ADAPTIVE SIMILARITY METRICS IN CASE-BASED REASONING

摘要

A similarity measure is a critical component in any case-based reasoning (CBR) system. It compares two cases with respect to their "features," with each feature using a separate "comparator." The results of the comparators are combined according to some rule to give an overall measure of the similarity between the given cases. Previous works have described a CBR framework that can easily be instantiated to provide a case-based reasoner for virtually any problem domain. This uses an "adaptive," or "reflective," software architecture wherein case features are associated with their comparators dynamically via run-time references to metadata. New instances of the framework are created simply by changing the metadata. No reprogramming is required. In this paper, we extend this concept to allow for dynamic selection also of feature-comparator combination rules. This makes the framework more adaptive by eliminating the need to reprogram it for each such new rule. The overall effect is that the entire similarity measure is described by metadata. The approach is illustrated via an example.