ニューラル・ネットワークの学習に及ぼす要因の影響 : -排他的論理和の場合-

概要

論文の詳細を見る
A neural network can be characterized by some factors such as its structure, initial values of the weights which are given to the learning process. Although the learning will converge rapidly if these factors are appropriately set up, it would not converge in the worst case. The influence of these factors on the learning has been investigated from a qualitative viewpoint. In this paper we try to quantitatively evaluate the influence of these factors on the number of learning. Applying various experimental design techniques to the neural network, we can grasp the degree of contribution to decreasing the number of learning as the coefficient of determination of factors. The result also indicates there exist interaction effects between factors. Furthermore, the regression equation with two explanatory variables well estimates the number of learning. This approach can be applicable to an optimum design of neural networks.