Theory-guided Empirical Speedup Learning of Goal Decomposition Rules

摘要

Speedup learning is the study of improving the problem-solving performance with experience and from outside guidance. We describe here a system that successfully combines the best features of Explanation-based learning and empirical learning to learn goal decomposition rules from examples of successful problem solving and membership queries. We demonstrate that our system can efficiently learn effective decomposition rules in three different domains. Our results suggest that theory-guided empirical learning can overcome the problems of purely explanation-based learning and purely empirical learning, and be an effective speedup learning method.