The influence of clustering techniques in the RBF networks generalization

This paper presents and evaluates clustering techniques for the training of radial basis function neural networks (Moody and Dark 1989; Broomhead and Lowe 1988). The clustering techniques define the centers of the radial basis functions used by these networks. Therefore, the main purpose is to verify the influence of different clustering techniques in the performance of RBF networks. The K-means (MacQueen 1967), widely used for the centers choice in RBF networks, is contrasted with others clustering techniques, such as, optimal adaptive K-means (Chinrungrueng and Sequin 1995), DHB (Duda and Hart), DHF (Ismail et al. 1984), AFB (Ismail and Kamel 1986) and ABF (Ismail and Kamel 1986). The authors of these techniques claim that they are more likely to converge to an optimal or near-optimal configuration. Initially, the algorithms and a complete description of each technique are presented. Finally, using these techniques the RBF performance in a pattern recognition task is evaluated.