On the convergence of EKF-based parameters optimization for neural networks

An algorithm based on the extended Kalman filter (EKF) for optimization of parameters in neural networks is presented and a convergence analysis of the estimated parameters values to the optimal ones is made. By using results on stochastic stability of EKF in filtering for discrete-time nonlinear systems, it is proved that the approximation error of the proposed learning method is locally exponentially bounded in mean square. Such training can be performed also in batch mode and outperforms well-known training methods, as shown by means of simulation results.