Generalized Policy Learning with Planning

Abstract

Generalized policy learning seeks to find policies that solve multiple tasks within a planning domain. We introduce methods to search for policies independently in a domain from empty initialized policies. As an extension, we also propose a problem setting to learn satisficing policies between domains. In an independent domain, we propose a score function to guide the policy search. Our approach, Policy-Guided Planning for Generalized Policy Generation (PG3), evaluates policies based on how well it can be used to plan. Empirically, we show that PG3 allows generalized policy learning to occur more efficiently than other baselines with PDDL-based problems and policies represented as lifted decision lists. Finally, our experiments show that policies independently learned are qualitiatively similar, prompting further investigation on the possibilities of further accelerating the policy search process.