"Ar ficial Intelligence (AI) Chapter 1"

Course Overview
The course focuses on intelligent agents and various AI-related concepts such
as:
Search and game-playing

Logical and planning systems

Probability and decision theory

Machine learning

Natural language processing

Robo cs

Philosophical issues in AI
Defini ons of AI
The document presents mul ple defini ons of Ar ficial Intelligence from different
researchers:
Bellman (1978): AI automates human-like thinking processes such as decision-making, problem-
solving, and learning.

Rich & Knight (1991): AI focuses on making computers perform tasks that humans currently do
be er.

Charniak & McDermo (1985): AI studies mental facul es through computa onal models.

Luger & Stubblefield (1993): AI is about automa ng intelligent behavior in computers.

AI is categorized into four main approaches:

1. Thinking humanly – Studying human cogni ve processes.

2. Thinking ra onally – Using logic to make correct decisions.

3. Ac ng humanly – Mimicking human-like responses.

4. Ac ng ra onally – Making the best possible decision in a situa on.

The Turing Test

Proposed by Alan Turing (1950) in his paper Compu ng Machinery and Intelligence.

Instead of asking, "Can machines think?", he reframed it as "Can machines behave intelligently?"

The Turing Test involves a human interrogator trying to dis nguish between a human and a machine
based on their responses. If the machine can successfully imitate human-like responses, it is
considered intelligent.
 "Ac ng Humanly – The Turing Test"

The Turing Test, a concept proposed by Alan Turing in his 1950 paper "Compu ng Machinery and
Intelligence."

1. The Core Idea of the Turing Test

Instead of directly asking “Can machines think?”, Turing reframes the ques on as “Can machines
behave intelligently?”
He introduces the Imita on Game, in which a human interrogator interacts with both a human and a
machine through a text-based conversa on.

If the interrogator cannot reliably dis nguish between the human and the machine, the machine is
considered intelligent.

2. Turing’s Predic ons (for the Year 2000)

Turing predicted that by 2000, a machine might have a 30% chance of fooling a layperson into
believing it is human in a 5-minute conversa on.

Although AI has advanced significantly, passing the Turing Test in a general sense remains a
challenge.

3. An cipa on of Arguments Against AI

Turing foresaw major objec ons that people would raise against AI for the next 50 years.

These include concerns about consciousness, emo ons, crea vity, and whether machines truly
"understand" what they do.

4. Key Components of AI (According to Turing)

Turing suggested that AI should include:

Knowledge (having informa on about the world)

Reasoning (making logical decisions)

Language Understanding (processing human language)

Learning (improving through experience)

5. Problems with the Turing Test

The Turing Test is not reproducible (results vary depending on the interrogator).

It is not construc ve, meaning it does not provide a way to build an intelligent machine.

It is not mathema cally analyzable, making it difficult to measure intelligence precisely.

Conclusion
Turing’s contribu ons to AI, his predic ons, and the limita ons of the Turing Test as a measure of
intelligence.
 "Thinking Humanly – Cogni ve Science"
How cogni ve science relates to Ar ficial Intelligence (AI) and how AI research tries to model human
thinking.

1. Cogni ve Revolu on (1960s)

Before the 1960s, behaviorism was the dominant psychological theory, which focused only on
observable behavior.
The cogni ve revolu on shi ed the focus to informa on-processing psychology, which studies how
the human brain processes informa on like a computer.

This change helped shape AI by inspiring researchers to understand and replicate human thought
processes.

2. Studying the Internal Ac vi es of the Brain

AI researchers and cogni ve scien sts want to understand how the brain processes knowledge and
makes decisions.

The key ques on: At what level should we study the brain?

Should we focus on high-level knowledge representa on (like concepts and ideas)?

Or should we study the low-level structure of the brain (neurons and circuits)?

3. Two Main Approaches to Understanding Human Thinking

There are two ways to study human thought for AI development:

1. Top-Down Approach (Cogni ve Science)

 Focuses on human behavior and thinking pa erns.

Researchers observe, predict, and test how humans behave in different situa ons.

AI can then be designed to mimic these behaviors.

2. Bo om-Up Approach (Cogni ve Neuroscience)

Studies the brain’s physical structure and func ons.

Uses neurological data (brain scans, neuron ac vity, etc.) to understand thought processes.

AI is built by simula ng neural ac vity (like in ar ficial neural networks).

4. Difference Between Cogni ve Science and AI

Cogni ve Science tries to understand human thinking and behavior.

Ar ficial Intelligence (AI) aims to replicate intelligent behavior, whether it works like a human brain
or not.

Over me, Cogni ve Science and AI have become separate fields, but they s ll influence each other.
Conclusion
The historical shi from behaviorism to cogni ve science and how it influenced AI. It also introduces
two ways of modeling intelligence—one based on behavior and the other on neuroscience. AI
con nues to evolve by taking inspira on from both approaches.

"Thinking Ra onally – Laws of Thought"

How logic and reasoning have influenced the development of Ar ficial Intelligence (AI). It focuses on
the laws of thought, which aim to establish rules for correct thinking.

1. Thinking Ra onally – What Does It Mean?

 Norma ve vs. Descrip ve Thinking:
 AI can be designed to follow norma ve (prescrip ve) rules, meaning it should think in a
logically correct way, rather than just copying how humans actually think (descrip ve
thinking).

 Aristotle and Logical Reasoning:

Aristotle was one of the first to study logic formally, asking:

"What are the correct ways to think and argue?"

He and other Greek philosophers developed formal logic, which sets rules for valid reasoning
and argumenta on.

 Early Logic and AI:

Greek philosophers created logical rules and nota on for reasoning.

Their work led to mathema cal logic, which forms the founda on of modern AI and
computer science.

2. Problems with the Logic-Based Approach

Although logical reasoning is useful for AI, it has limita ons:

 Not all intelligent behavior follows strict logic.

Humans o en use intui on, emo ons, and crea vity, which cannot always be expressed
in formal logic.

AI that relies only on logic may struggle with real-world uncertainty and human-like problem-solving.

 What is the purpose of thinking?

Even if AI can follow logical rules, how does it decide which thoughts are important?

Should AI analyze everything or focus only on solving problems efficiently?

Human intelligence involves priori za on and decision-making, which is not always purely
logical.

3. Connec on to Modern AI
Logical reasoning plays an important role in AI, especially in areas like mathema cal proofs, expert
systems, and automated decision-making.
However, modern AI also incorporates machine learning and probabilis c reasoning, which do not
rely strictly on formal logic.

AI researchers have found that a balance between logic and other forms of intelligence (such as
learning from experience) is necessary for crea ng truly intelligent systems.

Conclusion
How formal logic shaped AI development but also highlights its limita ons. While AI systems can use
logic to make decisions, real-world intelligence requires more than just following strict rules—it also
needs flexibility, learning, and adaptability.

"Ac ng Ra onally"
The principle of ra onality in AI, which means that AI systems should act in a way that maximizes
goal achievement based on the available informa on.

1. What Does "Ac ng Ra onally" Mean?

Ra onal behavior means doing the right thing in a given situa on.

The right thing is defined as the ac on that maximizes success based on what is known.

AI systems are designed to act ra onally rather than just mimicking human thought processes.

2. Ra onality vs. Thinking

Ra onal ac on does not always require thinking.

Example: A blinking reflex is a ra onal ac on to protect the eyes but does not involve complex
thought.

However, thinking should serve ra onal ac on when needed.

AI should use reasoning only when it helps achieve be er results.

3. Aristotle’s Perspec ve on Ra onality

Aristotle, in Nicomachean Ethics, argued that every ac on aims toward some good.

This idea connects to AI, where ra onal agents are programmed to pursue beneficial outcomes
based on defined goals.
4. Connec on to AI
AI systems, such as intelligent agents, are designed to act ra onally by:

Gathering informa on about the environment.

Making decisions based on that informa on.

Taking ac ons that maximize success.

Ra onal agents are widely used in AI, including:

Self-driving cars (deciding the best route).

Recommenda on systems (sugges ng movies, products, etc.).

Game-playing AI (choosing the best move to win).

Conclusion
AI should focus on ra onal ac ons rather than just mimicking human thought. Ra onality in AI
means maximizing success based on available knowledge, and thinking should be used only when
necessary to achieve that goal.
 "Ra onal Agents"
The concept of agents in AI, par cularly ra onal agents, which are AI systems designed to make the
best possible decisions based on their environment.

1. What is an Agent?
An agent is any en ty that perceives informa on from its environment and takes ac ons based on
that informa on.

In AI, an agent can be:

A robot naviga ng an environment.

A so ware program making stock market predic ons.

A self-driving car responding to traffic signals.

2. What is a Ra onal Agent?

A ra onal agent is an agent that chooses the best ac on based on its knowledge to achieve a goal.

This course (as men oned in the slide) is about designing ra onal agents that can:

Perceive their surroundings.

Make decisions based on past and present observa ons.

Take the most effec ve ac on to achieve their objec ve.

3. Mathema cal Representa on of an Agent
An agent is represented as a func on that maps percept histories (everything the agent has sensed
so far) to ac ons.

Mathema cally:

f : P* A
P* = The set of all possible percept histories (everything the agent has seen/heard/sensed).

A = The set of possible ac ons the agent can take.

This means that for any given history of inputs, the agent decides on the best ac on to take next.

4. Goal of AI: Finding the Best Agent

For any specific task or environment, AI researchers try to design an agent that has the highest
performance.

Example:
In a chess game, an AI agent should pick the move that gives the best chance of winning.

In an autonomous vehicle, the AI should choose ac ons that ensure safe driving while reaching the
des na on efficiently.

5. The Challenge: Computa onal Limita ons

Perfect ra onality is not possible because:

Real-world problems are too complex.

Computers have limited processing power.

Time constraints make it impossible to evaluate all possible ac ons.

Instead of perfect ra onality, AI focuses on bounded ra onality, meaning it designs the best possible
program given the available machine resources.

Conclusion
AI is about designing ra onal agents that make the best possible decisions based on what they
perceive. While perfect ra onality is impossible, AI focuses on making agents that perform well
within real-world constraints.

"AI Prehistory"
The historical founda ons of Ar ficial Intelligence (AI) and how different fields have influenced its
development.
1. AI's Roots in Mul ple Disciplines
AI did not emerge from a single field; instead, it evolved through contribu ons from various
disciplines:

A. Philosophy (Logic & Reasoning)

Key Idea: How do humans think? Can thinking be represented formally?

Philosophers like Aristotle and Descartes explored logic and ra onality, forming the basis of AI
reasoning.

They viewed the mind as a physical system that follows rules of logic.

B. Mathema cs (Formal Representa on & Algorithms)

Key Idea: Can intelligence be represented mathema cally?

AI relies on formal logic for decision-making and algorithms to process informa on.

Key mathema cal concepts in AI:

Computa on Theory – What problems can be solved by a machine?

Decidability & Complexity – Can a problem be solved within a reasonable me?

Probability – How should AI handle uncertainty?

C. Psychology (Learning & Cogni on)

Key Idea: How do humans learn and think?

AI studies learning, percep on, and decision-making, inspired by human cogni on.

Experimental psychology (such as psychophysics) helps in designing AI models that mimic human
intelligence.

D. Linguis cs (Language Processing)

Key Idea: Can AI understand and generate language?

AI applies grammar rules and syntax to process human language.

This led to Natural Language Processing (NLP), used in AI chatbots, voice assistants, etc.

E. Neuroscience (Brain Structure & Func on)

Key Idea: How does the brain process informa on?

AI takes inspira on from neurons and brain structures to build neural networks.

Modern AI, like deep learning, is based on principles of brain ac vity and learning mechanisms.

F. Control Theory (Automa on & Stability)

Key Idea: How can AI control systems op mally?

AI applies control theory to robo cs, self-driving cars, and automated systems.

Focuses on stability and homeostasis (keeping systems balanced and adap ve).

Conclusion
AI has evolved from many different disciplines:

Philosophy gave AI its logical founda ons.

Mathema cs provided algorithms and formal reasoning.

Psychology & Neuroscience helped AI understand learning and cogni on.

Linguis cs contributed to AI's ability to process language.

Control Theory introduced stability and automa on principles.

This interdisciplinary approach is what makes AI such a powerful and evolving field today.

"A Brief History of AI"

A meline of key events in the history of Ar ficial Intelligence (AI), highligh ng breakthroughs,
setbacks, and major trends in AI research.
1. Early AI Founda ons (1940s – 1950s)
1943 – McCulloch & Pi s:

Developed a Boolean circuit model of the brain, which was the first a empt to create ar ficial
neurons.

This work laid the founda on for neural networks and modern AI.

1950 – Alan Turing’s Paper:

In Compu ng Machinery and Intelligence, Turing introduced the Turing Test to determine if a
machine can think like a human.

This paper inspired much of early AI research.

2. Early AI Programs & The Birth of AI (1952 – 1956)

1952-1969 – Early AI Experiments:

AI began to show promise in solving problems without human interven on (e.g., early robo cs).

The first AI programs were created, such as:

Samuel’s Checkers Program – A self-learning AI that played checkers.

Newell & Simon’s Logic Theorist – An AI that could prove mathema cal theorems.

Gelernter’s Geometry Engine – An AI system that could solve geometric problems.

1956 – The Dartmouth Conference:

This was the official birth of Ar ficial Intelligence as a field.

The term "Ar ficial Intelligence" was coined.

Researchers believed AI could be solved within a few decades, which turned out to be overly
op mis c.

3. AI’s First Challenges (1965 – 1979)

1965 – Robinson’s Logical Reasoning Algorithm:

A breakthrough in logical reasoning that helped AI solve complex problems using formal logic.

1966-1974 – The Discovery of Computa onal Complexity:

AI researchers realized that many problems were computa onally hard (NP-hard) and could not be
solved efficiently.

Interest in neural networks declined because computers at the me were too weak to support them.

1969-1979 – Knowledge-Based Systems Development:

AI researchers shi ed focus to expert systems, which used stored knowledge to solve problems (e.g.,
medical diagnosis).

4. The Rise and Fall of Expert Systems (1980 – 1993)

1980-1988 – The Boom of Expert Systems:

AI became commercially successful with expert systems, used in industries such as healthcare and
finance.

Large companies invested heavily in AI.

1988-1993 – The "AI Winter":

Many expert systems failed due to high costs and maintenance issues.
AI funding dropped sharply, leading to a period called the "AI Winter", where research slowed down.

5. AI’s Resurgence & Modern Era (1985 – Present)

1985-1995 – Neural Networks Make a Comeback:

Researchers rediscovered ar ficial neural networks, which became more powerful with be er
hardware.

1988 – The Rise of Probabilis c AI:

AI started using probability and decision-theore c methods for handling uncertainty.

This was important for areas like speech recogni on and robo cs.

1990s-Present – AI Becomes More Technical & Diverse:

AI research became more complex and technical, leading to advancements in:

Gene c Algorithms (GAs) – AI that evolves solu ons like natural selec on.

Ar ficial Life (ALife) – Simula ng life-like behaviors in AI.

So Compu ng – Using fuzzy logic and probabilis c methods for AI decision-making.

This led to modern AI breakthroughs, including deep learning, self-driving cars, and advanced
robo cs.

Conclusion
The history of AI has had periods of rapid growth and setbacks. While early AI efforts were overly
op mis c, modern AI has become more prac cal, data-driven, and powerful, leading to the
intelligent systems we use today.

"State of the Art in AI"

The capabili es and limita ons of AI at the me the slides were created (1998). It presents a list of
tasks and ques ons whether AI could perform them at that me.
1. What AI Could Do in 1998?
The following tasks were par ally possible or under research in 1998:

Play a decent game of table tennis

AI could control robo c arms to play simple games.

However, real- me reflex-based sports were s ll a challenge.

Drive along a curving mountain road

Some early autonomous vehicle research had basic self-driving capabili es.

AI could follow predefined routes but struggled with unpredictable environments.

Play a decent game of bridge

AI had already defeated humans in chess (Deep Blue vs. Kasparov, 1997).

Bridge, however, involves hidden informa on and teamwork, making it harder for AI.

Discover and prove a new mathema cal theorem

AI had been used to verify mathema cal proofs.

Some systems, like automated theorem provers, could generate proofs.

Give competent legal advice in a specialized area of law

AI-assisted legal research was emerging.

Expert systems could suggest legal ac ons in narrow domains.

Translate spoken English into spoken Swedish in real me

Basic speech-to-text and transla on models existed but were not fluent.

Real- me spoken transla on was a major challenge.

2. What AI Could NOT Do in 1998?

These tasks were not possible for AI at that me:

Drive in the center of Cairo

AI struggled with dense, unpredictable traffic.

Self-driving cars needed significant advances in computer vision and decision-making.

Write an inten onally funny story

AI lacked crea vity and humor understanding.

Humor is subjec ve and depends on cultural context and human emo on.

3. AI Progress Since 1998 (Today’s Perspec ve)

Self-driving cars (Tesla, Waymo) can navigate complex environments.

AI language models (GPT-4, Bard) can write stories, jokes, and crea ve content.

AI legal advisors can now analyze contracts and suggest legal ac ons.

Google Translate & AI voice assistants allow real- me spoken transla on.

AI has discovered new mathema cal theorems and even assisted in scien fic research.

Conclusion
AI’s capabili es in 1998 and raises ques ons about what AI can and cannot do. Since then, AI has
made significant progress, solving many of these challenges, but some tasks (like true crea vity and
humor) remain difficult.