On training radial basis function neural networks using optimal fuzzy clustering

The major issues in developing radial basis functions neural networks are the determination of the appropriate number of hidden nodes and the kernel parameter values. Both of them are directly related to the underlying structure of the training data. To discover this structure we propose a new training algorithm that uses, in sequence, hierarchical fuzzy clustering and optimal clustering. The result is a network topology with a small number of nodes without significant loss of the accurate modeling performance. To verify the efficiency of the method we test three well-known cluster validity indices. Finally, the simulation results demonstrate the modeling capabilities of the proposed method.