ABSTRACT

Reasoning from prior experience depends upon having a large memory of prior cases and a system for retrieving them when they are relevant. Often, relevance means similarity to the current situation on the basis of abstract or thematic features other than the features used to initially describe the current situation. To retrieve cases relevant to some new situation, a system must be able to describe the new situation in abstract terms and use that description as a search key or as a means to judge the appropriateness of prior cases. Typically, the abstract description process has been considered as separate from the memory search process. This paper presents a scheme for integrating the feature extraction and memory search processes and argues in favor of such an approach on methodological and efficiency grounds. It presents a program that exploits parallelism to control some of the high processing costs associated with feature extraction and memory search.