A Fast Orthogonal Search Algorithm For Radial Basis Function Neural Networks

| This paper presents a fast orthogo-nalization process to train a Radial Basis Function (RBF) neural network. The traditional methods for connguring the RBF weights is to use some matrix inversion or iterative process. These traditional approaches are either time consuming or computationally expensive, and often do not converge to a solution. The goal of this paper is rst to use a fast orthogonalization process to nd the nodes of the RBF network which produce the most improvement on a target function, and then to nd the weights for these nodes. Several applications of RBF networks using this fast orthogonal search technique has been conducted and a classiication problem is presented. The problem involves classi-cation of human chromosomes, which is a highly complicated 30 dimensional and 24 class problem. Experimental results will be presented to show that the fast orthogonal search technique not only outperforms the traditional technique, but it also uses much less time and eeort. I. Introduction The growth of neural networks has been heavily innu-enced by the Radial Basis Function (RBF) neural networks. The application of the RBF network can be found in pattern recognition 1, 2], function approximation 3, 4], signal processing 4, 5], system equalization 6] and more. The two most important parameters of a RBF node, the center and the covariance matrix, have been researched throughly 2, 6]. A major issue handled by these researchers is the reduction of the number of nodes. This reduction involves clustering of the input samples without any consideration of the target function, or the convergence of the weights. The weights (the most signiicant component of any neural network) of the RBF network were left untouched by most of the researchers. This oversight is not ignorance but a conndence on the traditional approaches. For RBF weights, the traditional approaches only work when the training samples are well behaved. In real life, the training samples are not well behaved causing major problems on nding the RBF weights. The Issue of this paper is to nd a set of most signiicant nodes and their weights for a given network, using a technique which considers both the structure of the input parameter space and the target function to which the network will be trained. The traditional approach to design an RBF network is to rst select a set of network parameters (number of nodes, node centers, node covariances) and then nd the …