Learning from Physical Analogies: A Study in Analogy and the Explanation Process

摘要

To make programs that understand and interact with the world as well as people do, we must duplicate the kind of flexibility people exhibit when conjecturing plausible explanations of the diverse physical phenomena they encounter. This process often involves drawing upon physical analogies - viewing the situation and its behavior as similar to familiar phenomena, conjecturing that they share analogous underlying causes, and using the plausible interpretation as a foothold to further understanding, analysis, and hypothesis refinement. This thesis investigates analogical reasoning and learning applied to the task of constructing qualitative explanations of observed physical phenomena. Primary emphasis is placed on two central questions. First, how are analogies elaborated to sanction new inferences about a novel situation. This problem is addressed by contextual structure-mapping, a knowledge-intensive adaptation of Gentner's structure-mapping theory. It presents analogy elaboration as a map and analyze cycle, in which two situations are placed in correspondence, followed by problem solving and inference production focused on correspondence inadequacies. These ideas are illustrated via PHINEAS, a program which uses similarity to posit qualitative explanations for time-varying descriptions of physical behaviors. It builds upon existing work in qualitative physics to provide a rich environment in which to describe and reason with theories of the physical world. Keywords: Artificial intelligence; Theses. (kt)