MODEL-BASED NEURAL ALGORITHMS FOR PARAMETER-ESTIMATION

Recently theta-adaptive neural networks (TANN) were developed for para meter estimation for systems with nonlinear parametrization [1]. This paper discusses the estimation problem when additional information is available about the model structure. In particular, algorithms are pre sented for recursive parameter estimation for systems with partial non linear parametrization and for systems where the nonlinearities appear in an additive manner in the regression equation. Training procedures are developed for the neural networks which ensure the stability of t he algorithms. It is shown how the training procedures can be modified to ensure stability in the presence of a bounded disturbance. The com plexity of the neural networks needed to perform the identification ta sks is greatly reduced compared to the TANN algorithm proposed in [1]. Simulation results are presented which demonstrate the capabilities o f the algorithms. (C) Elsevier Science Inc. 1998.