論理プログラミング環境におけるEBLの有効性計算

抄録

Explanation-Based Learning (EBL) fails to accelerate problem solving in some problem domains. An EBL system therefore needs to evaluate the utility of EBL to a given problem domain for determining whether it adopts EBL. Conventional EBL systems empirically evaluate the utility through test experiments solving a great number of test examples, which results in high evaluation cost. These empirical methods prohibit us from estimating the utility before conducting the test experiment. This paper presents a formal framework in which the utility of EBL, implemented in a logic programming, is computed analytically without any test experiments. We represent the utility of EBL as a function of two variables, the number and the distribution of test examples, and predict the utility on subsequent problems by analyzing the function. The utility function is determined by analyzing the trace of problem solving on training examples, not test examples. First, we define a standard EBL procedure including both problem solving and how to add learned rules to a rule base. After only training examples were solved, our method can compute the utility of EBL by evaluating computational cost both of an EBL system and a non-learning system assuming a distribution of test examples. Since our method needs no experimental result on test examples and no execution of learning procedures like EBG, it can efficiently predict the utility of EBL before test experiment. The utility of EBL obtained by our method is a function of the number of test examples, and it approaches to the utility evaluated empirically thus far as the number increases. Finally we show examples of computing the utility of EBL in a Mitchell's SAFE-STACK example. As a result, we have very interesting results that EBL deteriorates problem solving even in simple domain theory such as safe_to_stack (X, Y).