An Introduction to AI Planning Ute Schmid Applied CS/Cognitive Systems Bamberg University , → Goals, Methods, Topics of AI , → Basic concepts of AI Planning , → Deductive Planning vs. State-based (Strips) Planning Remark: Planning is typically introduced in the last third of an introductory AI lecture. Basic knowlege about problem solving (search algorithms) and First Order Logic is presumed. An Introduction to AI Planning – p. 1
68
Embed
An Introduction to AI Planning - uni-bamberg.de · An Introduction to AI Planning Ute Schmid ... Planning is typically introduced in the last third of an introductory AI lecture.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
An Introduction to AI PlanningUte Schmid
Applied CS/Cognitive Systems
Bamberg University
↪→ Goals, Methods, Topics of AI
↪→ Basic concepts of AI Planning
↪→ Deductive Planning vs. State-based (Strips) Planning
Remark: Planning is typically introduced in the last third of an introductory AI lecture. Basic
knowlege about problem solving (search algorithms) and First Order Logic is presumed.
An Introduction to AI Planning – p. 1
AI and Knowledge Based Systems
Facts and Rules
Representation Formalisms(FOL−based, Production Rules)
Different Kinds of Knowledge(Domain−Specific, Common−sense)
Deduction (Theorem Proving),
Non−montonic/probabilistic reasoning
Abduction (Diagnosis)
Planning
Induction (Learning)
Inference AlgorithmKnowledge Base
An Introduction to AI Planning – p. 2
AI and Knowledge Based Systems
Facts and Rules
Representation Formalisms(FOL−based, Production Rules)
Different Kinds of Knowledge(Domain−Specific, Common−sense)
Deduction (Theorem Proving),
Non−montonic/probabilistic reasoning
Abduction (Diagnosis)
Planning
Induction (Learning)
Inference AlgorithmKnowledge Base
Broad View:AI Systems are Knowledge-Based Systems
An Introduction to AI Planning – p. 2
AI and Knowledge Based Systems
Facts and Rules
Representation Formalisms(FOL−based, Production Rules)
Different Kinds of Knowledge(Domain−Specific, Common−sense)
Deduction (Theorem Proving),
Non−montonic/probabilistic reasoning
Abduction (Diagnosis)
Planning
Induction (Learning)
Inference AlgorithmKnowledge Base
Narrow View:Knowledge-Based Systems is a sub-field of AI
Focus on: Knowledge Representation and DeductiveInference, Main Application: Expert-Systems
An Introduction to AI Planning – p. 2
AI and Knowledge Based Systems
Facts and Rules
Representation Formalisms(FOL−based, Production Rules)
Different Kinds of Knowledge(Domain−Specific, Common−sense)
Deduction (Theorem Proving),
Non−montonic/probabilistic reasoning
Abduction (Diagnosis)
Planning
Induction (Learning)
Inference AlgorithmKnowledge Base
↪→ I take the broader view
An Introduction to AI Planning – p. 2
Goals of AI
Engineering Methods for solving problems on a computer
algorithms/programs for inference, planning, learning, ...
Epistemology Modeling of cognitive processes
perception, reasoning, language understanding, ..
Formalisms Development of formalisms for describing and
evaluating problems and algorithms
calculi of logic, graph theory, complexity theory, ...
An Introduction to AI Planning – p. 3
Goals of AI
Engineering Methods for solving problems on a computer
algorithms/programs for inference, planning, learning, ...
Epistemology Modeling of cognitive processes
perception, reasoning, language understanding, ..
Formalisms Development of formalisms for describing and
evaluating problems and algorithms
calculi of logic, graph theory, complexity theory, ...
↪→ AI is inherently interdisciplinary!
An Introduction to AI Planning – p. 3
Goals of AI
Engineering Methods for solving problems on a computer
algorithms/programs for inference, planning, learning, ...
Epistemology Modeling of cognitive processes
perception, reasoning, language understanding, ..
Formalisms Development of formalisms for describing and
evaluating problems and algorithms
calculi of logic, graph theory, complexity theory, ...
Mathematics, Computer Science, Neurosciences,
Cognitive Psychology, Linguistics, Philosophy of Mind
An Introduction to AI Planning – p. 3
Main Topics of AI
Problem Solving/Planning
Inference/Deduction
Knowledge Representation
Machine Learning
An Introduction to AI Planning – p. 4
Main Topics of AI
Problem Solving/Planning
Inference/Deduction
Knowledge Representation
Machine Learning
General methods, applied in:Game Playing, Expert Systems, Tutor Systems, Diagnosis,Automatic Theorem Proving, Program Synthesis, CognitiveRobotics, Multi-Agent-Systems, ...
An Introduction to AI Planning – p. 4
Main Topics of AI
Problem Solving/Planning: Seminar Topic
Inference/Deduction
Knowledge Representation
Machine Learning: Lecture
General methods, applied in:Game Playing, Expert Systems, Tutor Systems, Diagnosis,Automatic Theorem Proving, Program Synthesis, CognitiveRobotics, Multi-Agent-Systems, ...
An Introduction to AI Planning – p. 4
AI Planning
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
An Introduction to AI Planning – p. 5
AI Planning
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
↪→ AI planning (“action planning”) deals with thedevelopment of representation languages for planningproblems and with the development of algorithms forplan construction.
An Introduction to AI Planning – p. 5
AI Planning
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
Alternative to Planning:Reinforcement Learning(Learn policies for selecting a suitable action to executein a given state)
An Introduction to AI Planning – p. 5
Applications
System Controlautonomous systemsvirtual agents
An Introduction to AI Planning – p. 6
Applications
Process Control:Production of physical goodsConstruction and configurationWorkflow managementMission planningProject planning
An Introduction to AI Planning – p. 6
Blocks World
Running example:
A
C
E
B
D
F
An Introduction to AI Planning – p. 7
Blocks World
Running example:
A
C
E
B
D
F
Objects:A, B, C, D, E, F, Table
An Introduction to AI Planning – p. 7
Blocks World
Running example:
A
C
E
B
D
F
Objects:A, B, C, D, E, F, TableRelations:clear: {C, E, F}ontable: {A, E, D}on:{(C, A), (B, D), (F, B)}
An Introduction to AI Planning – p. 7
States, Goals, Actions
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
An Introduction to AI Planning – p. 8
States, Goals, Actions
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
(partial) description of a situationC is on A and B is on D and F is on B and C is clear and E isclear and F is clear and A is on the table and E is on the tableand D is on the table
A
C
E
B
D
F
An Introduction to AI Planning – p. 8
States, Goals, Actions
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
Simple goal: D is clearConjunctive goal: D is clear and D is on A(disjunctive, quantified, ...)
An Introduction to AI Planning – p. 8
States, Goals, Actions
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
put block E on block F, put block C on the table
An Introduction to AI Planning – p. 8
States, Goals, Actions
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
put block E on block F, put block C on the table
Operator Scheme: define actions over variablesput(?x, ?y): put a block on another blockputtable(?x): put a block on the table
An Introduction to AI Planning – p. 8
States, Goals, Actions
A plan is a sequence of actions for transforming agiven state into a state which fulfills a predefined setof goals.
put block E on block F, put block C on the table
Operator Scheme: define actions over variablesput(?x, ?y): put a block on another blockputtable(?x): put a block on the table
↪→ operators have application conditionse.g., to put a block somewhere, it must be clear
An Introduction to AI Planning – p. 8
Planning Formalism
A planning formalism must provide
a language to represent states, goals and actions.
an algorithm for constructing a sequence of actionswhich transforms an initial state into a goal state.
An Introduction to AI Planning – p. 9
Planning Formalism
A planning formalism must provide
a language to represent states, goals and actions.
Typically a subset of FOL for states and goals
Actions are generated via instantiation of operatorschemes
an algorithm for constructing a sequence of actionswhich transforms an initial state into a goal state.
An Introduction to AI Planning – p. 9
Planning Formalism
A planning formalism must provide
a language to represent states, goals and actions.
Typically a subset of FOL for states and goals
Actions are generated via instantiation of operatorschemes
an algorithm for constructing a sequence of actionswhich transforms an initial state into a goal state.
Deductive Planning: Theorem Proving
State-Based Planning: Search in State-Space
An Introduction to AI Planning – p. 9
Deductive Planning
Based on FOL, Situation Calculus (Green, 1967)
An Introduction to AI Planning – p. 10
Deductive Planning
Based on FOL, Situation Calculus (Green, 1967)
States are represented as conjunction of facts:
Initial state s1
clear(a, s1).clear(b, s1).on(b, c, s1).ontable(a, s1).ontable(c, s1). AC
Operator Schemes are represented as Effect Axioms:on(X, Y, put(X, Y, S))← clear(X, S) ∧ clear(Y, S)“In a situation where blocks X and Y are clear, a new situation
where on(X, Y)) holds can be reached by putting X on Y.”
Rewrite to clausal form:¬clear(X, S) ∨ ¬clear(Y, S) ∨ on(X, Y, put(X, Y, S))
Proof that the goal can be reached by resolution:S = put(a, b, s1)
puttable(x) puttable(x’)puttable(A) put(B, C) put(B, C)
will become inconsistentfor z = B
Plan: Path from the initial state (leaf node) to the root (goalstate)
An Introduction to AI Planning – p. 23
Current Approaches
State-based approaches dominate the fieldGraphplan, Sat-Planning
Special aspects:Include (temporal/resource) constraintsPlanning under uncertainty (conformant planning)...
An Introduction to AI Planning – p. 24
Planning vs. Problem Solving
Problem solving:domain specific representations
search algorithms (such as A*) must be adapted toproblem
An Introduction to AI Planning – p. 25
Planning vs. Problem Solving
Problem solving:domain specific representations
search algorithms (such as A*) must be adapted toproblem
Planning:domain independent representation language
general-purpose algorithms for a given language
An Introduction to AI Planning – p. 25
Planning vs. Human Problem SolvingPlanning algorithms are (should be)complete (return a plan, if one exists) andcorrect (terminate with a legal action sequence or reportfailure)
An Introduction to AI Planning – p. 26
Planning vs. Human Problem SolvingPlanning algorithms are (should be)complete (return a plan, if one exists) andcorrect (terminate with a legal action sequence or reportfailure)
For complex problems, large parts of the search treemight be generated until a solution is found.
An Introduction to AI Planning – p. 26
Planning vs. Human Problem SolvingPlanning algorithms are (should be)complete (return a plan, if one exists) andcorrect (terminate with a legal action sequence or reportfailure)
For complex problems, large parts of the search treemight be generated until a solution is found.
Humans typically use greedy search strategies involvingheuristics: try to reach a state which is more similar tothe goal state in each step (experimental data, e.g. Hobbits
and Orcs, Greeno, 1974)↪→ efficient but incomplete(as the General ProblemSolver, Newell & Simon, 1972)
An Introduction to AI Planning – p. 26
Planning vs. Human Problem SolvingHobbits and Orcs (Greeno, 1974)
Subjects have problems with the trans-formation from state (6) to (7). Here 2and not only 1 passenger must be trans-ported back to the left river bank. That is,there must be created a situation whichis further removed from the goal statethan the situation before.
b H H H O
O O b H H O O
H O
b H H O O
H O b
O O
H H H O
b O O O
H H H b H H H O O
O
b
H H H O
O O
b H H H O O O
H H H O Ob
O
H H H
b O O O
H H O O
b H O
H H H O O Ob
(5) (6)
(7) (8)
(9) (10)
(11) (12)
(3) (4)
(2)(1)
An Introduction to AI Planning – p. 26
Planning vs. Human Problem Solving
Humans learn from problem solving experience (do notrun into the same dead-ends again)