Unscented transform-based dual adaptive control of nonlinear MIMO systems

Abstract

The paper proposes a multilayer perceptron neural network controller for dual adaptive control of a class

of stochastic MIMO nonlinear systems subject to functional uncertainty. The neural network parameters are adjusted in real-time using the Unscented Kalman filter algorithm and

no pre-operational training phase is required. Dual adaptive control aims to strike a compromise between the two

control characteristics of caution and probing, leading to an improved overall performance. The system is evaluated through numerical simulations and Monte Carlo analysis. The resulting performance of the dual adaptive controller is not only consistently superior to non-dual adaptive control schemes, but also surpasses the performance of similar controllers that are based on Extended Kalman filter estimators. This reflects the enhanced accuracy of the Unscented Kalman filter estimator, despite being a local estimation method. In addition, unlike use of other estimators, the proposed approach neither requires the computation of complex Jacobian matrices as part of the design, nor the evaluation of such matrices in real-time. This renders the proposed controller inherently amenable and practical for real-time implementation.