CellNet Co-Ev: Evolving better pattern recognizers using competitive co-evolution

Abstract

A model for the co-evolution of patterns and classifiers is presented. The CellNet system for generating binary classifiers is used as a base for experimentation. The CellNet system is extended to include a competitive co-evolutionary Genetic Algorithm, where patterns evolve as well as classifiers; This is facilitated by the addition of a set of topologically-invariant camouflage functions, through which images may disguise themselves. This allows for the creation of a larger and more varied image database, and also artificially increases the difficulty of the classification problem. Application to the CEDAR database of hand-written characters yields both an increase in reliability and an elimination of over-fitting relative to the original CellNet project. © Springer-Verlag Berlin Heidelberg 2004.