Neural approximations for state-space parametric identification of nonlinear systems

In the paper, the problem of designing a nonlinear parametric identifier for nonlinear discrete-time systems and noisy measurement channels is addressed. By generalizing the classical least-squares method we compute the estimation law off- line by solving a functional optimization problem. Convergence results of the estimation errors are stated and the approximate solution of the above problem is addressed by means of a feedforward neural network. A min-max technique is proposed to determine the weight coefficients of the "neural" identifier so as to estimate the system parameters to any given degree of accuracy, thus guaranteeing the boundedness of the estimation error.