Performance of neural network-based controller in the presence of bounded uncertainty

This paper examines the performance of a neural controller providing asymptotic tracking of a reference model output for a first-order time-varying plant in the presence of disturbance, noise, and unmodeled dynamics. The neural controller structure consists of feedback and filter components formulated in the form of a 3-layer feedforward network whose parameters are trained by the static backpropagation method. The number of parameters are chosen by an ad hoc procedure. Once training has been completed, and the parameters are fixed, nonlinear simulation results demonstrate the robustness of the neural network-based controller.