m-GFAM: An elegant approach to genetically optimize Fuzzy ARTMAP Neural Network Architectures

Adaptive Resonance Theory (ART) neural network architectures, such as Fuzzy ARTMAP (FAM), have solved successfully a variety of classification problems. However, FAM suffers from an inherent problem that of creating larger architectures than it is necessary to solve the problem at hand (referred to as the ART category proliferation problem). This problem is especially amplified for classification problems which have noisy data, and/or data, belonging to different labels, that significantly overlap. In this paper we introduce m-GFAM (modified genetically engineered Fuzzy ARTMAP), which is produced by evolving a population of FAM architectures. Our results demonstrate that m-GFAM successfully addresses the category proliferation problem by creating a small size trained ART structure that exhibits good generalization. Our experiments show that m-GFAM outperforms other ART architectures that have addressed the category proliferation problem before.