Data construction method for basis selection in RBF networks

Feedforward neural networks have demonstrated an ability to learn arbitrary nonlinear mappings. Knowledge of such mappings can be of use in the identification and control of unknown or nonlinear systems. One such network, the Gaussian radial basis function (RBF) network has received a great deal of attention recently. In RBF networks, however, the problems of determination of the appropriate number of Gaussian basis functions and existence of the overlapped basis functions remain two critical issues. In order to overcome the mentioned problems, a systematic procedure, namely Data Construction Method (DCM), was proposed in this paper. A numerical example of function approximation was provided for illustration and validation. The obtained results show that DCM is a useful technique to improve the learning performance of RBF networks.