TRAINING ELMAN NEURAL NETWORK FOR DYNAMIC SYSTEM IDENTIFICATION USING AN ADAPTIVE LOCAL SEARCH ALGORITHM

Recurrent neural networks, especially for Elman Neural Network, have attracted the attention of researchers in the fields of Dynamic System Identification (DSI) since they took the memory unit through the context delay. In this paper, we propose an Adaptive Local Search (ALS) algorithm to train Elman Neural Network (ENN) for Dynamic Systems Identification (DSI) from a new angle instead of traditional Back Propagation (BP) based gradient descent technique. Experimental results show that the proposed algorithm has greatly effective performances in the identification of linear and nonlinear dynamic systems in comparison with BP based algorithms. The results also demonstrate that the proposed algorithm is an alternative means of training ENN when the gradient-based methods fail to find an acceptable solution. So the proposed algorithm can be regarded as a new tool or identification approach to identify dynamical systems for the auto-control systems.