Evolving Cellular Automata as Pattern Classifier Niloy Ganguly 1 , Pradipta Maji 2 , Sandip Dhar 2 , Biplab K. Sikdar 2 , and P. Pal Chaudhuri 2 1 Computer centre, IISWBM, Calcutta, West Bengal, India 700073, n ganguly@hotmail.com 2 Department of Computer Science & Technology, Bengal Engineering College (D U), Howrah, West Bengal, India 711103 {pradipta,sdhar,biplab,ppc}@cs.becs.ac.in Abstract. This paper reports a high speed, low cost pattern classifier based on the sparse network of Cellular Automata. High quality of classi- fication of patterns with or without noise has been demonstrated through theoretical analysis supported with extensive experimental results. 1 Introduction The internetworked society has been experiencing a explosion of data that is acting as an impediment in acquiring knowledge. The meaningful interpretation of these data is increasingly becoming difficult. Consequently, researchers, prac- titioners, entrepreneurs from diverse fields are assembling together to develop sophisticated techniques for knowledge extraction. Study of data classification models form the basis of such research. A classification model comprises of two basic operations - classification and prediction. The evolving CA based classifier proposed in this paper derives its strength from the following features: – The special class of CA referred to as Multiple Attractor Cellular Automata (MACA) is evolved with the help of genetic algorithm to arrive at the desired model of CA based classifier. – In the prediction phase the classifier is capable of accommodating noise based on distance metric. – The classifier employs the simple computing model of three neighborhood Additive CA having very high throughput. Further, the simple, regular, modular and local neighborhood sparse network of Cellular Automata suits ideally for low cost V LSI implementation. The Cellular Automata (CA) preliminaries follows in the next section. 2 Cellular Automata Preliminaries The fundamentals of Cellular Automata we deal with is reported in the book [1]. The classifier reported in this work has been developed around a specific class of CA referred to as Multiple Attractor CA (MACA). S. Bandini, B. Chopard, and M. Tomassini (Eds.): ACRI 2002, LNCS 2493, pp. 56–68, 2002. c  Springer-Verlag Berlin Heidelberg 2002