Fast diagonal recurrent neural networks for identification and control of nonlinear systems

In this paper we apply recurrent neural networks (RNN) to the identification and control of nonlinear systems. The RNN are used to model unknown nonlinear functions of the system. We consider specific structures of RNN in which the hidden layer is composed of self-recurrent neurones, each feedbacking its output only into itself. We also consider for this class of RNN a fast learning algorithm derived through minimisation of a generalised criterion, which gives rise to new control schemes. We then define the identification and control tasks and detail the underlying structural assumptions that are required in order to develop the partially connected recurrent neural networks adaptive control systems. We present the learning scheme and discuss the form and properties of the resulting algorithms and the convergence speed of the control algorithm.