Unifying Classical Planning Approaches

State space and plan space planning approaches have traditionally been seen as fun-damentally different and competing approaches to domain-independent planning. Wepresent a plan representation and a generalized algorithm template, called UCP, for unify-ing these classical planning approaches within a single framework. UCP models planningas a process of refining a partial plan. The alternative approaches to planning are castas complementary refinement strategies operating on the same partial plan representation.UCP is capable of arbitrarily and opportunistically interleaving plan-space and state-spacerefinements within a single planning episode, which allows it to reap the benefits of both.We discuss the coverage, completeness and systematicity of UCP. We also present somepreliminary empirical results that demonstrate the utility of combining state-space andplan-space approaches. Next, we use the UCP framework to answer the question “whichrefinement planner is best suited for solving a given population of problems efficiently?”Our approach involves using subgoal interaction analysis. We provide a generalized ac-count of subgoal interactions in terms of plan candidate sets, and use it to develop a set ofguidelines for choosing among the instantiations of UCP. We also include some prelimi-nary empirical validation of our guidelines. In a separate appendix, we also describe howthe HTN planning approach can be integrated into the UCP framework. This research is supported in part by NSF research initiation award (RIA) IRI-9210997, NSF young investi-gator award (NYI) IRI-9457634 and ARPA/Rome Laboratory planning initiative grants F30602-93-C-0039 andF30602-95-C-0247. We thank Laurie Ihrig and Amol Dattatreya Mali for helpful comments.