Learning tracking controllers for unknown dynamical systems using neural networks

A tracking control architecture based on artificial neural networks is described. The architecture comprises two separate multilayer feedforward networks with tapped delay lines. One network is trained to represent the forward dynamic model of an unknown dynamical system, while another network learns to act as a feedforward controller to control the response of the unknown system to follow desired trajectories. During a learning process, the error backpropagation method and the delta rule are used to adjust the connection strengths in the networks. To adjust the connection strengths of the network that acts as a controller, the error signals are backpropagated through the network that models the unknown system to be controlled. The tracking control of a servomotor is used as an example, and simulation results are presented to demonstrate the effectiveness of the approach