RCE CLASSIFIERS: THEORY AND PRACTICE

Abstract

Restricted Coulomb Energy (RCE) classifiers, as described by Scofield et al. (1988), are shown to have a conceptual relationship with hyperspherical classifiers developed in the 1960s by Batchelor (1974). These classifiers are also shown to share similarities with networks of localized receptive fields and with psychological models of concept formation. Next, the performance of some RCE classifiers is examined. The ability of a trained RCE classifier to generalize to new instances is compared with that of several well-known classifiers. Then, four previously unexamined aspects of RCE classifiers are investigated empirically: (1) the influence of potential wells on training rate, (2) the influence of potential wells on storage requirement, (3) the influence of potential wells on generalization to new instances, and (4) rejection of an instance from an unknown class. Modifications of a traditional RCE classifier improve average correct classification at generalization from 83.2 to 90.7% without significant change in computational cost. By comparison, a nearest-neighbor performs at 93% and a feed-forward multilayer neural network at 88.4% on the same data. Surprisingly, when the improved RCE network is compared with its underlying adaptive nearest-neighbor component, one finds that the incorporation of potential wells into the RCE classifier does not reduce training time or instance storage requirement, nor does it improve generalization to new instances.