Diffusion LMS with component-wise variable
step-size over sensor networks
ISSN 1751-9675
Received on 30th January 2015
Revised on 13th August 2015
Accepted on 27th August 2015
doi: 10.1049/iet-spr.2015.0033
www.ietdl.org
Wei Huang
✉
, Xi Yang, Duanyang Liu, Shengyong Chen
College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou 310023, People’s Republic of China
✉ E-mail: huangwei@zjut.edu.cn
Abstract: In this study, the authors propose a novel component-wise variable step-size (CVSS) diffusion distributed
algorithm fo r estimating a specific parameter over sensor networks. The novel ty of the CVSS algorithm is that step -
sizes vary from each other on different components at each iteration. They derive the steady-state value of global
mean-square devi ation (MSD) and relative MSD (RMSD). In the numerical simulations, they compare the proposed
CVSS algorithm with several other least mean square (LMS) algorithms. Results show that, when compared with these
other algorithms, the CVSS algori thm can effectively reduce steady-state value and speed up convergence rate of
RMSD while not sacrificing the convergence rate of MSD. Results also reveal that the proposed CVS S algorithm can
achieve reduced difference of steady-state values of relative estimation error on various components.
1 Introduction
Distributed estimation over sensor networks is to estimate some
parameter of interest in noisy environment using the data collected
from the network [1–5]. Compared with centralised stra tegies for
parameter estimation, distributed strategy has the advantage of saving
communication cost and enhancing the robustness of networks while
still achieving accur ate estima tion. Actually, distributed estimation has
wide applications in a wide range of fields such as target localisation
[6], environmental monitoring [7] and cognitive radio [8] etc.
Currently, the LMS-type incremental [4, 9] and diffusion [1, 5,
10–12] distributed algorithms are widely studied. In this paper, we
mainly focus on the diffusion algorithm due to its enhanced
adaptation performance and wide applications. In most previous
works, it was assumed that the step-size is spatially and temporally
fixed. Fixed step-size (FSS) in the adaptation process can hardly
deliver satisfying performances when estimating a specified
parameter. Saeed et al. [13] observed that variable step-size (VSS)
diffusion LMS algorithm can deliver higher convergence rate of
global mean-square deviation (MSD) than the diffusion LMS
algorithm with FSS.
Assume the real parameter is w
o
and w is an estimate of w
o
. The
ith components of w
o
and w are denoted by w
o
i
and w
i
, respectively.
Let us define the absolute weight error on component i as
Dw
i
=|w
i
− w
o
i
|. If the order of magnitude of the first component
is larger than that of the second component, it is usually regarded
that the first component is more accurately estimated than the
second component even if the absolute weight errors on these two
components are equal. Thus, it is also meaningful to evaluate
estimation accuracy from the perspective of relative weight error.
The relative weight error on component i can be defined as
Dw
i
=
w
i
− w
o
i
w
o
i
. (1)
Previous studies, which assumed identical step-sizes on all
components in the adaptation process, did not take into account
relative weight error.
Therefore, in this paper, we propose a novel diffusion LMS
algorithm with component-wise VSS (CVSS) for distributed
estimation in the sensor network. The novelty of our proposed
algorithm is that step-sizes on all components in the adaptation
process are not only time varying but also vary from each other. We
investigate global MSD and relative MSD (RMSD) in this paper.
Here, RMSD is used to measure relative estimation error between
estimate and the real parameter. Detailed analysis for convergence,
stability and steady state of the proposed CVSS algorithm is
performed. We compare the CVSS algorithm with several other
LMS algorithms, including the FSS algorithm and several kinds of
VSS algorithms with identical step-sizes on all components. When
approximately the same MSD are achieved by tuning relevant
parameters in each algorithm, the superiority of the CVSS algorithm
over other algorithms under study are highlighted as follows:
† The CVSS algorithm achieves lower RMSD.
† The CVSS algorithm achieves higher convergence rate of RMSD.
† The CVSS algorithm achieves smaller difference of relative
weight errors on various components.
The rest of this paper is organised as follows. In Section 2, we
describe the data model and the FSS algorithm, followed by the
introduction for a kind of VSS algorithm in Section 3. In Section
4, we present the description of the proposed CVSS algorithm,
followed by the mean stability, mean-square convergence and
steady-state analysis in Section 5. In Section 6, we present the
results of numerical simulations for comparing the CVSS
algorithm with other algorithms, followed by the comparison of
computational complexities of all algorithms under study in
Section 7. Finally, conclusions are drawn in Section 8.
Notations: In this paper, we use boldface letters and normal letters to
denote random quantities and deterministic quantities, respectively.
The absolute value of a scalar is denoted by | · |. Besides, the
notation ( · )
T
stands for transposition of matrices/vectors and the
notation E[ · ] stands for expectation operation. The notation N
k
denotes the neighbours of nodes k, that is, the nodes having a
direct link with node k, including the node k itself. Other notations
not appearing here will be defined later if necessary.
2 Problem formulation and preliminaries
2.1 Problem formulation and the FSS algorithm
We consider a connected sensor network with N nodes in a
geographical region. Each node k has access to time realisations,
IET Signal Processing
Research Article
IET Signal Process., pp. 1–9
1
&
The Institution of Engineering and Technology 2015