A Stable Neural Network Controller for Linear Saturated Systems Containing Hysteresis

Abstract This paper introduces a design procedure for detemunation of weight matrices of an MLP-based controller to locally stabilize a class of multi-input nonlineaz systems consisting of a lineaz system together with a nonlineaz saturated actuator with or without hysteresis. No assumption is made on the stability of the open-loop system. Two different design problems aze considered: 1) designing a neural network which maximizes the guazanteed domain of stability of the overall system when the actuator bounds aze given, and 2) designing a neural network which guazantees the stability of the overall system in apre-specified region in state-space with a minimum actuator bound. The proposed controller has excellent advantage over the classical ones, regazding its guaranteed domain of stability. The design procedure replaces eazlier recursive neural network learning algorithms with a simple optimization problem.