Diffusion bias-compensated least mean square algorithms over multi-agent networks

Lu Fan, Lijuan Jia*, Tang Tang, Shunshoku Kanae

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

1 Citation (Scopus)

Abstract

We study the problem of parameter estimation over multi-agent networks, where all nodes are corrupted by both input and output noise. In that case, traditional least-mean square (LMS) algorithm gives biased results. Fortunately, a bias-compensated LMS (BCLMS) method is proposed with capability to remove the noise-induced bias and get unbiased estimations. However, each node performing bias compensation independently causes an increased variance of local estimates. We propose a diffusion BCLMS algorithm based on the study that the cooperation among all nodes can significantly reduce the local increased estimation variance as well as improve performance of the network. Simulations indicate that the proposed diffusion BCLMS algorithms outperform non-cooperative BCLMS method with lower local mean-square deviation and estimation variance.

Original languageEnglish
Title of host publicationProceedings of the 37th Chinese Control Conference, CCC 2018
EditorsXin Chen, Qianchuan Zhao
PublisherIEEE Computer Society
Pages4241-4246
Number of pages6
ISBN (Electronic)9789881563941
DOIs
Publication statusPublished - 5 Oct 2018
Event37th Chinese Control Conference, CCC 2018 - Wuhan, China
Duration: 25 Jul 201827 Jul 2018

Publication series

NameChinese Control Conference, CCC
Volume2018-July
ISSN (Print)1934-1768
ISSN (Electronic)2161-2927

Conference

Conference37th Chinese Control Conference, CCC 2018
Country/TerritoryChina
CityWuhan
Period25/07/1827/07/18

Keywords

  • Bias-compensated Method
  • Diffusion Strategy
  • Increased Estimation Variance
  • Parameter estimation

Fingerprint

Dive into the research topics of 'Diffusion bias-compensated least mean square algorithms over multi-agent networks'. Together they form a unique fingerprint.

Cite this