A learning algorithm for recurrent neural networks and its application to nonlinear identification

A new learning algorithm is presented for a supervised learning of recurrent neural networks without using a gradient method. First, fictitious teacher signals for the outputs of each hidden unit are algebraically determined by an error backpropagation (EBP) method. Then, the weight parameters are determined by using an exponentially weighted least squares (EWLS) method. This is called EBP-EWLS algorithm which is an extension of the algorithm for a multilayer neural network. The algorithm is applied for identification of a nonlinear system to show the effectiveness of the proposed method and a new idea for nonlinear identification.