Enhancing Search for Satisficing Temporal Planning with Objective-driven Decisions Patrick Eyerich

advertisement
Enhancing Search for
Satisficing Temporal Planning with
Objective-driven Decisions
J. Benton
Patrick Eyerich
Subbarao Kambhampati
g-value plateaus in Temporal Planning
 Common temporal planning objective function
(:metric (minimize (total-time)))
 Makespan as the evaluation function is inefficient for satisificing
search
 g-value plateaus
 Leads to worst case cost-variance between search operations
 The usual approach: Use a Surrogate Search
 Choose a surrogate evaluation function that allows for scalability,
improving the cost-variance between search states
 Objective Function ≠ Evaluation Function
 We want to improve “keeping track” of objective function
2
Temporal Fast Downward
 Temporal Fast Downward (TFD)
Objective Function
Corresponding
Evaluation Function
Surrogate
Evaluation Function
3
Temporal Fast Downward Search
5
@ end eff
@ end eff
3
@ start
2
4
@ end eff
6
@ end eff
@ start
@ end eff
2
4
Temporal Fast Downward Search
…
5
@ end eff
@ end eff
3
@ start
2
4
@ end eff
6
@ end eff
@ start
@ end eff
2
5
Find the Better Path
 Force consideration of better-makespan path
 Should maintain surrogate evaluation function’s
scalability
 Our idea: Determine whether operators are
useful according to makespan and force their
expansion
6
Useful Operators
 Related to Wehrle et al.’s useless actions
 At parent state s





Remove operator o from the domain
Find heuristic value for ,
Apply operator o to generate
Find heuristic value for ,
If
then operator is possibly
useful
 Its degree of usefulness is
7
Makespan-Usefulness Example
Get all trucks to
An optimal plan
8
Makespan-Usefulness Example
9
Lookahead on Useful Operators
 Force expansion of most makespan-useful
state before other states




Remove ‘best’ node from queue
Analyze for child states for makespan-usefulness
Expand state given by most useful operator
Evaluate each resulting grandchild state
according to the surrogate evaluation function
and push into queue
10
Useful Operator Lookahead
…
5
@ end eff
@ end eff
3
@ start
2
4
@ end eff
6
@ end eff
@ start
@ end eff
2
11
Empirical Evaluation
 4 Anytime search variations
 TFD
 TFD with useful lookahead,
 TFD with lazy evaluation followed by TFD with
useful lookahead (and without lazy evaluation),
 TFD with lazy evaluation followed by TFD without
lazy evaluation,
 Makespan heuristic using STN
 30 minute timeout
 Compared on IPC score
12
Results Over Time
13
Results Over Time
14
At 30 Minutes
15
Quality Change
16
Summary
 Used notion of operator usefulness
 Lookahead on most useful operator
 Use in combination with surrogate search
 Shown to improve plan quality in some
domains
 Continues to help when combined with
a portfolio-like approach
17
Future Work
 Lookahead more than one step
 k-deep local lookaheads on most useful
operators combined with best-first search
 Use relaxed solutions
 YAHSP-style lookahead but stop when no
makespan-useful operators are applicable
18
Download