Spectrum-based design of sinusoidal RBF neural networks

This paper introduces and describes the spectrum-based design of radial basis function (RBF) neural networks. The RBF networks used in this paper work with damped sinusoidal nonlinear activation functions. The concept of the associated data spectrum is introduced, and it is shown how to apply this spectrum to find the number of hidden neurons and their internal parameters for a neural network solution of the related data processing problem. A time series prediction application is presented. The relation of the proposed method to the support vector machine method and the application of the method to select appropriate basis functions for a problem with given data are discussed.