Summary
International Symposium on Nonlinear Theory and its Applications
2008
Session Number:C2L-E
Session:
Number:C2L-E4
Weight Distribution Criterion of Neural Networks for modeling Chaotic Attractors
Guoqiu Zhang, Yi Zhao, Hong Hu,
pp.-
Publication Date:2008/9/7
Online ISSN:2188-5079
DOI:10.34385/proc.42.C2L-E4
PDF download (358.8KB)
Summary:
In this paper, we describe a new weight distribution criterion to investigate generalization of large neural networks for modeling chaotic attractors. The technique of minimum description length is first employed to estimate the optimal neural network as the standard. Weight distribution criterion then reveals that large networks whose effective structures are consistent with this standard usually exhibit good performance while others whose effective structures are distinct from this standard perform poor. We illustrate our approach to the two kinds of time series: the Rossler system and normal human pulse data.