Minimal Error Rate Classification in a Non-stationary Environment via a Modified Fuzzy ARTMAP Network

This paper investigates the feasibility of the fuzzy ARTMAP neural network for statistical classification and learning tasks in an on-line setting. The inability of fuzzy ARTMAP in implementing a one-to-many mapping is explained. Thus, we propose a modification and a frequency measure scheme which tend to minimise the misclassification rates. The performance of the modified network is assessed with noisy pattern sets in both stationary and non-stationary environments. Simulation results demonstrate that modified fuzzy ARTMAP is capable of learning in a changing environment and at the same time, of producing classification results which asymptotically approach the Bayes optimal limits. The implications of taking time averages, rather than ensemble averages, when calculating performance statistics are also studied.