Learning Instance-Specific Macros

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, result in improved average solv-ing performance. Instance-specific macro learning, in con-trast, aims to build a predictor that can be used to estimate, for each planning instance, a subset of the previously identi-fied macros that is best for a specific problem instance. Based on off-line measures of the correlation between problem in-stance features and planner performance in macro-augmented domains, we build two such predictors using a standard ma-chine learning approach. Online, the features of the current problem instance are measured and a predictor selects the macros that should be added to the domain. Our empirical results over four standard planning domains demonstrate that our predictors perform as well as the perfect predictor on all domains and never perform worse than any of the other macro-based approaches tested, including an existing macro-learning system. However, a simpler, non-instance-specific method that chooses the best-on-average macro subset over the learning problems performs just as well as our predictors