IDENTIFICATION OF PERIODIC TIME SERIES BY WEIGHT DISTRIBUTION PROJECTION OF NEURAL NETWORKS
Neural networks as black-box models have been widely employed to model the time series for the purpose of prediction or classification while this paper puts an insight into the internal structure of such black boxes. We investigate the connectivity information (i.e. weights) of trained multi-layer neural networks and thus provide a novel technique,weight distribution projection, to identify the periodic property of the time series. We examine the relationship between the proposed approach and the neural network training algorithms so as to ensure the capability of the weight distribution projection with appropriate training function. Results demonstrate that when the network is well named by Bayesian regularization algorithm its weight distribution projection can clearly exhibit the periodic characterization of the periodic data while the same projection with other training algorithms, such as Levenberg-Marquardt algorithm fails to do that.
Weight distribution Neural networks Nonlinear dynamics
TONGFENG WENG GUOQIU ZHANG YI ZHAO
Shenzhen Graduate School, Harbin Institute of Technology, Shenzhen, China
国际会议
The Second International Conference on Information & Systems Sciences(ICISS2008)(第二届信息与系统科学国际会议)
大连
英文
40-45
2008-12-18（万方平台首次上网日期，不代表论文的发表时间）