Training of a three-layer dynamical recurrent neural network for nonlinear input-output mapping

A three-layer dynamical neural network with feedback and recurrent connections is proposed for nonlinear input-output mapping applications. A simple-to-implement distributed learning scheme is developed, and convergence properties of the training procedure are established. Application of the network architecture and the learning scheme to the identification of the dynamics of a nonlinear system is made, and a performance evaluation is given.<<ETX>>