Performance evaluation of self generating radial basis function for function approximation

In this paper, we analyze the learning ability of various design methods for radial basis function network, especially for function approximation problem. We consider the three methods; the k-means clustering algorithm by Moody and Darken (1989), the orthogonal least squares method by S. Chen et al. (1991), and the maximum absolute error selection (MAE) method by the authors (1993). We compare the learning ability through several function approximation problems. We show that MAE method requires the least number of basis functions to achieve the specified model error among these three methods.