Learning to Choose Instance-Specific Macro Operators

The acquisition and use of macro actions has been shown to be effective in improving the speed of AI planners. Current macro acquisition work focuses on finding macro sets that, when added to the domain, can improve the average solving performance. In this paper, we present Instance-specific macro learning. This kind of macro filtering depends on building a predictor that can be used to estimate, for each planning instance, the best subset of a previously collected set of macros to speed up the planning. Learning the predictor is done off-line based on the observed correlation between problem instance features and planner performance in macroaugmented domains. Our empirical results over five standard planning domains demonstrate that our predictors perform as well as the non-instance-specific method that chooses the best-on-average macro, and that there is a chance of improving the performance significantly using instance specific macros. Also in this paper, we tackle the problem of choosing macros from a large set of initial macros as well. We show that in this case approximate methods can produce macro sets that are comparable to the best macro sets.