Structure and Parameter Learning Algorithm of Jordan Type Recurrent Neural Networks

Nonlinear-system identification is very useful in a great variety of disciplines such as automatic control, mechanical diagnostics, and financial market prediction. Among state-of-the-art techniques, recurrent neural networks (RNN's) intrigue researchers by its temporal operation nature. However, time-consuming process and unsolved local minimum problem in training form a barrier for interested people. To overcome such a barrier, this paper proposes a structure and parameter learning method for recurrent neural networks in identifying both nonlinear autoregressive system (NAR) and nonlinear autoregressive system with exogeneous input (NARX). This learning scheme is then applied to model the dynamics of a Van der Pol oscillator and piezoelectric hysteresis. It is shown that the algorithm is effective in building RNN's with good generalization capability via cross validation.