Evolutionary computation for the automated design of category functions for fuzzy ART

Abstract

Fuzzy Adaptive Resonance Theory (ART) is a classic unsupervised learning algorithm. Its performance on a particular clustering problem is sensitive to the suitability of the category function for said problem. However, classic Fuzzy ART employs a fixed category function and thus is unable to benefit from the potential to adjust its category function. This paper presents an exploration into employing evolutionary computation for the automated design of category functions to obtain significantly enhanced Fuzzy ART performance through tailoring to specific problem classes. We employ a genetic programming powered hyper-heuristic approach where the category functions are constructed from a set of primitives constituting those of the original Fuzzy ART category function as well as additional hand-selected primitives. Results are presented for a set of experiments on benchmark classification tasks from the UCI Machine Learning Repository demonstrating that tailoring Fuzzy ART's category function can achieve statistically significant superior performance on the testing datasets in stratified 10-fold cross-validation procedures. We conclude with discussing the results and placing them in the context of being a first step towards automating the design of entirely new forms of ART.