A new clustering and training method for radial basis function networks

In view of the drawbacks of traditional learning algorithms for the radial basis function networks (RBFN), including improper selection of RBF centers, oversize problem of the network in the clustering stage and ill-condition in the training stage, a new clustering and training algorithm is proposed based on constructing an augmented vector consisting of both input and output, combined with the singular value decomposition (SVD) for selecting the significant basis function centers and for training the RBFN using an SVD-based recursive least squares (RLS) method so as to avoid the ill-conditioned problem. The new algorithm is superior to the RLS in convergence rate and mean square errors of training. The effectiveness and superiority of the proposed method are demonstrated via simulation examples.