Response of a feedback system with a neural network controller in the presence of disturbances

A neural network controller is presented as an approach to the reduction of the effect of disturbances on the performance of a feedback control system with a nonlinear plant. A backpropagation neural network was trained and subsequently used as a model-based controller for a one-dimensional robot arm. The simulation results obtained show that the neural network controller can perform quite well on a highly nonlinear system even in the presence of high levels of disturbances. A neural network controller trained with noisy data can adapt to the presence of disturbances better than a neural network controller trained with clean data. The neural network controller trained with noisy data was also found to perform better than a conventional model-based controller in the presence of disturbances. In the absence of disturbances, the former also matches the performance of the latter.<>