1993 Volume 113 Issue 6 Pages 394-401
In this paper, we propose a method to construct a suboptimal controller for a nonlinear system using neural networks. First we roughly approximate a nonlinear plant by a linear model. To derive the optimal control law for the approximated linear system, we introduce teh Luenberger observer and optimal feedback control law which minimizes a quadratic cost function. We consider two kinds of compensators based on neural networks. One is to compensate the observer and the other is to adjust the linear feedback controller. The neural networks are basically used to add the nonlinearlity of the plant to the approximated linear model.
After training the network, nonlinear observer could work well to estimate the state of the nonlinear plant. To represent the dynamical structure of the plant, we introduce a recurrent neural networl for the compensator of the observer. Finally, simulation results show the effectiveness of the proposed method for nonlinear optimal control problems.