On certainty equivalence neural network controllers

Neural networks have been used as system identifiers and controllers for nonlinear adaptive control. By combining neural networks based controllers with a linear robust adaptive controller, in a multiple models, switching and tuning (MMST) framework, some recent results have established boundedness of signals when such neural networks are used for a class of discrete-time nonlinear systems. It was required that a neural network "certainty equivalence" controller should generate a control input that renders the output of the chosen model identical to the desired output for the plant. This is not a trivial task when the model is non-affine in the current control input variable, and in such a case iterative optimization routines have to be resorted to. This paper examines the issues involved in determining such an input, relaxes some of the conditions, and proposes an improved scheme in which neural network controllers can be used while assuring boundedness of signals.