Abstract
One of the representative algorithms for learning of an artificial neural networks (ANN) is the back propagation method. But, it is sometimes difficult to converge and learn efficiently, because it minimizes the error function of the outputs for each learning data set in order and not simultaneously. In order to minimize the error functions for all learning data sets simultaneously, a new learning algorithm is proposed using the extended Bayesian method. However, a dificult but important problem in this method is to determine optimally the number of hidden layer units Lm for ANN and the parameter λ2 for the extended Bayesian method. Thus, the determination method of optimal λ2 and Lm by Akaike Bayesian information criteria is proposed. As a result of having applied it to consolidation and seismic problems, it is comparatively clear that the proposed method has faster convergence and higher learning capability.
