Smoothing the control action for NARMA-L2 controllers

There have been many approaches to the use of neural networks in control systems, such as the NARMA-L2 Controller, introduced by Narendra and Mukhopadhyay [1997]. One disadvantage of this controller is that it often produces chattering in the control action. In this paper we investigate the addition of linear feedback to the NARMA-L2 controller in order to smooth the control action. We show that under certain circumstances this linear feedback can cause problems of stability and increased steady state errors, especially when the plant model is inaccurate. These problems can be reduced significantly by the use of appropriate design strategies, which are presented in this paper.