2001 年 121 巻 3 号 p. 662-672
This paper presents a FIR-type neural network to identify nonlinear systems as a nonlinear identification model. We derive a learning algorithm of this neural network based on the gradient descent method and the conjugate gradient method. In this learning algorithm, we derive the algorithm which updates both weight coefficients contained in FIR-synapses and gains of sigmoid functions in each unit of each layer. Also, we present a generalized sigmoid function as the nonlinear function contained in each unit in order to designate the upper and lower limits of the function. It is shown that the accuracy of the identification model based on FIR-type neural networks is superior to that of the identification model based on ordinary backpropagation neural networks and the conjugate gradient algorithm is superior to the gradient descent algorithm in point of the convergence and the learning cycle.
J-STAGEがリニューアルされました! https://www.jstage.jst.go.jp/browse/-char/ja/