Utilization of hierarchical structure stochastic automata for neural network learning

The backpropagation method was proposed by Rumelhart and McClelland (1986). It has contributed a lot to the literature of neural network computing. However, it has become clear that the BP method has several limitations. One of the most important limitations is that it often fails to find a global minimum of total error function of neural networks. In order to overcome this limitation, we have proposed a hybrid algorithm which combines the random optimization method with the modified backpropagation method. This hybrid algorithm has been successfully applied to several actual problems such as forecast of stock price, construction of an intelligent computer gaming system, etc. However, even this hybrid algorithm sometimes needs a large number of calculation to escape from local minima of the total error function of neural networks. In this paper, learning performance of hierarchical structure stochastic automata, is utilized to accelerate convergence of this hybrid algorithm. Several computer simulation results confirm our ideas