Neural-Network Augmentation of Existing Linear Controllers

Published in Journal of Guidance Control and Dynamics, Vol. 28, No. 1, January-February 2005.Presented as Paper 2001-4163 at the Guidance, Navigation, and Control Conference, Montreal, Canada, 7 August 2001; received 28 July 2003; revision received 8 April 2004; accepted for publication 9 April 2004. Copyright copyright 2004 by Manu Sharma and Anthony J. Calise. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission.A method to augment existing linear controllers with a multilayer neural network is presented. The neural network is adapted online to ensure desired closed-loop response in the face of parametric plant uncertainty; no off-line training is required. The benefit of this scheme is that the neural-network output is simply added to the nominal control signal, thereby preserving the existing control architecture. Furthermore, the nominal control signal is only modified if the desired closed-loop response is not met. This method applies to a large class of modern and classical linear controllers. Stability guarantees are provided via Lyapunov-like analysis, and the efficacy of this scheme is illustrated through two numerical examples