Explanation Based Learning as Constrained Search

Abstract

This chapter discusses explanation-based learning as constraint search. In place of the typical Explanation-based generalization (EBG), especially when an inadequate domain theory is involved, a different inference method can be used—one that is not truth preserving (and does not do proofs) but still uses explanations to build a space of plausible generalizations. This approach has two advantages. It does not require complete domain theories. Plausible reasoning allows inference despite the incomplete knowledge. It also creates the opportunity to use multiple explanations. Differing explanations will constrain search in different ways and to varying amounts. If each of the explanations captures some different, useful information, then it is likely that one of the variety of explanations will be useful for each instance. An example of such a system is the MOB system. MOB is a failure-driven supervised learning system whose task is the pronunciation of novel English words. Each training instance is a correct word/pronunciation pair. When MOB is given a pair that contains a word that it fails to pronounce correctly, it attempts to learn the correct pronunciation. The success of MOB using EBL techniques in an inherently incomplete domain suggests that this combination of theory-based explanation and generalization with empirical utility testing addresses a few problems of incomplete domain theories.