Powell’s method applied to learning neural control of three unknown dynamic systems

This paper describes a direct neural network learning controller that is capable of improving its performance in the control of a non-linear system whose dynamics are unknown. This controller is able to improve its performance without having to identify a model of the plant, which is a necessity for most existing neural network controllers. This characteristic is obtained with a gradient-free neural network learning algorithm, Powell’s method. The performance of this new controller in the control of three non-linear systems, a pendulum, a double pendulum and a robot, was evaluated by simulations and experiments. The new controller has shown fast learning and small tracking error in the control of these systems.