using the range of the has as member feature, which is force), Disciple-COG
determines the minimal generalizations of US 1943 as follows:

fsingle-state forceg \ fforceg ¼ fsingle-state forceg

The reason the lower bound cannot contain any instance is that the learned rule will be
used by Disciple-COG in other scenarios (such as Afghanistan 2001–2002), where the
instances from WWII Europe 1943 do not exist, and Disciple-COG would not know how
to generalize them.

13:54:21,
.013
 12.4. Disciple-COG: Center of Gravity Analysis 383

The plausible upper bound condition is the maximal generalization of the specific
condition and is generated in a similar way. In particular, the maximal generalization of
Allied Forces 1943 is given by the following expression:

fobjectg \ fmultimember forceg ¼ fmultimember forceg

Also, the maximal generalization of US 1943 is:

fobjectg \ fforceg ¼ fforceg

As Disciple-COG learns new rules from the expert, the interaction between the expert and
Disciple-COG evolves from a teacher–student interaction toward an interaction where
both collaborate in solving a problem. During this mixed-initiative problem-solving phase,
Disciple-COG learns not only from the contributions of the expert, but also from its own
successful or unsuccessful problem-solving attempts, which lead to the refinement of the
learned rules.
As indicated in Figure 12.38, Disciple-COG applied Rule 4 to reduce the task, “Analyze
the strategic COG candidates for a member of European Axis 1943,” generating an example
that is covered by the plausible upper bound condition of the rule. This reduction was
accepted by the expert as correct. Therefore, Disciple-COG generalized the plausible lower
bound condition to cover it. For instance, European Axis 1943 is a multimember force, but it
is not an equal partners multistate alliance. It is a dominant partner multistate alliance
dominated by Germany 1943, as can be seen in Figure 10.22 (p. 323). As a consequence,
Disciple-COG automatically generalizes the plausible lower bound condition of the rule to
cover this example. The refined rule is shown in the left-hand side of Figure 12.40. This
refined rule is then generating the task reduction from the bottom part of Figure 12.38.
Although this example is covered by the plausible lower bound condition of the rule, the
expert rejects the reduction as incorrect. This shows that the plausible lower bound
condition is not less general than the concept to be learned, and it would need to be
specialized.
This rejection of the reduction proposed by Disciple-COG initiates an explanation
generation interaction during which the expert will have to help the agent understand
why the reduction step is incorrect. The explanation of this failure is that Finland 1943 has
only a minor military contribution to European Axis 1943 and cannot, therefore, provide the
center of gravity of this alliance. The actual failure explanation (expressed with the terms
from the object ontology) has the form:

Finland 1943 has as military contribution military contribution of Finland 1943 is minor
military contribution

Based on this failure explanation, Disciple-COG generates a plausible version space for an
Except When condition and adds it to the rule, as indicated on the right-hand side of
Figure 12.40. In the future, this rule will apply only to situations where the main condition
is satisfied and the Except When condition is not satisfied.

12.4.6 Experimental Results

The most remarkable feature of Disciple-COG is that it was developed to be actually
used to teach military personnel and it has been indeed used as a learning and a

13:54:21,
.013
384 Chapter 12. Disciple Agents

Rule4 after Positive Example 2 Rule4 aer Negative Example 3

IF IF
Analyze the strategic COG candidates for a Analyze the strategic COG candidates for a
member of ?O1. member of ?O1.

Question: Which is a member of ?O1? Question: Which is a member of ?O1?

Answer: ?O2 Answer: ?O2
Main Condition Main Condition
Plausible Upper Bound Condition Plausible Upper Bound Condition
?O1 is multimember force ?O1 is multimember force
has as member ?O2 has as member ?O2
?O2 is force ?O2 is force
Plausible Lower Bound Condition Plausible Lower Bound Condition
?O1 is multistate alliance ?O1 is multistate alliance
has as member ?O2 has as member ?O2
?O2 is single-state force ?O2 is single-state force
Except-When Condition
THEN
Analyze the strategic COG candidates for ?O2. Plausible Upper Bound Condition
?O2 is force
has as military contribution ?O3
?O3 is minor military contribution
Plausible Lower Bound Condition
?O2 is single-state force
has as military contribution ?O3
?O3 is minor military contribution
THEN
Analyze the strategic COG candidates for ?O2.

Figure 12.40. Rule refined with additional examples.

decision-support assistant in courses or individual lectures at the U.S. Army War College,
Air War College, Joint Forces Staff College, U.S. Army Intelligence Center, and other
civilian, military, and intelligence institutions (Tecuci et al., 2002a; 2002b). In particular,
successive versions of Disciple-COG have been used for elective courses and have been
part of the U.S. Army War College curriculum, uninterruptedly, since 2001, for a decade.
The textbook Agent-Assisted Center of Gravity Analysis (Tecuci et al., 2008b) provides a
detailed presentation of this agent, the embodied theory for COG determination that is
consistent with the joint military doctrine, and the use of this agent for the education
of strategic leaders. It includes a CD with lecture notes and the last version of the agent
(see lac.gmu.edu/cog-book/).
Each year, after being used in one or two courses, Disciple-COG was evaluated by the
students. The following, for instance, describes the evaluation results obtained in one of
these courses taught at the U.S. Army War College.
Each military student used a copy of the trained Disciple-COG agent as an intelligent
assistant that helped him or her to develop a center of gravity analysis of a war scenario.
As illustrated in Section 12.4.2, each student interacted with the scenario elicitation
module that guided him or her to describe the relevant aspects of the analyzed scenario.
Then the student invoked the autonomous problem solver (which used the rules learned

13:54:21,
.013
12.4. Disciple-COG: Center of Gravity Analysis 385

by Disciple-COG) and the report generator, obtaining a center of gravity analysis

report. This report contained the center of gravity candidates found by Disciple-
COG, together with the justifications for their identification as candidates, and
the justifications for the results of their testing (i.e., their elimination or their
preservation as likely centers of gravity). These justifications are generated based
on the rules learned by Disciple-COG and are intended to help the students learn
how to identify and test the center of gravity candidates for war scenarios. The
students were asked to study and evaluate the justifications generated by Disciple-
COG and to finalize the report. Figure 12.41 summarizes the results of their evalu-
ations. For instance, out of the 110 justifications generated for all the analyzed
scenarios, 76 were considered correct, 30 acceptable, and only 4 incorrect. Moreover,
most of the time the students have found these justifications to be complete and
easy to understand.
The use of Disciple-COG extended over four three-hour sessions. At the end, the
students were asked to evaluate a wide range of aspects related to the usability and
utility of the three Disciple-COG modules used and of Disciple-COG as a whole. The
students were presented with statements on various aspects of Disciple-COG and were
asked to express their level of agreement with these statements by using a five-point
scale (strongly disagree, disagree, neutral, agree, and strongly agree). Figure 12.42
includes some of the global evaluation results, showing that the students considered
Disciple-COG easy to learn and use and its use as assignment well suited to the
course’s learning objectives. Disciple-COG helped them to learn to perform a strategic
COG analysis of a scenario and should be used in future versions of this course.
Finally, a system such as Disciple-COG could be used in other U.S. Army War College
courses.
The evaluation results in Figure 12.42 are consistent with the results obtained in other
similar courses. For example, Figure 12.43 presents some of the evaluation results
obtained in a course at the Air War College.
This demonstrates that the Disciple approach allowed the development of an agent that
has been found to be useful for a complex military domain.

The justification is…

correct. 76

acceptable. 30

incorrect. 4

complete. 62
relatively complete. 43

significantly incomplete. 5

easy to understand. 71

relatively understandable. 36

difficult to understand. 3
0 20 40 60 80 100

Figure 12.41. Subjective evaluation of the justifications generated by Disciple-COG.

13:54:21,
.013
386 Chapter 12. Disciple Agents

8 8 The use of Disciple is an 8 Disciple helped me to learn

It was easy to learn
7 7 assignment that is well suited to 7 to perform a strategic COG
how to use Disciple
6 6 the course's learning objectives 6 analysis of a scenario
5 5 5
4 4 4
3 3 3
2 2 2
1 1 1
0 0 0
Strongly Disagree

Disagree

Neutral

Agree

Strongly Agree

Strongly Disagree

Disagree

Neutral

Agree

Strongly Agree

Strongly Disagree

Disagree

Neutral

Agree

Strongly Agree
It was easy to Disciple should be used in A system like Disciple could be
8 8 8
7 use Disciple 7
future versions of this course 7 used in other USAWC courses
6 6 6
5 5 5
4 4 4
3 3 3
2 2 2
1 1 1
0 0 0

Strongly Disagree

Disagree

Neutral

Agree

Strongly Agree
Strongly Disagree

Disagree

Neutral

Agree

Strongly Agree

Strongly Disagree

Disagree

Neutral

Agree

Strongly Agree

Figure 12.42. Global evaluation results from a COG class experiment at the U.S. Army War
College.

8 The use of Disciple is an 8

Disciple helped me to learn
7 assignment that is well suited to 7 to perform a strategic COG
6 the course's learning objecves 6 analysis of a scenario
5 5
4 4
3 3
2 2
1 1
0 0
Strongly Disagree Neutral Agree Strongly Strongly Disagree Neutral Agree Strongly
Disagree Agree Disagree Agree

8 Disciple should be 8 A system like Disciple

7 used in future versions 7 could be used in other
of this course Air War College courses
6 6
5 5
4 4
3 3
2 2
1 1
0 0
Strongly Disagree Neutral Agree Strongly Strongly Disagree Neutral Agree Strongly
Disagree Agree Disagree Agree

Figure 12.43. Global evaluation results from a COG class experiment at the Air War College.

13:54:21,
.013
 12.5. Disciple-VPT 387

12.5 DISCIPLE-VPT: MULTI-AGENT COLLABORATIVE

PLANNING

12.5.1 Introduction
While most of this book has focused on the development of agents for evidence-
based reasoning, the previous sections of this chapter have shown the generality of
the knowledge representation, reasoning, and learning methods of the Disciple
approach by describing other types of Disciple agents. This section takes a further
step in this direction by presenting a different type of Disciple architecture, called
Disciple-VPT (virtual planning team). Disciple-VPT (Tecuci et al., 2008c) consists of
virtual planning experts that can collaborate to develop plans of actions requiring
expertise from multiple domains. It also includes an extensible library of virtual
planning experts from different domains. Teams of such virtual experts can be rapidly
assembled from the library to generate complex plans of actions that require their
joint expertise. The basic component of the Disciple-VPT tool is the Disciple-VE
(virtual experts) learning agent shell that can be taught directly by a subject matter
expert how to plan, through planning examples and explanations, in a way that is
similar to how the expert would teach an apprentice. Copies of the Disciple-VE shell
can be used by experts in different domains to rapidly populate the library of virtual
experts of Disciple-VPT.
A virtual planning expert is defined as a knowledge-based agent that can rapidly
acquire planning expertise from a subject matter expert and can collaborate with other
virtual experts to develop plans that are beyond the capabilities of individual virtual
experts.
In this section, by planning, we mean finding a partially ordered set of elementary
actions that perform a complex task (Ghallab et al., 2004).
A representative application of Disciple-VPT is planning the response to emergency
situations, such as the following ones: a tanker truck leaking toxic substance near a
residential area; a propane truck explosion; a biohazard; an aircraft crash; a natural
disaster; or a terrorist attack (Tecuci et al., 2007d). The U.S. National Response Plan
(DHS, 2004) identifies fifteen primary emergency support functions performed by federal
agencies in emergency situations. Similarly, local and state agencies undertake these
functions responding to such emergencies without or before any federal assistance is
provided. Each such function defines an expertise domain, such as emergency manage-
ment; police operations; fire department operations; hazardous materials handling; health
and emergency medical services; sheltering, public works, and facilities; and federal law
enforcement. In this case, the library of Disciple-VPT will include virtual experts corres-
ponding to these domains.
The next section presents the general architecture of Disciple-VPT and discusses the
different possible uses of this general and flexible tool. Section 12.5.3 describes a sample
scenario from the emergency response planning area, which is used to present the
features of Disciple-VPT. Section 12.5.4 presents the architecture of the Disciple-VE
learning agent shell, which is the basis of the capabilities of Disciple-VPT, including its
learning-oriented knowledge representation. Section 12.5.5 presents the hierarchical task
network (HTN) planning performed by the Disciple virtual experts. After that, Section
12.5.6 presents a modeling language and methodology developed to help a subject matter

13:54:21,
.013
 388 Chapter 12. Disciple Agents

expert explain to a Disciple-VE agent how to plan, by using the task reduction paradigm.
Section 12.5.7 discusses how a Disciple-VE agent can perform complex inferences as part
of a planning process. The next two sections, 12.5.8 and 12.5.9, present the teaching and
learning methods of Disciple-VE, first for inference tasks and then for planning tasks.
Section 12.5.10 presents the organization of the library of virtual experts of Disciple-VPT.
After that, Section 12.5.11 presents Disciple-VPT’s approach to multi-agent collaboration.
Section 12.5.12 discusses the development of two virtual experts, one for fire operations
and the other for emergency management. Section 12.5.13 presents some evaluation
results, and Section 12.5.14 summarizes our research contributions.

12.5.2 The Architecture of Disciple-VPT

Figure 12.44 presents the end-user’s view of the three major components of Disciple-VPT:

VE Assistant, an agent that supports the user in using Disciple-VPT

VE Library, an extensible library of virtual planning experts

VE Team, a dynamically assembled team of virtual experts selected from the VE Library

The user interacts with the VE Assistant to specify a situation and the profiles of several
human experts who may collaborate to plan the achievement of various goals in that
situation. Next, a team of virtual planning experts with similar profiles is automatically
assembled from the VE Library. This VE Team then simulates the planning performed by
the human experts, generating plans for achieving various goals in the given situation.
The goal of a system such as Disciple-VPT is to allow the development of collaborative
planners for a variety of applications by populating its library with corresponding virtual

VE Library

Disciple-VE
Disciple-VE
Disciple-VE
Disciple-VE
Disciple-VE
Disciple-VE
Disciple-VE
Disciple-VE
Disciple-VE

User
VE Assistant

VE Team
Disciple-VE Disciple-VE

DISTRIBUTED
KNOWLEDGE BASE

KB

Disciple-VE Disciple-VE
KB KB

KB KB KB

Disciple-VE Disciple-VE

Figure 12.44. Overall architecture of Disciple-VPT.

13:54:21,
.013
 12.5. Disciple-VPT 389

experts. For instance, planning the response to emergency situations requires virtual
experts for emergency management, hazardous materials handling, federal law enforce-
ment, and so on. Other application areas, such as planning of military operations, require
a different set of virtual experts in the VE Library. Moreover, for a given type of task and
application area, different multidomain planning systems can be created by assembling
different teams of virtual experts.
There are many ways in which a fully functional Disciple-VPT system can be used for
training or actual planning assistance. For instance, in the context of emergency response
planning, it can be used to develop a wide range of training scenarios by guiding the user
to select between different scenario characteristics. Disciple-VPT could also be used to
assemble teams of virtual planning experts who can demonstrate and teach how people
should plan the response to various emergency situations. Another approach is to assem-
ble combined teams that include both people and virtual experts. The team members will
then collaborate in planning the response to the generated emergency scenario. In a
combined team, human responders can play certain emergency support functions by
themselves or can play these functions with the assistance of corresponding virtual
experts. During the training exercise, a responder who has a certain emergency support
function will learn how to perform that function from a corresponding virtual expert with
higher competence. The responder will also learn how to collaborate with the other
responders or virtual experts who perform complementary support functions.
The Disciple-VPT approach to expert problem solving extends significantly the applic-
ability of the classical expert systems (Buchanan and Wilkins, 1993; Durkin, 1994; Awad,
1996; Jackson, 1999; Awad and Ghaziri, 2004). Such an expert system is limited to a narrow
expertise domain, and its performance decreases dramatically when attempting to solve
problems that have elements outside its domain of expertise. On the contrary, a Disciple-
VPT type system can efficiently solve such problems by incorporating additional virtual
experts. Because many expert tasks actually require collaboration with other experts, a
Disciple-VPT–type system is more suitable for solving real-world problems.
The next section introduces in more detail a scenario from the emergency response
planning area that informed the development of Disciple-VPT.

12.5.3 The Emergency Response Planning Problem

Emergency response planning was introduced in the previous sections. The following is a
sample emergency situation that will be used to present Disciple-VPT:

Workers at the Propane bulk storage facility in Gainsville, Virginia, have been
transferring propane from a train car to fill one of two 30,000 gallon bulk
storage tanks. A fire is discovered in the fill pipe at the bulk tank, and a large
fire is developing. The time is 15:12 on a Wednesday in the month of May. The
temperature is 72 degrees and there is a light breeze out of the west. The
roads are dry and traffic volume is moderate. The fire department is summoned
to the scene five minutes after the fire started. The facility is located in a
rapidly growing area 2,000 feet from an interstate highway and 200 feet from
two heavily traveled U.S. highways. New shopping centers have popped up in
the area, including grocery stores, large box building supply facilities, and large
box retail facilities. As always, these facilities are accompanied by fast
food restaurants and smaller retail stores. Residential concentrations include

13:54:21,
.013
 390 Chapter 12. Disciple Agents

approximately 2,400 residents. The Local Emergency Operations Plan has all
the required components, including public information, communications, and
sheltering. Shelters utilize schools managed by the Red Cross. The Virginia
Department of Transportation provides highway services.

Planning the appropriate response to this emergency situation requires the collaboration
of experts in fire department operations, emergency management, and police operations.
The generated plan will consist of hundreds of partially ordered actions.
One group of actions deals with the arrival of resources, such as fire units, emergency
management services units, police units, as well as individuals with different areas of
expertise (e.g., emergency manager, safety officer, highway supervisor, planning officer,
training logistics officer, public information officer).
Another group of actions deals with the establishment of the structure of the Incident
Command System (ICS) and the allocation of resources based on the evaluation of the
situation. The structure of the ICS follows the standard U.S. National Incident Manage-
ment System (FEMA, 2007). The National Incident Management System establishes
standard incident management processes, protocols, and procedures so that all local,
state, federal, and private-sector emergency responders can coordinate their responses,
share a common focus, and more effectively resolve events. Its main components are the
unified command, the command staff, and the general staff. The structure and organiza-
tion of these components depend on the current situation. For example, in the case of the
preceding scenario, the unified command includes representatives from the fire depart-
ment, police department, highway department, and propane company. The command
staff includes a safety officer, a public information officer, and a liaison officer. The general
staff includes an operation section, a planning section, a logistics section, and a finance
and administration section. Each of these sections is further structured and staffed.
Yet other groups of actions deal with the various activities performed by the compon-
ents of the Incident Command System. For instance, in the case of the preceding scenario,
the fire management group may perform the cooling of the propane tank with water. The
evacuation branch may evacuate the Gainsville hot zone. The emergency manager may
arrange for transportation, sheltering, and emergency announcements to support the
evacuation. The Gainsville perimeter control branch implements the perimeter control
for the Gainsville hot zone. The Gainsville traffic control branch implements the traffic
control to facilitate the evacuation of the Gainsville hot zone. The Gainsville command
establishes rapid intervention task forces to respond if the propane tank explodes.
One difficulty in generating such a plan, apart from the fact that it involves many
actions, is that the actions from the preceding groups are actually performed in parallel.
The goal of Disciple-VPT is to provide a capability for rapid and low-cost development of
virtual planning experts to be used in this type of multidomain collaborative planning.
Moreover, the plans generated by the system should be more comprehensive than those
produced by a collaborative team of humans and should be generated much faster and
cheaper than currently possible. The next section introduces the Disciple-VE learning
agent shell, which is the basis of Disciple-VPT.

12.5.4 The Disciple-VE Learning Agent Shell

The concept of a learning agent shell was introduced in Section 3.2.3. The Disciple-VE
learning agent shell is an extension of a Disciple shell that incorporates capabilities of

13:54:21,
.013
12.5. Disciple-VPT 391

learning for planning to enable rapid development of knowledge-based planners, as will

be discussed in the following sections.
As discussed in Chapter 4, the general problem-solving paradigm of a Disciple agent is
problem reduction and solution synthesis (Nilsson, 1971; Powell and Schmidt, 1988;
Tecuci, 1988; Durham, 2000). In the context of planning, this approach reduces to
hierarchical task network (HTN) planning, where an initial complex task is reduced to a
partially ordered set of elementary actions (Tate, 1977; Allen et al., 1990; Nau et al., 2003;
Ghallab et al., 2004). In the case of Disciple-VE, planning tasks are integrated with
inference tasks, which significantly increases the power of HTN planning.
As with other Disciple agents, the knowledge base of Disciple-VE contains two main
types of knowledge: an object ontology and a set of reasoning rules. A fragment of the
object ontology for emergency response planning is shown in Figure 12.45. For example,
the major fire emergency concept represents all the incidents that are major fire emergen-
cies. One such instance is Gainsville incident. As shown in Figure 12.45, this information is
represented as “Gainsville incident instance of major fire emergency.”
As discussed in Section 5.3, the instances and concepts are organized into generaliza-
tion hierarchies, as illustrated in Figure 12.45. These structures are not strict hierarchies,
meaning that a concept may be a subconcept of several concepts (e.g., propane is both a
chemical substance and a hazardous substance).
The instances and concepts may have features representing their properties and
relationships, such as “Gainsville incident is caused by fire1,” and “fire1 is fueled by gas
propane f1.” The bottom part of Figure 12.46 shows all the features of fill pipe1 in the
interface of the Association Browser of Disciple-VE.
Each feature, such as is fueled by, is characterized by a domain (in this case, fire) and a
range (hazardous substance). The features are also organized into a generalization hier-
archy. For example, the top part of Figure 12.46 shows (a rotated view of) a fragment of the
feature hierarchy, in the interface of the Hierarchical Browser of Disciple-VE. In this
hierarchy, the feature has as part (shown in the left-hand side of Figure 12.46) is more
general than has as member, which, in turn, is more general than has as supervisor.
Together, the object hierarchy and the feature hierarchy represent the object ontology
of a Disciple-VE agent. Thus, the object ontology is a hierarchical representation of the
objects from the application domain, representing the different kinds of objects, the proper-
ties of each object, and the relationships existing between objects.
In general, the object ontology does not contain all the relevant concepts and instances
from the application domain and is therefore incomplete. Also, the representation of a
given concept or instance may not include all its relevant features, being itself incomplete.
Such an object ontology will have to be extended by the agent during the planning and
learning process.
As discussed previously, and also illustrated in the following sections, the object
ontology plays a crucial role in Disciple, being the basis of knowledge representation,
user–agent communication, planning, knowledge acquisition, and learning.
The representation of concepts in Disciple-VE is discussed in Section 7.2. For example,
the concept “fire fueled by gas propane” is represented by the pair {?O1, ?O2}, where ?O1
is a fire fueled by ?O2, and ?O2 is gas propane, as indicated by the following expression.

?O1 instance of fire

is fueled by ?O2
?O2 instance of gas propane

13:54:21,
.013
 object

... subconcept of ...

event equipment component
substance
... ... subconcept of
BLEVE subconcept of
subconcept of
valve
...
water pipe fire hydrant
emergency chemical substance
subconcept of
hazardous substance
... ... subconcept of emergency relief valve
subconcept of ... shutoff valve
subconcept of
explosive discharge pipe
major emergency instance of opened
fill pipe has as gate status
propane toxic substance shut off valve1
corrosive liquid
subconcept of has as operating status
subconcept of ...
.013

instance of damaged
container
major fire emergency gas propane liquid propane
has as part
13:54:21,

subconcept of
fire instance of fill pipe1 ...
has as content tank
instance of
gas propane f1 is connected to
subconcept of
Gainsville incident instance of
is fueled by propane tank
is located in is situated in
is caused by
... instance of fire engine water tank

large has as size fire1 is impinging on

propane tank1

Figure 12.45. Fragment of the object ontology from the emergency planning area.
12.5. Disciple-VPT 393

Figure 12.46. Feature hierarchy (top) and instance description (bottom).

In general, a concept may be a conjunctive expression. For example, the following

expression represents the concept “fire fueled by gas propane where the fire is
not small.”

?O1 instance of fire

is fueled by ?O2
?O2 instance of gas propane
Except-When
?O1 instance of fire
has as size small

13:54:21,
.013
 394 Chapter 12. Disciple Agents

The reasoning rules of Disciple-VE are expressed with the elements of the object
ontology. Reduction rules indicate how general planning or inference tasks can be reduced
to simpler tasks, actions, or solutions. Synthesis rules indicate how solutions of simpler tasks
can be combined into solutions of complex tasks, or how actions can be combined into
partially ordered plans for more complex tasks.
The next section introduces the type of hierarchical task network planning performed
by Disciple-VE and the associated elements that are represented into its knowledge base.

12.5.5 Hierarchical Task Network Planning

The planning paradigm used by Disciple-VE is HTN planning (Ghallab et al., 2004),
extended to facilitate agent teaching and learning and mixed-initiative planning.
The goal of an HTN planner is to find a partially ordered set of elementary actions that
perform a complex task by successively decomposing the task into simpler and simpler tasks,
down to the level of elementary actions. HTN planning is the planning approach that has
been used for practical applications more than any other approach because it is closer to
how human experts think when solving a planning problem.
We will illustrate the HTN planning process performed by a Disciple-VE agent with
the abstract example from Figure 12.47. In this example, Planning Task 1 is reduced
to Planning Task 2 and Planning Task 3. This means that by performing Planning
Task 2 and Planning Task 3, one accomplishes the performance of Planning Task 1.
Because Planning Task 2 is reduced to Action 1 and Action 2, and Planning Task 3 is
reduced to Action 3, a plan for performing Planning Task 1 consists of Action 1, Action 2,
and Action 3.
There are two types of reductions: task decomposition and task specialization. Task
decomposition means breaking a task into a partially ordered set of subtasks and/or actions.
Task specialization means reducing a task to a more detailed task or to an action.
The tasks or actions in a decomposition can be partially ordered. For instance, in
Figure 12.47, Action 2 has to be performed after Action 1 has been performed. Notice also

S1 Planning Task 1 Goal 1 S4

Question 1/Answer 1

Abstract Task 2 Abstract Task 3

Preconditions 2 Preconditions 3
Planning Planning
S1 Goal 2 S3 S3 Goal 3 S4
Task 2 Task 3

Question 2/Answer 2 Question 3/Answer 3

Abstract Task 4 Abstract Task 5 Abstract Task 6

Preconditions 4 Preconditions 5 Preconditions 6

Delete Add Delete Add Delete Add
S1 Action 1 S2 Action 2 S3 Action 3 S4
Effects 1 Effects 1 Effects 2 Effects 2 Effects 3 Effects 3
Resources 1 Duration 1 Resources 2 Duration 2 Resources 3 Duration 3

Figure 12.47. Hierarchical task network planning example.

13:54:21,
.013
12.5. Disciple-VPT 395

that there is no order relation between Planning Task 2 and Planning Task 3. This means
that these tasks may be performed in parallel or in any order. Stating that Planning
Task 3 is performed after Planning Task 2 would mean that any subtask or action of
Planning Task 3 has to be performed after all subtasks and actions of Planning Task 2.
Formulating such order relations between the tasks significantly increases the efficiency of
the planning process because it reduces the number of partial orders that it has to
consider. On the other hand, it also reduces the number of generated plans if the tasks
should not be ordered.
Planning takes place in a given world state. A world state is represented by all the objects
present in the world together with their properties and relationships at a given moment of
time. For instance, the bottom part of Figure 12.45 shows a partial representation of a
world state where fire1, which is situated in fill pipe1, is impinging on propane tank1. As will
be discussed in more detail in Section 12.5.10, each world state is represented by Disciple-
VE as a temporary state knowledge base.
The states are changed by the performance of actions. Abstract representations of
actions are shown at the bottom of Figure 12.47. An action is characterized by name,
preconditions, delete effects, add effects, resources, and duration. An action can be per-
formed in a given world state Si if the action’s preconditions are satisfied in that state. The
action’s execution has a duration and requires the use of certain resources. The resources are
objects from the state Si that are uniquely used by this action during its execution. This
means that any other action that would need some of these resources cannot be executed in
parallel with it. As a result of the action’s execution, the state Si changes into the state Sj, as
specified by the action’s effects. The delete effects indicate what facts from the initial state Si
are no longer true in the final state Sj. The add effects indicate what new facts become true in
the final state Sj.
An action from the emergency planning area is shown in Figure 12.48. The action’s
preconditions, name, delete, and add effects are represented as natural language phrases
that contain instances, concepts, and constants from the agent’s ontology. The action’s
duration can be a constant, as in this example, or a function of the other instances from the
action’s description. Resources are represented as a list of instances from the ontology.
The starting time is computed by the planner.

Figure 12.48. An example of an action.

13:54:21,
.013
 396 Chapter 12. Disciple Agents

A goal is a representation of a partial world state. It specifies what facts should be true
in a world state so that the goal is achieved. As such, a goal may be achieved in several
world states.
A task is characterized by name, preconditions, and goal. A task is considered for
execution in a given world state if its preconditions are satisfied in that state. Successful
execution of the task leads to a new world state in which the task’s goal is achieved. Unlike
actions, tasks are not executed directly, but are first reduced to actions that are executed.
Figure 12.49 shows a task reduction tree in the interface of the Reasoning Hierarchy
Browser of Disciple-VE. The initial task, “Respond to the Gainsville incident,” is reduced to
five subtasks. The second of these subtasks (which is outlined in the figure) is successively
reduced to simpler subtasks and actions.
Figure 12.50 shows the reduction of the initial task in the interface of the Reasoning
Step Editor, which displays more details about each task. As in the case of an action, the
task’s name, preconditions, and goal are represented as natural language phrases that
include instances and concepts from the agent’s ontology. Notice that none of the visible
“After” boxes is checked, which means that Sub-task (1), Sub-task (2), and Sub-task (3) are
not ordered.
The single most difficult agent training activity for the subject matter expert is to make
explicit how he or she solves problems by using the task reduction paradigm, an activity that
we call modeling an expert’s reasoning. To cope with this problem, we have developed an
intuitive modeling language, a set of modeling guidelines, and a set of modeling modules
that help the subject matter experts to express their reasoning (Bowman, 2002). However,
planning introduces additional complexities related to reasoning with different world
states and with new types of knowledge elements, such as preconditions, effects, and
goals. For these reasons, and to facilitate agent teaching by a subject matter expert, we
have extended both the modeling approach of Disciple and the classical HTN planning
paradigm (Ghallab et al., 2004), as discussed in the next section.

12.5.6 Guidelines for HTN Planning

To teach the agent how to plan, the expert first has to show the agent an example in the
form of a planning tree such as the ones in Figures 12.47 and 12.50. The expert formulates
the initial task (e.g., Planning Task 1 in Figure 12.47) and then follows a systematic
procedure to develop a detailed plan of actions that perform the initial task. This follows
a task reduction paradigm where the initial task is successively reduced to simpler and
simpler tasks, down to the level of elementary actions. The partially ordered set of these
elementary actions represents the plan for performing the initial task.
As presented in Figure 12.47 and illustrated in Figures 12.49 and 12.50, the task
reduction process is guided by questions and answers, as if the expert is asking himself
or herself how to reduce the current task. Consider, for instance, a task that may be
reduced (i.e., performed) in different ways. Then the question should be related to the
factors that determine the reduction strategy to choose. Therefore, the answer will help the
expert to choose the strategy and define the reduction. If there is only one way to reduce
the current task, then no question/answer pair is necessary.
Thus, to develop a planning tree, follow the modeling guidelines discussed in Section
4.12. The other guidelines, discussed in the previous chapters, are also applicable, but
there are additional planning-specific guidelines, which will be discussed in the following.

13:54:21,
.013
.013
13:54:21,

Figure 12.49. An example of a task reduction tree in the Reasoning Hierarchy Browser.
398 Chapter 12. Disciple Agents

Figure 12.50. An example of a task reduction step in the Reasoning Step Editor.

Guideline 12.1. Use a plausible task ordering when specifying a

task decomposition
Notice that Planning Task 1 from Figure 12.47 has to be performed in the initial state S1.
Planning Task 2 is also performed in the state S1 and its preconditions have to be satisfied in
that state. Similarly, Action 1 has to be performed in the state S1. However, this action changes
the world state from S1 to S2. Therefore, Action 2 has to be performed in the state S2 and its
preconditions have to be satisfied in that state. Moreover, it also changes the world state to S3.
What is the state in which Planning Task 3 is performed? Because there is no order
relationship between Planning Task 2 and Planning Task 3, Planning Task 3 can, in
principle, be performed either in the state S1, or S2, or S3. In reality, some order relation-
ships may be determined by the resources used by the elementary actions. For instance, if
both Action 2 and Action 3 need the same resource, they cannot be executed in parallel.

13:54:21,
.013
12.5. Disciple-VPT 399

When showing a planning example to the agent, the expert does not need to consider
all the previously discussed possible ordering relations, which would be very difficult in
the case of the complex problems addressed by Disciple-VE. Instead, the expert has to
consider one possible order, such as the one in which Planning Task 3 is performed after
Action 2, in state S3. This allows both the expert and the agent to have a precise
understanding of the state in which each task is performed, which is necessary in order
to check that its preconditions are satisfied. Thus, when specifying a decomposition of a
task into subtasks and/or actions, the expert has to describe the subtasks and actions in a
plausible order, even though they can also be performed in a different order.

Guideline 12.2. Specify the planning tree in a top-down and

left-to-right order
Allowing the process specified by Guideline 12.1 required the extension of the HTN
planning paradigm, as discussed in the following. Let us consider the decomposition of
Planning Task 1 from Figure 12.47 into Planning Task 2 and Planning Task 3, by
following the top-down and left-to-right process. At the time the expert has to specify
this decomposition, he or she knows that Planning Task 2 has to be performed in the
state S1, but the expert does not know the state in which Planning Task 3 has to be
performed. This state can be determined only after the entire subplan for Task 2 has
been specified. In other words, Task 3 can be only specified after the entire subplan for
Task 2 has been specified. In order to resolve this contradiction, we have introduced the
notion of abstract task.
An abstract task is a simplified specification of a task that does not depend on the actual
state in which the task is going to be executed. As such, an abstract task does not have any
precondition, and does not refer to any specific objects or their properties.
An abstract task can be reduced to a concrete task if certain preconditions are satisfied.
In principle, the same abstract task may be reduced to different concrete tasks. Therefore,
the abstract task is not a characteristic of a given concrete task.

Guideline 12.3. Define preconditions when reducing an abstract

task to a concrete task
When reducing an abstract task to a concrete task, formulate the preconditions to identify
those instances and constants from the current world state that are referred to in the name
of the concrete task, but are not referred to in the previous elements of the task reduction
step that includes this concretion.
To illustrate this guideline, let us consider again the reduction step from Figure 12.50
and the pane labeled “Sub-task (3).” Notice that the concrete task includes the following
instances: “Gainsville command” and “Gainsville incident.” Each of these instances appear in
the elements listed under Sub-task (2). For example, “Gainsville command” appears in the
“Goal” part. Therefore no preconditions are required to make the concretion from the
abstract task to the concrete task shown in the pane labeled “Sub-task (3).” However, the
expert may still wish (and is allowed) to specify preconditions that identify the instances
that appear in the concrete task, as was actually done in this example.
With the introduction of the abstract tasks, the expert can now reduce Planning Task 1 from
Figure 12.47 to Abstract Task 2 and Abstract Task 3. Then he or she can continue with the

13:54:21,
.013
 400 Chapter 12. Disciple Agents

reduction of Abstract Task 2 to Planning Task 2, a reduction performed in state S1. Precondition
2 represents the facts from the state S1 that are required in order to make this reduction.
After that, the expert continues with the reduction of Planning Task 2 to Action 1 and
Action 2. Thus Planning Task 2 is actually performed by executing Action 1 and Action 2,
which changes the world state from S1 to S3. At this point, the expert can specify the goal
achieved by Planning Task 2. This goal is an expression that depends on the effects of
Action 1 and Action 2, but is also unique for Task 2, which is now completely specified.
Next the expert can continue with planning for Abstract Task 3 in the state S3.

Guideline 12.4. Specify the goal of the current task to enable the
specification of the follow-on tasks
The goal of a task represents the result obtained if the task is successfully performed. The
main purpose of the goal is to identify those instances or facts that have been added by the
task’s component actions and are needed by its follow-on tasks or actions. Thus, specify
this goal to include these instances or facts.
To illustrate this guideline, let us consider the Sub-task (2) pane in Figure 12.50. Notice
that the two instances from the “Goal” part (“Gainsville command” and “Gainsville ICS”) are
used in the follow-on expressions of the reduction from Figure 12.50.
The Reasoning Hierarchy Browser, shown in Figure 12.49, and the Reasoning Step
Editor, shown in Figure 12.50, support the modeling process. The Reasoning Hierarchy
Browser provides operations to browse the planning tree under development, such as
expanding or collapsing it step by step or in its entirety. It also provides the expert with
macro editing operations, such as deleting an entire subtree or copying a subtree and
pasting it under a different task. Each reduction step of the planning tree is defined by
using the Reasoning Step Editor, which includes several editors for specifying the com-
ponents of a task reduction step. It has completion capabilities that allow easy identifica-
tion of the names from the object ontology. It also facilitates the viewing of the instances
and concepts from the expressions being edited by invoking various ontology viewers.
An important contribution of Disciple-VE is the ability to combine HTN planning with
inference, as described in the following section.

12.5.7 Integration of Planning and Inference

As illustrated in Figure 12.47, each planning operation takes place in a given world state, and
the actions, through their effects, change this state. The planning process is complex and
computationally expensive because one has to keep track of these various world states.
However, some operations do not involve the change of the world state, but reasoning about
a given state. Let us consider the top-level reasoning steps from Figure 12.49, where the task,
”Respond to the Gainsville incident,” is reduced to five subtasks. The third of these subtasks,
“Gainsville command evaluates the Gainsville incident and determines the needed incident
action plan,” is further reduced to two inference actions (not shown in Figure 12.49):

Inference: Gainsville command evaluates the situation created by the Gainsville incident.
Inference: Gainsville command determines the incident action plan for overpressure
situation with danger of BLEVE.

13:54:21,
.013
12.5. Disciple-VPT 401

The first of these inference actions has as result “overpressure situation with danger of
BLEVE in propane tank1 caused by fire1.” BLEVE is the acronym for boiling liquid
expanding vapors explosion.
From the perspective of the planning process, an inference action simulates a complex
inference process by representing the result of that process as the add effect of the inference
action. An inference action is automatically reduced to an inference task. The inference
task is performed in a given world state to infer new facts about that state. These facts are
represented as the add effects of the corresponding inference action and added into the
world state in which the inference action is performed.
The inference process associated with an inference task is also performed by using the
task reduction paradigm, but it is much simpler than the planning process because all the
reductions take place in the same world state. An abstract example of an inference tree is
shown in Figure 12.51.
An inference task is performed by successively reducing it to simpler inference tasks, until
the tasks are simple enough to find their solutions. Then the solutions of the simplest tasks are
successively combined, from the bottom up, until the solution of the initial task is obtained.
This task reduction and solution synthesis process is also guided by questions and
answers, similarly to the planning process (and as discussed in Section 4.2). Figure 12.52
shows the top part of the inference tree corresponding to the following inference task:

Inference: Gainsville command determines the incident action plan for overpressure
situation with danger of BLEVE.

This task is first reduced to two simpler inference tasks:

Determine what can be done to prevent the overpressure situation with danger of
BLEVE to evolve in a BLEVE1.
Determine how to reduce the effects in case the overpressure situation with danger
of BLEVE does evolve in a BLEVE1.

The first of these subtasks is successively reduced to simpler and simpler subtasks, guided
by questions and answers, as shown in Figure 12.52.
Notice that an inference tree no longer needs to use elements such as abstract tasks,
preconditions, actions, effects, resources, or duration. The teaching process is also much
simpler than in the case of the planning process. Therefore, we will first present how the
expert can teach Disciple-VE to perform inference tasks. Then we will present how the

Inference Task 1 Solution 1

World
State S
Question/Answer Question/Answer

Inference Task 2 Solution 2 Inference Task 3 Solution 3

Question/Answer Question/Answer Question/

Answer

Inference Task 4 Solution 4 Inference Task 5 Solution 5

Question/ Question/
Answer Answer

Figure 12.51. Abstract inference tree.

13:54:21,
.013
 Reducon
Step
.013
13:54:21,

Figure 12.52. Inference reduction tree.

 12.5. Disciple-VPT 403

teaching and learning methods for inference tasks have been extended to allow the expert
also to teach the agent how to perform planning tasks.

12.5.8 Teaching Disciple-VE to Perform Inference Tasks

Disciple-VE learns inference tasks by employing the methods discussed in Chapters 9
and 10. Therefore, in the following discussion, we will illustrate this process by referring to
these chapters.
Let us consider again the fragment of the inference tree shown in Figure 12.52. During
the teaching process, the subject matter expert builds this inference tree. Each step in the
tree consists of a task, a question, its answer, and one or several subtasks. From each of
these steps, the agent learns a general task reduction rule. Table 9.1 defines the problem of
learning these rules.
Let us consider the third reduction step from the task reduction tree in Figure 12.52, a
step also shown in Figure 12.53. From this task reduction step, Disciple-VE learned the
rule shown in Figure 12.54. This is an IF-THEN rule that preserves the structure and
natural language patterns from the example. Indeed, the IF task, the question/answer pair,
and the THEN task are generalizations of the corresponding elements from the example
where the instances and constants have been replaced with variables. In addition, the rule
contains a main condition. An instance of the rule is considered a correct reduction if the
corresponding variable values satisfy the main condition.
The rule in Figure 12.54 is only partially learned because, instead of a single applicabil-
ity condition, it contains a plausible version space for it. The plausible lower bound of the
applicability condition is the set of the tuples of the rule variable values that are less
general than the corresponding elements of the Lower Bound column and satisfy the
relationships from the Relationship table. For example, any value of ?O4 should be an
instance of shut off valve, which has as operating status ?S1, which should have the value
damaged. Moreover, this value of ?O4 should be the value of the relationship has as part of
an instance of ?O5, which should be a fill pipe and should have the relationships indicated
in the relationship table, and so on. The plausible upper bound of the applicability
condition is interpreted in a similar way, using the concepts from the Upper
Bound column.
Rule learning is accomplished through a mixed-initiative process between the expert
(who knows why the reduction is correct and can help the agent to understand this) and

Figure 12.53. Example of a reduction step for an inference task.

13:54:21,
.013
404 Chapter 12. Disciple Agents

Figure 12.54. The rule learned from the example in Figure 12.53.

the Disciple-VE agent (which is able to generalize the task reduction example and its
explanation into a general rule, by using the object ontology as a generalization lan-
guage). The learning method is the one in Table 9.2 and will be illustrated in the
following.
The first step of the learning method is Mixed-Initiative Understanding (explanation
generation). The question and its answer from the task reduction step represent the
expert’s reason (or explanation) for performing that reduction. Therefore, understanding
the example by Disciple-VE means understanding the meaning of the question/answer
pair in terms of the concepts and features from the agent’s ontology. This process is
difficult for a learning agent that does not have much knowledge because the experts
express themselves informally, using natural language and common sense, and often omit
essential details that they consider obvious. The question/answer pair from the example in
Figure 12.53 is:

13:54:21,
.013
12.5. Disciple-VPT 405

Question: Can we turn gas propane f1 off from fire1?

Answer: No because shut off valve1 is damaged.
We can expect a person to assume, without being told, that the reason we are
considering turning the gas propane f1 off from fire1 is because it is fueling fire1. We
can also expect the person to assume that we are considering shutting off valve1 because it
is part of fill pipe1 with the propane. However, an automated agent is not able to make
these assumptions and has to be helped to get an as complete understanding of the
example as possible. For instance, a more complete explanation of the example from
Figure 12.53 consists of the facts shown in Figure 12.55.
The process of identifying such explanation pieces is the one described in Section 9.5.
The quality of the learned rule depends directly on the completeness of the found explan-
ation. However, there is no requirement that the found explanation be complete, and in
fact this rarely occurs. The agent will continue to improve the rule while using it in
reasoning (as described in Chapter 10), when it will be easier to discover the missing
explanation pieces.
The next step of the rule-learning method is Example Reformulation (see Table 9.2),
which consists in transforming the example from Figure 12.53 and its explanation from
Figure 12.55 into an equivalent IF-THEN rule, by replacing each instance or constant with
a variable and restricting the variables to those values, as illustrated in Figure 12.56. This
expression is an instance of the general rule to be learned. The goal of the rule-learning
process is to determine which values of the variables from the condition lead to correct
task reduction steps. That is, Disciple-VE has to learn the concept that represents the set of
instances of the rule’s variables for which the corresponding instantiation of the rule is
correct (i.e., the applicability condition of the rule).
First Disciple-VE generalizes the applicability condition from Figure 12.56 to an initial
plausible version space condition, as described in Section 9.7 and illustrated in the
following.
The plausible upper bound condition is obtained by replacing each variable value with
its maximal generalization, based on the object ontology.
Let us consider the value fire1 of the variable ?O2. The most general concept from the
object ontology that is more general than fire1 is object. However, the possible values for
?O2 are restricted by the features of fire1 identified as relevant as part of the explanation of
the example from Figure 12.55. As indicated in Figure 12.56, ?O2 should have the features
is fueled by and is situated in. This means that the values of ?O2 have to be part of the
domains of these features. Thus:

most general generalizationðfire1Þ ¼ object \ Domainðis fueled byÞ

\ Domainðis situated inÞ
¼ object \ fire \ object ¼ fire

fire1 is fueled by gas propane f1

fire1 is situated in fill pipe1 has as part shutoff valve1
shutoff valve1 has as operating status damaged
the value is specifically damaged

Figure 12.55. Explanation of the example from Figure 12.53.

13:54:21,
.013
406 Chapter 12. Disciple Agents

IF the task is
Determine whether we can prevent the ?O1
by extinguishing ?O2

Q: Can we turn ?O3 off from ?O1?

A: No because ?O4 is ?S1
Condition
?O1 is BLEVE1
?O2 is fire1
is fueled by ?O3
is situated in ?O5
?O3 is gas propane f1
?O4 is shutoff valve1
has as operating status ?S1
?O5 is fill pipe1
has as part ?O4
?S1 is damaged

THEN
Determine whether we can prevent the ?O1
by extinguishing ?O2 when we cannot turn
?O3 off from ?O2

Figure 12.56. Specific inference rule covering only the initial example.

Applying a similar procedure to each variable value from the condition in Figure 12.56 one
obtains the plausible upper bound condition shown in Figure 12.57.
The plausible lower bound condition is obtained by replacing each variable value with
its minimal generalization that is not an instance, based on the object ontology. The
procedure is similar to the one for obtaining the plausible upper bound condition.
Therefore:

least general generalizationðfire1Þ ¼

fire \ Domainðis fueled byÞ \ Domainðis situated inÞ ¼

fire \ fire \ object ¼ fire

The reason the lower bound cannot contain any instance is that the learned rule will be
used by Disciple-VE in other scenarios where the instances from the current scenario
(such as fire1) do not exist, and Disciple-VE would not know how to generalize them. On
the other hand, we also do not claim that the concept to be learned is more general than
the lower bound.
Notice that the features from the explanation of the example significantly limit the size
of the initial plausible version space condition and thus speed up the rule-learning
process. This is a type of explanation-based learning (DeJong and Mooney, 1986; Mitchell

13:54:21,
.013
12.5. Disciple-VPT 407

generalization

generalization
Maximal

Minimal
?O1 is BLEVE1
?O2 is fire1
is fueled by ?O3
is situated in ?O5
?O3 is gas propane f1
?O4 is shutoff valve1
has as operating status ?S1
?O5 is fill pipe1
has as part ?O4
?S1 is damaged

Figure 12.57. Generation of the initial plausible version space condition.

et al., 1986) except that the knowledge base of Disciple-VE is incomplete and therefore
rule learning requires additional examples and interaction with the expert.
After the rule was generated, Disciple-VE analyzes it to determine whether it was
learned from an incomplete explanation (Boicu et al., 2005). To illustrate, let us consider
again the process of understanding the meaning of the question/answer pair from
Figure 12.53, in terms of the concepts and features from the agent’s ontology. In the
preceding, we have assumed that this process has led to the uncovering of implicit
explanation pieces. However, this does not always happen. Therefore, let us now assume
that, instead of the more complete explanation pieces considered in the preceding, the
identified explanation pieces of the example are only those from Figure 12.58. In this case,
the learned rule is the one from Figure 12.59.

13:54:21,
.013
408 Chapter 12. Disciple Agents

shutoff valve1 has as operating status damaged

the value is specifically damaged

Figure 12.58. Incomplete explanation of the example from Figure 12.53.

Figure 12.59. Rule learned from the example in Figure 12.53 and the explanation in
Figure 12.58.

The variables from the IF task of a rule are called input variables because they are
instantiated when the rule is invoked in problem solving. The other variables of the rule are
called output variables.
During the problem-solving process, the output variables are instantiated by the agent
with specific values that satisfy the rule’s applicability condition. In a well-formed rule, the
output variables need to be linked through explanation pieces to some of the input
variables of the rule. Therefore, one rule analysis method consists of determining whether
there is any output variable that is not constrained by the input variables. For instance, in
the case of the rule from Figure 12.59, Disciple-VE determined that the variables ?O3, ?O4,
and ?S1 are not constrained and asks the expert to guide it to identify additional explan-
ation pieces related to their corresponding values (i.e., gas propane f1, shutoff valve1,
and damaged).
If the rule passes the structural analysis test, Disciple-VE determines the number of its
instances in the current knowledge base and considers that the rule is incompletely
learned if this number is greater than a predefined threshold. In such a case, the agent
will attempt to identify which variables are the least constrained and will attempt to
constrain them further by interacting with the expert to find additional explanation pieces.

13:54:21,
.013
 12.5. Disciple-VPT 409

Following such a process, Disciple-VE succeeds in learning a reasonable good rule from
only one example and its explanation, a rule that may be used by Disciple-VE in the planning
process. The plausible upper bound condition of the rule allows it to apply to situations that
are analogous with the one from which the rule was learned. If the expert judges this
application to be correct, then this represents a new positive example of the rule, and the
plausible lower bound condition is generalized to cover it, as discussed in Section 10.1.3.1.
Otherwise, the agent will interact with the expert to find an explanation of why the
application is incorrect and will specialize the rule’s conditions appropriately, as discussed
in Section 10.1.4. Rule refinement could lead to a complex task reduction rule, with
Except-When conditions that should not be satisfied in order for the rule to be applicable.

12.5.9 Teaching Disciple-VE to Perform Planning Tasks

12.5.9.1 Why Learning Planning Rules Is Difficult
Figure 12.60 compares the learning of inference rules with the learning of planning rules. The
left-hand side of Figure 12.60 shows an inference step and a planning step, while the right-
hand side shows the rules that would be learned from these steps. In the case of an inference
step, Disciple-VE learns a rule by generalizing the expressions from the examples to patterns
and by generating a plausible version space for the applicability condition of the rule.
The learning of the planning rule is much more complex, not just because it involves
the learning of several applicability conditions, but mainly because these conditions have
to be learned in different states of the world. Indeed, Condition 11g and Condition 2g are
learned in the state S1, but Condition 3g has to be learned in the state S3. However, the
state S3 is only known after the entire reduction tree for Planning Task 2 has been
developed. What this means is that Disciple-VE would start learning the rule in the state
S1, will then continue with the planning and inference corresponding to the subtree of
Planning Task 2, and only after that can resume and finalize the learning of the rule. But
this is impractical for two main reasons. First, it leads to the starting of learning many
complex planning rules, with the associated management of temporary representations for
these rule fragments. Second, these incompletely learned rules cannot be used in problem
solving. Thus, in the case of a planning tree that contains recursive applications of a task
reduction step, Disciple-VE would start learning a new rule for each application, although
these rules will end up being identical.

12.5.9.2 Learning a Set of Correlated Planning Rules

The main source of difficulty for learning a planning rule from the planning example in
Figure 12.60 is the need first to develop the entire planning tree for Planning Task 2. We
have discussed a similar difficulty in Section 12.5.6, in the context of modeling the expert’s
planning process. In that case, the expert could not specify the reduction of Planning Task
1 into Planning Task 2 and Planning Task 3 before completing the entire planning for
Planning Task 2. The solution found to that problem was to introduce the notion of an
abstract task. This notion will also help overcome the difficulty of learning planning rules,
as will be explained in the following.
Rather than learning a single complex planning rule from a task reduction example,
Disciple-VE will learn a set of simpler planning rules that share common variables, as
illustrated in the right part of Figure 12.61.

13:54:21,
.013
 Inference Step Example Inference Rule
Si Inference Task 1 Inference Task 1g

Question/Answer Question g/Answer g Condition g

Inference Task 2 Inference Task 3 Inference Task 2g Inference Task 3g

Planning Step Example Planning Rule

S1 Planning Task 1 Goal 1 S4 Planning Task 1g Goal 1g

.013

Question 11/Answer 11 Question 11g/Answer 11g Condition 11g

13:54:21,

Abstract Task 2 Abstract Task 3 Abstract Task 2 Abstract Task 3

Precondition 2g Condition 2g Precondition 3g Condition 3g

Precondition 2 Precondition 3
Planning S3 Planning Planning Planning
S1 Goal 2 Goal 3 S4 Goal 2g Goal 3g
Task 2 Task 3 Task 2g Task 3g

Figure 12.60. Learning inference rules versus learning planning rules.

 Planning Task Reduction Rule
Planning Task 1g Goal
Planning Step Example Synthesis
Question 11g/Answer 11g Condition 11g Rule
S1 Planning Task 1 Goal 1 S4
Abstract Task 2 Abstract Task 3 Goal 1g
Question 11/Answer 11
Shared Variables
Abstract Task 2 Abstract Task 3
Concretion Rule Goal Concretion Rule Goal
Precondition 2 Precondition 3 Mapping Abstract Task 3 Mapping
Abstract Task 2
Planning Planning Rule Rule
S1 Goal 2 S3 Goal 3 S4
Task 2 Task 3 Precondition 2g Goal 2g Precondition 3g Goal 3g
Condition 2g Condition 3g

Planning Task 2g Planning Task 3g

Planning Task Reduction Rule

Planning Step Example Planning Task 2g Goal
.013

Synthesis
S1 Planning Task 2 Goal 2 S3 Question 21g/Answer 21g Condition 21g Rule
Goal 2g
13:54:21,

Question 21/Answer 21 Abstract Task 4 Abstract Task 5

Abstract Task 4 Abstract Task 5 Shared Variables

Precondition 4 Precondition 5 Action Concretion Rule Action Concretion Rule

S1 Delete Add S2 Delete Add Abstract Task 4 Abstract Task 5

Action 1 Effects 1 Effects 1 Action 2 Effects 2 Effects 2 S3
Resources 1 Duration 1 Resources 2 Duration 2 Precondition 4g Condition 4g Precondition 5g Condition 5g

Delete Add Delete Add

Action 1g Effects 1g Effects 1g Action 2g Effects 2g Effects 2g
Resources 1g Duration 1g Resources 2g Duration 2g

Figure 12.61. Learning correlated planning rules.

412 Chapter 12. Disciple Agents

These rules will not be learned all at once, but in the sequence indicated in Figure 12.62.
This sequence corresponds to the sequence of modeling operations for the subtree of
Planning Task 1, as discussed in Section 12.5.6.
First the expert asks himself or herself a question related to how to reduce Planning
Task 1. The answer guides the expert to reduce this task to two abstract tasks. From this
reduction, the agent learns a planning task reduction rule (see Figure 12.62a), by using the
method described in Section 12.5.9.4. Next the expert reduces Abstract Task 2 to Planning
Task 2, and the agent learns a task concretion rule (see Figure 12.62b) by using the method
described in Section 12.5.9.5. After that, the expert continues specifying the reduction tree
corresponding to Planning Task 2, and the agent learns rules from the specified planning
step, as indicated previously. During the development of this planning tree, the agent
may apply the preceding rules, if their conditions are satisfied, and may refine them based
on the expert’s feedback. After the entire subtree corresponding to Planning Task 2 is

Reduction Example Planning Task Reduction Rule

Planning Task 1 Planning Task 1g

S1
(a) Question 11/Answer 11 Question 11g/Answer 11g Condition 11g

Abstract Task 2 Abstract Task 3 Abstract Task 2 Abstract Task 3

Concretion Example Concretion Rule

Abstract Task 2 Abstract Task 2

(b) Precondition 2g
Precondition 2
S1
Planning Task 2 Condition 2g

Planning Task 2g Shared

Goal Example
Variables
Planning Task 2 Goal 2 S3 Goal Mapping Rule
(c)
Goal 2g
S1 Planning Task 1 Goal 1 S4

Question/Answer

Abstract Task 2 Abstract Task 3

Preconditions 2 Preconditions 3
Planning Planning
S1 Goal 2 S3 S3 Goal 3 S4
Task 2 Task 3

Question/Answer Question/Answer

Abstract Task 4 Abstract Task 5 Abstract Task 6

Preconditions 4 Preconditions 5 Preconditions 6

Delete Add
S1 Action 1 Delete Add S2 Action 2
Delete Add
S3 Action 3 S4

Concretion Rule
Effects Effects Effects Effects Effects Effects
Resources Duration Resources Duration Resources Duration

Concretion Example
Abstract Task 3
Abstract Task 3
Precondition 3g
(d)
Precondition 3 Condition 3g
S3
Planning Task 3
Planning Task 3g
Goal Example
Planning Task 3 Goal 3 S4
Goal Mapping Rule
(e)
S1 Planning Task 1 Goal 1 S4

Question/Answer

Goal 3g
Abstract Task 2 Abstract Task 3

Preconditions 2 Preconditions 3
Planning Planning
S1 Goal 2 S3 S3 Goal 3 S4
Task 2 Task 3

Question/Answer Question/Answer

Abstract Task 4 Abstract Task 5 Abstract Task 6

Preconditions 4 Preconditions 5 Preconditions 6

Delete Add
S1 Action 1 Delete Add S2 Action 2
Delete Add
S3 Action 3 S4
Effects Effects Effects Effects Effects Effects
Resources Duration Resources Duration Resources Duration

Goal Example Goal Synthesis Rule

(f)
Planning Task 1 Goal 1 S4 Goal 1g

Figure 12.62. The sequence of learning correlated planning rules.

13:54:21,
.013
12.5. Disciple-VPT 413

developed, the agent can learn the goal mapping rule corresponding to Goal 2, as
described in Section 12.5.9.4. The learning of the concretion rule for Abstract Task
3 and of the goal mapping rule for Goal 3 is done as described previously. After that,
Disciple-VE learns the goal synthesis rule corresponding to Goal 1, as described in
Section 12.5.9.4.
The preceding illustration corresponds to a reduction of a planning task into planning
subtasks. However, a planning task can also be reduced to elementary actions, as
illustrated at the bottom part of Figure 12.61. In this case, Disciple-VE will learn more
complex action concretion rules instead of task concretion rules, as discussed in
Section 12.5.9.6. In the following sections, we will present the aforementioned learning
methods.

12.5.9.3 The Learning Problem and Method for a Set of Correlated

Planning Rules
The problem of learning a set of correlated planning rules is presented in Table 12.5, and
the corresponding learning method is presented in Table 12.6. We will illustrate them by
using the top task reduction from Figure 12.49 (shown also in Figure 12.50).

12.5.9.4 Learning Correlated Planning Task Reduction Rules

The method for learning a correlated planning reduction rule is presented in Table 12.7.
This method is similar to the method of learning an inference rule presented in Table 9.2
and Section 12.5.8, except for the insertion of Step 3 in Table 12.7, which adds to the set V
the variables from the learned rule and their values in the example.
To illustrate the method in Table 12.7, let us consider the top reduction in Figure 12.49
(shown also in Figure 12.50). In that reduction, the top-level task is reduced to five abstract
tasks. The reduction is justified by the following question/answer pair:
Question: What kind of incident is Gainsville incident?
Answer: Gainsville incident is a major fire emergency because it involves a large propane tank
on fire.

Table 12.5 The Learning Problem for Correlated Planning Rules

GIVEN

A sequence of reduction and synthesis steps called SE that indicate how a specific planning task
is reduced to its immediate specific subtasks and/or actions and how its goal is synthesized from
their goals/effects.

A knowledge base that includes an ontology and a set of rules.

A subject matter expert who understands why the given planning steps are correct and may
answer the agent’s questions.

DETERMINE

A set of reduction, concretion, goal, and/or action rules called SR that share a common space of
variables, each rule being a generalization of an example step from SE.

An extended ontology (if needed for example understanding)

13:54:21,
.013
414 Chapter 12. Disciple Agents

Table 12.6 The Learning Method for Correlated Planning Rules

Let SE be a sequence of reduction and synthesis steps that indicate how a specific planning task T is
reduced to its immediate specific subtasks and/or actions, and how its goal is synthesized from
their goals/effects.
1. Initialize the set V of shared variables and their values in SE: V Ф
2. Learn a planning task reduction rule from the reduction of T to the abstract tasks ATi and
update the set V (by using the method described in Table 12.7 and Section 12.5.9.4).
3. For each abstract task ATi do
If ATi is reduced to a concrete Task Ti
Then 3.1. Learn a planning task concretion rule and update set V (using the
method from Section 12.5.9.5).
3.2. Develop the entire subtree of Ti (this may lead to the learning of new rules by
using the methods from Tables 9.2 and 12.6).
3.3. Learn the goal mapping rule for Ti (using the method from Section 12.5.9.4).
Else if ATi is reduced to an elementary action Ai
Then 3.1. Learn an action concretion rule and update the set V (using the method from
Section 12.5.9.6).
4. Learn the goal synthesis rule for T (by using the method described in Section 12.5.9.4).

As part of example understanding, Disciple-VE will interact with the expert to find the
following explanation pieces, which represent an approximation of the meaning of the
question/answer pair in the current world state:

Gainsville incident instance of major fire emergency

Gainsville incident is caused by fire1
fire1 is situated in propane tank1 instance of propane tank
propane tank1 has as size large

Continuing with the steps from Table 12.7, Disciple-VE will learn the rule from the
left-hand side of the pane in Figure 12.63.
The final list of shared variables is shown in the right-hand side of this pane. The right-
hand side of the pane shows also the goal produced by the goal synthesis rule. This rule
generalizes the expression representing the goal associated with the IF task by replacing its
instances and constants with the corresponding variables from the list of shared variables.
Similarly, the goal mapping rule generalizes the goals of the THEN tasks.

12.5.9.5 Learning Correlated Planning Task Concretion Rules

The method of learning a correlated planning task concretion rule is similar to the method
of learning a correlated planning reduction rule presented in Table 5.7 and Section
12.5.9.4. To illustrate it, let us consider again the reduction from Figure 12.50. The Sub-
Task (3) pane includes a concretion step, which is shown again in Figure 12.64. The rule
learned from this concretion example is shown in Figure 12.65.
As part of Mixed-Initiative Understanding (see Table 12.7), what needs to be
understood are the preconditions of the concretion step. The approximation of their
meaning is:

13:54:21,
.013
12.5. Disciple-VPT 415

Table 12.7 The Learning Method for a Correlated Planning Task Reduction Rule

Let E be a reduction of a specific planning task T to one or several abstract tasks ATi, reduction
taking place in state Sk, and let V be the set of shared variables and their values.
(1) Mixed-Initiative Understanding (Explanation Generation)
Determine the meaning of the question/answer pair from the example E, in the context of the
agent’s ontology from the state Sk, through mixed-initiative interaction with the subject matter
expert. This represents a formal explanation EX of why the example E is correct. During this
process, new objects and features may be elicited from the expert and added to the ontology. This
is done in order to better represent the meaning of the question/answer pair in terms of the
objects and features from the ontology.
(2) Example Reformulation
Generate a variable for each instance and each constant (i.e., number, string, or symbolic
probability) that appears in the example E and its explanation EX. Then use these variables to
create an instance I of the concept C representing the applicability condition of the rule R to be
learned. C is the concept to be learned as part of rule learning and refinement. Finally, reformulate
the example as a very specific IF-THEN rule with I as its applicability condition. The elements of the
rule are obtained by replacing each instance or constant from the example E with the
corresponding variable.
(3) Updating of Shared Variables and Values
Add to the set V the new variables and their values from the condition C.
(4) Analogy-based Generalizations
Generate the plausible upper bound condition of the rule R as the maximal generalization of I in
the context of the agent’s ontology.
Generate the plausible lower bound condition of the rule R as the minimal generalization of I that
does not contain any specific instance.
(5) Rule Analysis
If there is any variable from the THEN part of a rule that is not linked to some variable from the IF
part of the rule, or if the rule has too many instances in the knowledge base, then interact with the
expert to extend the explanation of the example and update the rule if new explanation pieces are
found. Otherwise, end the rule-learning process.

Gainsville incident has as ICS Gainsville ICS has as ICS unified command Gainsville command

The Rule Analysis step takes the value of the set V into account to determine the unlinked
output variables. In particular, an output variable from the concrete task does not need to
be linked to input variables if it is part of the input value of V.

12.5.9.6 Learning a Correlated Action Concretion Rule

If the reduction of a planning task includes actions, then Disciple-VE learns also correlated
action concretion rules, as illustrated at the bottom part of Figure 12.61. The learning
method is similar to that for learning a correlated task concretion rule, except that the
resulting rule has additional action components such as Delete Effects, Add Effects,
Resources, and Duration.
Let us consider the abstract task from Figure 12.48 and its concretion. The action
concretion rule learned from this example is shown in Figure 12.66.

13:54:21,
.013
 416 Chapter 12. Disciple Agents

Figure 12.63. Planning reduction rule learned from the reduction in Figure 12.49.

Figure 12.64. A task concretion example.

12.5.10 The Virtual Experts Library

The previous sections have presented the capability of a Disciple-VE agent shell to acquire
planning expertise rapidly from a subject matter expert. This capability makes possible the
development of a wide range of planning agents that can collaborate in performing
complex tasks. These agents are maintained in the VE Library, where their knowledge
bases (KBs) are hierarchically organized, as illustrated in Figure 12.67. In this illustration,
there are four expertise domains, D1, D2, D3, and D4, and nine virtual experts, each
associated with a KB from the bottom of the hierarchy. Each virtual expert agent VE is a
customization of the Disciple-VE shell. The three leftmost virtual experts are experts in the

13:54:21,
.013
12.5. Disciple-VPT 417

Figure 12.65. The planning concretion rule learned from the example in Figure 12.64.

Figure 12.66. The correlated action concretion rule learned from the example in Figure 12.48.

13:54:21,
.013
418 Chapter 12. Disciple Agents

KB-0

KB-12 KB-34

KB-D1 KB-D3
KB-D2

KB-D4
VE KB-B1 VE KB-A1 VE KB-B3 VE KB-I3

VE KB-I1
VE KB-B2 VE KB-I2 VE KB-A2 VE KB-B4

Figure 12.67. The organization of the VE Library.

domain D1 with different levels of expertise: basic, intermediary, and advanced. In

addition to receiving their specific KBs (e.g., KB-B1), they all inherit general knowledge
about the domain D1, knowledge represented in KB-D1. They also inherit knowledge from
the higher-level KBs, KB-12 and KB-0. These higher-level KBs contain general knowledge,
useful to many agents, such as ontologies for units of measure, time, and space.
Traditional knowledge engineering practice builds each KB from scratch, with no
knowledge reuse, despite the fact that this is a very time-consuming, difficult, and error-
prone process (Buchanan and Wilkins, 1993; Durkin, 1994; Awad, 1996; Jackson, 1999;
Awad and Ghaziri, 2004). On the contrary, the hierarchy of KBs from the VE Library offers
a practical solution to the problem of knowledge reuse, speeding up the process of
building a new virtual expert. Consider, for example, developing a new virtual expert for
the D3 domain. This expert will already start with a KB composed of KB-D3, KB-34, and
KB-0. Thus, the VE Library can also be regarded as a knowledge repository for the new
virtual experts to be developed.
The updating of each KB from the hierarchical repository (e.g., KB-12) is the responsi-
bility of a team consisting of a knowledge engineer and one or several subject matter
experts. They use the specialized browsers and editors of the Disciple-VE shell. The left-
hand side of Figure 12.68 shows the interface of the Object Browser, which displays the
objects in a tree structure. The objects that are inherited from an upper-level KB (such as
information measure or length measure) are displayed with a gray background. The
right-hand side of Figure 12.68 shows the interface of the Object Viewer, which displays
additional information about the object selected in the Object Browser (e.g., fire engine
company E501), such as its direct superconcepts and its features. The top part of
Figure 12.46 shows the interface of the Hierarchical Browser, which displays the hier-
archical relationships between objects or features in a graph structure. The bottom
part of Figure 12.46 shows the interface of the Association Browser, which displays
an object and its relationships with other objects. Additional tools include the Object
Editor, the Feature Editor, and the Rule Editor. All these tools have been introduced in
Chapters 5 and 6.

13:54:21,
.013
12.5. Disciple-VPT 419

Figure 12.68. The interface of the Object Browser and Object Viewer.

To allow the KBs from the hierarchy to be updated and extended separately, the
Disciple-VPT system maintains multiple versions for each KB. Let us assume that each
KB from Figure 12.67 has the version 1.0. Let us further assume that the management team
for KB-0 decides to make some changes to this KB that contains units of measure. For
instance, the team decides to include the metric units, to rename “gallon” as “US gallon,”
and to add “UK gallon.” As a result, the team creates version 2.0 of KB-0. However, the
other knowledge bases from the library (e.g., KB-12) still refer to version 1.0 of KB-0. The
management team for KB-12 is informed that a higher version of KB-0 is available. At this
point, the team can decide whether it wants to create a new version of KB-12 that inherits
knowledge from version 2.0 of KB-0. The KB update process uses the KB updating tool of
Disciple-VE. This tool creates version 2.0 of KB-12 by importing the knowledge from
version 1.0 of KB-12, in the context of version 2.0 of KB-0. Even though the version 2.0
of KB-12 has been created, Disciple-VPT still maintains KB-0 version 1.0 and KB-12
version 1.0, because these versions are used by KB-D1 version 1.0 and by other KBs from
the repository. The management team for KB-D1 may now decide whether it wants to
upgrade KB-D1 to the new versions of its upper-level KBs, and so on. Because of the
version system, each KB from the library maintains, in addition to its version, the versions
of the other KBs from which it inherits knowledge.
Another important knowledge management functionality offered by Disciple-VPT is
that of splitting a KB into two parts, a more general one and a more specific one. This
allows a KB developer first to build a large KB and then to split it and create a hierarchy
of KBs.

13:54:21,
.013
 420 Chapter 12. Disciple Agents

When a virtual expert is extracted from the VE Library and introduced into a VE Team
(see Figure 12.44), all the KBs from which it inherits knowledge are merged into a shared
KB in order to increase the performance of the agent. Let us consider the Intermediate
agent from the domain D3 (see Figure 12.67). In this case, KB-D3, KB-12, KB-34, and KB-0
are all merged into the Shared KB of this agent. As a consequence, the structure of the KBs
of this agent during planning is the one from Figure 12.69. Notice that, in addition to the
Shared KB, there are three other types of KBs, Domain KB, Scenario KB, and State KB, all
hierarchically organized. Domain KB is the KB of this Intermediate agent from the domain
D3, knowledge independent of any particular scenario. Each scenario is represented into a
different KB called Scenario KB. For example, there would be a Scenario KB for the
propane tank fire scenario described in Section 12.5.3, and a different Scenario KB for a
scenario involving red-fuming nitric acid spilling from a truck parked near a residential
area (Tecuci et al., 2008c). Moreover, under each scenario KB there is a hierarchy of State
KBs. KB-S1 represents the state obtained from SKB-M after the execution of an action
which had delete and/or add effects. As additional actions are simulated during planning,
their delete and add effects change the state of the world. KB-S11, KB-S12, and KB-S13 are
the states corresponding to three alternative actions. The entire description of the state
corresponding to KB-S11 is obtained by considering the delete and add effects in the states
KB-S11 and KB-S1, and the facts in the scenario SKB-M.

12.5.11 Multidomain Collaborative Planning

Each virtual agent from the VE Library is expert in a certain expertise domain. However,
these expertise domains are not disjointed, but overlapping, as illustrated in Figure 12.70.

KB-0

KB-12 KB-34 Shared KB

KB-D3

VE-D3 KB-I3 Domain KB

Scenario KBs

SKB-1 ... SKB-M ... SKB-N

State KBs
KB-S1

KB-S11 KB-S12 KB-S13

Figure 12.69. The organization of an agent’s knowledge base during planning.

13:54:21,
.013
 12.5. Disciple-VPT 421

D1
.Ti
.Tm

D3 D2

Figure 12.70. Sample coverage of expertise domains.

In this illustration, the planning task Tm belongs only to D2 and can be performed only by a
virtual expert from that domain. Ti is a task common to D1 and D2 and can, in principle, be
performed either by a virtual expert in D1 or by a virtual expert in D2. In general, a virtual
expert will cover only a part of a given expertise domain, depending on its level of
expertise. For instance, the virtual expert library illustrated in Figure 12.67 includes three
virtual experts from the domain D1, a basic one, an intermediate one, and an advanced
one, each covering an increasingly larger portion of the domain. Therefore, whether a
specific virtual expert from the domain D2 can generate a plan for Tm and the quality of the
generated plan depend on the expert’s level of expertise.
A virtual expert has partial knowledge about its ability to generate plans for a given task,
knowledge that is improved through learning. For instance, the virtual expert knows that it
may be able to generate plans for a given task instantiation because that task belongs to its
expertise domain, or because it was able to solve other instantiations of that task in the
past. Similarly, it knows when a task does not belong to its area of expertise. The virtual
experts, however, do not have predefined knowledge about the problem-solving capabil-
ities of the other experts from a VE Team or the VE Library. This is a very important feature
of Disciple-VPT that facilitates the addition of new agents to the library, or the improve-
ment of the existing agents, because this will not require taking into account the know-
ledge of the other agents.
The task reduction paradigm facilitates the development of plans by cooperating virtual
experts, where plans corresponding to different subtasks of a complex task may be
generated by different agents. This multi-agent planning process is driven by an auction
mechanism that may apply several strategies. For instance, the agents can compete for
solving the current task based on their prior knowledge about their ability to solve that
task. Alternatively, the agents may actually attempt to solve the task before they bid on it.

12.5.12 Basic Virtual Planning Experts

We have developed two basic virtual experts: a fire department operations expert and an
emergency management expert. The development of these virtual experts was guided by
the toxic substance leaking scenario described in Tecuci et al. (2008c) and by the propane
tank on fire scenario described in Section 12.5.3.
First we have worked with a subject matter expert to model the plan generation process
for these two scenarios by using the task reduction paradigm, as illustrated in Figure 12.50.
Besides the development of the reasoning trees, another result of this modeling process
was the development of the modeling methodology presented in Section 12.5.6. Yet
another result of this modeling process was the identification of the object concepts that
need to be present in the ontology of Disciple-VE so that it can perform this type of
reasoning. Based on this specification of the object ontology, and by using the ontology

13:54:21,
.013
 422 Chapter 12. Disciple Agents

development modules of Disciple-VE, we have developed an object ontology consisting of

410 concepts, 172 feature definitions, 319 generic instances, and 944 facts. Fragments of
this ontology were presented in Figures 12.45 and 12.46.
Although we have worked with both aforementioned scenarios to develop the modeling
trees and the ontology, the teaching of the fire expert and of the emergency management
expert was based only on the scenario of the propane tank on fire. As a result of the
teaching process, the virtual fire expert learned eighty-one planning rules, and the virtual
emergency management expert learned forty-seven planning rules. The two developed
virtual experts share the object ontology.

12.5.13 Evaluation of Disciple-VPT

To perform a preliminary evaluation of the Disciple-VPT system, we have developed a
scenario pattern based on the scenario from Section 12.5.3. Then we have asked a fire
department operations expert and an emergency management expert to define a specific
scenario based on this pattern, by providing the missing elements. For example, the two
experts decided on the day and time of the incident, the position of the fire with respect to
the propane tank, the estimated amount of propane in the tank, the available resources
(fire engine companies, truck engine companies, hazardous materials [hazmat] units,
police personnel, county personnel, etc.), and the time that they are arriving at the scene
of the incident. After that, both the team of human experts and Disciple-VPT independ-
ently generated plans to respond to this situation.
The plan generated by Disciple-VPT consisted of eighty-nine partially ordered elemen-
tary actions. A fragment of this plan is shown in Figure 12.71. This plan was evaluated by
the aforementioned experts and by the expert with whom we have developed the agents.
Then each expert filled out a questionnaire. The questionnaire included statements about
various characteristics of the generated plan and about the Disciple-VPT system. The
experts were asked to indicate whether they strongly agree (SA), agree (A), are neutral (N),
disagree (D), or strongly disagree (SD) with these statements.
The top part of Table 12.8 presents some of the results of the evaluation of the plan
generated by Disciple-VPT, which the human experts considered good and easy to under-
stand. One of the experts disagreed with the way Disciple-VPT ordered some of the actions.
However, this reflected a disagreement between the evaluating experts and not a planning
error, the generated order being that taught by the expert who instructed Disciple-VPT.
The evaluation of the hierarchical plan generation process (see the middle of
Table 12.8) shows that it significantly improves the understandability of this process.
Finally the experts agreed or strongly agreed that Disciple-VPT has many potential
applications in emergency response planning, from developing exercises for training, to
actual training, and to its use as planning assistant.
A limitation of this evaluation is that it is based on the opinion of only three experts.
More experts are needed in order for the results to have statistical significance.

12.5.14 Final Remarks

This section presented a major extension of the Disciple theory and methodology for the
development of knowledge-based agents by subject matter experts, with limited assistance
from knowledge engineers.

13:54:21,
.013
 Start Dura-
Id Action Result Resources
Time tion
Suburbane Propane Company facility manager, a person, Add: Suburbane Propane Company facility manager arrived at the 1.0 min
1 Suburbane Propane Company facility manager 0.0 s
arrives at the scene of the Gainsville incident scene and is available to take required actions 0.0 s
ALS unit M504, an ALS unit, arrives at the scene of the Add: ALS unit M504 arrived at the scene and is available to take 5.0 min
2 ALS unit M504, paramedic 504a, and paramedic 504b 0.0 s
Gainsville incident required actions 0.0 s
fire engine driver E504, fire fighter E504b, fire engine
fire engine company E504, a fire engine company, arrives Add: fire engine company E504 arrived at the scene and is available company E504, fire engine E504, deluge nozzle E504, 5.0 min
3 0.0 s
at the scene of the Gainsville incident to take required actions water hose E504, fire officer E504, and fire fighter 0.0 s
E504a
… … … … …
Delete: fire officer E504 is no longer available
fire officer E504 assumes the command of the incident Add: Gainsville ICS, the incident command system for the
5.0 min
48 command system for the Gainsville incident, as fire Gainsville incident, is created and fire officer E504 assumes ICS fire officer E504 15.0 s
0.0 s
department representative in the ICS unified command command as fire department representative in the Gainsville
command
Gainsville command evaluates the situation created by the Add: overpressure situation with danger of BLEVE in propane 5.0 min
49 Gainsville command 30.0 s
Gainsville incident tank1 is caused by fire1 15.0 s
.013

Add: The plan is to apply cooling water, to evacuate people from

Gainsville command determines the abstract incident action 5.0 min
50 1.0 mi around propane tank1, to perform traffic control, perimeter Gainsville command 30.0 s
plan for overpressure situation 45.0 s
control, and to establish rapid intervention task forces
… … … … …
13:54:21,

fire engine company E504, fire engine driver E504, fire

fire engine company E504 sets up water hose E504 to apply Delete: water hose E504 is available 8.0 min 3.0 min
54 fighter E504b, fire fighter E504a, water hose E504, and
water to propane tank1 Add: water hose E504 is assigned 30.0 s 0.0 s
fire engine E504
Delete: deluge nozzle E525 that belongs to fire engine E525 is no
fire engine company E525 drops off deluge nozzle E525 for fire fighter E525a, fire officer E525, deluge nozzle 8.0 min 2.0 min
55 longer available
fire engine company E504 E525, fire engine driver E525, and fire fighter E525b 30.0 s 0.0 s
Add: deluge nozzle E525 is assigned
… … … … …

fire engine company E525 establishes continuous water fire fighter E525a, fire fighter E525b, fire officer E525, 14.0 min 7.0 min
63 Add: fire hydrant1 is assigned to fire engine company E504
supply from fire hydrant1 for fire engine company E504 fire engine driver E525, and fire engine company E525 30.0 s 0.0 s

… … … … …

Figure 12.71. Fragment of a plan generated by Disciple-VPT.

424 Chapter 12. Disciple Agents

Table 12.8 Evaluation Results

Generated Plan SA A N D SD

The generated plan is easy to understand. 3

The generated plan consists of a good sequence of actions. 2 1

The objectives of the actions are clear from the generated plan. 3

The generated plan has a good management of the resources. 1 1 1

The level of detail in the plan is appropriate. 3

Hierarchical Plan Generation Process SA A N D SD

The hierarchical task-reduction structure makes the logic of the plan

2 1
clear.

The hierarchical task-reduction structure makes the goals of the plan

1 2
clear.

The questions, the answers, and the preconditions help understand

2 1
the logic of the plan generation process.

The preconditions, effects, duration, and resources make the

1 2
specification of the plan’s actions clear.

The hierarchical task-reduction structure may be used to teach new

3
persons how to plan.

Usability of Disciple-VPT SA A N D SD

Disciple-VPT could be used in developing and carrying out exercises

2 1
for emergency response planning.

Disciple-VPT could be used for training the personnel for emergency

2 1
response planning.

Disciple-VPT could be used as an assistant to typical users, guiding

3
them how to respond to an emergency situation.

First, we have extended the Disciple approach to allow the development of complex
HTN planning agents that can be taught their planning knowledge, rather than having it
defined by a knowledge engineer. This is a new and very powerful capability that is not
present in other action planning systems (Tate, 1977; Allen et al., 1990; Nau et al., 2003;

13:54:21,
.013
12.5. Disciple-VPT 425

Ghallab et al., 2004). This capability was made possible by several major developments
of the Disciple approach. For instance, we have significantly extended the knowledge
representation and management of a Disciple agent by introducing new types of know-
ledge that are characteristic of planning systems, such as planning tasks and actions
(with preconditions, effects, goal, duration, and resources) and new types of rules
(e.g., planning tasks reduction rules, concretion rules, action rules, goal synthesis rules).
We have introduced state knowledge bases and have developed the ability to manage
the evolution of the states in planning. We have developed a modeling language and
a set of guidelines that help subject matter experts express their planning process.
We have developed an integrated set of learning methods for planning, allowing the
agent to learn general planning knowledge starting from a single planning example
formulated by the expert.
A second result is the development of an integrated approach to planning and infer-
ence, both processes being based on the task reduction paradigm. This improves the
power of the planning systems that can now include complex inference trees. It also
improves the efficiency of the planning process because some of the planning operations
can be performed as part of a much more efficient inference process that does not require
a simulation of the change of the state of the world.
A third result is the development and implementation of the concept of library of virtual
experts. This required the development of methods for the management of a hierarchical
knowledge repository. The hierarchical organization of the knowledge bases of the virtual
experts also serves as a knowledge repository that speeds up the development of new
virtual experts that can reuse the knowledge bases from the upper levels of this hierarchy.
A fourth result is the development of the multidomain architecture of Disciple-VPT,
which extends the applicability of the expert systems to problems whose solutions require
knowledge of more than one domain.
A fifth result is the development of two basic virtual experts, a basic fire expert and a
basic emergency management expert, that can collaborate to develop plans of actions that
are beyond their individual capabilities.
Finally, a sixth result is the development of an approach and system that has high
potential for supporting a wide range of training and planning activities.

13:54:21,
.013
 Design Principles for Cognitive
13 Assistants

This book has presented an advanced approach to developing personal cognitive assist-
ants. Although the emphasis in this book has been on cognitive assistants for evidence-
based hypothesis analysis, the Disciple approach is also applicable to other types of tasks,
as was illustrated by the agents presented in Chapter 12. Moreover, the Disciple approach
illustrates the application of several design principles that are useful in the development of
cognitive assistants in general. In this chapter, we review these principles, which have
been illustrated throughout this book. Each of the following sections starts with the
formulation of a principle and continues with its illustration by referring back to previous
sections of the book.

13.1 LEARNING-BASED KNOWLEDGE ENGINEERING

Employ learning technology to simplify and automate the knowledge engineering process.

It is generally accepted that knowledge engineering is very difficult, involving many

creative tasks. One way to simplify this process significantly is to automate as much of
the knowledge engineering process as possible. As discussed in Section 3.3.6, the
approach taken with Disciple is to replace each knowledge base development activity of
the knowledge engineer (e.g., modeling the problem-solving process, ontology develop-
ment, rule learning, rule refinement) with an equivalent activity that can be performed
directly by a subject matter expert and the Disciple agent, with limited or no support from
the knowledge engineer (see Figure 3.19, p. 107).
Consider, for example, the modeling of the problem-solving process. A knowledge
engineer would need to instruct a subject matter expert how to express his or her
reasoning in the divide-and-conquer analysis and synthesis framework. Then the expert
and the agent can model the solutions of new problems by themselves. In this process, the
agent will support the expert in various ways. For example, the agent may employ
previously learned rules to suggest likely reductions of the current problem or hypothesis.
Or it may learn and reuse reasoning patterns to suggest reductions to the expert.
Now consider the development and testing of the reasoning rules, which the knowledge
engineer does through interviews with the subject matter expert, as discussed in Section
3.1.4. This time-consuming and error-prone task is reduced to several tasks that the
subject matter expert and the Disciple agent can easily perform, as discussed in Sections
9.4 and 10.1.2. In particular, the expert provides examples and helps the agent to
understand them, and the agent generates the reasoning rules. Then the agent employs

426

13:56:23,
.014
 13.4. Knowledge Base Structuring 427

the rules in problem solving and the expert critiques the reasoning process which, in turn,
guides the agent in refining the rules.

13.2 PROBLEM-SOLVING PARADIGM FOR USER–AGENT

COLLABORATION

Use a problem-solving paradigm that is both natural for the human user and appropriate for the
automated agent.

Cognitive assistants, by their very nature, need to be based on problem-solving paradigms

that facilitate user–agent collaboration. On one hand, the employed problem-solving
paradigm needs to be natural enough for the human users. On the other hand, it has to
be formal enough to be automatically applied by the agents.
As discussed in Chapter 4, Disciple agents employ a divide-and-conquer approach
where complex problems or hypotheses, expressed in natural language, are successively
reduced to simpler and simpler ones, guided by corresponding questions and answers.
The typical questions are those from Rudyard Kipling’s well-known poem, “I Keep Six
Honest . . .”: “What?” “Why?” “When?” “How?” “Where?” and, “Who?”
While the analysis part of the problem-solving strategy is very natural, its synthesis part
may be challenging, as pointed out by Toffler (1984), and also experienced by us with the
development of many Disciple agents, such as Disciple-COG (see Sections 4.1 and 12.4).
Therefore, in Disciple-EBR, we have developed a simplified version of solution synthesis for
evidence-based reasoning that is actually very easy to use, as discussed in Sections 4.3 and 4.4.

13.3 MULTI-AGENT AND MULTIDOMAIN PROBLEM SOLVING

Use a problem-solving paradigm for the agent that facilitates both collaboration between users
assisted by their agents and the solving of problems requiring multidomain expertise.

Many existing or potential applications of cognitive assistants are cross-domain, requiring the
collaboration not only between a user and his or her assistant, but also among several users,
each with his or her own area of expertise, as illustrated by the emergency response planning
domain addressed by Disciple-VPT (see Section 12.5). The problem reduction strategy
employed by the Disciple agents can reduce a multidisciplinary problem to subproblems
that may be solved by different experts and their agents. Then, the domain-specific solutions
found by individual users may be combined to produce the solution of the multidisciplinary
problem. With such an approach, each agent supports its user, not only in problem solving,
but also in collaboration and sharing of information with the other users.

13.4 KNOWLEDGE BASE STRUCTURING FOR

KNOWLEDGE REUSE

Structure the knowledge base to facilitate knowledge reuse.

Knowledge base development is a very complex activity, and knowledge reuse can signifi-
cantly simplify it. Moreover, knowledge reuse facilitates the communication with other
agents that share the same knowledge.

13:56:23,
.014
 428 Chapter 13. Design Principles for PLAs

Disciple agents facilitate the reuse of knowledge through two types of knowledge struc-
turing. First, the agent’s knowledge is structured into an ontology that defines the concepts
of the application domain and a set of problem-solving rules expressed with these concepts.
The ontology is the more general part, being applicable to an entire domain. Therefore,
when developing a new application, parts of the ontology can be reused from the previously
developed applications in that domain.
A second type of knowledge structuring is the organization of the knowledge repository
of the agent as a three-level hierarchy of knowledge bases, as discussed in Section 3.3.5
(see Figure 3.17, p. 104). The top of the knowledge repository is the Shared KB, which
contains general knowledge for evidence-based reasoning applicable in all the domains.
Under the Shared KB are Domain KBs, each corresponding to a different application
domain. Finally, under each Domain KB are Scenario KBs, each corresponding to a
different scenario. Therefore, when developing the KB for a new scenario, the agent reuses
the corresponding Domain KB and the Shared KB. Similarly, when developing a new
Domain KB, the agent reuses the Shared KB.

13.5 INTEGRATED TEACHING AND LEARNING

Use agent teaching and learning methods where the user helps the agent to learn and the agent
helps the user to teach it.

Learning the elements of the knowledge base is a very complex process. In the Disciple
approach, this process is simplified by a synergistic integration of teaching and learning, as
discussed in Chapters 9 and 10. In particular, the user helps the agent to learn by providing
representative examples of reasoning steps, as well as hints that guide the agent in
understanding these examples. But the agent also helps the user to teach it by presenting
attempted solutions to problems for the user to critique, as well as attempted explanations
of an example, from which the user can select the correct ones.

13.6 MULTISTRATEGY LEARNING

Use multistrategy learning methods that integrate complementary learning strategies in order to
take advantage of their strengths to compensate for each other’s weaknesses.

No single learning strategy is powerful enough to learn the complex reasoning rules
needed by the agent, but their synergistic combination is, as illustrated by the methods
employed by the Disciple agents. In particular, a Disciple agent employs learning from
explanations and analogy-based generalization to generate a plausible version space rule
from a single problem-solving example and its explanation (see Section 9.4, p. 258, and
Figure 9.9, p. 259). It then employs learning by analogy and experimentation to generate
additional examples of the learned rules that are to be critiqued by the user, and it
employs empirical induction from examples, as well as learning from failure explan-
ations, to refine the rule based on the new examples and explanations (see Section 10.1.2
and Figure 10.1, p. 295). This results in a powerful method that enables the agent to learn
very complex rules from only a few examples and their explanations, obtained in a
natural dialogue with the user.

13:56:23,
.014
 13.8. Modeling, Learning, and Problem Solving 429

13.7 KNOWLEDGE ADAPTATION

Use methods that allow continuous adaptation of agent’s knowledge.

Due to the complexity of the real world and to its dynamic nature, an agent’s knowledge
elements will always be approximate representations of real-world entities. Therefore,
improving and even maintaining the utility of a cognitive assistant depends on its capacity
to adapt its knowledge continuously to better represent the application domain.
The rule-learning methods of the Disciple agents continually improve the rules based
on their failures and successes. But these improvements are done in the context of an
existing ontology, which may itself evolve. Therefore, when the ontology undergoes
significant changes, the previously learned rules need to be relearned. This process can
be done automatically, if each rule maintains minimal generalizations of the examples and
the explanations from which it was learned, as was discussed in Section 10.2.

13.8 MIXED-INITIATIVE MODELING, LEARNING,

AND PROBLEM SOLVING

Use mixed-initiative methods where modeling, learning, and problem solving mutually support
each other to capture the expert’s tacit knowledge.

Figure 13.1 illustrates the synergistic integration of modeling, learning, and problem
solving as part of the reasoning of a Disciple agent. These activities mutually support each
other. For example, problem solving generates examples for learning to refine the rules,
and the refined rules lead to better problem solving..
Modeling provides learning with the initial expert example for learning a new rule. But
the previously learned rules or patterns also support the modeling process by suggesting
possible reductions of a problem or hypothesis, as discussed, for instance, in Section 3.3.2.
Finally, modeling advances the problem-solving process with the creative solutions
provided by the user. But problem solving also supports the modeling process by provid-
ing the context for such creative solutions, thus facilitating their definition.

Creative
solution Problem
Context for Solving
creative solution Generated
example
Mixed-
Modeling
Initiative
Reasoning Refined
rule
Expert
example Learning

Rule-based
guidance

Figure 13.1. Integrated modeling, learning, and problem solving.

13:56:23,
.014
 430 Chapter 13. Design Principles for PLAs

13.9 PLAUSIBLE REASONING WITH PARTIALLY

LEARNED KNOWLEDGE

Use reasoning methods that enable the use of partially learned knowledge.

As already indicated, an agent’s knowledge elements will always be approximate repre-

sentations of real-world entities. This is more so for a learning agent that incrementally
improves its knowledge. Since most of the knowledge of a learning agent is partially
learned, the agent should be able to use partially learned knowledge in problem solving.
Otherwise, it would not be able to function as an assistant until it is “fully” developed.
A Disciple agent not only uses partially learned knowledge in problem solving, but also
assesses the plausibility (or confidence level) of a problem or hypothesis reduction based
on its position with respect to the plausible version space conditions of the corresponding
reduction rule, as discussed in Section 7.7 (see Figure 7.14, p. 216). The agent can also
assess the plausibility of an entire reasoning tree by considering the confidence levels of its
component reductions.

13.10 USER TUTORING IN PROBLEM SOLVING

Employ approaches to user tutoring that allow the agent to teach its problem-solving paradigm
easily and rapidly to its users, facilitating their collaboration.

A cognitive assistant collaborates with its user in problem solving. This requires the user
not only to understand the agent’s reasoning but also to contribute to it. A subject matter
expert teaches a Disciple agent similarly to how the expert would teach a student, through
problem-solving examples and explanations. Then, when a user employs a Disciple agent,
he or she can easily learn from its explicit reasoning, as illustrated by the Disciple-COG
agent discussed in Section 12.4. Alternatively, the agent can behave as a tutoring system,
guiding the student through a series of lessons and exercises, as illustrated by TIACRITIS
(Tecuci et al., 2010b).
Personalized learning, which was identified as one of the fourteen Grand Challenges for
Engineering in the Twenty-first Century (NAE, 2008), is a very important application of
cognitive assistants.

13.11 AGENT ARCHITECTURE FOR RAPID AGENT

DEVELOPMENT

Employ learning agent shells that allow rapid agent prototyping, development, and customization.

As discussed in Section 3.2.3, a learning agent shell enables rapid development of an agent
because all the reasoning modules already exist. Thus one only needs to customize some
of the modules and to develop the knowledge base. For example, the customizations
performed to develop Disciple-COG consisted in developing a report generator and a
simplified interface for the problem solver.
As discussed in Section 3.2.4, a learning agent shell for evidence-based reasoning further
speeds up the development of an agent. First, the agent shell was already customized for

13:56:23,
.014
 13.12. Design Based on an Agent Life Cycle 431

evidence-based reasoning. Second, part of the knowledge base is already defined in the
shell, namely the Shared KB for evidence-based reasoning (EBR KB).

13.12 DESIGN BASED ON A COMPLETE AGENT LIFE CYCLE

Design the agent by taking into account its complete life cycle, to ensure its usefulness for as long a
period of time as possible.

This involves the incorporation of methods that support the various stages in the life cycle
of the agent. For example, Figure 13.2 illustrates a possible life cycle of a Disciple agent
that was discussed in Section 3.2.4.
The first stage is shell customization, where, based on the specification of the type of
problems to be solved and the agent to be built, the developer and the knowledge engineer
may decide that some customizations or extensions of the Disciple shell may be necessary
or useful. The next stage is agent teaching by the subject matter expert and the knowledge
engineer, supported by the agent itself, which simplifies and speeds up the knowledge
base development process. Once an operational agent is developed, it is used for training
of end-users, possibly in a classroom environment. The next stage is field use, where copies

Disciple Shell customization

1 Developer and knowledge engineer
Knowledge integration
Agent teaching and optimization
Disciple
Disciple
2 6 Disciple
Disciple
Subject matter expert Knowledge engineer
and knowledge engineer and subject matter expert
Training of end-user
3 5 After action review and learning

Disciple Disciple

End-user End-user
4

Libraries Cognitive
Knowledge Repositories
Massive Databases
assistant GLOBAL
Disciple KNOWLEDGE BASE
SEARCH
AGENT
SEARCH
SEARCH
AGENT
AGENT

Disciple
Field use

Reasoning
assistance
Disciple Learning
Disciple
Collaboration Information
assistance sharing

Figure 13.2. Possible life cycle of a Disciple agent.

13:56:23,
.014
432 Chapter 13. Design Principles for PLAs

of the agent support users in their operational environments. At this stage, each agent
assists its user both in solving problems and in collaborating with other users and their
cognitive assistants. At the same time, the agent continuously learns patterns from this
problem-solving experience by employing a form of nondisruptive learning. However,
because there is no learning assistance from the user, the learned patterns will not include
a formal applicability condition. It is during the next stage of after action review and
learning, when the user and the agent analyze past problem-solving episodes, that the
formal applicability conditions are learned based on the accumulated examples. In time,
each cognitive assistant extends its knowledge with additional expertise acquired from its
user. This creates the opportunity of developing a more competent agent by integrating
the knowledge of all these agents. This can be accomplished by a knowledge engineer,
with assistance from a subject matter expert, in the next stage of knowledge integration and
optimization. The result is an improved agent that may be used in a new iteration of a
spiral process of development and use.

13:56:23,
.014
References

Allemang, D., and Hendler, J. (2011). Semantic Web for the Working Ontologist: Effective Modeling in
RDFS and Owl, Morgan Kaufmann, San Mateo, CA.
Allen, J., Hendler, J., and Tate, A., (eds.) (1990). Readings in Planning, Morgan Kaufmann, San
Mateo, CA.
Anderson, T., Schum, D., and Twining, W. (2005). Analysis of Evidence, Cambridge University Press,
Cambridge, UK.
Awad, E. M. (1996). Building Expert Systems: Principles, Procedures, and Applications, West, New
York, NY.
Awad, E. M., and Ghaziri, H. M. (2004). Knowledge Management, Pearson Education International,
Prentice Hall, Upper Saddle River, NJ, pp. 60–65.
Basic Formal Ontology (BFO) (2012). Basic Formal Ontology. www.ifomis.org/bfo (accessed August
31, 2012).
Betham, J. (1810). An Introductory View of the Rationale of the Law of Evidence for Use by Non-
lawyers as Well as Lawyers (vi works 1–187 (Bowring edition, 1837–43, originally edited by James
Mill circa 1810).
Boicu, C. (2006). An Integrated Approach to Rule Refinement for Instructable Knowledge-Based Agents.
PhD Thesis in Computer Science, Learning Agents Center, Volgenau School of Information
Technology and Engineering, George Mason University, Fairfax, VA.
Boicu, C., Tecuci, G., and Boicu, M. (2005). Improving Agent Learning through Rule Analysis, in
Proceedings of the International Conference on Artificial Intelligence, ICAI-05, Las Vegas, NV, June
27–30. lac.gmu.edu/publications/data/2005/ICAI3196Boicu.pdf (accessed April 12, 2016)
Boicu, M. (2002). Modeling and Learning with Incomplete Knowledge, PhD Dissertation in
Information Technology, Learning Agents Laboratory, School of Information Technology and
Engineering, George Mason University. lac.gmu.edu/publications/2002/BoicuM_PhD_Thesis.pdf
(accessed November 25, 2015)
Boicu, M., Tecuci, G., Bowman, M., Marcu, D., Lee, S. W., and Wright, K. (1999). A Problem-Oriented
Approach to Ontology Creation and Maintenance, in Proceedings of the Sixteenth National Confer-
ence on Artificial Intelligence Workshop on Ontology Management, July 18–19, Orlando, Florida,
AAAI Press, Menlo Park, CA. lac.gmu.edu/publications/data/1999/ontology-1999.pdf (accessed
November 25, 2015)
Boicu, M., Tecuci, G., Marcu, D., Bowman, M., Shyr, P., Ciucu, F., and Levcovici, C. (2000). Disciple-
COA: From Agent Programming to Agent Teaching, in Proceedings of the Seventeenth International
Conference on Machine Learning (ICML), Stanford, CA, Morgan Kaufman, San Mateo, CA, lac.gmu
.edu/publications/data/2000/2000_il-final.pdf (accessed November 25, 2015)
Bowman, M. (2002). A Methodology for Modeling Expert Knowledge That Supports Teaching Based
Development of Agents, PhD Dissertation in Information Technology, George Mason University,
Fairfax, VA. lac.gmu.edu/publications/data/2002/Michael%20Bowman-Thesis.pdf (accessed
November 25, 2015)
Bresina, J. L., and Morris, P. H. (2007). Mixed-Initiative Planning in Space Mission Operations, AI
Magazine, vol. 28, no. 1, pp. 75–88.

433

13:56:23,
.015
434 References

Breuker, J., and Wielinga, B. (1989). Models of Expertise in Knowledge Acquisition, in Guida, G., and
Tasso, C. (eds.), Topics in Expert Systems Design, Methodologies, and Tools, North Holland,
Amsterdam, Netherlands, pp. 265–295.
Buchanan, B. G., and Feigenbaum, E. A. (1978). DENDRAL and META-DENDRAL: Their Applications
Dimensions, Artificial Intelligence, vol. 11, pp. 5–24.
Buchanan, B. G., and Shortliffe, E. H. (eds.) (1984). Rule-Based Expert Systems: The MYCIN Experi-
ments of the Stanford Heuristic Programming Project, Addison-Wesley, Reading, MA.
Buchanan, B. G., and Wilkins, D. C. (eds.) (1993). Readings in Knowledge Acquisition and Learning:
Automating the Construction and Improvement of Expert Systems, Morgan Kaufmann, San
Mateo, CA.
Buchanan, B. G., Barstow, D., Bechtal, R., Bennett, J., Clancey, W., Kulikowski, C., Mitchell, T., and
Waterman, D. A. (1983). Constructing an Expert System, in Hayes-Roth, F., Waterman, D.,
and Lenat, D. (eds.), Building Expert Systems, Addison-Wesley, Reading, MA, pp. 127–168.
Carbonell, J. G. (1983). Learning by Analogy: Formulating and Generalizing Plans from Past Experi-
ence, in Michalski, R. S., Carbonell, J. M., and Mitchell, T. M., Machine Learning: An Artificial
Intelligence Approach, Tioga, Wellsboro, PA, pp. 137–162.
Carbonell, J. G. (1986). Derivational Analogy: A Theory of Reconstructive Problem-Solving and
Expertise Acquisition, in Michalski, R. S., Carbonell, J. G., and Mitchell, T. M. (eds.), Machine
Learning: An Artificial Intelligence Approach, vol. 2, Morgan Kaufmann, San Mateo, CA,
pp. 371–392.
Chaudhri, V. K., Farquhar, A., Fikes, R., Park, P. D., and Rice, J. P. (1998). OKBC: A Programmatic
Foundation for Knowledge Base Interoperability, in Proceedings of the Fifteenth National Confer-
ence on Artificial Intelligence (AAAI-98), AAAI Press, Menlo Park, CA, pp. 600–607.
Clancey, W. (1985). Heuristic Classification, AI Journal, vol. 27, pp. 289–350.
Clausewitz, C. von (1832 [1976]). On War, translated and edited by Howard, M., and Paret, P.
Princeton University Press, Princeton, NJ.
Cohen, L. J. (1977). The Probable and the Provable, Clarendon Press, Oxford. UK.
Cohen, L. J. (1989). An Introduction to the Philosophy of Induction and Probability, Clarendon Press,
Oxford, UK.
Cohen, M. R., and Nagel, E. (1934). An Introduction to Logic and Scientific Method, Harcourt, Brace,
New York, NY, pp. 274–275.
Cohen, P., Schrag, R., Jones, E., Pease, A., Lin, A., Starr, B., Gunning, D., and Burke, M. (1998). The
DARPA High-Performance Knowledge Bases Project, AI Magazine, vol. 19, no. 4, pp. 25–49.
Cooper, T., and Wogrin, N. (1988). Rule-based Programming with OPS5, Morgan Kaufmann, San
Mateo, CA.
Cross, S. E., and Walker, E. (1994). DART: Applying Knowledge-based Planning and Scheduling to
Crisis Action Planning, in Zweben, M., and Fox, M. S. (eds.), Intelligent Scheduling, Morgan
Kaufmann, San Mateo, CA, pp. 711–729.
Cyc (2008). OpenCyc Just Got Better – Much Better! www.opencyc.org (accessed August 22, 2008).
Cyc (2016). The Cyc homepage, www.cyc.com (accessed February 3, 2016).
Dale, A. I. (2003). Most Honourable Remembrance: The Life and Work of Thomas Bayes, Springer-
Verlag, New York, NY.
David, F. N. (1962). Games, Gods, and Gambling, Griffin, London, UK.
David, P. A., and Foray, D. (2003). Economic Fundamentals of the Knowledge Society, Policy Futures
in Education. An e-Journal, vol. 1, no. 1, Special Issue: Education and the Knowledge Economy,
January, pp. 20–49.
Davies, T. R., and Russell, S. J. (1990). A Logical Approach to Reasoning by Analogy, in Shavlik, J.,
and Dietterich, T. (eds.), Readings in Machine Learning, Morgan Kaufmann, San Mateo, CA,
pp. 657–663.
DeJong, G., and Mooney, R. (1986). Explanation-based Learning: An Alternative View, Machine
Learning, vol. 1, pp. 145–176.
Department of Homeland Security (DHS) (2004). National Response Plan.
Desai, M. (2009). Persistent Stare Exploitation and Analysis System (PerSEAS), DARPA-BAA-09-55,
https://www.fbo.gov/index?s=opportunity&mode=form&id=eb5dd436ac371ce79d91c84ec4e91341
&tab=core&_cview=1 (accessed April 13, 2016)

13:56:23,
.015
References 435

DOLCE (2012). Laboratory for Applied Ontology, www.loa-cnr.it/DOLCE.html (accessed August 31,
2012)
Drucker, P. (1993). Post-Capitalist Society, HarperCollins, New York.
Durham, S. (2000). Product-Centered Approach to Information Fusion, AFOSR Forum on Information
Fusion, Arlington, VA, October 18–20.
Durkin, J. (1994). Expert Systems: Design and Development, Prentice Hall, Englewood Cliffs, NJ.
Dybala, T. (1996). Shared Expertise Model for Building Interactive Learning Agents, PhD Dissertation,
School of Information Technology and Engineering, George Mason University, Fairfax, VA.
lac.gmu.edu/publications/data/1996/Dybala-PhD-abs.pdf (accessed April 12, 2016)
Echevarria, A. J. (2003). Reining in the Center of Gravity Concept. Air & Space Power Journal, vol.
XVII, no. 2, pp. 87–96.
Eco, U., and Sebeok, T. (eds.) (1983). The Sign of Three: Dupin, Holmes, Peirce, Indiana University
Press: Bloomington.
Eikmeier, D. C. (2006). Linking Ends, Ways and Means with Center of Gravity Analysis. Carlisle
Barracks, U.S. Army War College, Carlisle, PA.
Einstein, A. (1939). Letter from Albert Einstein to President Franklin D. Roosevelt: 08/02/1939. The
letter itself is in the Franklin D. Roosevelt Library in Hyde Park, NY. See the National Archives copy
in pdf form at media.nara.gov/Public_Vaults/00762_.pdf (accessed November 16, 2014).
EXPECT (2015). The EXPECT homepage, www.isi.edu/ikcap/expect/ (accessed May 25, 2015).
Farquhar, A., Fikes, R., and Rice, J. (1997). The Ontolingua Server: A Tool for Collaborative Ontology
Construction, International Journal of Human–Computer Studies, vol. 46, no. 6, pp. 707–727.
Federal Rules of Evidence (2009). 2009–2010 ed. West Publishing, St. Paul, MN.
Feigenbaum, E. A. (1982). Knowledge Engineering for the 1980s, Research Report, Stanford Univer-
sity, Stanford, CA.
Feigenbaum, E. A. (1993). Tiger in a Cage: The Applications of Knowledge-based Systems, Invited
Talk, AAAI-93 Proceedings, www.aaai.org/Papers/AAAI/1993/AAAI93-127.pdf. (accessed April 13,
2016)
Fellbaum, C. (ed.) (1988). WordNet: An Electronic Lexical Database. MIT Press, Cambridge, MA.
FEMA (Federal Emergency Management Agency) (2007). National Incident Management System.
www.fema.gov/national-incident-management-system (accessed April 13, 2016).
Ferrucci, D., Brown, E., Chu-Caroll, J., Fan, J., Gondek, D., Kalynapur, A. A., Murdoch, J. W., Nyberg,
E., Prager, J., Schlaefer, N., and Welty, C. (2010). Building Watson: An Overview of the DeepQA
Project, AI Magazine, vol. 31, no. 3, pp. 59–79.
Filip, F. G. (1989). Creativity and Decision Support System, Studies and Researches in Computer and
Informatics, vol. 1, no. 1, pp. 41–49.
Filip, F. G. (ed.) (2001). Informational Society–Knowledge Society, Expert, Bucharest.
FM 100–5. (1993). U.S. Army Field Manual 100–5, Operations, Headquarters, Department of the
Army, Washington, DC.
FOAF (2012). The Friend of a Friend (FOAF) project. www.foaf-project.org/ (accessed August 31,
2012)
Forbes (2013). www.forbes.com/profile/michael-chipman/ (accessed September 2, 2013)
Forbus, K. D., Gentner, D., and Law, K. (1994). MAC/FAC: A Model of Similarity-Based Retrieval,
Cognitive Science, vol. 19, pp. 141–205.
Friedman-Hill, E. (2003). Jess in Action, Manning, Shelter Island, NY.
Gammack, J. G. (1987). Different Techniques and Different Aspects on Declarative Knowledge, in
Kidd, A. L. (ed.), Knowledge Acquisition for Expert Systems: A Practical Handbook, Plenum Press,
New York, NY, and London, UK.
Gentner, D. (1983). Structure Mapping: A Theoretical Framework for Analogy, Cognitive Science,
vol. 7, pp. 155–170.
Geonames (2012). GeoNames Ontology–Geo Semantic Web. [Online] www.geonames.org/ontology/
documentation.html (accessed August 31, 2012)
GFO (2012). General Formal Ontology (GFO). [Online] www.onto-med.de/ontologies/gfo/ (accessed
August 31, 2012)
Ghallab, M., Nau, D., and Traverso, P. (2004). Automatic Planning: Theory and Practice, Morgan
Kaufmann, San Mateo, CA.

13:56:23,
.015
436 References

Giarratano, J., and Riley, G. (1994). Expert Systems: Principles and Programming, PWS, Boston, MA.
Gil, Y., and Paris, C. (1995). Towards model-independent knowledge acquisition, in Tecuci, G., and
Kodratoff, Y. (eds.), Machine Learning and Knowledge Acquisition: Integrated Approaches,
Academic Press, Boston, MA.
Giles, P. K., and Galvin, T. P. (1996). Center of Gravity: Determination, Analysis and Application.
Carlisle Barracks, U.S. Army War College, Carlisle, PA.
Goodman, D., and Keene, R. (1997). Man versus Machine: Kasparov versus Deep Blue, H3 Publica-
tions, Cambridge, MA.
Gruber, T. R. (1993). A Translation Approach to Portable Ontology Specification. Knowledge Acquisi-
tion, vol. 5, pp. 199–220.
Guizzardi, G., and Wagner, G. (2005a). Some Applications of a Unified Foundational Ontology in
Business, in Rosemann, M., and Green, P. (eds.), Ontologies and Business Systems Analysis, IDEA
Group, Hershey, PA.
Guizzardi, G., and Wagner, G. (2005b). Towards Ontological Foundations for Agent Modeling
Concepts Using UFO, in Agent-Oriented Information Systems (AOIS), selected revised papers of
the Sixth International Bi-Conference Workshop on Agent-Oriented Information Systems.
Springer-Verlag, Berlin and Heidelberg, Germany.
Hieb, M. R. (1996). Training Instructable Agents through Plausible Version Space Learning, PhD
Dissertation, School of Information Technology and Engineering, George Mason University, Fair-
fax, VA.
Hobbs, J. R., and Pan, F. (2004). An Ontology of Time for the Semantic Web, CM Transactions on
Asian Language Processing (TALIP), vol. 3, no. 1 (special issue on temporal information process-
ing), pp. 66–85.
Horvitz, E. (1999). Principles of Mixed-Initiative User Interfaces, in Proceedings of CHI '99, ACM
SIGCHI Conference on Human Factors in Computing Systems, Pittsburgh, PA, May. ACM Press,
New York, NY. research.microsoft.com/~horvitz/uiact.htm (accessed April 13, 2016)
Humphreys, B. L., and Lindberg, D.A.B. (1993). The UMLS Project: Making the Conceptual Connec-
tion between Users and the Information They Need, Bulletin of the Medical Library Association,
vol. 81, no. 2, p. 170.
Jackson, P. (1999). Introduction to Expert Systems, Addison-Wesley, Essex, UK.
Jena (2012). Jena tutorial. jena.sourceforge.net/tutorial/index.html (accessed August 4, 2012).
JESS (2016). The rule engine for the JAVA platform, JESS webpage: www.jessrules.com/jess/down
load.shtml (accessed February 3, 2016)
Joint Chiefs of Staff (2008). Joint Operations, Joint Pub 3-0, U.S. Joint Chiefs of Staff, Washington, DC.
Jones, E. (1998). HPKB Year 1 End-to-End Battlespace Challenge Problem Specification, Alphatech,
Burlington, MA.
Kant, I. (1781). The Critique of Pure Reason, Project Gutenberg, www.gutenberg.org/ebooks/4280
(accessed, August 19, 2013)
Keeling, H. (1998). A Methodology for Building Verified and Validated Intelligent Educational
Agents – through an Integration of Machine Learning and Knowledge Acquisition, PhD Dissertation,
School of Information Technology and Engineering, George Mason University, Fairfax, VA.
Kent, S. (1994). Words of Estimated Probability, in Steury, D. P. (ed.), Sherman Kent and the Board of
National Estimates: Collected Essays, Center for the Study of Intelligence, CIA, Washington, DC.
Kim, J., and Gil, Y. (1999). Deriving Expectations to Guide Knowledge Base Creation, in Proceedings
of AAAI-99/IAAI-99, AAAI Press, Menlo Park, CA, pp. 235–241.
Kneale, W. (1949). Probability and Induction, Clarendon Press, Oxford, UK. pp. 30–37.
Kodratoff, Y., and Ganascia, J-G. (1986). Improving the Generalization Step in Learning, in Michalski,
R., Carbonell, J., and Mitchell, T. (eds.), Machine Learning: An Artificial Intelligence Approach,
vol. 2. Morgan Kaufmann, San Mateo, CA, pp. 215–244.
Kolmogorov, A. N. (1933 [1956]). Foundations of a Theory of Probability, 2nd English ed., Chelsea,
New York, NY, pp. 3–4.
Kolmogorov, A. N. (1969). The Theory of Probability, in Aleksandrov, A. D., Kolmogorov, A. N., and
Lavrentiev, M. A. (eds.), Mathematics: Its Content, Methods, and Meaning, vol. 2, MIT Press,
Cambridge, MA, pp. 231–264.
Langley, P. W. (2012). The Cognitive Systems Paradigm, Advances in Cognitive Systems, vol. 1, pp. 3–13.

13:56:23,
.015
References 437

Laplace, P. S. (1814). Théorie Analytique des Probabilités, 2nd édition, Paris, Ve. Courcier, archive.
org/details/thorieanalytiqu01laplgoog (accessed January 28, 2016)
Le, V. (2008). Abstraction of Reasoning for Problem Solving and Tutoring Assistants. PhD Dissertation
in Information Technology. Learning Agents Center, Volgenau School of IT&E, George Mason
University, Fairfax, VA.
Lempert, R. O., Gross, S. R., and Liebman, J. S. (2000). A Modern Approach to Evidence, 3rd ed., West
Publishing, St. Paul, MN, pp. 1146–1148.
Lenat, D. B. (1995). Cyc: A Large-scale Investment in Knowledge Infrastructure, Communications of
the ACM, vol. 38, no. 11, pp. 33–38.
Loom (1999). Retrospective on LOOM. www.isi.edu/isd/LOOM/papers/macgregor/Loom_Retrospec
tive.html (accessed August 4, 2012)
MacGregor, R. (1991). The Evolving Technology of Classification-Based Knowledge Representation
Systems, in Sowa, J. (ed.), Principles of Semantic Networks: Explorations in the Representations of
Knowledge, Morgan Kaufmann, San Francisco, CA, pp. 385–400.
Marcu, D. (2009). Learning of Mixed-Initiative Human-Computer Interaction Models, PhD Disserta-
tion in Computer Science. Learning Agents Center, Volgenau School of IT&E, George Mason
University, Fairfax, VA.
Marcus, S. (1988). SALT: A Knowledge-Acquisition Tool for Propose-and-Revise Systems, in Marcus,
S. (ed.), Automating Knowledge Acquisition for Expert Systems, Kluwer Academic., Norwell, MA,
pp. 81–123.
Masolo, C., Vieu, L., Bottazzi, E., Catenacci, C., Ferrario, R., Gangemi, A., and Guarino, N. (2004).
Social Roles and Their Descriptions, in Dubois, D., Welty, C., and Williams, M-A. (eds.), Principles
of Knowledge Representation and Reasoning: Proceedings of the Ninth International Conference
(KR2004), AAAI Press, Menlo Park, CA, pp. 267–277.
McDermott, J. (1982). R1: A Rule-Based Configurer of Computer Systems, Artificial Intelligence
Journal, vol. 19, no. 1, pp. 39–88.
Meckl, S., Tecuci, G., Boicu, M., and Marcu, D. (2015). Towards an Operational Semantic Theory of
Cyber Defense against Advanced Persistent Threats, in Laskey, K. B., Emmons, I., Costa, P. C. G.,
and Oltramari, A. (eds.), Proceedings of the Tenth International Conference on Semantic Technolo-
gies for Intelligence, Defense, and Security – STIDS 2015, pp. 58–65, Fairfax, VA, November 18–20.
lac.gmu.edu/publications/2015/APT-LAC.pdf (accessed January 12, 2016)
Michalski, R. S. (1986). Understanding the Nature of Learning: Issues and Research Directions, in
Michalski, R. S., Carbonell, J. G., and Mitchell T. (eds.), Machine Learning, vol. 2, Morgan
Kaufmann, Los Altos, CA, pp. 3–25.
Michalski, R. S., and Tecuci, G. (eds.) (1994). Machine Learning: A Multistrategy Approach, vol. IV,
Morgan Kaufmann, San Mateo, CA. store.elsevier.com/Machine-Learning/isbn-9781558602519/
(accessed May 29, 2015)
Minsky, M. (1986). The Society of Mind, Simon and Schuster, New York, NY.
Mitchell, T. M. (1978). Version Spaces: An Approach to Concept Learning. PhD Dissertation, Stanford
University, Stanford, CA.
Mitchell, T. M. (1997). Machine Learning. McGraw-Hill, New York, NY.
Mitchell, T. M., Keller, R. M., and Kedar-Cabelli, S. T. (1986). Explanation-Based Generalization:
A Unifying View, Machine Learning, vol. 1, pp. 47–80.
Murphy, P. (2003). Evidence, Proof, and Facts: A Book of Sources. Oxford University Press, Oxford, UK.
Musen, M. A. (1989). Automated Generation of Model-based Knowledge Acquisition Tools, Morgan
Kaufmann., San Francisco, CA.
NAE (National Academy of Engineering) (2008). Grand Challenges for Engineering. www.engineer
ingchallenges.org/cms/challenges.aspx (accessed April 13, 2016)
Nau, D., Au, T., Ilghami, O., Kuter, U., Murdock, J., Wu, D., and Yaman, F. (2003). SHOP2: An HTN
Planning System, Journal of Artificial Intelligence Research, vol. 20, pp. 379–404.
Negoita, C. V., and Ralescu, D. A. (1975). Applications of Fuzzy Sets to Systems Analysis, Wiley,
New York, NY.
Nilsson, N. J. (1971). Problem Solving Methods in Artificial Intelligence, McGraw-Hill, New York, NY.
Nonaka, I., and Krogh, G. (2009). Tacit Knowledge and Knowledge Conversion: Controversy and
Advancement in Organizational Knowledge Creation Theory, Organization Science, vol. 20, no. 3

13:56:23,
.015
438 References

(May–June), pp. 635–652. www.ai.wu.ac.at/~kaiser/birgit/Nonaka-Papers/tacit-knowledge-and-

knowledge-conversion-2009.pdf (accessed April 13, 2016)
Noy, N. F., and McGuinness, D. L. (2001). Ontology Development 101: A Guide to Creating Your First
Ontology. Stanford Knowledge Systems Laboratory Technical Report KSL-01–05 and Stanford
Medical Informatics Technical Report SMI-2001-0880, March, Stanford, CA.
NRC (National Research Council) (1996). National Research Council: National Science Education
Standards. National Academy Press, Washington, DC. www.nap.edu/openbook.php?recor
d_id=4962 (accessed April 13, 2016)
NRC (National Research Council) (2000). National Research Council: Inquiry and the National
Science Education Standards, National Academy Press, Washington, DC. www.nap.edu/cata
log.php?record_id=9596 (accessed April 13, 2016)
NRC (National Research Council) (2010). Preparing Teachers: Building Evidence for Sound Policy,
National Academies Press, Washington, DC. www.nap.edu/catalog.php?record_id=12882 (accessed
April 13, 2016)
NRC (National Research Council) (2011). A Framework for K-12 Science Education: Practices, Cross-
cutting Concepts, and Core Ideas. www.nap.edu/catalog.php?record_id=13165 (accessed April 13,
2016)
Obrst, L., Chase, P., and Markeloff, R. (2012). Developing an Ontology of the Cyber Security Domain,
in Proceedings of the Seventh International Conference on Semantic Technologies for Intelligence,
Defense, and Security – STIDS, October 23–26, Fairfax, VA.
OKBC (Open Knowledge Based Connectivity) (2008). OKBC homepage. www.ksl.stanford.edu/soft
ware/OKBC/ (accessed August 4, 2012)
O’Keefe, R. M., Balci, O., and Smith, E. P. (1987). Validating Expert Systems Performance, IEEE
Expert, no. 2, vol. 4, pp. 81–90.
Oldroyd, D. (1986). The Arch of Knowledge: An Introductory Study of the History of the Philosophy and
Methodology of Science, Routledge Kegan & Paul, London, UK.
Ontolingua (1997). Ontolingua System Reference Manual. www-ksl-svc.stanford.edu:5915/doc/
frame-editor/index.html (accessed August 4, 2012).
Ontolingua (2008). The Ontolingua homepage. www.ksl.stanford.edu/software/ontolingua/
(accessed August 4, 2012).
OWLIM (2012). OWLIM family of semantic repositories, or RDF database management systems.
www.ontotext.com/owlim (accessed August 4, 2012).
Pan, F., and Hobbs, J. R. (2004). Time in OWL-S. Proceedings of the AAAI Spring Symposium
on Semantic Web Services, Stanford University, Stanford, CA, AAAI Press, Palo Alto, CA,
pp. 29–36.
Pan, F., and Hobbs, J. R. (2012). A Time Zone Resource in OWL. www.isi.edu/~hobbs/timezoneho
mepage.html (accessed August 31, 2012)
Pease, A. (2011). Ontology: A Practical Guide, Articulate Software Press, Angwin, CA. www.ontology
portal.org/Book.html (accessed April 13, 2016)
Peirce, C. S. (1898 [1992]). Reasoning and the Logic of Things, edited by Ketner, K., Harvard University
Press, Cambridge, MA.
Peirce, C. S. (1901 [1955]). Abduction and Induction, in Philosophical Writings of Peirce, edited by
Buchler, J., Dover, New York, NY, pp. 150–156.
Pellet (2012). OWL 2 Reasoner for Java. clarkparsia.com/pellet/ (accessed August 4, 2012).
Plotkin, G. D. (1970). A Note on Inductive Generalization, in Meltzer, B., and Michie, D. (eds.),
Machine Intelligence 5, Edinburgh University Press, Edinburgh, UK, pp. 165–179.
Pomerol, J. C., and Adam, F. (2006). On the Legacy of Herbert Simon and His Contribution to
Decision-making Support Systems and Artificial Intelligence, in Gupta, J. N. D., Forgionne, G. A.,
and Mora, M. T. (eds.), Intelligent Decision-Making Support Systems: Foundations, Applications
and Challenges, Springer-Verlag, London, UK, pp. 25–43.
Powell, G. M., and Schmidt, C. F. (1988). A First-order Computational Model of Human Operational
Planning, CECOM-TR-01–8, U.S. Army CECOM, Fort Monmouth, NJ.
Protégé (2000). The Protégé Project. protege.stanford.edu (accessed April 13, 2016)
Protégé (2015). Protégé ontology editor and knowledge base framework, homepage. protege.stanford
.edu (accessed May 25, 2015).

13:56:23,
.015
References 439

Puppe, F. (1993). Problem Classes and Problem Solving Methods, in Systematic Introduction to
Expert Systems: Knowledge Representations and Problem Solving Methods, Springer Verlag, Berlin
and Heidelberg, Germany, pp. 87–112.
Ressler, J., Dean, M., and Kolas, D. (2010). Geospatial Ontology Trade Study, in Janssen, T., Ceuster,
W., and Obrst, L. (eds.), Ontologies and Semantic Technologies for Intelligence, IOS Press, Amster-
dam, Berlin, Tokyo, and Washington, DC, pp. 179–212.
Rooney, D., Hearn, G., and Ninan, A. (2005). Handbook on the Knowledge Economy, Edward Elgar,
Cheltenham, UK.
Russell, S. J., and Norvig, P. (2010). Artificial Intelligence: A Modern Approach, Prentice Hall, Upper
Saddle River, NJ, pp. 34–63.
Schneider, L. (2003). How to Build a Foundational Ontology – the Object-centered High-level
Reference Ontology OCHRE, in Proceedings of the 26th Annual German Conference on AI, KI
2003: Advances in Artificial Intelligence, Springer-Verlag, Heidelberg, Germany, pp. 120–134.
Schreiber, G., Akkermans, H., Anjewierden, A., de Hoog, R., Shadbolt, N., Van de Velde, W., and
Wielinga, B. (2000). Knowledge Engineering and Management: The Common KADS Methodology,
MIT Press, Cambridge, MA.
Schum D. A. (1987). Evidence and Inference for the Intelligence Analyst (2 vols), University Press of
America, Lanham, MD.
Schum, D. A. (1989). Knowledge, Probability, and Credibility, Journal of Behavioral Decision Making,
vol. 2, pp. 39–62.
Schum, D. A. (1991). Jonathan Cohen and Thomas Bayes on the Analysis of Chains of Reasoning, in
Eells, E., and Maruszewski, T. (eds.), Probability and Rationality: Studies on L. Jonathan Cohen’s
Philosophy of Science, Editions Rodopi, Amsterdam, Netherlands, pp. 99–145.
Schum, D. A. (1999). Marshaling Thoughts and Evidence during Fact Investigation, South Texas Law
Review, vol. 40, no. 2 (summer), pp. 401–454.
Schum, D. A. (1994 [2001a]). The Evidential Foundations of Probabilistic Reasoning, Northwestern
University Press, Evanston, IL.
Schum, D. A. (2001b). Species of Abductive Reasoning in Fact Investigation in Law, Cardozo Law
Review, vol. 22, nos. 5–6, pp. 1645–1681.
Schum, D. A. (2011). Classifying Forms and Combinations of Evidence: Necessary in a Science of
Evidence, in Dawid, P., Twining, W., and Vasilaki. M. (eds.), Evidence, Inference and Inquiry,
British Academy, Oxford University Press, Oxford, UK, pp. 11–36.
Schum, D. A., and Morris, J. (2007). Assessing the Competence and Credibility of Human Sources
of Evidence: Contributions from Law and Probability, Law, Probability and Risk, vol. 6,
pp. 247–274.
Schum, D. A., Tecuci, G., and Boicu, M. (2009). Analyzing Evidence and Its Chain of Custody:
A Mixed-Initiative Computational Approach, International Journal of Intelligence and Counter-
intelligence, vol. 22, pp. 298–319. lac.gmu.edu/publications/2009/Schum%20et%20al%20-%
20Chain%20of%20Custody.pdf (accessed April 13, 2016)
Shafer, G. (1976). A Mathematical Theory of Evidence, Princeton University Press, Princeton, NJ.
Shafer, G. (1988). Combining AI and OR, University of Kansas School of Business Working Paper
No. 195, April.
Simon, H. (1983). Why Should Machines Learn? in Michalski, R. S., Carbonell, J. G., and Mitchell,
T.M. (eds.), Machine Learning, vol. 1, Morgan Kaufmann, Los Altos, CA, pp. 25–38.
Simonite, T. (2013). Bill Gates: Software Assistants Could Help Solve Global Problems, MIT Technol-
ogy Review, July 16. www.technologyreview.com/news/517171/bill-gates-software-assistants-
could-help-solve-global-problems/ (accessed April 13, 2016)
Siri (2011). Apple’s Siri homepage. www.apple.com/ios/siri/ (accessed April 12, 2016)
Strange, J. (1996). Centers of Gravity & Critical Vulnerabilities: Building on the Clausewitzian Foundation
so That We Can All Speak the Same Language, Marine Corps University Foundation, Quantico, VA.
Strange, J., and Iron, R. (2004a). Understanding Centers of Gravity and Critical Vulnerabilities, Part 1:
What Clausewitz (Really) Meant by Center of Gravity. www.au.af.mil/au/awc/awcgate/usmc/cog1.pdf
(accessed May 25, 2015)
Strange, J., and Iron, R. (2004b). Understanding Centers of Gravity and Critical Vulnerabilities, Part 2:
The CG-CC-CR-CV Construct: A Useful Tool to Understand and Analyze the Relationship between

13:56:23,
.015
440 References

Centers of Gravity and Their Critical Vulnerabilities. www.au.af.mil/au/awc/awcgate/usmc/cog2.pdf

(accessed May 25, 2015)
Tate, A. (1977). Generating Project Networks, in Proceedings of IJCAI-77, Boston, MA, Morgan
Kaufmann, San Francisco, CA, pp. 888–893.
Tecuci, G. (1988). Disciple: A Theory, Methodology and System for Learning Expert Knowledge,
Thèse de Docteur en Science, University of Paris-South. lac.gmu.edu/publications/1988/Tecu
ciG_PhD_Thesis.pdf (accessed April 13, 2016)
Tecuci, G. (1992). Automating Knowledge Acquisition as Extending, Updating, and Improving a
Knowledge Base, IEEE Transactions on Systems, Man and Cybernetics, vol. 22, pp. 1444–1460.
lac.gmu.edu/publications/1992/TecuciG_Automating_Knowledge_Acquisition.pdf (accessed April
13, 2016)
Tecuci, G. (1993). Plausible Justification Trees: A Framework for the Deep and Dynamic Integration
of Learning Strategies, Machine Learning Journal, vol. 11, pp. 237–261. lac.gmu.edu/publications/
1993/TecuciG_Plausible_Justification_Trees.pdf (accessed April 13, 2016)
Tecuci, G. (1998). Building Intelligent Agents: An Apprenticeship Multistrategy Learning Theory,
Methodology, Tool and Case Studies, Academic Press, London, UK. lac.gmu.edu/publications/
1998/TecuciG_Building_Intelligent_Agents/default.htm (accessed April 13, 2016)
Tecuci, G., and Keeling, H. (1999). Developing an Intelligent Educational Agent with Disciple,
International Journal of Artificial Intelligence in Education, vol. 10, no. 3–4. lac.gmu.edu/publica
tions/1999/TecuciG_Intelliget_Educational_Agent.pdf (accessed April 13, 2016)
Tecuci, G., and Kodratoff, Y. (1990). Apprenticeship Learning in Imperfect Theory Domains, in
Kodratoff, Y., and Michalski, R. S. (eds.), Machine Learning: An Artificial Intelligence Approach,
vol. 3, Morgan Kaufmann, San Mateo, CA, pp. 514–551. lac.gmu.edu/publications/data/1990/
apprenticeship_1990.pdf (accessed April 13, 2016)
Tecuci, G., and Kodratoff, Y. (eds.) (1995). Machine Learning and Knowledge Acquisition: Integrated
Approaches, Academic Press, London, UK. lac.gmu.edu/publications/1995/TecuciG_MLKA_Inte
grated_Approaches.pdf (accessed April 13, 2016)
Tecuci, G., and Michalski, R. S. (1991). A Method for Multistrategy Task-Adaptive Learning Based
on Plausible Justifications, in Birnbaum, L., and Collins, G. (eds.), Machine Learning: Proceedings
of the Eighth International Conference, June, Chicago, IL, Morgan Kaufmann, San Mateo,
CA, pp. 549–553. lac.gmu.edu/publications/1991/TecuciG_Multistrategy_Learning_Method.pdf
(accessed April 13, 2016)
Tecuci, G., Kedar, S., and Kodratoff, Y. (guest eds.) (1994). Knowledge Acquisition Journal, vol. 6, no. 2
(special issue on the integration of machine learning and knowledge acquisition), pp. 89–214.
Tecuci, G., Boicu, M., Wright, K., Lee, S. W., Marcu, D., and Bowman, M. (1999). An Integrated Shell
and Methodology for Rapid Development of Knowledge-based Agents, in Proceedings of the
Sixteenth National Conference on Artificial Intelligence (AAAI-99), July 18–22, Orlando, FL, AAAI
Press, Menlo Park, CA, pp. 250–257. lac.gmu.edu/publications/data/1999/ismrdkba.pdf (accessed
April 13, 2016)
Tecuci, G., Boicu, M., Wright, K., Lee, S. W., Marcu, D. and Bowman, M. (2000). A Tutoring Based
Approach to the Development of Intelligent Agents, in Teodorescu, H. N., Mlynek, D., Kandel, A.,
and Zimmermann, H. J. (eds.), Intelligent Systems and Interfaces, Kluwer Academic Press, Boston,
MA. lac.gmu.edu/publications/data/2000/2000_Disciple-Planning.pdf (accessed April 13, 2016)
Tecuci, G., Boicu, M., Bowman, M., and Marcu, D., with commentary by Burke, M. (2001). An
Innovative Application from the DARPA Knowledge Bases Programs: Rapid Development of a
Course of Action Critiquer, AI Magazine, vol. 22, no. 2, pp. 43–61. lac.gmu.edu/publications/2001/
TecuciG_Disciple_COA_IAAI.pdf (accessed April 13, 2016)
Tecuci, G., Boicu, M., Marcu, D., Stanescu, B., Boicu, C., Comello, J., Lopez, A., Donlon, J., and
Cleckner, W. (2002a). Development and Deployment of a Disciple Agent for Center of Gravity
Analysis, in Proceedings of the Eighteenth National Conference of Artificial Intelligence and the
Fourteenth Conference on Innovative Applications of Artificial Intelligence, AAAI-02/IAAI-02,
Edmonton, Alberta, Canada, AAAI Press/MIT Press, New York, NY, and Cambridge, MA,
pp. 853–860. lac.gmu.edu/publications/data/2002/dddacga.pdf (accessed April 13, 2016)
Tecuci, G., Boicu, M., Marcu, D., Stanescu, B., Boicu, C., and Comello, J. (2002b). Training and Using
Disciple Agents: A Case Study in the Military Center of Gravity Analysis Domain, AI Magazine,

13:56:23,
.015
References 441

vol. 24, no. 4, pp. 51–68. lac.gmu.edu/publications/2002/TecuciG_Disciple_COG_IAAI.pdf

(accessed April 13, 2016)
Tecuci, G., Boicu, M., Ayers, C., and Cammons, D. (2005a). Personal Cognitive Assistants for Military
Intelligence Analysis: Mixed-Initiative Learning, Tutoring, and Problem Solving, in Proceedings of
the First International Conference on Intelligence Analysis, May 2–6, McLean, VA, MITRE Corpor-
ation, Bedford, MA. lac.gmu.edu/publications/data/2005/Tecuci-Disciple-LTA.pdf (accessed April
13, 2016)
Tecuci, G., Boicu, M., Boicu, C., Marcu, D., Stanescu, B., and Barbulescu, M. (2005b). The Disciple-
RKF Learning and Reasoning Agent, Computational Intelligence, vol. 21, no. 4, pp. 462–479. lac.gmu
.edu/publications/2005/TecuciG_Disciple_RKF_CI.pdf (accessed April 13, 2016)
Tecuci, G., Boicu, M., and Cox, M. T. (guest eds.) (2007a). AI Magazine, vol. 28, no. 2 (special issue
on mixed-initiative assistants). www.aaai.org/ojs/index.php/aimagazine/issue/view/174/showToc
(accessed May 29, 2015)
Tecuci, G., Boicu, M., and Cox, M. T. (2007b). Seven Aspects of Mixed-Initiative Reasoning: An
Introduction to the Special Issue on Mixed-Initiative Assistants, AI Magazine, vol. 28, no. 2 (special
issue on mixed-initiative assistants), pp. 11–18. lac.gmu.edu/publications/2007/BoicuM_AIMaga
zine_Intro.pdf (accessed April 13, 2016)
Tecuci, G., Boicu, M., Marcu, D., Boicu, C., Barbulescu, M., Ayers, C., and Cammons, D. (2007c).
Cognitive Assistants for Analysts, Journal of Intelligence Community Research and Development
(JICRD). Also published in Auger, J., and Wimbish, W. (eds.) (2007). Proteus Futures Digest:
A Compilation of Selected Works Derived from the 2006 Proteus Workshop, 1st ed., Proteus
Management Group, Carlisle Barracks, PA, pp. 303–329. lac.gmu.edu/publications/2007/TecuciG_
Cognitive_Assistants.pdf (accessed April 13, 2016)
Tecuci, G., Boicu, M., Hajduk, T., Marcu, D., Barbulescu, M., Boicu, C., and Le, V. (2007d). A Tool for
Training and Assistance in Emergency Response Planning, in Proceedings of the Hawaii Inter-
national Conference on System Sciences, HICSS40, January 3–6, Hawaii, IEEE Computer Society
Press. lac.gmu.edu/publications/2007/Disciple-VPT%20Hawaii.pdf (accessed April 13, 2016)
Tecuci, G., Boicu, M., Marcu, D., Boicu, C., and Barbulescu, M. (2008a). Disciple-LTA: Learning,
Tutoring and Analytic Assistance, Journal of Intelligence Community Research and Development
(JICRD), July. lac.gmu.edu/publications/2008/Disciple-LTA08.pdf (accessed April 13, 2016)
Tecuci, G., Boicu, M., and Comello, J. (2008b). Agent-Assisted Center of Gravity Analysis, CD with
Disciple-COG and Lecture Notes used in courses at the U.S. Army War College and Air War
College. George Mason University Press, Fairfax, VA. lac.gmu.edu/cog-book/ (accessed April 13,
2016)
Tecuci, G., Boicu, M., Marcu, D., Barbulescu, M., Boicu, C., Le, V., and Hajduk, T. (2008c). Teaching
Virtual Experts for Multi-Domain Collaborative Planning, Journal of Software, vol. 3, no. 3
(March), pp. 38–59. lac.gmu.edu/publications/2008/TecuciG_Disciple_VE_JS.pdf (accessed
April 13, 2016)
Tecuci, G., Schum, D. A., Boicu, M., Marcu, D., and Hamilton, B. (2010a). Intelligence Analysis as
Agent-Assisted Discovery of Evidence, Hypotheses and Arguments, in Phillips-Wren, G., Jain, L.C.,
Nakamatsu, K., and Howlett, R.J., (eds.), Advances in Intelligent Decision Technologies, SIST 4,
Springer-Verlag, Berlin and Heidelberg Germany, pp. 1–10. lac.gmu.edu/publications/2010/
Tecuci-Discovery-in-motion-imagery.pdf (accessed April 13, 2016)
Tecuci, G., Schum, D. A., Boicu, M., Marcu, D., Hamilton, B., and Wible, B. (2010b). Teaching
Intelligence Analysis with TIACRITIS, American Intelligence Journal, vol. 28, no. 2 (December),
pp. 50–65. lac.gmu.edu/publications/2010/Tiacritis-AIJ.pdf (accessed April 13, 2016)
Tecuci, G., Marcu, D., Boicu, M., Schum, D. A., and Russell, K. (2011a). Computational Theory and
Cognitive Assistant for Intelligence Analysis, in Proceedings of the Sixth International Conference on
Semantic Technologies for Intelligence, Defense, and Security – STIDS, November 16–18, Fairfax, VA,
pp. 68–75. ceur-ws.org/Vol-808/STIDS2011_CR_T9_TecuciEtAl.pdf (accessed May 29, 2015)
Tecuci, G., Schum, D. A., Boicu, M., and Marcu, D. (2011b). Introduction to Intelligence Analysis:
A Hands-on Approach with TIACRITIS, 2nd ed., Learning Agents Center, George Mason University,
Fairfax, VA (1st ed., 2010).
Tecuci, G., Schum, D. A., Marcu, D., and Boicu, M. (2014). Computational Approach and Cognitive
Assistant for Evidence-based Reasoning in Intelligence Analysis, International Journal of Intelligent

13:56:23,
.015
442 References

Defence Support Systems, vol. 5, no. 2, pp. 146–172. lac.gmu.edu/publications/2014/Disciple-CD-

IJIDSS.pdf (accessed April 13, 2016)
Tecuci, G., Marcu, D., Boicu, M., and Schum, D. A. (2015). COGENT: Cognitive Agent for Cogent
Analysis, Proceedings of the 2015 AAAI Fall Symposium “Cognitive Assistance in Government and
Public Sector Applications, pp. 58–65, Arlington, VA, November 12–14. lac.gmu.edu/publications/
2015/Cogent-overview.pdf (accessed January 31, 2016)
Tecuci, G., Schum, D. A., Marcu, D., and Boicu, M. (2016). Intelligence Analysis as Discovery
of Evidence, Hypotheses, and Arguments: Connecting the Dots, Cambridge University Press,
New York, NY.
Toffler, A. (1984). Science and Change, foreword to Prigogine, Ilya, and Stengers, Isabelle, Order out
of Chaos: Man’s New Dialogue with Nature, Bantam, New York, NY, pp. xi–xxvi.
TopBraid Composer (2012). TopBraid Composer Ontology Development Tool. www.topquadrant.com/
products/TB_Composer.html (accessed August 4, 2012).
Toulmin, S. E. (1963). The Uses of Argument, Cambridge University Press, Cambridge, UK.
Turing, A. (1950). Computing Machinery and Intelligence, Mind, vol. 59, pp. 433–460.
Turoff, M. (2007). Design of Interactive Systems, in Emergency Management Information Systems
Tutorial, 40th Hawaii International Conference on System Sciences, HICSS-40, Hawaii, January 3.
UMLS (2012). Unified Medical Language System, U.S. National Library of Medicine. www.nlm.nih.gov/
research/umls/ (accessed August 4, 2012)
UNESCO (2005). Toward Knowledge Societies, unesdoc.unesco.org/images/0014/001418/141843e.pdf
(accessed October 1, 2011)
Van Gelder, T. J. (2007). The Rationale for Rationale, Law, Probability and Risk, vol. 6, pp. 23–42.
Van Melle, W., Scott, A. C., Bennett, J. S., and Peairs, M. (1981). The EMYCIN Manual, Report No.
HPP-81-16, Computer Science Department, Stanford University, Stanford, CA.
Veloso, M. (1994). Planning and Learning by Analogical Reasoning, Springer Verlag, Berlin, Germany.
W3C (2015). Semantic Web. www.w3.org/standards/semanticweb/ (accessed April 13, 2016)
Walton, D. (2004). Abductive Reasoning, University of Alabama Press, Tuscaloosa, AL.
Warden, J. A. III. (1993). Strategic Warfare: The Enemy as a System, in Mitchum, A. U. (ed.), Concepts
in Airpower for the Campaign Planner, Air Command and Staff College, Maxwell AFB, AL.
Waterman, D. A., and Hayes-Roth, F. (eds.) (1978). Pattern-Directed Inference Systems, Academic
Press, Orlando, FL.
Wigmore, J. H. (1913). The Problem of Proof, Illinois Law Review, vol. 8, no. 2, pp. 77–103.
Wigmore, J. H. (1937). The Science of Judicial Proof: As Given by Logic, Psychology, and General
Experience and Illustrated in Judicial Trials, 3rd ed., Little, Brown, Boston, MA.
Winston, P. H. (1980). Learning and Reasoning by Analogy, Communications of the ACM, vol. 23,
pp. 689–703.
WordNet (2012). WordNet: A Database for English, Princeton University, Princeton, NJ. wordnet.
princeton.edu/ (accessed August 4, 2012)
Zadeh, L. (1983). The Role of Fuzzy Logic in the Management of Uncertainty in Expert Systems, Fuzzy
Sets and Systems, vol. 11, pp. 199–227.

13:56:23,
.015
Index

abduction, 26 analogy-based generalization, 265–70

abductive inference 26 analysis
see abductive reasoning definition, 35
abductive learning, 223 analysis and synthesis, 113–24
abductive reasoning, 5–8, 26, 57 inquiry-driven, 113–22
Doyle’s interpretation, 8 for center of gravity determination, 116
Peirce’s interpretation, 7 for evidence-based reasoning, 122
abstract task, 399 problem solving, 113–15
abstraction symbolic integration, 113
pattern, 330 analytic task, 36
of probabilistic solution, 330 classification, 36
of reasoning, 329–31 critiquing, 36
of reasoning tree, 330–1 diagnosis, 36
of statement, 329–30 intelligence analysis, 36
access, 135 interpretation, 36
accuracy, 49 monitoring, 36
of demonstrative tangible evidence, 134 prediction, 36
action, 395 Anderson, Terence, 5
add effects, 395 Apple, 35
delete effects, 395 approximate reasoning, 21
example, 395 arch of knowledge, 5, 26
resources, 395 argument
Adam, Federic, 32 direction, 54
agent force, 54
cognitive, 31 Aristotle, 6
cognitive assistant, 33–4 artificial intelligence, 29–33
intelligent, 30, 34 definition, 29
knowledge-based, 30, 33–4 assumption, 18
agent development assumption-based reasoning, 143
ontology, 106 authenticity, 49
optimization, 107 of demonstrative tangible evidence, 134
rapid prototyping, 105 of real tangible evidence, 134
rule learning and ontology refinement, 106 authoritative record, 137
specification, 105 Awad, Elias M., 88, 178–9, 389, 418
use and personalization, 106
agent teaching and learning Bacon, Francis, 16
for center of gravity analysis, 380–3 Baconian inference
for course of action critiquing, 355–60 different tests of hypotheses, 16
for emergency response planning, 403–17 eliminating hypotheses, 16
for workaround planning, 345–6 Baconian probability system, 16–20, 55
Allemang, Dean, 33, 40, 89 compared with others, 23–5
Allen, James, 391, 425 intersection, 20
analogical problem solving negation, 20
based on explanation similarity, 265 union, 20
analogy criterion, 266–70 Bayes, Thomas, 10

447

13:58:13,
448 Index

Bayes’ rule, 10 representation, 203

and assumptions, 18 concept elicitation, 176–9
Bayesian network learning, 223 conditional probability, 11
Bayesian probability system, 11–13, 55 connecting the dots, 46–55
and assumptions, 18 definition, 47
compared with others, 23–5 example, 47
Belief Functions, 11–16, 55 Cooper, Thomas A., 88
and assumptions, 18 credibility, 135
compared with others, 23–5 versus believability, 135
believability, 61 critical reasoning, 48
versus credibility, 135 Cross, Stephen E., 34
of evidence, 48 customer, 55
example, 49 cyber insider threat, 64–7
of source, 48, 54 definition, 64
believability assessment, 133–9 Cyc, 88, 90
believing as if. See assumption
benefit of the doubt, 19 Dale, Andrew I., 10
Betham, Jeremy, 27 DARP, 34
BFO, 90 data, 1
big data problem, 50 David, F. N., 10
Boicu, Cristina, 41, 338 David, Paul A., 33
Boicu, Mihai, 39, 41, 338, 352, 364, 407 Davies, Todd R., 225
Bowman, Michael, 41, 338, 352, 396 decision trees induction, 223
Bresina, John L., 34 deduction, 25
Breuker, Joost, 35 deduction, induction, and abduction, 50
browsing of reasoning tree, 331 deductive inference 26
BRU generalization, 246 see deductive reasoning
Buchanan, Bruce G., 83–4, 87, 389, 418 deductive reasoning, 26
Deep Blue, 34
candidate elimination algorithm, 298 defensibile argument, 54
Carbonell, Jaime G., 225 DeJong, Gerald, 258, 406
card sort method, 179 DENDRAL, 34
case-based reasoning and learning, 223 Desai, Mita, 68
causal network of relations, 225, 265 design principle, 426–32
certification, 88 complete life cycle, 431
chain of custody, 138–9 employ learning technology, 426
chain of reasoning, 4 knowledge adaptation, 429
Chaudhri, Vinay K., 90 knowledge reuse, 427
Chipman, Michael A., 34 mixed-initiative, 429
Clancey, William, 35 multi-agent and multidomain problem solving,
class. See concept 427
clause generalization, 245 multistrategy learning, 428
Clausewitz, Carl von, 365 plausible reasoning, 430
CLIPS, 88 problem solving for collaboration, 427
clustering, 223 rapid development, 430
Cohen, Jonathan L., 16–18 teaching and learning, 428
Cohen, Morris R., 6 user tutoring, 430
Cohen, Paul, 346, 349 direct subconcept of
collaborative autonomous agents, 70 definition, 157
combinatorial explosion, 51 disbelief. See lack of belief
Comello, Jerome, 376 Disciple agent, 426
COMINT, 133–4 Disciple-CD, 35–6, 41
CommonKADS, 89 Disciple-COA, 36, 89, 348–64
competence, 3, 49, 135 Disciple-COG, 34, 91, 364–85, 427
competing hypotheses, 59 Script Editor, 377
computational theory of evidence-based Disciple-EBR, 41, 49, 53, 55, 91, 257, 309, 427
reasoning, 64 Association Browser, 165
concept Expression Editor, 282
definition, 156 Feature Browser, 167, 175, 189
as feature value, 163 Feature Editor, 175
formal representation language, 243–5 Hierarchical Browser, 165, 182

13:58:13,
Index 449

Object Browser, 165, 175, 180 in cybersecurity, 27

Object Editor, 192–3 in forensics, 27
Object Viewer, 165 in intelligence analysis, 27
Rule Browser, 254 in law, 27
Disciple-LTA, 36 in medicine, 27
Disciple-VE, 390–4 in natural sciences, 27
Reasoning Hierarchy Browser, 396 evidential synergism, 52
Reasoning Step Editor, 396 example, 50–1
Disciple-VPT, 37, 387–425, 427 Except-When plausible version space condition,
architecture, 388–9 215, 251, 254–5, 300
VE Assistant, 388 EXPECT, 88
VE Library, 388, 416–20 expert problem
VE Team, 388 center of gravity analysis, 364–7
Disciple-WA, 37, 89, 338–48 course of action critiquing, 348–51
divide-and-conquer. See analysis and synthesis emergency response planning, 389–90
DOLCE, 90 workaround planning, 338–41
domain understanding, 83–5, 94, 176–9 expert system, 33–5
dots definition, 33–4
evidential, 47–8 vs knowledge-based system, 34
combinations of, 50 expert system shell, 88–9
idea, 48, 50, 52, 54 definition, 88
Doyle, Arthur Conan, 6 explanation, 266
drill-down analysis, 143 with comparison, 283
Drucker, Peter, 33 with fixed value, 280
Durham, Susan, 391 with function, 280
Durkin, John, 34, 389, 418 generation, 262–4
Dybala, Tom, 40, 338 incompleteness, 263
explanation piece, 261
Echevarria, Antulio J., 365 association, 261
Eco, Umberto, 8 correlation, 261
Eikmeier, Dale C., 365 generalization, 261
Einstein, Albert, 21 property, 261
elicitation script, 376–9 relation, 261
for feature, 377 specific value, 261
illustration, 367–9 explanation-based learning, 223–5
EMYCIN, 89
enumerative probability, 9–11 fact, 2
aleatory, 9 failure explanation, 254
epistemology, 2 Farquhar, Adam, 90
equivocal testimonial evidence, 136 Feigenbaum, Edward A., 38
evaluation Fellbaum, Christiane D., 90
of Disciple-COA, 360–4 Ferrucci, David, 34
of Disciple-COG, 383–5 Filip, Florin G., 32–3
of Disciple-VPT, 422 flash of insight, 7
of Disciple-WA, 346–8 FOAF, 90
event, 2 Foray, Dominique, 33
evidence, 1, 5 Forbes, 34
ambiguity, 23 Forbus, Kenneth D., 225
dissonance, 23 foundational ontology. See upper ontology
imperfect believability, 23 frame, 15
incompleteness, 23 frame of discernment, 15
inconclusiveness, 23 Friedman-Hill, Ernest, 90
versus Fuzzy probability system, 20–3, 55
event, 48 compared with others, 23–5
fact, 2 conjunction, 22
information, 2 disjunction, 22
knowledge, 2 negation, 22
evidence collection task, 60
evidence in search of hypotheses, 56–7 Galilei, Galileo, 6
evidence-based assessment, 122–4 Galvin, T. P., 365
evidence-based reasoning, 25–9, 46–76 game of chance, 9

13:58:13,
450 Index

Gammack, J.G., 176, 180 define feature-specific names, 196

Ganascia, Jean-Gabriel, 237, 265, 298 define similar siblings, 186
Gates, Bill, 40 group similar siblings, 187
generalization, 18 naming conventions, 188
of concepts with negation, 247 n-ary relations, 196
definition, 234 single subconcepts, 187
formal definition, 243–7 planning
inductive, 18 goal specification, 400
maximal, 266–7, 300 plausible task ordering, 398
minimal, 268, 298, 300 preconditions specification, 399
definition, 234 top-down, left-to-right specification, 399
of two concepts, 236 rule and hypothesis learning
of two concepts avoid learning from overly specific
definition, 236 examples, 286
generalization hierarchy, 157–8 avoid learning without explanations, 288
generalization rule, 229–33 domains and ranges of features, 286
climbing generalization hierarchies, 231 extend ontology before learning, 286
definition, 229 guide explanation generation, 288
dropping conditions, 231 identify entities before learning, 285
extending discrete sets, 232 recognize concepts in tree, 288
extending intervals, 231 rule refinement
extending ordered sets of intervals, 231 assess similar hypotheses, 321
extending symbolic probabilities, 232 extend ontology for failure explanations, 321
turning constants into variables, 230 Guizzardi, G., 90
turning occurrences of a variable into different
variables, 230 hands-on exercise
using feature definitions, 233 abstraction of reasoning, 331–4
using inference rules, 233 argumentation browsing, 76–81
using substitutions, 247 believability analysis, 140–3
generic instance, 104, 183, 188 evidence search, 130–3
and explanation selection, 263 feature hierarchy development, 189–92
genetic algorithms and evolutionary hypothesis analysis, 124–30
computation, 223 instance definition, 192–5
Gentner, Dendre, 225 modeling and formalization, 144–6
Geonames, 90 object hierarchy development, 180–6
GFO, 90 ontology browsing, 165–8
Ghallab, Malik, 387, 394, 396, 425 pattern reuse, 146–7
Ghaziri, Hassan M., 389, 418 rule and hypothesis learning, 275–9
Giarratano, Joseph, 34, 88 rule refinement, 319–21
Gil, Yolanda, 90, 360 Hayes-Roth, Frederick, 202
Giles, P. K., 365 Hendler, Jim, 33, 40, 89
giving benefit of the doubt. See assumption heuristic, 62
goal, 396 Hieb, Michael R., 41, 338
Goodman, David, 34 hierarchical task network planning, 394–403
Gruber, Tom R., 155 integration with inference, 403
guideline Hobbs, Jerry R., 90
abstraction Holmes, Sherlock, 6, 8, 47, 51
define context-dependent names, 334 Horvitz, Eric, 32
knowledge base HUMINT, 48–9, 139
one KB in memory, 111 Humphreys, B.L., 90
save succesive KB versions, 111 hypothesis
modeling examples, 29
define reduction considering future learning, 271–6
changes, 149 basic steps, 273
identify instances and constants, 148 refinement, 316–17
learn and reuse patterns, 149 testing, 59
reduction tree development, 148 hypothesis-driven evidence collection, 59
use specification to structure modeling, 147
ontology development IBM, 34
categories as concepts, 195 imaginative reasoning, 48, 57
concept naming, 189 IMINT, 48, 133–4

13:58:13,
Index 451

Incident Command System, 390 knowledge base

incremental inductive generalization, 299 development, 257
indicator Domain KB, 104, 270, 420, 428
almost certain indicator, 121 maintenance, 257
likely indicator, 121 Scenario KB, 104, 270, 420, 428
probability computation, 121 Shared KB, 104, 420, 428
very likely indicator, 121 State KB, 420
individual. See instance knowledge engineer, 256, 426
induction, 25 definition, 83
inductive inference 26 knowledge engineering, 33–41, 426
see inductive reasoning definition, 33
inductive learning from examples, 223–4, 238–42 knowledge society, 40
problem definition, 239 Kodratoff, Yves, 39, 237, 265, 298
inductive reasoning, 26 Kolmogorov, Andrey N., 10, 14
inference Kolmogorov’s axioms of probability, 10
action, 401 Krogh, Georg von., 39
engine, 207
network, 53 lack of belief, 15
task, 401 Langley, Patrick W., 31
inferential catastrophes, 49 Laplace, Pierre, 9
inferential force or weight, 61 law, 5
information, 2 Anglo-American, 5
inheritance, 162–3 laws of large numbers, 10
default, 162 Le, Vu, 41, 338
multiple, 162 learner
input variable, 408 aggressive, 241, 266
inquiry-based learning in science, 70–6 cautious, 240, 268
instance dual-strategy, 241
definition, 157 learning
instance of, 157 with evolving ontology, 309–20
instance-based learning, 223 with incomplete representation language,
integrated teaching and learning, 39 242
intelligence analysis, 56–64 learning agent shell, 90–1, 390
intelligent agent, 30–2 definition, 90
interview with expert learning agent shell for evidence-based reasoning,
structured, 177 91–3
unstructured, 84, 177 definition, 91
dichotomous questions, 177 learning apprentice, 223
multiple-choice questions, 177 learning by analogy, 39, 223, 225–6
ranking scale questions, 177 learning from examples, 39
Intuit, 34 learning from explanations, 39
Iron, Richard, 365 learning planning rules, 417
isa. See subconcept of action concretion rules, 415–17
correlated rules, 409–13
Jackson, Peter, 389, 418 difficulty, 409
Java, 41 method, 413
Jena, 89–90 problem definition, 413
JESS, 89 task concretion rules, 414–15
Jones, Eric, 338, 347, 349–50, 361–2 task reduction rules, 413–14
Lempert, R. O., 134
Kant, Immanuel, 35 Lenat, Douglas B., 90, 360
Keeling, Harry, 39, 41, 338 likelihood
Keene, Raymond, 34 versus likeliness, 122
Kent, Sherman, 23 likelihood ratio, 12
Kim, Jihie, 360 likeliness
Kipling, Rudyard, 113, 427 versus likelihood, 122
Kneale, W., 6 Lindberg, D. A. B., 90
knowledge Locke, John, 6
as justified true belief, 2–4 Loom, 90, 343
repository, 104–5 lower bound condition
reuse, 418, 427 as analogy criterion, 268–70

13:58:13,
452 Index

MacGregor, Robert, 90, 343, 360 Norvig, Peter, 30

machine learning Noy, Natalya F., 174
definition, 223
main plausible version space condition, 215, 251 O’Keefe, R. M., 88
MAPGEN, 34 object feature, 158–60
Marcu, Dorin, 41, 338 domain, 158
Marcus, Sandra, 37, 89 generalization hierarchy, 160
MASINT, 48, 133 partially learned, 159
Masolo, C., 90 range, 158
McDermott, John, 34 objectivity, 4, 49, 135
McGuinness, Deborah L., 174 Obrst, Leo, 90, 174
Meckl, Steven, 29 observational sensitivity, 49, 135
Michalski, Ryszard S., 39, 222, 227 OCHRE, 90
Minsky, Marvin, 222 odds, 10
missing evidence, 137 OKBC, 90
versus negative evidence, 137 Oldroyd, David, 5, 26
Mitchell, Tom M., 212, 222, 238, 258, 298, 318, 407 Ontolingua, 90
mixed evidence, 14, 138 ontological commitment, 155
example, 138 ontology, 155–65
mixed-initiative definition, 155
interaction, 252 of evidence, 134
problem solving, 39 maintenance, 197
reasoning, 32–3, 262 matching, 163–5
definition, 32 of problem-solving tasks, 35–7
in explanation generation, 264 version, 311
modeling the problem-solving process, 113–22 ontology design and development, 174–81
center of gravity analysis, 369–74 for center of gravity analysis, 376–8
course of action critiquing, 351–74 for course of action critiquing, 352–5
emergency response planning, 394–6 for emergency response planning, 357–90
workaround planning, 341–3 for workaround planning, 343–4
modeling-based ontology specification, 100–1, operation with assumption
179–81, 286 define, 146
Mooney, Ray, 258, 406 define with justification, 146
more general than delete, 146
definition, 157 operation with case study
Morris, Jon, 25 end, 80
Morris, Paul H., 34 run, 78
multi-INT fusion, 61 operation with explanation
multistrategy learning, 39, 223, 226–7 define with comparison, 284
Murphy, Peter, 5 define with function, 283
Musen, Mark A., 89 generate with fixed value, 280
MYCIN, 34 guide generation, 279
operation with hypothesis
Nagle, Ernest, 6 associate search criteria, 132
naïve Bayes learning, 223 browse analysis of hypothesis, 78
narrative, 55 convert between modeled and formalized, 146
n-ary feature, 160–1 define explanations, 275
NASA, 34 display with problem viewer, 279
National Science Education Standards, 70 insert intermediary hypothesis, 145
Nau, Dana, 366, 425 introduce into abstract reasoning tree, 334
negative evidence, 136 modify abstraction, 334
versus missing evidence, 137 move hypotheses, 145
negative example, 223 remove from abstract reasoning tree, 334
definition, 227 specify, 145
negative exception specify by instantiating pattern, 147
definition, 228 specify reduction by reusing pattern, 147
Negoita, Constantin V., 21 update name of elementary hypothesis, 129
neural networks, 223 operation with item of evidence
Newton, Isaac, 6 assess, 127
Nilsson, Nils J., 391 associate to hypothesis, 127
Nonaka, Ikujiro, 39 define, 143

13:58:13,
Index 453

operation with knowledge base positive example, 223

close, 110 definition, 227
create, 110 positive exception
load and select, 108 definition, 227
save, 109 possibility function, 21
operation with ontology posterior odds, 12
add direct subconcept or instance, 194 posterior probability, 11
add direct superconcept, 194 Powell, Gerald M., 391
change feature domain, 190 prior probability, 11
change feature range, 190 probabilistic reasoning, 25
define concept or instance, 182 probabilistically equivocal testimonial
define feature, 189, 193 evidence, 136
define generic instance, 183 probability scale
define specific instance, 182 Baconian, 17
delete concept or instance, 186 conventional, 15
rename concept or instance, 186 degree of support or belief, 15
view feature definition, 167 problem reduction rule, 204
view object, 167 production system architecture, 202–3
operation with reasoning tree propose-and-revise, 89
define instances and constants, 145 Protégé, 89–90
delete node, 145 protocol analysis, 178
recognize concepts, 289 pure evidence, 13–14
replace generated step with modeling step, 320
select synthesis function, 129 R1, 34
specify node, 145 Ralescu, Dan A., 21
operation with rule rapid prototyping, 93–100
delete with Rule Browser, 321 Rationale system, 23
display with Rule Viewer, 278 reduction
learning steps, 278 indicator, 121
refine with negative example, 320 necessary and sufficient condition, 120
refine with positive example, 319 scenario, 120
view with Rule Browser, 320 sufficient condition, 120
OPS, 88 reduction example
ordered symbolic probability scale, 61 generalized, 270–1
output variable, 408 reformulation, 264–5
overgeneralization, 242 understanding, 260–2
OWLIM, 90 reduction rule, 204–7, 394
reformulation rule
Pan, Feng, 90 definition, 229
Paris, Cecile, 88 relative frequency, 9
partially learned knowledge, 212–16 relevance, 53, 61
reasoning, 215–16 argument, 4, 54
representation, 212–15 of event, 50
concepts, 212–13 reliability
features, 159, 213 of demonstrative tangible evidence, 134
hypotheses, 214 repeatable process, 9
rules, 214 report generation, 374, 430
pattern learning, 95–8 requirements specification, 85, 93
versus rule learning, 102 Ressler, J., 90
Peirce, Charles S., 6–8 Riley, Gary, 34, 88
Pellet, 89–90 Rooney, David, 32
personalized learning, 430 Roosevelt, Franklin D., 21
planning, 387 rote learning, 223
collaborative, 420–1 rule analysis, 270
multidomain, 420–1 rule and ontology matching, 207–12
plausible explanation, 262 rule induction, 223
plausible lower bound condition, 298, 406 rule learning, 252–71
plausible upper bound condition, 405 basic steps, 260
plausible version space, 213, 300 features, 317
Plotkin, G. D., 237 illustration, 253–7
Pomerol, Jean-Charles, 32 for inference tasks, 403–9

13:58:13,
454 Index

rule learning (cont.) subconcept of, 157

method overview, 258–60 subject matter expert, 83, 256, 426
versus pattern learning, 102 Subjective Bayesian. See Bayesian probability
problem definition, 257–8 system
rule refinement, 294–309 substance-blind, 133
through condition specialization SUMO, 90
illustration, 305–8 support vector machines, 223
method, 307 synthesis
with Except-When conditions definition, 35
basic steps, 305 problem-level, 116
illustration, 300–5 reduction-level, 116
features, 317 synthesis function
illustration, 253–7 almost certain indicator, 207
method overview, 295–6 likely indicator, 207
with negative examples, 300–9 max, 120, 207
summary, 308 min, 120, 207
with positive examples, 296–300 on balance, 124, 207
illustration, 296–9 very likely indicator, 207
method, 298–300 synthesis rule, 204–7, 394
summary, 300 synthetic task, 36–7
problem definition, 294 assignment, 37
rule regeneration debugging, 37
on demand, 310–12 design, 37
illustration, 312–16 planning, 37
method, 316–20 scheduling, 37
problem definition, 309–11
Russell, Stuart J., 30, 225 tacit knowledge, 39
taking for granted. See assumption
SALT, 89 tangible evidence, 48, 133–5
Schmidt, C.F., 391 types, 134
Schneider, Luc, 90 task, 396
Schreiber, Guus, 33, 35–6, 89 decomposition, 394
Schum, David A., 6, 11, 16, 20, 22, 25, 29, 46, 55, 61, goal, 400
133, 135, 138 specialization, 394
Sebeok, Thomas, 8 Tate, Austin, 391, 424
Semantic Web, 40, 90, 174 TECHINT, 133
Shafer, Glenn, 13–14, 23 Tecuci, Gheorghe, 29–30, 32–4, 36–7, 39–40, 64,
Shortliffe, Edward H., 34 91–2, 116, 148, 155, 171–2, 212, 227, 237,
siblings, 186–7 262, 338, 346, 360, 362, 374, 384, 387, 391,
definition, 186 420–1, 430
SIGINT, 48, 133–4 term generalization, 245
Simon, Herbert, 222 testimonial evidence, 48, 135–7
Simonite, T., 40 based on opinion, 136
Siri, 35 completely equivocal, 136
software assistant, 40 obtained at second hand, 136
solution synthesis rule unequivocal based upon direct observation, 136
problem-level, 206–7 testing, 256–7
reduction-level, 205, 207 TIACRITIS, 35–6, 41
specialization Toffler, Alvin, 120, 427
minimal TopBraid Composer, 89–90
definition, 235 Toulmin, Stephen E., 23
of two concepts, 237 transitivity, 161
of two concepts trifles, 47
definition, 237 truthfulness. See veracity
specialization rule, 229–34 Turing, Alan, 39
definition, 229 Turoff, Murray, 32
specific instance, 104, 183, 188 type. See subconcept of
statistical process, 9
Stavrou, Angelos, 64 U.S. National Academy of Engineering, 40
Strange, Joseph, 365, 369 U.S. National Incident Management System, 390
subclass of. See subconcept of U.S. National Research Council, 71–2, 75–6

13:58:13,
Index 455

U.S. National Response Plan, 387 W3C, 33, 40

UFO, 90 Wagner, G., 90
UMLS, 90 Walker, E., 34
underspecialization, 242 Walton, Douglas, 43
understandability, 135 Warden III, John A., 365
unequivocal testimonial evidence, 136 Waterman, Donald A., 202
UNESCO, 33 Watson, 34
universal ontology. See upper ontology weight of evidence
upper bound condition as degree of support, 13
as analogy criterion, 266–8 what-if scenario, 143
upper ontology, 90 Whewell, William, 6
user instance, 176 wide-area motion imagery, 68–70
utility ontology, 90 forensic analysis, 68–9
real-time analysis, 68
validation, 88 Wielinga, Bob, 35
Van Gelder, Tim, 23 Wigmore, John H., 23, 53
Van Melle, W., 89 Wilkins, David C., 389, 418
Veloso, Manuela, 225 Winston, Patrick H., 225, 265
veracity, 4, 49, 135 Wogrin, Nancy, 88
verbal assessments of probability, 22–3 WordNet, 90
versus numerical, 22 world state
verification, 88 definition, 395
virtual planning expert
definition, 387 Zadeh, Lofti, 8, 20
illustration, 422

13:58:13,