Andrew Coles, Maria Fox, Amanda Smith
This paper explores issues encountered when performing online management of large collections of macro-actions generated for use in planning. Existing approaches to managing collections of macro-actions are designed for use with offline macro-action learning, pruning candidate macro-actions on the basis of their effect on the performance of the planner on small training problems. In this paper we introduce macro-action pruning techniques based on properties of macro-actions that can be discovered online, whilst solving only the problems we are interested in. In doing so, we remove the requirement for additional training problems and offline filtering. We also show how search-time pruning techniques allow the planner to scale well to managing large collections of macro-actions. Further, we discuss the properties of macro-actions that allow the online identification of those that are likely to be useful in search. Finally, we present results to demonstrate that a library of macro-actions managed using the techniques described can give rise to a significant performance improvement across a collection of domains with varied structure.
Subjects: 1.11 Planning; 12. Machine Learning and Discovery
Submitted: Jun 26, 2007