Functional adaptive dual control of a class of nonlinear MIMO systems

Abstract

This paper proposes three novel neural network controllers for dual adaptive control of a class of functionally uncertain, nonlinear, multiple-input/multiple-output stochastic systems. Both Gaussian radial basis function and sigmoidal multilayer perceptron neural networks are considered for approximation of the unknown dynamic functions. Control and estimation are effected through optimization of a stochastic cost function which elicits the dual

effects of caution and probing, resulting in control laws that take into consideration the interactions between estimation and control, leading to improved control performance. The Kalman filter is used for estimation of the weights of the radial basis function network, while the extended and unscented Kalman filter are used for the multilayer perceptron case. The performances of the three schemes are compared and evaluated through extensive Monte Carlo simulations and statistical significance tests.