2017 IEEE International Conference on Computational Science and Engineering (CSE) and IEEE International Conference
on Embedded and Ubiquitous Computing (EUC)
An Estimation of Distribution Algorithm Based
Load-balanced Clustering of Wireless Sensor
Networks
Dongbin Jiao, Liangjun Ke(Corresponding Author), Weibo Yang
State Key Laboratory for Manufacturing Systems Engineering
Xi’an Jiaotong University
Xi’an, P. R. China 710049
Email: {dbjiao, yang.wei.bo.09}@stu.xjtu.edu.cn, keljxjtu@xjtu.edu.cn
Abstract—The load-balanced clustering is a most significant
problem for WSNs with unequal load of the sensor nodes but
it is known to be an NP-hard problem. This paper introduces
a new model for the problem in which the objective function
is to maximize the overall minimum lifetime of the cluster
heads. To solve this model, we propose a novel estimation of
distribution algorithm based dynamic clustering approach (EDA-
MADCA). In EDA-MADCA, a new vector encoding is introduced
for representing a complete clustering solution, and a probability
matrix model is constructed to guide the individual search.
In addition, EDA-MADCA merges the EDA based exploration
and the local search based exploitation within the memetic
algorithm (MA) framework. A minimum-lifetime-based local
search (MLLS) strategy is presented to avoid invalid search and
enhance the local exploitation of the EDA. Experiment results
demonstrate that EDA-MADCA can prolong network lifetime, it
outperforms the existing DECA algorithm in terms of various
performance metrics.
I. I NTRODUCTION
Wireless sensor networks (WSNs) is an advanced informa-
tion acquisition technology. It has got a considerable develop-
ment in recent years. A WSN typically is equipped with a large
number of small, low power and inexpensive communication
devices called sensor nodes, which are randomly or manually
densely deployed in an unattended target area and harsh
environments. A significant class of wireless sensor networks
applications is the monitoring area, such as continuous sens-
ing, event identification (ID), event detection, local control
of actuators and location sensing [1]. The application needs
to deploy plenty of sensor nodes for continuous sensing,
data aggregating and communicating. These sensor nodes
periodically collect and process local sensing information, and
finally send it to the remote processing center which is called
base station (BS) or sink.
Generally, the sensor nodes are deployed in desolate area
or vile environment, transmitting or receiving data is over
the wireless medium, the replacement or recharging of the
embedded batteries is a very difficult and impractical process
once these nodes have been deployed. Therefore, energy is
a very precious resource for WSNs and has to be managed
wisely so as to prolong the network lifetime for the duration
978-1-5386-3221-5/17 $31.00 © 2017 IEEE
DOI 10.1109/CSE-EUC.2017.35
Jing Li
State Key Laboratory of Astronautic Dynamics
Xi’an Satellite Control Center
Xi’an, P. R. China 710043
Email: carol lee 0727@sina.com
of a specific mission [2]. For the sake of tackling the chal-
lenging issue, various technologies have been studied such as
low-power radio communication hardware [3], energy-aware
medium access control (MAC) layer protocols [4]. However,
hierarchical clustering [5], [6] has been testified to be an
effective technique to conserve sensor energy [7], [8] and also
be a promising solution to schedulable tasks .
In hierarchical clustering architecture, the sensor nodes are
divided into different clusters, each cluster consists of cluster
head (CH) and cluster membership. The set of cluster heads
forms backbone of a WSN, providing a scalable solution to
organizational networking tasks, and acting as local controllers
of network workings [9]. In each cluster, various missions
are managed by cluster head, such as data receiving, data
aggregation, transmission, authentication, task assignment [5].
However, as the leader of a cluster, cluster head has more
energy consumption than cluster membership, it signifies that
cluster head bears some extra work load due to various
activities compared to their member sensor nodes. These activ-
ities include receiving sensed data, sending control message,
data aggregation and transmitting data to the base station. In
addition, in hierarchical clustering of WSNs, cluster heads
are needed to cover a large area of interest without reducing
the service quality of the system [10]. But the sensor nodes
and cluster heads may be not “well distributed”, some cluster
heads may be overloaded owing to connecting too many sensor
nodes and long-haul communication with the base station.
Such overload not only increases communication delay and
decreases performance of overall network, but also shortens
the network lifetime.
Therefore, load-balanced distribution of cluster heads is one
of the critical issues of WSNs, it can efficiently make use of
scarce energy resources in battery operated sensor nodes. This
problem is also known Load-Balanced Clustering Problem
(LBCP), it has been proved to be NP-hard [10], and hence too
computationally expensive to find out proper size of cluster for
a large-scale WSN by exact algorithms.
In this paper, the problem of load-balanced cluster-
ing is briefly expressed as a WSN with n sensor n-
151
odes S = (s1 , s2 , . . . , sn ) and m cluster heads CH = the cluster heads and the base station and the cluster heads
(ch1 , ch2 , . . . , chm ) (The notation is summarized in Table I.) are chose randomly which may lead to energy inefficiency.
form the appropriate size of cluster of the sensor nodes around Fuzzy method is used in some of research works for load-
the cluster heads in order to reduce the overall energy utiliza- balanced clustering of the WSNs. A fuzzy logic based cluster-
tion and improve the entire network lifetime, this process is head election is proposed in [18] to reduce the energy con-
not static but rather dynamically adjusted to the cluster sizes sumption and prolong the network lifetime. In this approach,
according to the remaining energy of cluster heads. three parameters, i.e., energy, concentration and centrality, are
However, since the number of all possible clusterings is used as input variables. Compared to LEACH, the approach
very huge, it is impracticable to enumerate. Sampling is a achieved better results. However, it requires centralized means
reasonable way to deal with such a kind of problem. Esti- for electing cluster head, hence the algorithm is difficult to
mation of distribution algorithm (EDA) [11] is a sampling scalability.
based optimization tool which has been used for various Singh et al. [19] propose a novel energy-aware cluster head
academic and engineering problems [12]. Here, we propose selection in WSNs using particle swarm optimization (PSO),
an estimation of distribution algorithm (EDA) based dynamic whereas it does not take into account the cluster formation.
clustering approach for WSNs called EDA-MADCA. To the Wang et al. [20] propose an ant colony (ACO) based clustering
best of our knowledge, we are the first to employ EDA to routing algorithm in WSNs. For more surveys on evolutionary
address the load-balanced clustering problem in WSNs. algorithms for WSNs, they can be found in [21] and the
The remainder of this paper is structured as follows. Section references therein.
II reviews the related work. Section III describes a new model In load-balanced clustering problem (LBCP), the residual
for LBCP. Section IV presents EDA-MADCA that is tailored energy of cluster heads and sensor nodes is the utmost essential
for WSNs. Section V demonstrates experiments and simulation factors for balancing load and prolonging the network lifetime.
results. Finally, Section VI concludes the paper. This paper proposes a novel algorithm(i.e., EDA-MADCA) for
solving the problem.
II. R ELATED W ORK
Many metaheuristic algorithms have been proposed in III. A N EW M ODEL FOR LBCP
WSNs and here we survey some promising approaches related
A. Notations
to clustering and lifetime maximization.
Hussain et al. [13] propose a genetic algorithms (GA) The notations are listed in Table I.
based energy efficiency hierarchical clustering method which
increases the network lifetime and they adopt four fitness pa- TABLE I
M EANINGS OF THE NOTATIONS
rameters to define the fitness function. But the method ignores
load balancing between sensor nodes and cluster heads. Bari et
al. [14] present a GA-based approach for data routing by relay Notation Meaning
nodes in two-tiered WSNs. However, they did not consider n number of sensor nodes
m number of cluster heads
data communication between the sensor nodes and the cluster S set of sensor nodes
heads within each cluster. Kuila et al. [15] propose a GA CH set of cluster heads
based load-balanced clustering method for WSNs. The method BS base station
si ith node in S
forms the proper size of cluster which the maximum load of chi ith node in CH
each cluster head is minimized. The main drawback is that d Euclidian distance between si and chi
the cluster heads directly communicate with the base station R maximum communication range
roundmax maximum number of rounds
which may be impractical for large-scale network. Besides, l(chi ) chi  s lifetime
the remaining energy of sensor nodes and cluster heads were ci number of sensor nodes which can be assigned to chi
not considered per round. Esi (0) initial energy of si
Echi (0) initial energy of chi
Chakraborty et al. [16] present a differential evolution (DE) Eresidual (chi , r) residual energy of chi in rth round
based memetic algorithm which addresses routing problem- Ereq (chi , r) requested energy of chi in rth round
s with more than a thousand relay nodes. Its goal is to EDA (chi , r) aggregation energy of chi in rth round
ER (chi , r) energy depleted by chi during receiving packet in rth
accomplish the cluster heads that minimizes the maximum round
energy expenditure. Whereas, the method does not considers ET (chi , r) energy consumed during chi transmission to BS in
the cluster formation and inappropriate clustering may cause rth round
energy dissipation. Kuila et al. [17] propose a novel DE
based clustering algorithm for WSNs. A new fitness function
which takes into consideration energy expenditure of both
the cluster heads and sensor nodes is derived for prolonging B. Energy Model
the network lifetime. Meanwhile, the method embeds a local The radio energy dissipation model in [22] is used here.
improvement so as to accelerate convergence rate and get In this model, the energy consumption mainly occurs at the
better performance. However, it ignores the distance between transmitter, the power amplifier, and the receiver to run the

152
radio electronics. It adopts the free space and the multi-
path fading channel, depending on the distance between the
transmitter and receiver. The total energy expended to deliver
an l-bit packet from the transmitter to its receiver over a link
of distance d is shown below:
E(l, d) = ET (l) + ETamp (l, d) + ER (l) (1)
 of network lifetime is the time until the first cluster head
lEelec + lf s d2 + lEelec if d < dT H
E(l, d) = (2)
lEelec + lmp d4 + lEelec if d ≥ dT H
In Eq. (2), Eelec is the electronics energy which depends on
some factors [22], such as digital coding, modulation, filtering,
and spreading of the signal, but the amplifier energy may
be f s d2 or mp d4 depending on the distance between the
transmitter and the receiver and the acceptable bit-error rate,
dT H is the threshold distance. The energy consumption which
aggregates n̄ message signals of length l-bit was calculated
as:
EDA = n̄lEda (3)
where Eda is the energy required for data aggregation. The
radio channel is assumed to be symmetric.
C. Network Model
As in [17], we assume the WSNs scenario as follows:
• All the sensor nodes are randomly deployed along with
a few cluster heads within a square area.
• Cluster heads are chosen prior and the locations of sensor
nodes are known.
• The positions of all nodes (i.e., sensor nodes and cluster
heads) and base station are fixed once they are deployed.
• The message can be transmitted between a sensor node
and cluster head by wireless link.
• Each sensor node has a list of cluster heads and it can be
assigned to only one cluster head by single-hop within
their communication range per round.
• All sensor nodes have the same initial energy, and all
cluster heads have the same energy, but the energy of
cluster heads is more than the energy of sensor nodes.
The base station has no energy restriction.
• A sensor node is regarded as alive if only its energy
is larger than zero and at least one alive cluster head
is reachable within its communication range. In case a
sensor node can not find any cluster head within its
communication range, even though it may have some
remaining energy, it is still regarded as dead in the
network model.
The operation of data gathering is divided into rounds as
done in LEACH. Each round consists of two stages: set-
up and clustering. In the set-up phase, all sensor nodes and
cluster heads are assigned to a unique ID. The sensor nodes
broadcast message containing their ID by CSMA/CA MAC
protocol [23]. The cluster heads within their communication
range of the sensor nodes can collect the message and send
the local sensing information to the BS. During the clustering
153
phase, the BS receives the data gathering from cluster heads
by executing clustering algorithm. Then all cluster heads send
an announcement message to their member by single-hop. In
each cluster, the communication uses TDMA protocol [23].
D. Problem Representation
1) Definition of Lifetime: The WSNs lifetime is defined in
different ways in [24]. Among them, the popular definition
depletes all its energy. The lifetime l(chi ) of cluster head chi
is calculated as:
 
Eresidual (chi , r)
l(chi ) = (4)
Ereq (chi , r)
where the Eresidual (chi , r) and Ereq (chi , r) can be calculated
respectively as follows:
Eresidual (chi , r) = Eresidual (chi , r − 1) − Ereq (chi , r) (5)
Ereq (chi , r) = Emessage (chi , r) + Epacket (chi , r) (6)
Emessage (chi , r) = EmB (chi , r)+EmR (chi , r)+EmT (chi , r)
(7)
Epacket (chi , r) = ci ER (chi , r) + ci EDA (chi , r) + ET (chi , r)
(8)
The energy consumed by cluster head chi , denoted by
Ereq (chi , r), is composed of Emessage (chi , r) which is con-
sumed by broadcasting message and Epacket (chi , r) which
is consumed during delivering packet each round. They
are calculated by Eq. (2) and (3), respectively. Note that
Emessage (chi , r) is composed of three parts: a) EmB (chi , r),
energy consumption during cluster head chi broadcasts mes-
sage to its member sensor nodes. b) EmR (chi , r), energy
consumption during sensor node si replies with an acknowl-
edgment. c) EmT (chi , r), energy consumption during cluster
head chi sends an acknowledgment message to sensor node
si .
2) Objective Function: In [17], the objective function is
defined as the standard deviation of the cluster heads lifetime
and standard deviation of the average distance between cluster
heads and cluster membership in order to keep the load balance
between cluster heads and cluster membership. However, this
objective function ignores the distances between cluster heads
and base station. To optimize such an objective function, some
cluster heads far away from base station may be assigned
too many sensor nodes, which may lead to more energy
consumption and these cluster heads will die earlier.
In this paper, we notice that the remaining energy of each
cluster head is critical to prolong the network lifetime and a
possible way to predict the cluster head lifetime is to take
account of the remaining energy which was also considered
in [25]. This motivates us to build the objective function
as the predicted cluster head lifetime according to residual
energy Eresidual (chi , r) and request energy Ereq (chi , r) . Algorithm 1 Generate Cluster Head List
Each round, starting from each cluster head and ending at Input: n, m, S, CH, roundmax (see Table. I).
the BS, the maximum predicted lifetime L by a cluster head 1: while r ≤ roundmax do
is given as: 2: for i = 1 to n do
L = max min l(chi ) (9) 3: for j = 1 to m do
c∈Ω 1≤i≤m
4: if Eresidual (si , r) > 0 and Eresidual (chj , r) > 0
where Ω is the set of all possible clustering, variable c is
and d < R then
a clustering, (ch1 , ch2 , . . . , chm ) corresponds to c, l(chi ) is
5: count(si , r) ← count(si , r) + 1
calculated by Eq. (4). The goal is to maximize the overall
6: count(chj , r) ← count(chj , r) + 1
minimum l per round so as to extend the network lifetime.
7: ClusterHeadsList(i, count(si , r)) = j
This objective function takes into account not only the
8: end if
energy expenditure between sensor nodes and cluster heads but
9: end for
also the energy consumption between cluster heads and base
10: end for
station. By optimizing this objective function, it is helpful to
11: end while
adaptively adjust the distance between sensor nodes and cluster
Output: ClusterHeadList
heads, distance between cluster heads and base station, and the
size of clusters, thereby balancing the load of cluster heads and
reducing the data delivering overhead. As a result, the network
lifetime could be prolonged. B. Solution Representation

IV. E STIMATION OF D ISTRIBUTION A LGORITHM BASED An valid encoding is significance to solve LBCP. It is
DYNAMIC C LUSTERING A PPROACH helpful to avoid producing redundant solutions.
In EDA-MADCA, a solution is denoted by a cluster head
Estimation of distribution algorithms (EDAs) [11] are a
vector CH as show in Fig. 1. The elements in cluster heads
class of evolutionary algorithms (EAs). Compared to other
vector CH correspond to the IDs of cluster heads. The length
EAs, EDAs employ neither crossover nor mutation. They
of each solution is equal to the number of sensor nodes. In
build an explicit probabilistic model and sample promising
detail, the element chi means that sensor node si is assigned
candidate solutions according to the probabilistic model. The
to cluster head chi in clustering phase. Some elements may
probabilistic model will be updated with the generated solu-
have the same ID, which means that those sensor nodes belong
tions. The main steps of basic EDAs are described in [26].
to the same cluster head chi .
As a classic EDA, population-based incremental learning
(PBIL) [27] has already attracted significant attentions re-
cently. It is developed by combining genetic algorithms (GA) Sensor node sequence vector S1 S2 S3 S4 S5 S6 S7 S8 S9 S10

and competitive learning for maintaining the stochastic search Cluster head vector 3 3 1 2 2 1 1 3 3 1
space and using them to guide the search to explore promising
solutions by a probability vector constructed by using the elite
sub-population. Fig. 1. Solution of the example instance.
In our study, the EDA is incorporated in the Memetic
algorithm (MA) framework. Memetic algorithms [28] are As the length of each solution is equivalent to the number
widely employed as a cooperation between evolutionary or of the sensor nodes, adding or removing any sensor node will
any population-based method and individual learning or local change the length of a solution and re-clustering is required.
search procedures [12].
The idea of EDA-MADCA is to balance the energy con- C. Fitness Function
sumption among the cluster heads which transmit to the cluster Fitness function measures the quality of the population of
membership by adopting dynamically clustering, balancing the solutions. A well-defined fitness function is always problem
load of cluster heads and forming appropriate sized clusters so dependent and is helpful to increase the chance of searching
as to prolong the entire network lifetime. The EDA-MADCA promising area of the solution space. The fitness function is
is introduced in the subsections as follows. defined as follows:
A. Generation of Cluster Head List
L= max min l(chji ) (10)
In this network, since communication range of sensor nodes 1≤j≤P opsize 1≤i≤m

is limited, the sensor nodes only select neighboring cluster
heads to build wireless link. In other words, if a cluster head
is beyond the transmitting range, they can not communicate
with each other. Therefore, within its communication range,
each sensor node corresponds to some reachable cluster heads.
The generation of the cluster head list is shown in Algorithm
1.
From Eq. (10), we can see that the larger value of L, the higher
the fitness value is.
D. Design of Probability Model
In some sense, the probabilistic model describes the dis-
tribution of solution space. The standard PBIL uses a real-
valued probability vector and samples solutions according to

154
the probability vector. In general, the probabilistic model is Algorithm 2 Minimum Lifetime Based Local Search
built based on the characteristics of the solving problem. Input: elitesolution, clusterlif etime, ESsize , n, m (see
In light of the above cluster heads list and solution repre- Table I)
sentation, a single real-valued probability vector may be not 1: for i = 1 to ESsize do
good for delivering the information learnt before. Therefore, 2: [clustelif etimemin , Nmin ] =
the probability model is constructed as a probability matrix min(clusterlif etime(i, m))
P , where P = [pij ]n×m , pij = 1/count(si , r), i ∈ n, j ∈ m. 3: solution = elitesolution(i, n)
pij denotes the probability that the sensor node si is assigned 4: clustermin = find(solution == Nmin )
to the cluster head chj and modeled as a uniform distribution 5: lenmin = size(clustermin )
within its communication range. Note that if the sensor sensor 6: if lenmin = 0 then
si can not be assigned to the cluster head chj in this round, 7: sensornode = clustermin (unidrnd(lenmin ))
the probability pij = 0. 8: clusterhead ← 0
E. Selection of Elite Solutions 9: d←∞
10: if m > 1 then
At each iteration of the EDA-MADCA, to generate a new
11: for j = 1 to m do
solution, a cluster head vector CH is produced by the roulette
12: if j = Nmin then
strategy via sampling the searching space guided by the
13: d ← distance(sensornode, j)
probability matrix P . To explore the promising searching area,
14: if d < d then
the probability matrix P should be well adjusted by using
15: clusterhead ← j
some elite solutions [12]. The elite solutions form a superior
16: d ← d
sub-population. Let ESsize be the size of the sub-population.
17: end if
In our implementation, ESsize is set to θ · P opsize where
18: end if
θ ∈ [0, 1] is a parameter. Elite solutions are selected from the
19: end for
population according to the tournament selection strategy, then
20: end if
they are employed to update the probability matrix P at next
21: result = f ind(ClusterHeadsList(sensornode, :
iteration.
) == clusterhead)
F. Minimum-Lifetime-Based Local Search 22: temp = size(result)
23: if temp(2) = 0 then
In order to enhance the exploitation ability, we present a
24: solution(sensornode) ← clusterhead
MLLS strategy. The basic idea is to move some sensor nodes
25: end if
from the cluster head with the minimum lifetime to other
26: end if
cluster nodes, thereby delaying the first death of cluster heads.
27: for k to n do
Using each elite solution as input, it works as follows:
28: Count(solution(k)) ← Count(solution(k)) + 1
• Find the cluster head with the minimum lifetime and the 29: end for
set of sensor nodes assigned to the cluster head. 30: CalculateF itness
• Randomly select a sensor node from the set, then reassign 31: if F itnesselite (i) < F itness then
it to another cluster head which is closest in its candidate 32: elitesolution(i, n) = solution
cluster heads list except its previous cluster head. 33: F itnesselite (i) = F itness
• Evaluate the modified fitness value of the elite solution. 34: clusterlif etimeelite (i) = clusterlif etime
If the new fitness value is better than the previous one, 35: end if
the elite solution will be replaced by the new solution. 36: end for
Otherwise, the previous solution remains constant. Output: better elite solution
This process is described in Algorithm 2.
G. Updating Mechanism
After the superior sub-population have been modified, the α = 20%. Xij k is the following indicator function within the
the probability matrix P will be updated by using the histor- kth solution of the superior sub-population:
ical knowledge of searching and the statistics information of 
superior sub-population. The updating process is based on the 1 if chj appears before or in the ith position
k
Xij =
Hebbian-inspire rule [29] which is expressed as follows: 0 otherwise
(12)
1 
ES size

pij (g + 1) = (1 − α)pij (g) + α Xij k (11) In Eq. (11), the first term states the information inherited
ESsize j=1 from the parent and the second term states the statistics infor-
mation learnt from the superior sub-population. The parameter
where i = 1, 2, . . . n. α ∈ (0, 1) is the learning rate, implies
α controls the contribution of the parent when updating the
the maximum proportion of solutions to be chose. In our paper,
probability matrix P .

155
 V. E XPERIMENT R ESULTS AND A NALYSIS
In this section, we evaluate the performance of EDA-
MADCA by comparing it with DECA [17]. To the best of our
knowledge, DECA is the best performance in this network
model. To be fair comparison, we run EDA-MADCA and
DECA with the same parameters, simulation environment,
network topology, fitness function and performance metrics.
A. Experiment Setup
We evaluate the performance of EDA-MADCA via simu-
lations in MATLAB (version R2012b). All experiments were
run on a PC with Intel (R) Core (TM) i7, 3.60 GHz CPU,
8 G RAM and Windows 7. The values of simulation param-
eters are listed in Table II, which is similar to the ones in
[17], [22]. Specifically, sensor nodes are randomly deployed
within a square field of 200 m × 200 m, and the maximum
communication range R was set to 150 m. The number n of
sensor nodes was varied from 100 to 500 and the number m
of cluster heads was varied from 30 to 50. The initial energy
Esi (0) of each sensor node and Echi (0) of each cluster head
are 1 J and 5 J, respectively.
TABLE II
S IMULATION PARAMETER CONFIGURATION
Parameter Value
n 100 - 500
Field 200 × 200 m2
BS (100, 100), (200, 200), (100, 250)
m 30 - 50
Esi (0) 1.0 J
Echi (0) 5.0 J
R 150 m
Eelec 50 nJ/bit
f s 10 pJ/bit/m2
mp 0.0013 pJ/bit/m4
d0 87.0 m
EDA 5 nJ/bit
Packet size 4000 bits
Message size 200 bits
We take into account three scenarios to corresponding to
different locations of the base station in the network. In detail,
in Scenario 1, the base station is located at the center of
the sensing area, i.e., its coordinate is (100 m, 100 m). In
Scenario 2, the base station is located at the top right-
hand corner of the sensing area, i.e., at the (200 m, 200 m)
coordinate. In Scenario 3, the base station is located outside
the sensing area at (100 m, 250 m).
To perform EDA-MADCA and DECA, the parameters set-
tings of the two algorithms are shown in Table III. The two
algorithms are run 5 times independently on each data point.
B. Performance Measures
There are several metrics to evaluate the performance of
the clustering approach [24], [30]. In this study, the following
measures are employed to evaluate the performance of the
EDA-MADCA:
TABLE III
PARAMETER CONFIGURATION OF ALGORITHMS
EDA-MADCA Value DECA Value
P opsize 30 P opsize 30
Learning Rate (α) 0.2 Scaling F actor (F ) 0.5
θ 0.2 Cr 0.7
M ax Iteration (gmax )200 - 300 M ax Iteration (gmax )200 - 300
Running T imes 5 Running T imes 5
• F irst Dead Cluster Head (FDCH) as the number of
rounds until the first cluster head drains out of energy.
• Half Dead Cluster Head (HDCH) as the number of
rounds until the half number of cluster heads drains out
of energy.
• Last Dead Cluster Head (LDCH) as the number of
rounds until all cluster heads drain out of energy.
• Stability period as the time between the start of network
operation and FDCH.
• Instability period as the time from FDCH to LDCH.
• Balanced degree of energy consumption (BDEC) is
calculated as:
BDEC = (LDCH − F DCH)/LDCH. (13)
The BDEC measures the performance of the algorithm
to balance the consumed energy.
• N umber of alive sensor nodes per round as instan-
taneous measure the total number of sensor nodes which
have not drained out all energy.
C. Simulation Results
According to the performance metrics mentioned above, we
study the performance of EDA-MADCA and DECA from the
following aspects. For brevity, we only report part of results
here and similar results are obtained for different scale network
in three scenarios.
1) Stability Period and Network Lifetime: Fig. 2 compare
EDA-MADCA with DECA with respect to FDCH for different
scale networks in three scenarios. As can be seen that for more
than 83% data points (except two data points in Fig. 2) EDA-
MADCA can obtain better values of FDCH.
2) Instability Period and Number of Alive Sensor Nodes:
Fig. 3 illustrates the distribution of alive sensor nodes with
regards to the number of rounds for two algorithms, i.e.,
EDA-MADCA and DECA. It can be seen that EDA-MADCA
declines slower than DECA. In addition, EDA-MADCA can
lead to smaller number of inactive sensor nodes than DECA
for different scale networks in three scenarios.
3) Energy Efficiency: Energy efficiency is a critical issue
in WSNs. In this paper, we study energy efficiency of EDA-
MADCA and DECA in terms of BDEC, energy consumption
each round.
Fig. 4, EDA-MADCA can find smaller values of BDEC than
DECA on 9 out of 12 data points. It means that EDA-MADCA

156
 Scenaro - 1 Scenario - 2
600 600
EDA-MADCA EDA-MADCA
DECA DECA
500 500

First Dead Cluster Head(FDCH)

First Dead Cluster Head(FDCH)

400 400

300 300

200 200

100 100

0 0
0 200 300 400 500 0 200 300 400 500
Number of Sensor Nodes Number of Sensor Nodes

(a) (b) (c)

Fig. 2. First Dead Cluster Head (FDCH) of EDA-MADCA and DECA for 50 cluster heads in three scenarios.

(a) (b) (c)

Fig. 3. Alive sensor nodes of EDA-MADCA and DECA for different scale networks in three scenarios.

can balance better than DECA. In particular, EDA-MADCA in terms of various performance metrics. As future work, we
works prominently better than DECA when the network size will study multi-hop routing in WSNs.
is n = 400.
ACKNOWLEDGMENT
Fig. 5 clearly shows the energy consumption of EDA-
MADCA and DECA per round for different scale networks This work was supported by National Natural Science
in three scenarios. For three different scenarios, we observe Foundation of China (No. 61573277), the Open Research Fund
that EDA-MADCA consumes less energy than DECA each of the State Key Laboratory of Astronautic Dynamics under
round. It means that EDA-MADCA is more energy-efficient Grant 2016ADL-DW403. The authors would like to thank Mr.
compared with DECA, because the MLLS strategy considers Xuan Liang and Dr. Ke Shang for their kind help and valuable
that the balanced energy dissipation is helpful to improve the suggestions. The authors are also thankful to the anonymous
lifetime of cluster heads and guarantee the energy consumption referees.
of cluster heads evenly. It is noted that the energy dissipation
R EFERENCES
of EDA-MADCA does not vary with the location of the base
station. We also notice that similar results are obtained for [1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless
sensor networks: a survey,” Computer networks, vol. 38, no. 4, pp. 393–
other scale networks. 422, 2002.
[2] G. Gupta and M. Younis, “Load-balanced clustering of wireless sen-
VI. C ONCLUSION sor networks,” in Communications, 2003. ICC’03. IEEE International
Conference on, vol. 3. IEEE, 2003, pp. 1848–1852.
In this study, we have introduced a new model for LBCP [3] B. H. Calhoun, D. C. Daly, N. Verma, D. F. Finchelstein, D. D.
Wentzloff, A. Wang, S.-H. Cho, and A. P. Chandrakasan, “Design
in which the objective function is to maximize the overall considerations for ultra-low energy wireless microsensor nodes,” IEEE
minimum lifetime of the cluster heads in WSNs. We have pre- Transactions on Computers, vol. 54, no. 6, pp. 727–740, 2005.
sented EDA-MADCA, which combines EDA with local search [4] M. A. Yigitel, O. D. Incel, and C. Ersoy, “Qos-aware mac protocols for
wireless sensor networks: A survey,” Computer Networks, vol. 55, no. 8,
in the framework of memetic algorithm (MA), to optimize pp. 1982–2004, 2011.
the network lifetime in WSNs by load-balanced clustering. [5] A. A. Abbasi and M. Younis, “A survey on clustering algorithms for
This algorithm has some prominent features: It uses a valid wireless sensor networks,” Computer communications, vol. 30, no. 14,
pp. 2826–2841, 2007.
vector encoding to represent a clustering solution, and sets up a [6] A. Bari, A. Jaekel, and S. Bandyopadhyay, “Clustering strategies for
probability matrix model to guide the individual search. More- improving the lifetime of two-tiered sensor networks,” Computer Com-
over, it uses a minimum-lifetime-based local search (MLLS) munications, vol. 31, no. 14, pp. 3451–3459, 2008.
[7] J.-S. Liu and C.-H. R. Lin, “Energy-efficiency clustering protocol in
strategy to avoid invalid search. Experimental results confirms wireless sensor networks,” Ad Hoc Networks, vol. 3, no. 3, pp. 371–
that EDA-MADCA can extend network lifetime over DECA 388, 2005.

157
 (a)
Fig. 4. Blanced degree of energy consumption (BDEC) of EDA-MADCA and DECA for 30 cluster heads in three scenarios.
(a)
Fig. 5. Energy dissipation of EDA-MADCA and DECA for different scale networks in three scenarios.
[8] M. Younis, M. Youssef, and K. Arisha, “Energy-aware management for
cluster-based sensor networks,” Computer networks, vol. 43, no. 5, pp.
649–668, 2003.
[9] D. Wei, Y. Jin, S. Vural, K. Moessner, and R. Tafazolli, “An energy-
efficient clustering solution for wireless sensor networks,” IEEE trans-
actions on wireless communications, vol. 10, no. 11, pp. 3973–3983,
2011.
[10] C. P. Low, C. Fang, J. M. Ng, and Y. H. Ang, “Efficient load-
balanced clustering algorithms for wireless sensor networks,” Computer
Communications, vol. 31, no. 4, pp. 750–759, 2008.
[11] P. Larranaga and J. A. Lozano, Estimation of distribution algorithms: A
new tool for evolutionary computation. Springer Science & Business
Media, 2002, vol. 2.
[12] S.-Y. Wang and L. Wang, “An estimation of distribution algorithm-
based memetic algorithm for the distributed assembly permutation flow-
shop scheduling problem,” IEEE Transactions on Systems, Man, and
Cybernetics: Systems, vol. 46, no. 1, pp. 139–149, 2016.
[13] S. Hussain, A. W. Matin, and O. Islam, “Genetic algorithm for hierar-
chical wireless sensor networks,” Journal of Networks, vol. 2, no. 5, pp.
87–97, 2007.
[14] A. Bari, S. Wazed, A. Jaekel, and S. Bandyopadhyay, “A genetic
algorithm based approach for energy efficient routing in two-tiered
sensor networks,” Ad Hoc Networks, vol. 7, no. 4, pp. 665–676, 2009.
[15] P. Kuila, S. K. Gupta, and P. K. Jana, “A novel evolutionary approach for
load balanced clustering problem for wireless sensor networks,” Swarm
and Evolutionary Computation, vol. 12, pp. 48–56, 2013.
[16] U. K. Chakraborty, S. K. Das, and T. E. Abbott, “Energy-efficient routing
in hierarchical wireless sensor networks using differential-evolution-
based memetic algorithm,” in 2012 IEEE Congress on Evolutionary
Computation. IEEE, 2012, pp. 1–8.
[17] P. Kuila and P. K. Jana, “A novel differential evolution based clustering
algorithm for wireless sensor networks,” Applied soft computing, vol. 25,
pp. 414–425, 2014.
[18] I. Gupta, D. Riordan, and S. Sampalli, “Cluster-head election using
fuzzy logic for wireless sensor networks,” in 3rd Annual Communication
Networks and Services Research Conference (CNSR’05). IEEE, 2005,
pp. 255–260.
[19] B. Singh and D. K. Lobiyal, “A novel energy-aware cluster head
selection based on particle swarm optimization for wireless sensor
(b) (c)
(b) (c)
networks,” Human-Centric Computing and Information Sciences, vol. 2,
no. 1, pp. 2–13, 2012.
[20] G. Wang, Y. Wang, and X. Tao, “An ant colony clustering routing
algorithm for wireless sensor networks,” in Genetic and Evolutionary
Computing, 2009. WGEC’09. 3rd International Conference on. IEEE,
2009, pp. 670–673.
[21] A. M. Zungeru, L.-M. Ang, and K. P. Seng, “Classical and swarm
intelligence based routing protocols for wireless sensor networks: A sur-
vey and comparison,” Journal of Network and Computer Applications,
vol. 35, no. 5, pp. 1508–1536, 2012.
[22] W. B. Heinzelman, A. P. Chandrakasan, and H. Balakrishnan, “An
application-specific protocol architecture for wireless microsensor net-
works,” IEEE Transactions on Wireless Communications, vol. 1, no. 4,
pp. 660–670, 2002.
[23] P. Baronti, P. Pillai, V. W. Chook, S. Chessa, A. Gotta, and Y. F. Hu,
“Wireless sensor networks: A survey on the state of the art and the
802.15. 4 and zigbee standards,” Computer communications, vol. 30,
no. 7, pp. 1655–1695, 2007.
[24] I. Dietrich and F. Dressler, “On the lifetime of wireless sensor networks,”
ACM Transactions on Sensor Networks, vol. 5, no. 1, pp. 1–39, 2009.
[25] Y. Wu, S. Fahmy, and N. B. Shroff, “On the construction of a maximum-
lifetime data gathering tree in sensor networks: Np-completeness and
approximation algorithm,” in INFOCOM 2008. The 27th conference on
computer communications. IEEE, 2008, pp. 1013–1021.
[26] J. A. Lozano, Towards a new evolutionary computation: advances on
estimation of distribution algorithms. Springer Science & Business
Media, 2006, vol. 192.
[27] S. Baluja and R. Caruana, “Removing the genetics from the standard
genetic algorithm,” in Machine Learning: Proceedings of the Twelfth
International Conference, 1995, pp. 38–46.
[28] F. Neri and C. Cotta, “Memetic algorithms and memetic computing op-
timization: A literature review,” Swarm and Evolutionary Computation,
vol. 2, pp. 1–14, 2012.
[29] J. A. Lozano, “Analyzing the population based incremental learning
algorithm by means of discrete dynamical systems,” Complex Systems,
vol. 12, pp. 465–479, 2000.
[30] N. Sabor, M. Abo-Zahhad, S. Sasaki, and S. M. Ahmed, “An unequal
multi-hop balanced immune clustering protocol for wireless sensor
networks,” Applied Soft Computing, vol. 43, pp. 372–389, 2016.
158