Recurrent neural networks for signal processing trained by a new second order algorithm

A new second order algorithm based on Scaled Conjugate Gradient for training recurrent and locally recurrent neural networks is proposed. The algorithm is able to extract second order information performing two times the corresponding first order method. Therefore the computational complexity is only about two times the corresponding first order method. Simulation results show a faster training with respect to the first order algorithm. This second order algorithm is particularly useful for tracking fast varying systems.