Extending the use of plateau-escaping macro-actions in planning

Many fully automated planning systems use a single, domain independent heuristic to guide search and no other problem specific guidance. While these systems exhibit excellent performance, they are often out-performed by systems which are either given extra human-encoded search information, or spend time learning additional search control information offline. The benefit of systems which do not require human intervention is that they are much closer to the ideal of autonomy. This document discusses a system which learns additional control knowledge, in the form of macro-actions, during planning, without the additional time required for an online learning step. The results of various techniques for managing the collection of macro-actions generated are also discussed. Finally, an explanation of the extension of the techniques to other planning systems is presented.