Scalable Evolutionary Design of CA Pattern Classifier Joy Deep Nath, Pabitra Mitra, and Niloy Ganguly Department of Computer Science and Engineering, Indian Institute of Technology, Kharagpur, India {jnath,pabitra,niloy, }@iitkgp.ac.in Abstract. The paper reports a scalable evolutionary design for pat- tern recognition using Multiple Attractor Cellular Automata (MACA). MACA helps to impart non-linearity in the classifier using Hamming dis- tance based attractors. Isomorphism in MACA was exploited to make the method scalable to large classification problems involving non-linear boundaries. Extensive experimentation was performed on datasets with different topologies to establish the efficacy of the proposed method as compared to existing popular approaches like support vector machines. The classifier was applied to satellite image analysis problem. Exper- iments on different types of data sets were performed to discover the classifier’s feature selection capabilities. Key words: Cellular Automata, Pattern Classification, Genetic Algo- rithm, Feature Selection 1 Introduction The classification problem may be again viewed as partitioning the feature space and mapping the corresponding regions to different classes. Machine learning methods provide techniques to determine the boundaries of the partitions in the features space and hence help in learning the classes. The most general techniques are based on Euclidean metric nearest neighbor rule or linear discrimination which are essentially linear classifiers and are not suitable for all problems. A data dependent non-linear metric is more versatile as it helps in capturing and imparting non-linearity to classifiers inherently. Multiple attractor cellular automata (MACA), a special class of cellular automata has the inherent property of generating a non-linear partitioning of the feature space based on Hamming distance metrics [2]. In our present work, we show the results of a comprehensive set of experi- ments on MACA based classifier. We demonstrate that identification of isomor- phic MACA leads to low computational complexity as compared to the scheme in [2, 3]. Lowering the complexity has helped us in making the classifier scalable. We have studied the performance of our classification scheme on a number of real life application problems. A number of interesting phenomena were observed which provides insight for designing more efficient classifiers. The basic design of the MACA is stated in the next section.