Selective Learning of Macro-Operators with Perfect Causality

Abstract

The macro-operator learning means extracting sub-sequences from a worked example and generalizing them for the future problem solving. In general, the candidate of macro-operators extracted from a worked example are so many and they include a lot of useless ones. Therefore, if the learning system generates macro-operators from all the candidates, the amount of macro-operators will get explosively large. Thus, the major problem in the macro-operator learning is how the learning system selects the valid macro-operators out of many candidates. To cope with this problem, we suggest the method of learning macro-operators with Perfect Causality : the heuristics to select only the valid macro-operators. We developed PiL2 system that can acquire useful macro-operators selectively with Perfect Causality and generalize them with EBG (Explanation-Based Generalization) method. We made the experiment in the robot planning by using STRIPS as a problem solver.