Case-based reasoning (CBR) is a reasoning process that solves new problems by retrieving similar prior problem-solving episodes and adapting their solutions to fit the new situations. Learning by acquiring new cases is a fundamental part of case-based reasoning, and the process of learning by case acquisition has been a central focus of CBR research. However, little attention has been given to the acquisition of case adaptation knowledge. This paper discusses progress on a model of how case-based reasoners can learn to make better use of prior cases by learning how to adapt them to new circumstances more effectively, and by refining the similarity criteria they use to select relevant candidate cases.
Motivation for studying how adaptation knowledge is learned comes from the psychological literature. Experiments by Gentner & Toupin [1986] demonstrate a developmental shift in the similarity criteria used by children for analogical reasoning, and show that the shift is manifested in how they adapt stories to apply to new characters. Keane [1994] shows that when selecting analogues for an analogical problem-solving task, subjects favor analogues that are easier to apply to the new problem situation; they refine similarity criteria on the basis of adaptation knowledge.
Our research models how multiple types of learning can
improve the case adaptation process and examines how their
interrelationships contribute to the overall performance of a CBR
system. This paper first describes our task domain and the basic
structure of our testbed system. It then describes how the
system learns to improve its case adaptation process, discusses
the effects of case adaptation learning for an initial set of
test examples, and describes how adaptation learning determines
new similarity criteria. The paper closes by placing our
research in the context of other models and examining possible future
directions.