A Practical Solution of the EBL Utility Problem and Its On-Line Evaluation

Abstract

The utility problem has appeared through the intensive research on EBL (Explanation-Based Learning) in recent years. The problem means a trade-off between the benefit of learned knowledge and the cost to locate and apply the knowledge. One of the basic solutions is a search for a benefit and harmless subset of learned knowledge, however most of the searches are very expensive and thus need the off-line evaluation to ignore the learning cost. This paper describes an inexpensive method to search a locally optimal macro-ordering in EBL, macro-operator learning. The complexity of the search is restricted within linear order of the amount of learned macros, and the necessary information including statistics on given problems is obtained through the execution of a simple EBL procedure through partial problems. For verifying the effectiveness of our approach, we implemented an IS-EBL using the inexpensive estimation of the utility and made on-line evaluation with a non-learning system and a non-selective EBL in a map-coloring problem. As a result, we had good experimental results and clarified a class in which the IS-EBL works well.