LoopDISTILL: Learning Looping Domain-Specific Planners from Example Plans

Because general-purpose planning methods have difficulty with large-scale planning problems, researchers have resorted to hand writing domain-specific planners to solve them. An interesting alternative is to use example plans to demonstrate how to solve problems in a particular domain and to use that information to automatically learn domain-specific planners. In this paper, we present the LoopDISTILL algorithm for identifing repeated structures in observed plans and show how to convert looping plans into domain-specific planners, or dsPlanners. Looping dsPlanners are able to apply experience acquired from the solutions to small problems to solve arbitrarily large ones. We show that automatically learned dsPlanners are able to solve large-scale problems much more effectively than are state-of-the-art general-purpose planners and are able to solve problems many orders of magnitude larger than general-purpose planners can solve. Introduction General-purpose planners have traditionally had difficulty with large-scale planning problems, although many largescale problems have a repetative structure, because they do not capture or reason about such repetition. Instead, to solve large-scale problems, programmers have had to rely on the tedious and difficult process of hand writing special-purpose planners that may precisely encode the repeated structure. However, example plans are often available, and can demonstrate this structure. In previous work, we introduced the concept of automatically-generated domain-specific planning programs (or dsPlanners) and showed how to use example plans to learn non-looping dsPlanners, which can solve problems of limited size (Winner & Veloso 2003). Here, we present the novel LoopDISTILL algorithm for automatically identifing the repeated structure of example plans to learn looping dsPlanners. DsPlanners execute independently of a generalpurpose planning program and are very efficient; they return a solution plan in time that is linear in the size of the dsPlanner and of the problem, modulo state-matching effort. We show that looping dsPlanners can solve large-scale planning problems more quickly than can general-purpose planners and that they can solve much larger problems than Copyright c © 2007, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. can general-purpose planners. And because dsPlanners are learned directly from example plans, there is no need for tedious hand coding. We first discuss related work. We then define dsPlanners and explain how we use them to generate the solution plans for new problems. Then we discuss classes of loops, describe our algorithm for identifying loops in observed plans, and illustrate its behavior with examples. Next we present the results of using plans with identified loops as planners and compare this to using state-of-the-art general-purpose planners. Finally, we present our conclusions.