I. Introduction
NEURAL networks have been widely used to model unknown functions based on given training input–output patterns and have been found useful in applications such as pattern recognition, prediction, and control [15], [16], [23]. Among the various proposed network architectures, the radial basis function (RBF) network is a popular one due to its simple architecture and learning scheme [1], [3], [12], [21], [26], [29]–[31]. To build an RBF network for an application, one is confronted with two problems. What is the initial configuration of the desired network, and what are the kernel functions to be used? Traditionally, the number of hidden nodes is determined by trial-and-error and the initial values for weights and parameters are randomly assigned. Network structures with different numbers of hidden nodes have to be tried before one with satisfactory or the best performance is adopted. This may take a long time and may become a big burden on the user [8], [19]. Gaussian functions are also used as kernel functions, but they may have difficulties when the desired output has abrupt changes or constant values in certain intervals [18]. To overcome this difficulty, more nodes must be used in RBF networks, resulting in a high computational complexity.