A Self-Organized Multiple Model Approach for Neural-Adaptive Control of Jump Nonlinear Systems

Abstract

A stochastic approach is used to control a multi-modal class of jump nonlinear stochastic systems whose underlying functions are unknown and which can change arbitrarily in time. Gaussian radial basis function neural networks are used to set up a number of local models, each characterising the different nonlinear plant modes. Being unknown, these different modes are identified on-line during control operation without resorting to a separate estimation phase. This entails detecting the occurrence of a mode change during operation. Since no information on the number of possible modes is assumed known, a self-organizing scheme is used to allocate automatically an appropriate number of local models in real time. Function identification, mode change detection and control signal generation are all based on probabilistic techniques utilising concepts of Kalman filtering, the multiple model algorithm and dual control. Simulations are given to show the effectiveness of the system for tracking a reference input, despite jumps in the unknown plant dynamics.