A Task Decomposition Algorithm Based on the Distribution of Input Pattern Vectors for Classification Problems

This paper proposes an algorithm for decomposing a multi-class classification problem into a set of two-class classification problems. The algorithm divides a set of input pattern vectors corresponding to each class into subsets according to the distribution of the selected input pattern vectors. The distribution is represented by Gaussian mixture models which are estimated by EM algorithm with MDL criterion. In this paper, the algorithm applied for constructing a modular neural network. Experimental results showed that the algorithm simplifies multi-class classification problems efficiently.