Database Reference
In-Depth Information
have proposed an algorithm for distributed classification and estimation
in wireless sensor networks. The data is modeled as
y
i
=
θ
+
T
i
+
ν
i
where
y
i
's are the measurements at each sensor node,
θ
∈
R
is the com-
mon unknown parameter,
T
i
∈{
are the unknown discrete terms
which denotes the class label of each node, and
ν
i
's are zero mean iid
Gaussian random variables with finite variance. The goal of each node is
to estimate
θ
and
T
i
. Due to the existence of
θ
, the final estimate would
require a consensus algorithm over all nodes. The paper develops a max-
imum likelihood estimator to estimate the unknown parameter and infer
the class labels in the distributed setting using a gossip based protocol.
The paper further proposes an EM algorithm for the case in which the
T
i
's are assumed to be iid Bernoulli trials. Experimental results show
that the proposed methods have similar convergence rates compared to
existing methods but stronger robustness in various situations, for in-
stance when the offset of the misbehaving sensors is not known, or in
the presence of outliers.
Further reading
: There are a number of other papers in these areas
which we point out here. Sun and Qi [54] discuss the fact that there
exist a particular set of features and a particular classifier which has the
best performance, in terms of highest accuracy with the least number of
features used. The authors discuss a method of dynamic target classifi-
cation in which an optimal set of features and classifiers are determined
based on some minimal value of cost function. Experimental results show
that this approach can significantly reduce the computational time and
also achieve better classification accuracy.
Eyal
et al.
[26] present an asynchronous algorithm for distributed
data classification over arbitrary connected networks. They present a
generic algorithm converges for any connected topology, data and class
distribution. The paper presents examples of two specific instantiations
of the generic algorithm: (1) a distance based classification scenario akin
to the famous
k
-means clustering, and (2) a gaussian mixture model data
distribution with expectation maximization for learning latent factors.
Duarte and Hu [22] discuss the application of vehicle classification in
sensor networks. Each sensor in the WSN is equipped with a micro-
phone or a geophone. Upon detection of the presence of a vehicle in
the vicinity of the sensor, the on-board processor first extracts features
in the frequency domain using FFT. The next step is to use a local
classifier at each node to generate a preliminary hypothesis about the
observation using only the data present at that node. The authors have
experimented with 3 classifiers - a
k
-nn based classifier, a maximum like-
0
,
1
}
