Learning macro-actions genetically from plans

Despite recent progress in planning, many complex domains and even larger problems in simple domains remain hard and challenging. One way to achieve further improvement is to utilise knowledge acquired for the planner from the domain. Macro-actions are 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 works on macros utilise knowledge explicitly specific to the planners and the domains. But presumably any
particular properties are not likely to be common with different planners or wider range of domains. Therefore, a macro learning system that does not exploit any structural knowledge about planners and domains explicitly is of immense interest. This paper presents an offline system capable of learning macros genetically from plans. Given a planner, a domain, and necessary problems, our system generates macros, lifting from plans of some problems, under guidance from a genetic algorithm. It represents macros like regular actions, evaluates them individually by solving other problems, and suggests the best macro to be added to the domain permanently. Genetic algorithms are automatic learning methods that can learn properties of a system using no explicit knowledge
about it. Our system thus does not strive to discover or utilise any structural properties specific to a planner or a domain.