Machine-learning with cellular automata

Abstract

As the possibility of combining different classifiers into Multiple Classifier System (MCS) becomes an important direction in machine-learning, difficulties arise in choosing the appropriate classifiers to combine and choosing the way for combining their decisions. Therefore in this paper we present a novel approach - Classificational Cellular Automata (CCA), The basic idea of CCA is to combine different classifiers induced on the basis of various machine-learning methods into MCS in a non-predefined way. After several iterations of applying adequate transaction rules only a set of the most appropriate classifiers for solving a specific problem is preserved. We empirically showed that the superior results compared to AdaBoost ID3 are a direct consequence of self-organization abilities of CCA. The presented results also pointed out important advantages of CCA, such as: problem independency, robustness to noise and no need for user input. © Springer-Verlag Berlin Heidelberg 2005.