Adaptive Control Using a Recurrent Neural Network Observer

A new recurrent neural network approach for on line adaptive control is presented. The resulting control scheme matches both the conventional and neural control methods. Using input-output data, a modified recurrent Elman’s network is trained to model a general non-linear discrete time system. By assuming a linearisation of this neural model a time varying adaptive observer is derived. Therefore conventional control techniques, like pole placement or quadratic optimal control, can be applied. Using simultaneous on line training of the neural network and controller synthesis, the resulting algorithm is adaptive. In order to evaluate its performance, this technique is applied to a laboratory scaled pilot plant, an airstream heating, and compared with the performance of a conventional self-tuning adaptive controller (STAC). For this particular process the experimental results allows to conclude that this control strategy is a promising technique.