Abstract
In this study, we propose multilayer neural networks with adjustable intermediate elements and show their higher learning capability by computer simulation. A well-known backpropagation learning algorithm has a problem because of local minima. In the previous work, Yokoi et al. have proposed intermediate elements, which intervene between the two layers and function to perform feature detection and categorization, in order to avoid increasing of local minima and to increase the learning capability. They have demonstrated that the learning capability is increased due to intermediate elements. However, evaluation of intermediate elements under the optimum learning condition and adjusting of intermediate elements have not yet been discussed. The optimum learning condition is needed to adequately evaluate the best performance of neural networks. Adjusting of intermediate elements is expected to more increase the learning capability though the design of networks becomes complex. In this paper, we presented a Design of Experiments based optimization method to obtain the optimum learning condition such as the number of elements in hidden layers, a learning rate and momentum, and showed that the learning capability in the case of using adjustable intermediate elements is much higher than constant intermediate elements.
