Although most offline and online training algorithms based on gradient search techniques like backpropagation algorithm and its modifications or on Kalman filter approaches, it has been shown that these techniques are severely limited in their ability to find global solutions, they converge slowly, get local minimization too easily and courses oscillation. Global search techniques have been identified as a potential solution to this problem, but they are limited to offline training because of the long time of convergence. The paper is focused on presenting of applying online genetic algorithm to train recurrent artificial neural networks. Here; improvement are made on the real coding genetic algorithm by introducing a reserve elite chromosome. The new approach is tested on the Elman network (which generally suffer from very long training time) for several types of dynamic system plants. The simulation results show that the proposed algorithm is able to train ENN with less training data set in corresponding to Kalman filter training algorithm.
展开▼
