Systems identification using recurrent asymptotically stable neural networks

A training procedure for a class of neural networks that are asymptotically stable is presented. The training procedure is a gradient method which adapts the interconnection weights as well as the relaxation constants and the slopes of the activation functions used so as to minimize the error between the expected and obtained responses. A method for assuring that stability is maintained throughout the training procedure is also given. Such a network was used to identify the dynamic behavior of several nonlinear dynamical systems, a PUMA 560 robot and a boat based on collected rudder/heading data.<
>