Speedup of hypothetical reasoning by experience-based learning mechanism

While hypothetical reasoning is a useful knowledge system framework that is applicable to many practical problems, its crucial problem is its slow inference speed. The paper presents a speedup method for hypothetical reasoning systems that uses an experience-based learning mechanism which can learn knowledge from inference experiences to improve the inference speed in subsequent inference processes, sharing subgoals similar to those of the prior inference processes. This learning mechanism has common functions with existing explanation-based learning. However, unlike explanationbased learning, the learning mechanism described in the paper has a learning capability even at intermediate subgoals that appear in the inference process. Therefore, the learned knowledge is useful even in the case in which a new goal given to the system shares a subgoal that is similar to those learned in the prior inference. Since the amount of the learned knowledge becomes very large, the learning mechanism also has a function of learning selectively only at subgoals which substantially contribute to the speedup of the inference. In addition, the mechanism also includes a knowledge reformation function, which transforms learned knowledge into an efficient form to be used in subsequent inference.