Analogical Reasoning by Decomposing Explanation Structure

Abstract

Analogical inference systems have generally used similarity between precedent examples and a target, both of which are described using a knowledge representation such as a semantic network. To find similarity, such systems have to enumerate pairings between nodes in semantic networks. This enumeration causes problems since it results in a combinatorial explosion of pairings. In this paper we propose that analogical inference be a combination of generalizing examples and applying them to a target. This method infers the target without using similarity, but is still different from conventional learning methods in that it determines the generalization referring to the target. We use explanationbased learning to acquire a concept, which is extracted and generalized based on a partial structure of an explanation. This process need not enumerate pairings since it extracts an explanation from causal relations in examples. Although Winston's framework regards causal relations as the most important, it does not refer to any explanation explicitly, and it requires pairings, which leads to a combinatorial explosion. In contrast, our method enables efficient analogical inference and constructive generation of a result, rather than checking a given result or selecting one from given choices. We demonstrate our method by using the cup example and the geometric analogy.