Modified Fuzzy ARTMAP Approaches Bayes Optimal Classification Rtaes: An Empirical Demonstration

This report examines the feasability of the fuzzt ARTMAP neural network for classifying statistical data and analyses the results according to the Bayes decision criterion. A binary decision with a single observation classification problem is chosen to demonstrate and assess the performance of fuzzy ARTMAP. In this task, fuzzy ARTMAP is used to categorise two classes of Gaussian-distributed continuous-valued random variables autonomously and on-line. Various configurations of the task have been investigated by varying the source (mean) separation, prior probabilities and variances of the two Gaussian sources. The results illustrate the limitations of fuzzy ARTMAP in this content. This in turn leads to a modificatin to the algorithm of fuzzy ARTMAP. Together with a proposed category selection scheme, fuzzy ARTMAP is better able to approach the Bayes optimal classification rates for a binary decision domain.