Identification algorithm of neural network based on dynamic generalized objective function

To improve the identification accuracy and robustness to the peak and disorder noise of dynamic neural network learning algorithm, a new algorithm is presented whose objective function is constructed by combining a deterministic function to approximate the absolute value function with least square criteria, and recursive equations for weights training of output layer are derived using Gauss-Newton iterative algorithm without any simplification. Comparison with the Karayiannis method, the new algorithm has better robustness when disorder and peak noises exist in the training samples. Simulation results show the efficiency of the proposed method.