Adaptive control of dynamic systems using neural networks

An application of neural networks to the adaptive control of time varying dynamic systems is presented. The essence of the approach adopted was to use a neural network model of the inverse dynamics of a plant to control that plant. A control system is described which comprises a control section, a training section, an error prediction mechanism, and am adaptation assessment mechanism (a critic). The control section, which is a neural network (Φˆ^C), is placed in series with the controlled plant, Φˆ^C is a recurrent backpropagation, neural network. It is a direct copy of Φˆ, also a recurrent backpropagation network, which is positioned across the plant to learn its inverse dynamics during operation. Φˆ^C generates control signals to the plant, taking as its inputs the desired plant output and the feedback error. The latter is produced by the error prediction mechanism to compensate for discrepancies between the desired and actual outputs due to inverse modelling uncertainties and varying environments. The control performance assessment mechanism is employed to ensure that the performance of the control system improves when the weights of Φˆ^C are updated. Simulation results are presented to illustrate the applicability of the proposed scheme to adaptive tracking control problems