Learning Useful Macro-actions for Planning with N-Grams

Automated planning has achieved significant breakthroughs in recent years. Nonetheless, attempts to improve search algorithm efficiency remain the primary focus of most research. However, it is also possible to build on previous searches and learn from previously found solutions. Our approach consists in learning macro-actions and adding them into the planner's domain. A macro-action is an action sequence selected for application at search time and applied as a single indivisible action. Carefully chosen macros can drastically improve the planning performances by reducing the search space depth. However, macros also increase the branching factor. Therefore, the use of macros entails a utility problem: a trade-off has to be addressed between the benefit of adding macros to speed up the goal search and the overhead caused by increasing the branching factor in the search space. In this paper, we propose an online domain and planner-independent approach to learn 'useful' macros, i.e. macros that address the utility problem. These useful macros are obtained by statistical and heuristic filtering of a domain specific macro library. The library is created from the most frequent action sequences derived from an n-gram analysis on successful plans previously computed by the planner. The relevance of this approach is proven by experiments on International Planning Competition domains.