Introduction to Planning

Outline

What is planning?
Formal definitions of planning problems Basic planning algorithms Recent development

What is Planning?

Planning is a search problem that requires to find an efficient sequence of actions that transform a system from a given starting state to the goal state

Whats Given?

Initial state of the problem Goal state of the problem A finite set of actions:

pre-conditions: a finite set of conditions for the action to be performed

post-conditions: a finite set of conditions that will be changed after the action is performed cost

Whats Output?

A sequence of actions that meet the following criteria

every action matches the current system state can transform system from initial state to goal state

the total cost of the actions is below a specified value

Planning as a Real-World Problem

Planning problem has a wide range of applications in the real world

planning in daily life game world workflow management

Planning a Trip
Begin Preparation Airline Reservation

Rental Car Reservation

Hotel Reservation

Confirm Reservations

End

Towers of Hanoi

Problem Analysis

Optimal solution contains 2n-1 actions, where n is the number of disks Up to 6 possible actions in each system state The upper bound on search space to (2n-1) find an optimal solution will be 6

Sliding-Tile Puzzle

Planning in Workflow Management

Start Assembly Install Video Card
Install Network Card Install Internal Disk Install Motherboard

Insert Modem
Plug in CD

Examine Order

Gather Components Assembly Box and Motherboard

Plug in Battery Test End Assembly

Partial-Order Plan versus Total-Order Plan

Partial-order plan

consists partially ordered set of actions sequence constraints exist on these actions

plan generation algorithm can be applied to transform partial-order plan to total-order plan consists totally ordered set of actions

Total-order plan

Partial-Order Plan
Get brush Start Get ladder

Paint ceiling

Finish

Total-Order Plan
Start Get ladder Get brush Paint ceiling Finish

Start

Get brush

Get ladder

Paint ceiling

Finish

Features of Planning Problems

Large search space Action is associated with system states Restrictions on the action sequence Valid solution may not exist Optimization requirement

STRIPS Planning System

A tuple T = (P, O, I, G), where

P is a finite set of ground literals, the conditions O is a finite set of operators I is the initial state, a subset of P G is the goal state, a subset of P

STRIPS Operators

Each operator O has the following attributes

PC, a set of ground literals, defines the precondition of the operator D, a set of ground literals, defines the conditions that will be removed after the operation is executed A, a set of ground literals, defines the conditions that will be added after the operation is executed C, the cost of the operation

A Simple Example - Blocks World

A C B

C
Goal State On(A, B) On(B, C) On(C, Fl) Clear(A) Clear(Fl)

Initial State On(C, A) Clear(Fl) On(A, Fl) Clear(B) On(B, Fl) Clear(C)

Graphical Representation of Initial State

On(C, A) Clear(Fl) On(A, Fl) Clear(B) On(B, Fl) Clear(C)

Start

Graphical Representation of Goal State

Nil

finish

On(A, B)

On(B, C)

On(C, Fl)

Clear(A)

Clear(Fl)

Block World - Operator

Move(x, y, z)

Move block x that is above y to above z

PC: On(x,y), Clear(x), Clear(z) D: Clear(z), On(x, y) A: On(x,z), Clear(y), Clear(Fl)

Graphical Representation of Operator

A
Clear(y) On(x, z) Clear(Fl)

Operator

Move(x, y, z)

PC

On(x, y)

Clear(x)

Clear(z)

Forward Chaining

Search from the initial state Expand the search tree by finding the set of all applicable operators from the current state

applicable means the precondition of the operator is a

subset of current state

Update current state For every state that is reached, record the shortest path (or path with lowest cost) from the initial state to this state If the goal state is reached, stop the algorithm

Forward Chaining
Move(C, A, Fl) On(C, A) Clear(Fl) On(A, Fl) Clear(B) On(B, Fl) On(B, C) Clear(C) Clear(Fl) On(C, Fl) Clear(Fl) On(A, Fl) Clear(B) On(B, Fl) Clear(C) Clear(A)

Move(B, Fl, C)

On(C, A) Clear(B) On(A, Fl)

Backward Chaining

Search backward from the goal state Expand the search tree by finding the set of all applicable operators that can reach the current state

applicable means set A of the operator is a subset of

current state

Update the state If the initial state is reached, stop the algorithm The solution is a partial-ordered plan Constraints in action ordering may be violated

Backward Chaining
On(A, B) On(B, C) On(C, Fl) Clear(A) Clear(Fl)
On(B, C) On(C, Fl) Clear(A) Clear(Fl) On(A, Fl) Clear(B) Move(A, Fl, B)

Recent Developments

Plan Reuse Graphplan

Problem-specific planning
Evolutionary computation approach

Plan Reuse

Reuse old plans for new planning problems Consists of two steps

plan matching plan modification

generally, plan reuse is even harder than plan from scratch do better only when two problems are close enough plan matching could be the bottleneck

Research findings

Graphplan

Partial-order general planner Constructing a planning graph before search

plan graph contains all possible actions that can be taken in each time step actions that interfere with one another can coexist in the graph

More efficient than other general planners in some problems

Problem-specific Planning

Heuristics combined during search

heuristics is problem dependent cannot apply to other problems

Usually outperforms general planners in specific problems Example

in sliding-tile puzzle, accurate estimation of the distance between current state and goal state can speed up the search for a plan

Genetic Algorithms

Parallel search and optimization algorithm Inspired by the basic rule of natural selection, survival-of-the-fittest Non-deterministic algorithm

randomness is incorporated during implementation

Solution Formation

Candidate solution is encoded as a list of genes, called chromosome

1 0 1 1 0 1

Start with a population of randomly generated chromosomes Population is evolved in every generation

evaluate each chromosome select chromosomes to the next generation apply genetic operations

Genetic Operation

crossover
1 0 0 0 1 0 1 1

Crossover point 0 1 1 1

Before crossover

After crossover

1 0 1 1 1 1
0 0 0 1 0 1

Genetic Operation

Mutation
1 0 1 1 0 1

Before mutation After mutation

GA Procedure

Initialize the population

Evaluate each chromosome in the population While the stopping condition is not met

select chromosomes to the next generation

crossover, mutation evaluate newly generated chromosomes

End while

Evolutionary Computation Approach

Non-deterministic algorithm Starts from a set of randomized plan Plans are evolved during generations In each generation

evaluate the performance of the plan select the plans to next generation, based on performance crossover, mutation to create new plans

Evolutionary Computation Approach

Solution encoding

a list of floating-point number, each number maps to an action in the problem number of actions can vary during search

0.8 0.5 0.3 0.2 0.5

A4 A3 A2 A1 A2

Evolutionary Computation Approach

State-aware crossover

crossover points are selected based on the matching of states s1 s2 Match(s1, s2) = true

Parent 1
Parent 2 Child 1 Child 2

Evolutionary Computation Approach

Mutation

randomly select an action and replace it with a random floating-point number 0.8 0.5 0.3 0.2 0.5

Before mutation After mutation

0.8 0.5 0.6 0.2 0.5

Evolutionary Computation Approach

Fitness evaluation

consists of two parts: goal fitness fg and cost fitness fc goal fitness: how the solutions reach the goal state cost fitness: the total cost of the solution overall fitness = a * fg + b * fc

Experimental Results

Towers of Hanoi

can solve the 5 and 6-disk case with high probability state-aware crossover does not perform well can solve 3*3 problem case with high probability state-aware crossover perform much better since a looser definition of matching state is used

Sliding-tile puzzle

Performance scalability is not good Heuristics may be helpful to improve the

References

Artificial Intelligence: A Modern Approach. Stuart

Russell and Peter Norvig. Prentice Hall.

Artificial Intelligence: A New Symthesis. Nils Nilsson.

Kaufmann.

Intelligent Planning: A Decomposition and Abstraction Based Approach. Qiang Yang. Springer.
Internet-Based Workflow Management: Towards a Semantic Web. Dan C. Marinescu. Wiley, 2002.

References

Plan reuse versus plan generation: a theoretical and empirical analysis. B. Nebel, J. Koehler. Journal of Artificial
Intelligence 76 (1995), 427-454.

Fast planning through planning graph analysis. A. L. Blum,

M. L. Furst Journal of Artificial Intelligence 90 (1997), 281300. Planning as heuristic search. B. Bonet, H. Geffner. Journal of Artificial Intelligence 129 (2001), 5-33. What it is, What it could be, An introduction to the Special Issue
on Planning and Scheduling. D. McDermott, J. Handler. Journal of Artificial Intelligence 76 (1995), 1-16.

Concluding Words

Failing to plan is planning to fail --Effie Jones