データと先験的知識を融合したニューラルネットワークの学習法

抄録

Neural network has been widely applied to business problems such as marketing, management decision making, stock market and so on. In this kind of fields, the data are usually very noisy and simple application of back propagation is not possible. For nonlinear regression problems, a method has been proposed to realize the monotonicity of input-output relation by using constraint of coefficients. In this paper, we propose a method to the BP training with prior knowledge by using gradient reference points. The effectiveness of this algorithm is shown by numerical simulation.