A New Method for Faster Neural Networks Learning Introducing Functions of Synaptic Weights and Its Application to Nonlinear System Control

A New Method for Faster Neural Networks Learning Introducing Functions of Synaptic Weights and Its Application to Nonlinear System Control Obayashi Masanao, Member, Umesakol Kousuke, Non-member, Kobayashi Kunikazu, Non-member (Yamaguchi University) In this paper, a new method for faster neural networks learning is proposed. Characteristic of our method is that let neural networks have functions of synaptic weights instead of synaptic weights in order to improve the sensitivity of the criterion functions with respect to the synaptic weights. By constructing the functions of synaptic weights appropriately, the learning process can be significantly improved. By a simulation study of learning of controller parameters for a nonlinear crane system control, it is clarified that the speed of learning by the proposed method is much faster than that of the conventional method with moment.

[1]  Kunikazu Kobayashi,et al.  A new method for faster neural networks learning introducing functions of synaptic weights , 1999, ICONIP'99. ANZIIS'99 & ANNES'99 & ACNN'99. 6th International Conference on Neural Information Processing. Proceedings (Cat. No.99EX378).

[2]  Mingui Sun,et al.  An adaptive training algorithm for back-propagation neural networks , 1992, [Proceedings] 1992 IEEE International Conference on Systems, Man, and Cybernetics.

[3]  Tsu-Shuan Chang,et al.  An adaptive step size for backpropagation using linear lower bounding functions , 1995, IEEE Trans. Signal Process..

[4]  Kotaro Hirasawa,et al.  Computing Higher Order Derivatives in Universal Learning Networks , 1998, J. Adv. Comput. Intell. Intell. Informatics.

[5]  Marios M. Polycarpou,et al.  High-order neural network structures for identification of dynamical systems , 1995, IEEE Trans. Neural Networks.

[6]  Li-Qun Xu,et al.  Automatic learning rate optimization by higher-order derivatives , 1997, Proceedings of International Conference on Neural Networks (ICNN'97).