Time-optimal control by means of neural networks

The development of time-optimal controllers is often hindered by the complexity of the analytical design process. In order to avoid this difficulty, the numerical approximation of the highly nonlinear optimal control law by means of a neural network has been proposed, and an algorithm which allows the network to learn the required control actions by means of an iterative optimization process has been developed. In particular, this process involves the simultaneous minimization of both the time necessary to complete the control action as well as the final state error. It has the advantage of being very general, since no a-priori assumptions are made about the plant, and of being very flexible in that it permits the inclusion of problem-specific constraints. The performance of the technique has been investigated by applying it to both a second- and a fourth-order test plant, with very positive results being obtained in both cases