Learning Declarative Control Rules for ConstraintBased Planning

1
Learning Declarative Control Rules for
Constraint-Based Planning

• Yi-Cheng Huang
• Bart Selman
• Cornell University
• Henry Kautz
• University of Washington

2
Outline

• Overview of Planning
• Motivation
• Learning Framework
• Experimental Results
• Conclusion

3
Overview of Planning

• Planning - Find a sequence of actions that
  transform an initial state to a goal state
• State complete truth assignment to a set of
  variables (fluents)
• Action a partial function State State
• specified by three sets of variables preconditio
  n, add list, delete list

4
An Example
Goal
Initial
Initial
a
b
a
b
BOS
SFO
NYC
5
A Sample Action

• ( UnloadAirplane (?pln ?pkg ?airport )
• preconditions
• (in ?pkg ?pln) (at ?pln ?airport)
• effects
• (not (in ?pkg ?pln)) (at ?pkg
  ?airport)
• )

6
PLAN
1. LoadAirplane P pkg-a at BOS 2. FlyAirplane P
from SFO to NYC 3. LoadAirplane P pkg-b at
NYC 4. FlyAirplane P from NYC to SFO 5.
UnloadAirplane P pkg-a at SFO 5. UnloadAirplane
P pkg-b at SFO
Goal
Initial
Initial
a
b
a
b
BOS
SFO
NYC
7
Planning

• Domain-independent planning PSPACE-complete
  (Chapman 1987 Bylander 1991 Backstrom 1993)
• General focus on planning avoid search as much
  as possible.
• TLPlan use control knowledge to guild a
  forward-chaining planner (Bacchus Kabanza
  2000).
• Same level of control can be effectively used in
  Blackbox - a Constraint-Based Planner (Huang,
  Selman, Kautz 1999).

8
A Control Rule Example
Goal
Initial
a
a
a
BOS
SFO
NYC
Do NOT unload an object from an airplane if the
airport is not in the objects goal city
9
Motivation

• Control Rules used in TLPlan and Blackbox are
  hand-coded.
• Can we acquire domain knowledge automatically?
• Idea Learn control rules on a sequence of small
  problems solved by planner.

10
Learning Framework
Problem
Blackbox Planner
Plan Justification / Type Inference
ILP Learning Module / Verification
Control Rules
11
Target Concepts for Actions

• Action Select Rule indicate conditions under
  which the action can be performed immediately.
• Ex. Unload a package at its goal location.
• Action Reject Rule indicate conditions under
  which it must not be performed.
• Ex. Do not load a package at its goal location.

12
Heuristics for Extracting Examples

• Basic Assumption
• Plan found by planner on simple problems are
  optimal or near-optimal.
• Actions appear in an optimal plan must be
  selected.
• Actions that do not appear must be rejected.
• Definition
• real action action appears in the plan.
• virtual action action that its preconditions
  hold but does not appear in the plan.

13
Real Virtual Actions for UnloadAirplane
1. LoadAirplane P pkg-a at BOS 2.
UnloadAirplane P pkg-a at BOS 2. FlyAirplane P
from BOSto NYC 3. UnloadAirplane P pkg-a at
NYC 3. LoadAirplane P pkg-b at NYC 4.
UnloadAirplane P pkg-a at NYC 4.
UnloadAirplane P pkg-b at NYC 4. FlyAirplane P
from NYC to SFO 5. UnloadAirplane P pkg-a at
SFO 5. UnloadAirplane P pkg-b at SFO
14
Heuristics for Extracting Examples
15
ILP Rule Induction

• Based on Quinlans FOIL (Quinlan 1990 1996).
• ( ) action literals
• Literal
• Xi Xj
• P(X1,, Xn)
• goal (P(X1,, Xn))
• negation of the above

16
Reject Rule UnloadAirplane
UnloadAirplane (obj, plane, loc)
goal(at (obj, loc2)) (loc ! loc2)
17
Learning Time
18
Empirical Results
19
Conclusion

• Our system is simple and modular Learning time
  is short.
• Learned rules are useful on various domains.
• Learned rules are represented in logic form
  Learned rules can be used to other planning
  systems.