Optimized feedforward neural networks for on-line identification of nonlinear models

Optimization of a class of nonlinear approximators corresponding to feedforward neural networks is investigated for on-line identification of nonlinear models in high-dimensional settings. The parameters are optimized by minimizing a cost function, which consists of the summation of two terms: a fitting penalty term and a term related to changes in the parameters. The relative influence of the two terms on the overall minimization can be tuned, according to a proper scalar. The resulting algorithm has properties of convergence and robustness. Simulation results are performed to compare its performance with classical algorithms, such as back-propagation and learning based on the extended Kalman filter, used for adjusting parameters in neural-network identification of nonlinear models. The advantages of the proposed approach are shown.