Design of dynamic neural network based on the sensitivity analysis of model output - Details

Author：

Han, Hong-Gui (Han, Hong-Gui.) (Scholars：韩红桂) | Li, Miao (Li, Miao.) | Qiao, Jun-Fei (Qiao, Jun-Fei.) (Scholars：乔俊飞)

Indexed by：

EI Scopus PKU CSCD

Abstract：

The　capabilities　of　neural　networks　are　influenced　by　the　learning　algorithms　and　the　topologies.　Thus,　in　order　to　solve　the　problem　of　dynamic　topologies,　a　new　design　method　for　dynamic　structure　of　neural　network　is　proposed　in　this　paper.　The　dynamic　design　for　neural　network　is　based　on　the　sensitivity　analysis(SA)　of　the　model　output.　This　algorithm　can　delete　the　nodes　in　the　hidden　layer　whose　contribution　ratios　are　too　little;　and　add　new　nodes　to　the　hidden　layer　whose　ratios　are　too　large　relied　on　the　nearest　neighbor　interpolation.　Finally,　This　proposed　algorithm　is　used　to　track　the　nonlinear　functions　and　predict　the　nonlinear　systems,　the　results　demonstrate　the　good　effect　of　the　dynamic　feed-forward　neural　network(SAFNN).

Keyword：

Design Feedforward neural networks Sensitivity analysis Neural networks Topology

Author Community：

[ 1 ] [Han, Hong-Gui]College of Electronic and Control Engineering, Beijing University of Technology, Beijing 100124, China
[ 2 ] [Li, Miao]College of Electronic and Control Engineering, Beijing University of Technology, Beijing 100124, China
[ 3 ] [Qiao, Jun-Fei]College of Electronic and Control Engineering, Beijing University of Technology, Beijing 100124, China