SELF-ORGANIZING NEURAL-NETWORK-BASED CLUSTERING METHOD WITH FUZZY OUTPUTS

A pattern clustering method based on the Kohonen feature mapping algorithm and the backpropagation multilayer perceptron is described. The method comprises two phases. First, the Kohonen algorithm and a simple cluster labelling procedure is applied to the training data set to divide it into labelled clusters. The data clusters are then employed to train a three-layer perceptron using the error backpropagation technique. Thus the method is self-organizing by virtue of the Kohonen algorithm and naturally produces fuzzy outputs as a consequence of the backpropagation network. The results of using the proposed method on two standard clustering problems are presented. These show that the method has superior performance compared to crisp clustering networks such as the Kohonen feature map and the ART-2 network.