A new approach for finding the global minimum of error function of neural networks