Modified LMS and NLMS algorithms with non-negative weights

Non-negative constraints arise in certain system identification problems when the systems to be identified have only positive coefficients. This paper studies the stochastic behavior of modified LMS and NLMS algorithms, modified so as to only allow adaptation with positive coefficients. The mean and...

Full description

Saved in:
Bibliographic Details
Published inSignal processing Vol. 223; p. 109567
Main Authors Bershad, Neil J., Bermudez, José C.M.
Format Journal Article
LanguageEnglish
Published Elsevier B.V 01.10.2024
Subjects
Adaptive filters
Analysis
LMS algorithm
Stochastic algorithms
Stochastic algorithms
LMS algorithm
Adaptive filters
Analysis
Online AccessGet full text
ISSN0165-1684
DOI10.1016/j.sigpro.2024.109567

Cover

More Information
Summary:Non-negative constraints arise in certain system identification problems when the systems to be identified have only positive coefficients. This paper studies the stochastic behavior of modified LMS and NLMS algorithms, modified so as to only allow adaptation with positive coefficients. The mean and second moment behavior of these algorithms are analyzed for Gaussian inputs. Excellent agreement is demonstrated between the theory and Monte Carlo simulations for the LMS algorithm. •Convergence of LMS and NLMS algorithms under non-negative weight constraints.•At each iteration, the weight vector is projected onto the feasible space.•The new algorithm optimally projects the weight vector onto the feasible set.•Models are derived for the mean and mean-square behavior of the adaptive weights.•Simulation results are in excellent agreement with theoretical predictions.
ISSN:0165-1684
DOI:10.1016/j.sigpro.2024.109567