DISTILL: Learning Domain-Specific Planners by Example

An interesting alternative to domain-independent planning is to provide example plans to demonstrate how to solve problems in a particular domain and to use that information to learn domainspecific planners. Others have used example plans for case-based planning, but the retrieval and adaptation mechanisms for the inevitably large case libraries raise efficiency issues of concern. In this paper, we introduce dsPlanners, or automatically generated domain-specific planners. We present the DISTILL algorithm for learning dsPlanners automatically from example plans. DISTILL converts a plan into a dsPlanner and then merges it with previously learned dsPlanners. Our results show that the dsPlanners automatically learned by DISTILL compactly represent its domain-specific planning experience. Furthermore, the dsPlanners situationally generalize the given example plans, thus allowing them to efficiently solve problems that have not previously been encountered. Finally, we present the DISTILL procedure to automatically acquire one-step loops from example plans, which permits experience acquired from small problems to be applied to solving arbitrarily large ones.