A game-theoretic formulation on adaptive categorization in ART networks

Abstract

The concept of adaptive categorization is introduced to ART-type networks in this paper. Adaptive categorization capability also improves learning performance in self-organizing systems and online learning systems. Classical ART-types networks, however, have only fixed single size cluster formation in categorization, which is controlled by the scalar vigilance parameter. This categorization methodology usually cannot give satisfactory results as the data pattern space is not covered thoroughly by fixed boundary clusters. A game-theoretic formulation and analysis on the competitive clustering nature of ART-type networks are presented. A game-theoretic vigilance parameter adaptation algorithm is then proposed to form variable sized clusters so that the data pattern space is covered much thoroughly. Simulations are presented to demonstrate reliable categorizations obtained from variable sized clusters using game-theoretic vigilance parameter adaptation.