Locating the centers in an RBF network

In radial basis functions (RBF) networks both supervised and unsupervised methods are used to determine the location of centers. Though the supervised training of centers are reported to give longer training times it has been suggested that their generalising performance is better than the corresponding results when unsupervised methods are used in fixing the centers. A reduction in training time and better generalising ability dictates that we use the unsupervised methods to locate the centers and then fine tune them. Fine tuning them corresponds to finding a suitable step size to achieve a desired amount of reduction in the error. In this paper we have obtained bounds on the Hessian of the error function to get a suitable step size in the supervised training of the network