Learning Efficient Rules by Maintaining the Explanation Structure

Many learning systems suffer from the utility problem; that is, that time after learning is greater than time before learning. Discovering how to assure that learned knowledge will in fact speed up system performance has been a focus of research in explanationbased learning (EBL). One way to analyze the utility problem is by examining the differences between the match process (match search) of the learned rule and the problem-solving process from which it is learned. Prior work along these lines examined one such difference. It showed that if the search-control knowledge used during problem solving is not maintained in the match process for learned rules, then learning can engender a slowdown; but that this slowdown could be eliminated if the match is constrained by the original search-control knowledge. This article examines a second difference — when the structure of the problem solving differs from the structure of the match process for the learned rules, time after learning can be greater than time before learning. This article also shows that this slowdown can be eliminated by making the learning mechanism sensitive to the problem-solving structure; i.e., by reflecting such structure in the match of the learned rule.