Learning chaotic dynamics in recurrent RBF network

Recurrent neural network with feedback and self-connection seems suited for temporal dynamics which expresses the input-output relation depending on time. Two learning procedures of the recurrent networks for computing gradients of the error function have been proposed in the literatures. One is to use sensitivity equations, the other is to use adjoined equations. We propose a recurrent radial basis function (RBF) network and describe a procedure for finding the error gradients. We take advantage of the excellent function approximation capability of the RBF network