Learning Macro-Actions for Arbitrary Planners and Domains

摘要

Many complex domains and even larger problems in simple domains remain challenging in spite of the recent progress in planning. Besides developing and improving planning technologies, re-engineering a domain by utilising acquired knowledge opens up a potential avenue for further research. Moreover, macro-actions, when added to the domain as additional actions, provide a promising means by which to convey such knowledge. A macro-action, or macro in short, is a group of actions selected for application as a single choice. Most existing work on macros exploits properties explicitly specific to the planners or the domains. However, such properties are not likely to be common with arbitrary planners or domains. Therefore, a macro learning method that does not exploit any structural knowledge about planners or domains explicitly is of immense interest. This paper presents an offline macro learning method that works with arbitrarily chosen planners and domains. Given a planner, a domain, and a number of example problems, the learning method generates macros from plans of some of the given problems under the guidance of a genetic algorithm. It represents macros like regular actions, evaluates them individually by solving the remaining given problems, and suggests individual macros that are to be added to the domain permanently. Genetic algorithms are automatic learning methods that can capture inherent features of a system using no explicit knowledge about it. Our method thus does not strive to discover or utilise any structural properties specific to a planner or a domain.