AN IMPROVED RULE GENERATION METHOD FOR EVIDENCE-BASED CLASSIFICATION SYSTEMS

A new method is described for generating rules which attempt to optimi ze classification when class samples are not contiguous nor necessaril y segregated in feature space. The method combines well-known clusteri ng techniques (Leader and K-Means methods) with Stochastic Relaxation to minimize a combined cluster entropy function. Further, a technique is developed which is capable of determining the cluster weights which optimize classification performance and reflect the Boolean structure s of the associated convex clusters.