Conclusion and Future Work

Despite the great progress that AI planning has recently achieved, many benchmarks remain challenging for current planners. In this paper we presented techniques that automatically learn macro-operators from previous experience in a domain, and use them to speed up the search in future problems. We evaluated our methods on standard benchmarks from international planning competitions, showing significant improvement for domains where structure information can be inferred. We implemented our ideas in MACRO-FF, an extension of FF version 2.3. MACRO-FF participated in the classical part of the fourth international planning competition, competing in seven domains and taking first place in three of them.

Exploring our method more deeply and improving the performance in more classes of problems are major directions for future work. We also plan to extend our approach in several directions. Our learning method can be generalized from macro-operators to more complex structures such as HTNs. Little research focusing on learning HTNs has been conducted, even though the problem is of great importance.

We plan to explore how a heuristic evaluation based on the relaxed graphplan can be improved with macro-operators. As shown in this article, a macro added to an original domain formulation as a regular operator influences the results of the heuristic function. This is convenient (no changes are necessary in the planning engine), but limited only to STRIPS domains. For other subsets of PDDL, the relaxed graphplan algorithm can be extended with special capabilities to handle macros when no enhanced domain definition is provided.

The long-term goal of component abstraction is automatic reformulation of planning domains and problems. When a real-world problem is abstracted into a planning model, the corresponding formulation is expressed at an abstraction level that a human designer considers appropriate. However, choosing a good abstraction level could be a hard and expensive problem for humans. Hence methods that automatically update the formulation of a problem based on its structure would be helpful.