A Study of Explanation-Based Methods for Inductive Learning

摘要: This paper formalizes a new learning-from-examples problem: identifying correct concept definition from positive examples such that the is some specialization of target defined by domain theory. It describes an empirical study evaluates three methods for solving this explanation-based generalization (EBG), multiple example (mEBG), and method, induction over explanations (IOE). The demonstrates two existing (EBG mEBG) exhibit shortcomings: (a) they rarely identify definition, (b) are brittle in their success depends greatly on choice encoding theory rules. IOE, does not these shortcomings. method applies to construct training as mEBG, but forms employing similarity-based policy explanations. IOE has advantage explicit can be exploited aid learning process, dependence initial significantly reduced, concepts learned few examples. context implemented system, called Wyl2, which learns variety chess including “skewer” “knight-fork.”