CellNet Co-Ev: Evolving Better Pattern Recognizers Using Competitive Co-evolution

Taras Kowaliw¹, Nawwaf Kharma², Chris Jensen², Hussein Moghnieh², and Jie Yao²

¹Computer Science Dept., Concordia University, 1455 de Maisonneuve Blvd. W., Montreal, Quebec, Canada H3G 1M8
taras.kowaliw@utoronto.ca

²Electrical & Computer Eng. Dept., Concordia University, 1455 de Maisonneuve Blvd. W., Montreal, Quebec, Canada H3G 1M8
kharma@ece.Concordia.ca

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 coevolutionary 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.

LNCS 3103, p. 1090 ff.

Full article in PDF

lncs@springer.de
© Springer-Verlag Berlin Heidelberg 2004