A Stochastic Optimization Approach for Training the Parameters in Neural Networks

Recently, back-propagation method has often been applied to adapt artificial neural network for various pattern classification problems. However, an important limitation of this method is that it sometimes fails to find a global minimum of the total error function of neural network. In this paper, a hybrid algorithm which combines the modified back-propagation method and the random optimization method is proposed in order to find the global minimum of the total error function of neural network in a small number of steps. It is shown that this hybrid algorithm ensures convergence to a global minimum with probability 1 in a compact region of weight vector space. Further, several computer simulation results dealing with the problem of forcasting air pollution density, stock price, and etc. are given.