Learning and Exploiting Configuration Knowledge for a Portfolio-based Planner

In the recent years the field of automated plan generation has significantly advanced and several powerful domain-independent planners have been developed. However, no one of these systems clearly outperforms all the others in every known benchmark domain. It would then be useful to have a multi-planner system capable of automatically selecting and combining the most efficient planning technique(s) for each given domain. In this paper we propose a planner, called PbP (Portfolio-based Planner), which automatically configures a portfolio of existing planners, possibly using a useful set of macro-actions for each of them. The configuration relies on some knowledge about the performance of the planners in
the portfolio and the observed usefulness of sets of macro-actions, which is automatically generated by a statistical analysis considering a set of training problems for the domain under consideration. The configuration knowledge for the given domain consists of a promising combination of planners in the portfolio, each one with a (possibly empty) set of macro-actions, and additional information specializing their round-robin scheduling at planning time. PbP has two variants, one focusing on speed (PbP.s) and one on plan quality (PbP.q).
A preliminary version of PbP.s entered the learning track of the sixth IPC, and was the overall winner of this competition track. An experimental analysis presented in the paper confirms the effectiveness of PbP.s, indicates that PbP.q performs better than the IPC6 planners, shows that the learned configuration knowledge can be very useful for PbP.s/q, and
demonstrates that PbP.s/q can perform much better than the basic planners forming the portfolio.