木パターン上の決定リストの学習とその推論制御への応用

抄録

This paper introduces a new concept, a decision list over tree patterns (DLTP), which is a natural extension of a decision list, for dealing with tree structured objects. First, we show a theoretical result of the learnability of this class. We define the class, k-node-DLTP, which is a subclass of decision lists over tree patterns whose number of tuples is bounded by k+1. A hardness result on the learnability of this class is shown. Then we propose a practical learning algorithm based on an information theoretical evaluation function. This algorithm is an extension of Quinlan's ID3 algorithm. One of the most interesting application areas of this work is inference control. We can use decision lists over tree patterns for controlling the application of rules in SLD resolution process. We apply the proposed algorithm to learning strategies for applying rules and experimental results in several domains are presented. These experiments show the effectiveness of the proposed heuristic evaluation function for finding small size hypotheses. From the view point of Explanation Based Learning, our method can be regarded as a method for refining domain rules so as not to backtrack while solving training examples.