This paper presents a methodology for evolving populations of Radial Basis Function (RBF) networks, in order to optimize the accuracy of the corresponding model predictions. The method encodes possible non-symmetric fuzzy partitions of the input space as chromosomes and then uses the non-symmetric fuzzy means algorithm to deploy an RBF network for each partition. The chromosomes are evolved through the use of a specially designed Genetic Algorithm, thus resulting to improved RBF models. The proposed approach has been applied successfully to neural network training benchmark problems.