Lecture 11: Automated Planning

Lecture 11: Automated Planning¶

AIMA Chapter 11 — 1 hour¶

Learning Objectives¶

• Define classical planning (STRIPS)

• Apply forward and backward state-space search

• Use planning as satisfiability

• Handle nondeterminism and hierarchical planning

Classical Planning¶

• State: Set of ground atoms

• Actions: Preconditions, effects (add/delete lists)

• Goal: Set of literals to achieve

• Initial state: Given

STRIPS Representation¶

• Precondition: Conjunction of literals

• Effect: Add list, delete list

• Example: Load(c,t) — Pre: At(c,t) ∧ Cargo(c), Effect: In(c,t)

Example: Blocks World¶

• At(block, location)

• On(block1, block2)

• Clear(x)

• Actions: Move(b,from,to), Stack(b,on), Unstack(b,from)

Forward State-Space Search¶

• State: Current world state

• Actions: All applicable actions

• Goal test: Goal satisfied

• Problem: Branching factor large

Backward Search (Regression)¶

• State: Set of subgoals to achieve

• Actions: Whose effects achieve subgoal

• Goal: Subgoals empty

• Regressive: Work backward from goal

Planning as SAT¶

• Variables: At(Location, t), Have(Arrow, t), etc.

• Axioms: Actions imply preconditions

• Frame axioms: What stays same

• SAT solvers: Very efficient

Heuristics for Planning¶

• Relaxation: Ignore delete lists

• State abstraction: Merge states

• h: Distance in relaxed problem

Hierarchical Planning¶

• High-level actions: Abstract (e.g., “Travel”)

• Refinement: Decompose into primitive actions

• HTN: Hierarchical Task Networks

Nondeterministic Domains¶

• Sensorless: No observations

• Contingent: Use observations

• Online: Interleave planning and execution

Summary¶

• Classical: STRIPS, forward/backward search

• SAT: Encode as propositional

• Heuristics: Relaxation

• Hierarchical: HTN

• Nondeterministic: Contingent, online

References¶

• Russell & Norvig, AIMA 4e, Ch. 11

• Chapter PDF: chapters/chapter-11.pdf

• aima-python: planning.ipynb

Questions?¶

Next lecture: Quantifying Uncertainty (Chapter 12)