论文信息 - A new approach to real‐time training of dynamic neural networks

A new approach to real‐time training of dynamic neural networks

A fast, efficient and new way of real-time training of dynamic neural networks is presented in this paper. The proposed method would be suitable for training neural networks for real-time input–output modelling of uncertain dynamic systems. The training method proposed in this paper is unique in the sense that only the outer layer weights are updated; the hidden layer weights are chosen randomly at the beginning of the process, and left unchanged. The outer layer weights are estimated with an ordinary Kalman filter. This provides a learning rate which is an optimally time-varying vector. Results of computer simulation experiments, including a comparison with conventional backpropagation, are presented. Copyright © 2003 John Wiley & Sons, Ltd.

Fahmida N. Chowdhury

[1] Hideaki Sakai,et al. A real-time learning algorithm for a multilayered neural network based on the extended Kalman filter , 1992, IEEE Trans. Signal Process..

[2] Andrew Chi-Sing Leung,et al. On the Kalman filtering method in neural network training and pruning , 1999, IEEE Trans. Neural Networks.

[3] Giuseppe De Nicolao,et al. Regularization networks: fast weight calculation via Kalman filtering , 2001, IEEE Trans. Neural Networks.

[4] A. Calise,et al. On approximate NN realization of an unknown dynamic system from its input-output history , 2000, Proceedings of the 2000 American Control Conference. ACC (IEEE Cat. No.00CH36334).

[5] Kumpati S. Narendra,et al. Identification and control of dynamical systems using neural networks , 1990, IEEE Trans. Neural Networks.

[6] Youmin Zhang,et al. A Fast U-d Factorization-based Learning Algorithm with Applications to Nonlinear System Modeling and Identification , 2022 .