RCE networks: an experimental investigation

Abstract

The author briefly surveys the history of hyperspherical classifiers and restricted Coulomb energy (RCE) networks. The ability of a trained RCE classifier to correctly classify new instances is compared with that of several well-known classifiers. Two unexplored aspects of RCE network classifiers are experimentally examined: (1) the treatment of potential wells on the number of training epochs, storage requirements, and generalization; and (2) rejection of an instance from an unknown class. Modifications to a traditional RCE classifier improve average generalization from 83.2% to 90.7% with comparable computational cost. For comparison, a nearest-neighbor classifier performs at 93% and a feedforward network at 88.4% on the same data. When the improved RCE classifier is compared with its underlying adaptive nearest-neighbor classifier the results show that the incorporation of potential wells into the RCE classifier does not reduce training time, nor pattern storage, nor does it improve generalization to new instances. >