RBF neural networks with centers assignment via Karhunen-Loeve transform

A technique for assigning the Gaussian centers to a radial basis function neural network (RBFNN) based on the Karhunen-Loeve transform (KLT), is proposed by applying this technique to time series prediction, a significant performance improvement is obtained in comparison with usual prediction methods that use RBFNNs. For instance, by assigning the KLT scaled eigenvectors to the RBFNN centers yields lower prediction normalized mean squared error and requires less previous known samples than the usual technique that applies the own training set vectors to the centers. The present technique has also shown improved performance when compared with prediction based on RBFNNs that uses the K-means clustering algorithm.