RBF networks are widely used for the quasi non-parametric estimation of real-valued multidimensional functions through a finite set of samples. However, there are numerous ways to estimate the three types of parameters involved in RBF learning: The centres and the widths of the radial functions and their weights. While the literature often deals with the problem of weight computation, many assumptions are usually made on how to choose the center locations and widths of the kernels. This paper shows adaptive ways to evaluate these parameters, based on neural network and vector quantization techniques, and relying on different assumptions on the function to approximate. It also shows the parallel between approximation of functions and of probability densities by RBF networks, and how the techniques developed for one of these domains may be extended to the other one.

Top