Intelligent neural network design for nonlinear control using simultaneous perturbation stochastic approximation (SPSA) optimization

Abstract

Recently intelligent control systems using neural networks (NN) have been widely applied. NNs are used to approximate complicated mathematical functions of nonlinear systems. This paper considers the design of an intelligent NN controller for nonlinear systems where the neural network is trained with the simultaneous perturbation stochastic approximation (SPSA) algorithm instead of the classical training methods. The main contribution of the SPSA method that it requires only two objective function measurements per iteration regardless of the dimension of the optimization problem. The effectiveness of the proposed scheme is demonstrated by the adaptive control of the translational oscillator / rotational actuator (TORA) system. Results of numerical simulation substantiate that the suggested approach leads to a fast way of controller designs by providing acceptable performance.