You are on page 1of 12

Ad Hoc Networks 7 (2009) 665–676

Contents lists available at ScienceDirect

Ad Hoc Networks
journal homepage: www.elsevier.com/locate/adhoc

A genetic algorithm based approach for energy efficient routing


in two-tiered sensor networks
Ataul Bari *, Shamsul Wazed, Arunita Jaekel, Subir Bandyopadhyay
School of Computer Science, University of Windsor, 401 Sunset Avenue, Windsor, ON, Canada N9B 3P4

a r t i c l e i n f o a b s t r a c t

Article history: Higher power relay nodes can be used as cluster heads in two-tiered sensor networks to
Received 28 May 2007 achieve improved network lifetime. The relay nodes may form a network among them-
Received in revised form 16 April 2008 selves to route data towards the base station. In this model, the lifetime of a network is
Accepted 21 April 2008
determined mainly by the lifetimes of these relay nodes. An energy-aware communication
Available online 29 April 2008
strategy can greatly extend the lifetime of such networks. However, integer linear program
(ILP) formulations for optimal, energy-aware routing quickly become computationally
intractable and are not suitable for practical networks. In this paper, we have proposed
Keywords:
Sensor networks
an efficient solution, based on a genetic algorithm (GA), for scheduling the data gathering
Routing of relay nodes, which can significantly extend the lifetime of a relay node network. For
Genetic algorithm smaller networks, where the global optimum can be determined, our GA based approach
Data gathering is always able to find the optimal solution. Furthermore, our algorithm can easily handle
large networks, where it leads to significant improvements compared to traditional routing
schemes.
Ó 2008 Elsevier B.V. All rights reserved.

1. Introduction We consider a two-tiered network architecture, where


higher powered relay nodes act as cluster heads and sensor
A sensor network is an interconnection of tiny, low- nodes transmit their data directly to their respective clus-
cost, low-powered and multi-functional sensor nodes ter heads. However, the relay nodes are still battery oper-
[1,13,25]. Sensor nodes are usually powered by lightweight ated and hence, power constrained. Total depletion of the
batteries, and replacing or recharging these batteries is of- power of a relay node severely impacts the functionality
ten not feasible. Therefore, in many cases, the lifetime of a of the network, since all sensor nodes belonging to the
sensor network is over as soon as the battery power in crit- cluster of the depleted relay node will be unable to send
ical node(s) is depleted [2,13,17,25]. Recently, researchers their data to the base station and the entire cluster be-
have proposed the use of some special nodes, called relay comes inoperative. This may also put additional load on
nodes [3–5,10,18,21,22,28,29,47] within the network, for the surviving relay nodes, causing faster depletion of the
balanced data gathering, to achieve fault tolerance and to batteries of other relay nodes.
extend network lifetime. The relay nodes can be used as The lifetime of a relay node network can vary consider-
cluster heads in hierarchical sensor networks [4,28] and ably, with the actual routing scheme used [3–5,28]. Data
can be provisioned with higher energy [3,5,47] as com- communication from relay nodes to the base station may
pared to the sensor nodes. be either single-hop (cluster heads send data from their
own clusters directly to the base station) [24,25] or mul-
ti-hop (cluster heads transmit data from sensor nodes in
their own cluster, and also forward data from other relay
* Corresponding author. Tel.: +1 519 253 3000x4406.
E-mail addresses: bari1@uwindsor.ca (A. Bari), wazed@uwindsor.ca (S.
nodes, towards the base station) [28,29,31,47]. An example
Wazed), arunita@uwindsor.ca (A. Jaekel), subir@uwindsor.ca (S. Bandyo- of the multi-hop data transmission model (MHDTM) is
padhyay). shown in Fig. 1.

1570-8705/$ - see front matter Ó 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.adhoc.2008.04.003
666 A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676

As the transmit energy dissipation increases rapidly approach is also not suitable for use in conjunction with
with the distance between the source and the destination directional antennae, which has been shown to improve
nodes, the lifetime of a network is usually increased by the performance of wireless sensor and ad hoc networks
allowing the relay nodes to form an upper-tier network [12,33,52]. Therefore, in this paper, we focus on non-flow-
among themselves and using multi-hop routing, instead splitting routing schemes, using MHDTM to extend the life-
of using the single-hop data transmission model. The time of the relay node network. Traditional multi-hop
MHDTM is also the only choice for large networks, where schemes used for routing in sensor networks include
all relay nodes are not capable of transmitting their data
directly to the base station, due to their limited transmis- (i) Minimum transmission energy model (MTEM)
sion range. [24,25], where each relay node i transmits to its
In this paper, we have adopted a centralized approach nearest neighbor j, such that the relay node j is clo-
for determining an energy efficient routing scheme. Such ser to the base station than the relay node i.
an approach is appropriate when the exact positions of (ii) Minimum hop routing model (MHRM) [21,23],
the nodes can be predetermined, or when the nodes are where each relay node finds a path to the base sta-
relatively stationary after deployment. The centralized ap- tion that minimizes the number of hops.
proach has been adopted in a number of recent papers
[26,28,29,35,37,44,48] and can be used in different applica- Integer linear program (ILP) formulations to find the
tion areas, such as habitat monitoring, environment moni- optimal routing scheme, for the non-flow-splitting model,
toring, or building monitoring. In many such wireless have been proposed in the literature [3,5], and are able to
sensor networks (WSN) applications, the network is as- significantly extend the network lifetime, compared to
sumed to be relatively static during normal operation both MTEM and MHRM. But these formulations become
[14,18,21], i.e., there may be some initial movement of computationally intractable, even for moderate sized net-
nodes during the deployment phase, but the nodes are works and are not suitable for large sensor networks. How-
mostly stationary afterwards. Our proposed scheme is ever, they can be important in evaluating the performance
applicable for such models and is not intended for highly of heuristic approaches that attempt to develop energy
dynamic topologies. efficient routing schemes. In this paper, we consider two-
Although we have used a centralized approach, it is pos- tiered sensor networks and present a genetic algorithm
sible to implement a limited amount of dynamic rerouting (GA) based approach to determine a suitable routing strat-
using our model. This is useful in situations where a relay egy for upper-tier relay node networks. A preliminary ver-
node depletes all its energy, or becomes faulty. In this case, sion of this work appeared in [49]. We first compare the
the loads on the remaining relay nodes may become highly performance of our approach to optimal ILP based solu-
unbalanced, resulting in significant reduction of network tions for small networks, and show, through simulations,
lifetime. Our genetic algorithm based approach can be that our approach is able to quickly converge to the opti-
used to quickly compute a new routing schedule, based mal solution. We then show that our GA based technique
on the current network state. This updated schedule can is able to achieve significant improvements (nearly dou-
be broadcast to all nodes and used until the next update. bling the network lifetime), compared to traditional
Most papers dealing with routing in a network of relay schemes, for both small and large networks.
nodes adopt the ‘‘flow-splitting” model [18,31,38]. This
means that the flow of outgoing data from a single node,
is divided into a number of sub-flows and sent to different 2. Review
destination nodes simultaneously. This approach has a
number of limitations, including the requirement by the 2.1. Overview of genetic algorithm
relay nodes to perform complex routing functions and
costly packet level power control for nodes that are The genetic algorithm (GA) is a technique for random-
equipped with a single transmitter [28]. The flow-splitting ized search and optimization [19,20,27,41,42] and has
been applied in a wide range of studies in solving optimi-
zation problems, especially problems that are not well-
structured and interact with large numbers of possible
solutions. In this paper, we have used standard GA termi-
Direction of
data flow nology [20,27,41,42], as follows.
Sensor node A GA starts with a set of the randomly generated possi-
Relay node ble solutions, called a population. Each individual solution,
(cluster head) in the population, is known as a chromosome or an individ-
ual. Each chromosome may be represented as a simple
Radio link
string or an array of genes, which contain a part of the solu-
Base station tion. The values of genes are called alleles. The length of
each chromosome in a population should be the same. A
fitness function is provided to assign the fitness value for
each individual. This function is based on how close an
individual is to the optimal solution – the higher the fitness
Fig. 1. An example of a two-tiered sensor network. value, the closer is the solution to the optimal solution.
A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676 667

Two randomly selected chromosomes, known as par- Case (ii) The locations of the nodes are determined using
ents, can exchange genetic information in a process called a GPS system [21–23]. We assume that the nodes
recombination or crossover, to produce two new chromo- are typically stationary after deployment, so that
somes known as child or offspring. If both the parents share the GPS system needs to be operated for a very
a particular pattern in their chromosome, then the same short period, only once at the beginning or at
pattern will be carried over to the offsprings. To obtain a very infrequent intervals, to determine the loca-
good solution, mutation is often applied on randomly cho- tions of the network nodes. The assignment of
sen chromosomes, after the process of crossover. Mutation sensor nodes to clusters can be handled by exist-
helps to restore any lost genetic values when the popula- ing clustering techniques [4], and the average
tion converges too fast. Once the processes of crossover expected data rate for each relay node calculated
and mutation have occurred in a population, the chromo- accordingly (note that, if the average data gener-
somes for the next generation are selected. To ensure that ated by each sensor node is known and the dis-
the new generation is at least as fit as the previous gener- tribution of the sensor nodes into the clusters is
ation, some of the poorest performing individuals of the known, this is straightforward). The data rate
current generation can be replaced by the same number need not be uniform, but can vary from node to
of the best performing individuals from the previous gen- node. We can then compute, at some central
eration. This process is called elitism [15]. This entire cycle location, the routing decisions and broadcast
is repeated until the stopping criterion of the algorithm is the result to the entire network. Since the com-
met. The steps of a standard GA [11] are outlined in Algo- munication from each node as well as the com-
rithm 1. munication broadcast from the base station to
the nodes will be a single, small packet, the
Algorithm 1: Genetic algorithm
energy dissipated for sending/receiving these
begin two packets is insignificant, compared to the
Generate an initial population subsequent transmissions, and will not have
Compute the fitness of each individual
any substantial impact on the lifetime of the net-
while (not stopping criterion) do
Choose parents from population. work [21–23]. After receiving the message from
Perform crossover to produce offsprings. the base station, each sensor node knows which
Perform mutations. relay node is the cluster head for itself, the relay
Compute fitness of each individual. nodes can update their routing tables and the
Replace the parents by the corresponding offsprings in
new generation.
network is ready to start operating.
end
end A number of different metrics have been used in the
literature to measure the lifetime of a sensor network
[7,9,16,36,38]. In [7], the lifetime of a sensor network
has been defined as the minimum of (i) the time when
the percentage of nodes that are alive (i.e., nodes
2.2. Network model whose batteries are not depleted) drops below a spec-
ified threshold, (ii) the time when the size of the larg-
For our model, we have considered a two-tiered wire- est connected component of the network drops below a
less sensor network, with n relay nodes (acting as cluster specified threshold, and (iii) the time when the volume
heads), labeled as node numbers 1; 2; 3; . . . ; n and one base covered drops below a specified threshold. The work in
station, labeled as node number n þ 1. Let D be the set of [9] has focused on coverage and considered the net-
all sensor nodes, and Di ; 1 6 i 6 n, be the set of sensor work lifetime as the period during which the entire re-
nodes belonging to the ith cluster, which has relay node i gion can be covered. In [16], the authors have provided
as its cluster head. We assume that each sensor node be- a comprehensive survey on the definitions of network
longs to exactly one cluster i.e., D ¼ D1 [ D2 [ . . . [ Dn lifetime used in the literature and presented a general
and Di \ Dj ¼ ;, for i–j. and concise definition of the network lifetime. The
We assume that the routing schedule is computed by work of [36] has defined the lifetime of the network
some centralized entity (e.g., the base station), which is as the lifetime of the sensor node that dies first. In
not power constrained. There are two possible scenarios [38], a number of metrics are used to define the net-
for determining the positions of the relay nodes and the work lifetime, e.g., N-of-N lifetime (i.e., the mission fails
sensor nodes as follows: if any relay/gateway node dies), K-of-N lifetime (i.e., the
mission survives if a minimum of K relay/gateway
Case (i) Nodes are either placed at specified locations nodes are alive) and m-in-K-of-N lifetime (i.e., the mis-
(determined by a suitable placement strategy sion survives if all m supporting nodes and overall a
such as in [5,47]). In this case, routing decisions minimum of K relay/gateway nodes are alive) [38]. In
can be computed before the deployment of the this paper, we have used N-of-N lifetime to compare
network. The sensor nodes and the relay nodes the results in our experimental setup. However, our ap-
may be pre-configured with this information. As proach can be used with other metrics by simply mod-
soon as all the nodes are in place, the network ifying the way the fitness of a chromosome is
can start operating. calculated.
668 A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676

2.3. Routing in sensor networks using relay nodes Characteristic (i) Each relay node receives data from
the sensor nodes belonging to its own cluster, and can
The problem of routing in wireless sensor networks, un- also receive data from any number of other relay nodes.
der the ‘‘flow-splitting” model, has been extensively cov- Characteristic (ii) Each relay node i transmits data to
ered in the literature. In [29], Hou et al. attempted to one other node j (either another relay node or the base
maximize the lifetime of a sensor network by provisioning station), such that node j is within the transmission
relay and sensor nodes with additional energy. They have range of node i and j is either the base station, or j is a
formulated the problem as a mixed-integer non-linear pro- relay node that is closer to the base station than node i.
gram and proposed a heuristic algorithm, whose perfor- Characteristic (iii) The base station only receives data,
mance cannot be guaranteed. In [31], the authors there is no transmission from base station to any relay
formulated the lifetime optimization problem in terms of node.
an integer linear program (ILP), under the flow-splitting
model. In [18], Falck et al. addressed the issue of balanced In this section, we define our chromosome representa-
data gathering in sensor networks and proposed a linear tion, specify the initial population, describe the fitness
program (LP) formulation that enforces some balancing function, and the strategies for crossover and mutation
constraints in the data gathering schedule. In [21], Gupta [42].
and Younis focused on load balanced clustering and pro- We represent the chromosome, for a specific routing
posed a heuristic solution for the optimization problem. scheme, as a string of node numbers. The length of each
Routing without flow splitting (i.e., single-path routing) chromosome is always equal to the number of relay nodes.
has been studied in [3,5,8,28,51]. In [28], the authors have A routing scheme for a network with 6 relay nodes, and
presented a transformation algorithm to convert a multiple one base station, is shown in Fig. 2a and the corresponding
outgoing flow routing model to a single outgoing flow chromosome is shown in Fig. 2b. In this example, the value
routing model. In [51], the authors have investigated the of the gene in position 1 is 3, indicating that node 1 trans-
problem of maximizing the network lifetime by appropri- mits to node 3. Similarly, the value in position 3 is 7, indi-
ately placing nodes which are not energy constrained cating that node 3 transmits to node 7 (base station).
(e.g., connected to a wall outlet). In [8], the authors pro-
pose a formulation for constructing minimum-energy 3.1. Initial population and fitness function
data-aggregation trees, for a flat architecture. In [3,5], the
authors have proposed ILP formulations to maximize the Each chromosome in the initial population corresponds
lifetime of relay node networks. to a valid routing scheme. In our approach, the construc-
A number of papers have demonstrated the usefulness tion of this initial routing is based on the positions of the
of a GA based approach in sensor networks [34,39,40]. relay nodes, which is used by the base station to first create
The work of [40] focused on deriving an energy efficient a list, Ni ; 1 6 i 6 n, that contains all the one-hop neighbors
scheme that satisfy the required detection probability j, of i, such that the link i ! j; 8j 2 Ni can be used to route
[40] using a distributed GA. The work of [34] focused on data from i towards the destination (base station) through
covering the targets, using disjoint groups of sensor nodes j. Based on this information, possible routing schemes
(sensor covers), such that the lifetime of the network is (chromosomes) for the initial population are generated
maximized. The coverage problem has been transformed using a greedy approach, by randomly picking up a next-
into the Disjoint Set Cover problem [9] and solved using hop node j 2 Ni for each source node i. For our simulation
a GA. Both the above works have focused on energy effi- experiments, we used examples where the relay node dis-
cient coverage, rather than designing routing schemes for tribution was dense enough (i.e., each node has at least one
the data communication. The work of [39] focused on find- neighbour every 2p=3 angular sector) that this simple
ing an optimal traffic distribution to improve the lifetime greedy scheme will always succeed. However, we note that
of multi-sensor networks. This approach considered a the our GA does not depend on the particular algorithm
flow-splitting model for the data communication. used to generate the initial population, and any suitable
routing algorithm such as existing geographic routing
techniques, e.g., GPSR, PSGR, GeRaF, [32,50,53] can also
3. Genetic algorithm based routing be used for this purpose. It is intended only to ensure the
validity of the routes that are placed in the initial popula-
Given a collection of n relay nodes, numbered from 1 to
n, and a base station, numbered as n þ 1, along with their
locations, the objective of the GA is to find a schedule for
Source nodes
data gathering in a sensor network, such that the lifetime 2 1 2 3 4 5 6
of the network is maximized. Each period of data gathering 1
4 3
is referred to as a round [3,5,31], and the lifetime is mea-
sured by the number of rounds until the first relay node 3 4 7 7 3 7
runs out of power. In other words, we use the N-of-N met- 5
7
ric [38] to measure the network lifetime. We also assume Base 6 Destintion nodes
that the initial energy provisioned in each relay node is
Station
equal. Our non-flow-splitting routing model satisfies the
following characteristics: Fig. 2. Representation of network graph as chromosome.
A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676 669

tion. It neither attempts to find a ‘‘best” route nor does it Parent A Parent B
consider the energy consumption of any node in the path. 1 2 3 4 5 6 Source 1 2 3 4 5 6
The following informal analysis shows that the search
space for our problem is enormous. On an average, if each 3 4 7 7 3 7 Destintion 7
5 3 4 6 7
node has d valid one-hop neighbors, then the number of
n
feasible routings for a network with n nodes is Oðd Þ. In or- Crossover point Crossover point
der to select a ‘‘good” energy efficient routing scheme,
Child A Child B
from such a large number of possible solutions, within a
1 2 3 4 5 6 Source 1 2 3 4 5 6
reasonable amount of time, a heuristic search technique,
such as GA, is needed.
After generating each new individual, we need to eval- 3 4 7 7 6 7 Destintion 5 3 4 7 3 7
uate its fitness value. We compute the fitness value as the
lifetime of the network, represented by the total number of Fig. 3. Example of single point crossover.
rounds, until the first relay node runs out of battery power
[3]. shows an example of k-point crossover, where k ¼ 1 (also
As in [3], the value of the fitness function for an individ- known as single point crossover), with two parent chromo-
ual is computed as somes, Parent A and Parent B, in a network with 6 relay
nodes (numbered 1; . . . ; 6) and a base station (numbered
Einitial
Lnet ¼ ð1Þ 7). After the crossover, two new child chromosomes, Child
Emax C and Child D, are generated.
where Lnet is the network lifetime in terms of rounds and
Einitial is the initial energy of a relay node. We assume that 3.3. Mutation
the value of Einitial is known beforehand and is the same for
all relay nodes. Emax is the maximum energy dissipated by Mutation is usually applied after the process of cross-
any relay node in the individual, for the routing scheme de- over, to improve the fitness value of an individual. To per-
fined by the chromosome, in one round of data gathering. form the mutation over an individual, instead of randomly
We compute the total transmit energy, ET i ðbti ; di;j Þ, dissi- selecting a gene, as in standard GA, we select the node, say
pated in a round by each relay node i; 1 6 i 6 n, to transmit node i, in the chromosome, which dissipates the maximum
bti amount of data to another node (either a relay node or energy due to the receiving and/or transmitting of its data.
the base station) j; 1 6 j 6 n þ 1, using the following first We denote node i as the critical node as it decides the life-
order radio model [25] time of the network. The purpose of selecting node i for
mutation is to reduce the total energy dissipation by the
m node and hence, to increase the network lifetime. This
ET i ðbti ; di;j Þ ¼ a2 bti þ bbti di;j ð2Þ
can be done

where di;j is the Euclidian distance between node i and j, a2  either by replacing edge i ! j from critical node i to edge
is the transmit energy coefficient, b is the amplifier coeffi- i ! k where k is randomly chosen from nodes which are
cient and m is the path loss exponent, 2 6 m 6 4. Similarly, closer to node i than j and also closer to the base station
the receive energy, ERi ðbri Þ, dissipated in a round by each than i, or
relay node i; 1 6 i 6 n, is defined as  by diverting some incoming flow away from node i by
ERi ðbri Þ ¼ a1 bri ð3Þ randomly deleting an existing edge u ! i, and adding
an edge u ! v to some node v which is closer to the base
where bri is the number of bits received by relay node i in a station than u.
round and a1 is the receive energy coefficient.
Hence, we compute Ei , the total energy dissipated by Fig. 4 shows a portion of a network where i is the critical
each relay node i for one round of data gathering as node and B is the base station. As shown in Fig. 4a, chang-
Ei ¼ ET i þ ERi . So, our metric for energy dissipation takes ing the destination node for the critical node i, from j to k
into consideration both the transmit energy and the re- should be done in a way such that di;k < di;j and dk;B < di;B ,
ceive energy. Finally, we compute Emax ¼ maxðEi ; 8i; 1 6
i 6 nÞ.
q u q u
3.2. Selection and crossover
p p
Selection of individuals is carried out using the Rou- v v
lette-Wheel selection method [20,27], where the probabil-
i i
ity of being selected increases with the fitness value of the j k j k
individual chromosome.
To produce new offsprings from the selected parents,
we have randomly used the uniform crossover (with swap-
ping probability 0.5) or k-point crossover (k = 1, 2, or 3, se-
B B
lected randomly) for each crossover operation [20]. Fig. 3 Fig. 4. Redistribution of load on critical node as mutation.
670 A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676

where dx;y is the Euclidean distance from node x to node y. in set V  V 1 . In the former case, we are done. In the latter
Similarly, in Fig. 4b, we reduce the load on the critical node case, z is a node that is strictly closer to the base station
i, by diverting the traffic from node u, so that it is sent to compared to x. From node z, we now have to follow the
node v, instead of node i. The alternate destination node outgoing edges of GParent 2 until we again either reach the
v should be selected such that dv;B < du;B and v lies within base station or we reach a node p such that the outgoing
the transmission range of u. edge from p in GParent 2 is to a node q in set V 1 . By virtue
of property (iii), there is no cycle and node q is closer to
3.4. Validity of the crossover and the mutation operations the base station compared to either x or z. Each time we
are going from a node in set V 1 to a node in set V  V 1 or
It is convenient to construct a directed graph G ¼ ðV; EÞ vice-versa we are going closer to the base station and never
where V is a set consisting of all relay nodes and the base encountering the same relay node again. Since there are a
station. Set E consists of edges ði ! j j i; j 2 VÞ where i is a finite number of relay nodes, it must be always possible to
relay node and j is either the base station or is a relay node reach the base station. h
such that j is closer to the base station than node i and
The mutation operation deletes one edge, say x ! y,
within the transmission range of i. The base station has
from some node x and adds an edge x ! z to some other
no outgoing edge and all other nodes have one or more
node z which is either the base station or a node that is clo-
edges to some other nodes (relay nodes or the base station)
ser to the base station compared to x. The mutation scheme
as described in characteristic (ii). The following properties
clearly ensures that the graph after mutation is a OD1G
of the graph G are useful in establishing the validity of the
graph with every node except the base station having ex-
crossover mechanism:
actly 1 outgoing edge. The path from node z to the base sta-
tion only includes nodes which are all closer to the base
Property (i) The graph G cannot contain a cycle since the
station than x. A node p that earlier used node x as an inter-
existence of an edge x ! y means that node y is strictly
mediate node in the path from p to the base station will
closer to the base station than x.
continue to use x as an intermediate node in its path to
Property (ii) All paths in graph G, starting from any
the base station but will use the edge x ! z rather than
node x ultimately terminate at the base station.
x ! y. Such a path also has no cycles. This means that
Property (iii) There is a one-to-one correspondence
our crossover and mutation scheme guarantee to generate
between a chromosome and a subgraph G of graph G
valid solutions only.
such that the set of nodes of subgraph G is V and each
node in G, except the base station, has exactly one edge Example. We assume, in our discussions that the initial
to some other node in G. A graph where the outdegree population consists of valid chromosomes, so that accord-
of each node is at most 1 has been called a outdegree-1 ing to characteristic (ii), each node in the chromosome
graph (OD1G) in [43]. either transmits directly to the base station or to another
Property (iv) If there are two OD1G subgraphs G1 and node (within its transmission range) which is closer to the
G2 of graph G representing two chromosomes and there base station. This means that the node(s) closest to the
is a path P 1 (P 2 ) from node i to node j (node j to node k) base station must transmit directly to the base station in all
in G1 (G2 ) then the concatenation of paths P1 and P2 valid solutions.
gives a cycle-free path in G from node i to node k.
If we consider two parent chromosomes, Parent A and
Parent B, as shown in Fig. 5. We note that both of these
Crossover to create two child chromosomes from two chromosomes cannot be valid simultaneously, for the fol-
parent chromosomes corresponds to the process of lowing reason. In Parent A, the predecessors of node 7
(the base station) are node 3 and node 4. This means that
 selecting two OD1G subgraphs GParent 1 and GParent 2 of at least one of 3 or 4 must be closest to the base station
graph G representing the two parent chromosomes, and therefore must transmit directly to the base station
 identifying a subset V 1 of V, in all valid chromosomes (assuming Parent A is valid).
 forming a graph GChild 1 (GChild 2 ) with the edges for nodes However, this is not the case in Parent B. So, both parent
in V 1 taken from edges in GParent 1 (GParent 2 ) and the edges A and parent B cannot be valid simultaneously.
for nodes in V  V 1 taken from edges in GParent 2 (GParent 1 ).
3.5. Dynamic reconfiguration of the data communication
It is easy to see that the graph GChild 1 is a OD1G graph scheme
with every node except the base station having exactly 1
outgoing edge. The fitness function presented in Eq. (1) assumes that
Theorem 1. In the OD1G graph GChild 1 ðGChild 2 Þ resulting all relay nodes have the same energy restrictions. How-
from a crossover operation, each node corresponding to a ever, this assumption may not hold in cases where a heter-
relay node, has a path to the base station. ogenous set of relay nodes is used. Also, the energy
dissipation by different relay nodes under a fixed routing
Proof. Let x be a relay node that is in set V 1 . Starting from scheme is usually different. Therefore, it is natural that,
node x, if we follow the edges that appear in GParent 1 , we after the network has been operational for some time, dif-
ultimately reach either the base station or a relay node y ferent relay nodes will end up with different levels of resid-
in V 1 such that the edge from y in GParent 1 is to a node z ual energy, even when all relay nodes were deployed with
A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676 671

Parent A Parent B
1 2 3 4 5 6 Source 1 2 3 4 5 6

3 4 7 7 3 5 Destintion 5 3 4 6 6 7

Crossover point Crossover point

Fig. 5. Example of invalid chromosomes.

same initial energy. The residual energy may also vary due We have assumed that the average data generation rate
to the variation of the network conditions, e.g., when some of the sensor nodes and the allocation of sensor nodes to
nodes fail. We can prolong the lifetime of a network by the clusters are known. Using this information, the average
computing a new routing schedule that takes into account load on each relay node due to the data generated by the
the currently available nodes and their residual energy, at sensor nodes belonging to its own cluster is computed
some point of time after the initial deployment of the net- beforehand.
work. This can easily be done by modifying our initial fit- The performance of the genetic algorithm is greatly af-
ness function, given in Eq. (1), as follows: fected by a number of factors, such as the population size,
Lnet ¼ minðLi ; 8i; 1 6 i 6 nÞ: ð4Þ the rate of crossover and mutation, and the method of
replacement. A small population size may lead to prema-
where Lnet is the network lifetime in terms of rounds as be- ture convergence before reaching an acceptable solution.
fore, and the lifetime Li ; 8i; 1 6 i 6 n, in terms of rounds, for On the other hand, if the population size is too large, it
each relay node i is computed as follows: leads to unnecessary computations [42]. Similarly, as sta-
Eiavailable ted in [19], if the mutation rate is set too low, the GA will
Li ¼ ð5Þ converge earlier before the optimal solution is found, and
Ei
if it is too high, it may cause stability problem in the pop-
Here, Eiavailable is the current available energy of relay node i, ulation. The rate of crossover is closely related to the rate
and Ei is the energy dissipated by the relay node i, for the of mutation and affects the performance of GA in a similar
routing scheme defined by the chromosome, in one round fashion [19]. We have run a number of experiments with
of data gathering. different values of these parameters to determine the
Such re-computation of the data gathering schedule can optimal set for each of network size. In our experiments,
be performed multiple times at predetermined intervals, or we have observed that increasing the crossover (and the
can be triggered by a change in network conditions (e.g. mutation) rate allows the GA to reach the optimal solu-
node failure). Alternately, each relay node may periodically tion faster. Initial selection is performed by selecting indi-
broadcast the available residual energy, which can be used viduals having fitness values that are above a
to compute the new routes and broadcast only if it benefits predetermined threshold. We have also used an elitist
the network. Failed or depleted nodes can be removed [15] approach, where about 10% of the most promising
from the computation. Lifetime improvement after each individuals from the current generation are allowed to
rescheduling contributes to the total lifetime of the net- proceed to the next generation. Although we allow the
work. The rescheduling can be performed until the nodes GA to run for a maximum of 100 generations, we have ob-
drain out of power or no significant improvement on the served that the best solution is typically found within 20
lifetime can be achieved. generations.

4. Experimental results 4.1. Comparison for the achieved lifetime

For our experiments, we have assumed that the com- As pointed out in [40], due to the stochastic nature and
munication energy dissipation is based on the first order the complexity of the GA model, it is difficult to derive ana-
radio model, as discussed in Section 3. lytical models. Therefore, results are typically validated
using extensive simulations. We have also adopted this ap-
(1) The values for the constants are the same as in [25], proach in this paper.
as follows: The placement of relay nodes is an important issue
(a) a1 ¼ a2 ¼ 50 nJ=bit, and has been shown to be a difficult problem [45,46].
2
(b) b ¼ 100 pJ=bit=m and A number of relay node placement strategies have been
(c) the path loss exponent, m, varied from 2 to proposed in the literature, such as in [5,47]. Our ap-
6. proach does not depend on the choice of a particular
(2) The range of each sensor (relay) node is 40 m placement scheme and can be used in conjunction with
(200 m), as in [47], and any suitable scheme. The experiments were run several
(3) the initial energy of each relay node is 5 J, as in times with two different strategies for relay node
[47]. placement:
672 A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676

(i) placed at predetermined grid points, as outlined in 4000


[5], and
3500
(ii) locations selected based on sensor node
distributions. 3000
Genetic Algorithm
MTEM

Lifetimes in Rounds
MHRM

In either case, the placement ensures that all sensor 2500


nodes in the network are covered by the set of relay nodes.
2000
For the second case, we used the placement scheme out-
lined in [6], but any other appropriate placement strategy 1500
could have been used as well. We have simulated our
1000
method on upper-tier relay node networks of different
sizes. The number of sensor nodes was varied from 100 500
nodes to over 3000 nodes.
For the first set of experiments, we have varied the 0
10 20 30 40 50 60 70 80
number of relay nodes from 10 to 60, and compared the Network Size (# of Relay Nodes)
achieved lifetime with those obtained using MHRM and
MTEM. We have measured the lifetime of the network, as Fig. 7. Comparison of the lifetime achieved by different methods, when
the number of rounds, until the first relay node runs out base station is located at the corner and the data rate is fixed.

of battery power. The results, shown in Fig. 6, indicate that


our approach is able to improve network lifetime by al-
4500
most 200% compared to traditional routing schemes. For
smaller networks (up to 18 relay nodes), we were able to 4000
determine the optimal solution, using an ILP [3]. We have
3500 Genetic Algorithm
observed that, using GA, we were able to achieve the opti- MTEM
Lifetimes in Rounds
MHRM
mal lifetime with substantially less amount of time. 3000
For the second set of experiments, we have varied the
2500
number of relay nodes from 10 to 85. In this case, we have
assumed that each cluster generates data at a fixed rate 2000
(1000 bits/round). With this assumption, we measured 1500
the lifetime for two cases:
1000
(i) with the base station at the lower left corner of the 500
sensing area, and
(ii) with the base station located at the center of the 0
10 20 30 40 50 60 70 80
sensing area. Network Size (# of Relay Nodes)

We have compared the lifetimes with those obtained Fig. 8. Comparison of the lifetime achieved by different methods, when
base station is located at the center and the data rate is fixed.
using MHRM and MTEM, for both cases. Fig. 7 (Fig. 8) com-
pares the lifetimes achieved by the different methods, with
the base station placed at the corner (center) of the
network. For the third set of experiments, we also have varied the
number of relay nodes from 10 to 85. But In this case, we
have assumed that each cluster generates data at a random
4000 rate (500–2000 bits/round). As in the previous set of exper-
MTEM iments, we considered two different positions of the base
3500 MHRM
Genetic Algorithm station, and compared the achieved lifetimes with those
3000 obtained using MHRM and MTEM. Figs. 9 (Fig. 10) shows
the lifetimes achieved by the different methods with the
Lifetime in Rounds

2500 base station placed at the corner (center) of the network.


When the base station is located at the corner (center),
2000
we have calculated, with a confidence interval of 95%, that
1500 our GA based approach extends the network lifetime by
200–460% (204–260%), compared to traditional MTEM
1000 and MHRM routing schemes.
Finally, we have tested our method on a dense, large
500
network. The sensing area was taken as 480 m  480 m,
0 which included 312 relay nodes. The base station was
10 15 20 25 30 35 40 45 50 55 60 placed at the (0, 0) coordinate. The rate for receiving data
Network Size (# of Relay Nodes)
for each relay node was randomly set between 500 and
Fig. 6. Comparison of the lifetime achieved by GA solutions to that using 1500 bits/round, and the maximum number of iterations
conventional methods. was set to 300. Comparing the network lifetimes obtained
A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676 673

4500 6
10

4000
5
10
3500
Genetic Algorithm GA Time

Solution Time in sec.


Lifetimes in Rounds

MTEM CPLEX Time


3000 4
MHRM 10

2500
3
10
2000

1500 2
10

1000
1
10
500

0 0
10 20 30 40 50 60 70 80 10
10 11 12 13 14 15 16 17 18 19 20 21
Network Size (# of Relay Nodes) Network Size (# of Relay Nodes)
Fig. 9. Comparison of the lifetime achieved by different methods, when Fig. 11. Time required by the ILP and GA based approach for different size
base station is located at a corner and the data rate is random. networks.

5000

4500 exponent, m, and varied it from m ¼ 3 to m ¼ 6 (up from


4000
m ¼ 2, which was used in the experiments presented in
Genetic Algorithm
MTEM the previous sections). We also varied the size of the net-
Lifetimes in Rounds

3500 MHRM
works from 11 to 20 relay nodes. The number of sensor
3000 nodes were varied according the number of relay nodes,
2500
in order to keep the node density approximately the same,
for all cases. We have compared the lifetime of the network
2000
achieved by this genetic algorithm approach to MTEM and
1500 Direct Transmission Energy Model (DTEM), where each re-
1000 lay node transmits data directly to the base station [25].
The results of the experiments, with the path loss expo-
500
nent, m ¼ 3; 4 and 6 are shown in Figs. 12–14, respectively.
0 As before, we have measured the lifetime in terms of the
10 20 30 40 50 60 70 80
Network Size (# of Relay Nodes)
number of rounds before the first relay node runs out of
battery powers. The figures indicate that, with higher val-
Fig. 10. Comparison of the lifetime achieved by different methods, when ues of m, the lifetimes of the networks suffer considerable
base station is located at the center and the data rate is random. reduction under DTEM, as compared to the other two
schemes. Although the GA approach always performs bet-

by GA with MHRM and MTEM, in this network, we have


found that the improvement is more than 150% over these
two models. 4
x 10
2
4.2. Comparison of execution times 1.8
Genetic Algorithm
Direct Transmission
1.6
For smaller networks (up to 20 relay nodes), we were Minimum Transmission
Lifetimes in Rounds

able to determine the optimal solution, using an ILP [3]. 1.4


We were able to achieve the same optimal lifetime using 1.2
GA, but required much less time. For example, for a 18-
1
node network, our GA found the optimal solution in less
than 25 seconds, whereas the ILP in [3] required several 0.8
hours, using the solver ilog CPLEX [30]. The execution time 0.6
required by both methods for different network size is
0.4
shown in Fig. 11.
0.2
4.3. Comparison of the lifetimes in lossy environments 0
10 11 12 13 14 15 16 17 18 19 20 21
Network Size (# of Relay Nodes)
Finally, we have studied the impact of a lossy environ-
ment on the lifetime of a sensor network. To simulate a Fig. 12. Comparison of the lifetime achieved by different methods in
lossy environment, we increased the value of the path loss lossy environments, when the path loss exponent, m ¼ 3.
674 A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676

10000 conventional routing strategies, such as MTEM and MHRM.


9000
Experimental results clearly demonstrate that our ap-
Genetic Algorithm proach can significantly extend the lifetime of the network
8000 Direct Transmission
Minimum Transmission (by nearly 200% on average), compared to traditional rout-
Lifetimes in Rounds

7000 ing schemes that do not consider energy dissipation of the


6000 nodes. We are currently investigating the implementation
of the approach in a distributed manner and enhancing the
5000
GA for solving the combined problem of clustering and
4000 routing.
3000
Acknowledgement
2000

1000 The work of A. Jaekel and S. Bandyopadhyay have been


0 supported by research grants from the Natural Sciences
10 11 12 13 14 15 16 17 18 19 20 21 and Engineering Research Council of Canada (NSERC).
Network Size (# of Relay Nodes)

Fig. 13. Comparison of the lifetime achieved by different methods in


lossy environments, when the path loss exponent, m ¼ 4. References

[1] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless


sensor networks: a survey, Computer Networks 38 (2002) 393–422.
[2] K. Akkaya, M. Younis, A survey on routing protocols for wireless
500 sensor networks, IEEE Transactions on Mobile Computing 3 (3)
(2005) 325–349.
450 [3] A. Bari, A. Jaekel, S. Bandyopadhyay, Maximizing the lifetime of two-
Genetic Algorithm tiered sensor networks, in: the Proceeding of IEEE International
400 Direct Transmission Electro/Information Technology Conference (EIT 2006), MI, May
Minimum Transmission
2006, pp. 222–226.
Lifetimes in Rounds

350
[4] A. Bari, A. Jaekel, S. Bandyopadhyay, Optimal load balanced
300 clustering in two-tiered sensor networks, in: Proceedings of the
Third IEEE/CreateNet International Workshop on Broadband
250 Advanced Sensor Networks (BASENETS), CA, October 2006.
[5] A. Bari, A. Jaekel, S. Bandyopadhyay, Placement and routing of relay
200 nodes in two-tiered sensor networks, in: Proceedings of the
150 International Symposium on Broadband Access Technologies in
Metropolitan Area Networks (ISBAT2006), Niagara Falls, October
100 2006.
[6] A. Bari, A. Jaekel, S. Bandyopadhyay, Optimal placement of relay
50 nodes in two-tiered, fault tolerant sensor networks, in: 12th IEEE
Symposium on Computers and Communications (ISCC), 2007, pp.
0
10 11 12 13 14 15 16 17 18 19 20 21 159–164.
[7] D.M. Blough, P. Santi, Investigating upper bounds on network
Network Size (# of Relay Nodes)
lifetime extension for cell-based energy conservation techniques in
stationary ad hoc networks, in: Proceedings of the 8th ACM
Fig. 14. Comparison of the lifetime achieved by different methods in International Conference on Mobile Computing and Networking
lossy environments, when the path loss exponent, m ¼ 6. (ACM MobiCom 2002), September 2002, pp. 183–192.
[8] M. Borghini, F. Cuomo, T. Melodia, U. Monaco, F. Ricciato, Optimal
data delivery in wireless sensor networks in the energy and latency
domains, in: Proceedings of the First International Conference on
ter than MTEM, the differences between the two become Wireless Internet (WICON’05), 2005, pp. 138–145.
[9] M. Cardai, D.Z. Du, Improving wireless sensor network lifetime
less as the value of m increases. This is expected, since with through power aware organization, Wireless Networks (2005) 333–
higher values of m, the transmit energy dissipation in- 340.
creases sharply with the increase of transmit distance, [10] X. Cheng, D-Z. Du, L. Wang, B.B. Xu, Relay sensor placement in
wireless sensor networks, IEEE Transactions on Computers (2001).
leaving the GA approach with fewer choices. But it is <http://citeseer.ist.psu.edu/cheng01relay.html>.
important to note that, even with very high m, the GA ap- [11] Ming-Tsung Chen, Shian-Shyong Tseng, A genetic algorithm for
proach still performs better in comparatively larger net- multicast routing under delay constraint in WDM network with
different light splitting, Journal of Information Science and
works, as shown in the Fig. 14. Engineering 21 (2005) 85–108.
[12] R.R. Choudhury, X. Yang, R. Ramanathan, N.H. Vaidya, Using
directional antennas for medium access control in ad hoc
5. Conclusions
networks, in: ACM MOBICOM, 2002.
[13] C.-Y. Chong, S.P. Kumar, Sensor networks: evolution, opportunities,
In this paper, we have proposed an efficient method and challenges, Proceedings of the IEEE 91 (8) (2003) 1247–1256.
based on the GA to determine an energy efficient routing [14] S. Coleri, P. Varaiya, Optimal placement of relay nodes in sensor
networks, in: ICC, 2006.
scheme for relay node networks. Our algorithm quickly [15] L. Davis, Handbook of Genetic Algorithms, Van Nostrand Reinhold,
converges to the optimal solution (obtained using ILPs), New York, 1991.
for smaller networks. However, unlike formulations based [16] I. Dietrich, F. Dressler, On the lifetime of wireless sensor networks,
University of Erlangen, Department of Computer Science 7,
on ILP, our approach is efficient and is capable of handling Technical Report, 04/06, December 2006.
much larger networks. For large networks, we have com- [17] E.J. Duarte-Melo, M. Liu, Analysis of energy consumption and
pared the solutions based on GA to those obtained using lifetime of heterogeneous wireless sensor networks, in:
A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676 675

Proceedings of the IEEE Global Telecommunications Conference, October 04–06, 2006), ISLPED’06, ACM Press, New York, NY, 2006,
Taipei, Taiwan, vol. 1, 2002, pp. 21–25. pp. 191–196.
[18] E. Falck, P. Floren, P. Kaski, J. Kohonen, P. Orponen, Balanced data [41] I. Rechenberg, Evolutionsstrategie: Optimierung technischer System
gathering in energy-constrained sensor networks, 2004, in: S. nach Prinzipien der biologischen evolution, Frommann-Holzboog,
Nikoletseas, J.D.P. Rolim (Eds.), Algorithmic Aspects of Wireless Stuttgart, 1973.
Sensor Networks: First International Workshop (ALGOSENSORS [42] K. Sastry, D. Goldberg, G. Kendall, Genetic Algorithms, in: E. Burke,
2004, Turku, Finland, July 2004), Lecture Notes in Computer G. Kendall (Eds.), Introductory Tutorials in Optimization
Science, vol. 3121, Springer-Verlag, Berlin, Heidelberg, 2004, pp. and Decision Support Techniques, Kluwer, 2005, pp. 97–125
59–70. (Chapter 4).
[19] C. Gazen, C. Ersoy, Genetic algorithms for designing multihop [43] A. Sen, S. Murthy, S. Bandyopadyay, gStreams: A new technique for
lightwave network topologies, Artificial Intelligence in Engineering fast recovery with capacity efficient protection in WDM mesh
13 (1999) 211–221. networks, in: IEEE International Conference on Communications,
[20] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and 2007, pp. 2211–2217.
Machine Learning, Addison Wesley, Reading, MA, 1989. [44] T. Stathopoulos, L. Girod, J. Heideman, D. Estrin, K. Weeks,
[21] G. Gupta, M. Younis, Load-balanced clustering of wireless sensor Centralized Routing for Resource-Constrained Wireless Sensor
networks, in: IEEE International Conference on Communications, Networks (SYS 5), 2006. <http://research.cens.ucla.edu/pls/portal/
vol. 3, 2003, pp. 1848–1852. url/item/20A76F675835AB5DE0406180528D2E28>.
[22] G. Gupta, M. Younis, Fault-tolerant clustering of wireless sensor [45] J. Suomela, Relay placement in sensor networks, Masters thesis,
networks, in: Proceedings of IEEE WCNC, 2003, 1579–1584. University of Helsinki, Department of Computer Science,
[23] G. Gupta, M. Younis, Performance evaluation of load-balanced 2005.
clustering of wireless sensor networks, in: 10th International [46] J. Suomela, Computational Complexity of Relay Placement in Sensor
Conference on Telecommunications, vol. 2, 2003, pp. 1577–1583. Networks, SOFSEM 2006.
[24] W. Heinzelman, Application-specific protocol architectures for [47] J. Tang, B. Hao, A. Sen, Relay node placement in large scale wireless
wireless networks, Ph.D. Thesis, Massachusetts Institute of sensor networks, Computer Communications 29 (4) (2006) 490–501.
Technology, 2000. [48] Q. Wang, E.Y. Hua, Centralized transmission power scheduling in
[25] W. Heinzelman, A. Chandrakasan, H. Balakrishnan, Energy efficient wireless sensor networks, in: ACM IWCMC, 2007.
communication protocol for wireless micro-sensor networks, in: [49] S. Wazed, A. Bari, A. Jaekel, S. Bandyopadhyay, Genetic algorithm
Proceedings of the 33rd HICSS, Maui, Hawaii, 2000, pp. 3005–3014. based approach for extending the lifetime of two-tiered sensor
[26] W.B. Heinzelman, A.P. Chandrakasan, H. Balakrishnan, An networks, in: IEEE/ComSoc International Symposium on Wireless
application-specific protocol architecture for wireless microsensor Pervasive Computing (ISWPC), 2007.
networks, IEEE Wireless Communication 1 (4) (2002) 660–670. [50] Y. Xu, W-C. Lee, J. Xu, G. Mitchell, PSGR: priority-based stateless geo-
[27] J. Holland, Adaptation in Natural and Artificial Systems, University of routing in wireless sensor networks, in: IEEE International
Michigan Press, 1975. Conference on Mobile Adhoc and Sensor Systems Conference,
[28] Y.T. Hou, Y. Shi, J. Pan, S.F. Midkiff, Lifetime-optimal data routing in 2005.
wireless sensor networks without flow splitting, in: Workshop on [51] M. Yarvis, N. Kushalnagar, H. Singh, A. Rangarajan, Y. Liu, S. Singh,
Broadband Advanced Sensor Networks, San Jose, CA, 2004. Exploiting heterogeneity in sensor networks, in: Proceedings of the
[29] Y.T. Hou, Y. Shi, H.D. Sherali, S.F. Midkiff, On energy provisioning and INFOCOM 2005, 24th Annual Joint Conference of the IEEE
relay node placement for wireless sensor networks, in: IEEE Computer and Communications Societies, vol. 2, 2005, pp. 878–
International Conference on Sensor and Ad Hoc Communications 890.
and Networks (SECON), vol. 32, 2005. [52] S. Zhang, A. Datta, A directional-antenna based MAC protocol for
[30] ILOG CPLEX 9.1 Documentation. Available at the website <http:// wireless sensor networks, ICCSA, LNCS, vol. 3481, 2005, pp. 686–
www.columbia.edu/~dano/resources/cplex91_man/index.html>. 695.
[31] K. Kalpakis, K. Dasgupta, P. Namjoshi, Maximum lifetime data [53] M. Zorzi, R.R. Rao, Geographic random forwarding (GeRaF) for ad hoc
gathering and aggregation in wireless sensor networks, in: and sensor networks: energy and latency performance, IEEE
Proceedings of the IEEE International Conference on Networking, Transactions on Mobile Computing 2 (4) (2003).
2002.
[32] B. Karp, H.T. Kung, GPSR: greedy perimeter stateless routing for
wireless networks, in: Proceedings of MobiCom, 2000, pp. 243–254.
[33] Y.-B. Ko, V. Shankarkumar, N.H. Vaidya, Medium access control Ataul Bari obtained his B. Engg. degree in
protocols using directional antennas in ad hoc networks, in: IEEE mechanical engineering from the University
INFOCOM, 2000. of Rajshahi (Engg. College), Bangladesh. He
[34] C. Lai, C. Ting, R. Ko, An effective genetic algorithm for improving received his BCS[H] and M.Sc. from the Uni-
wireless sensor network lifetime, in: Proceedings of the 9th Annual versity of Windsor in 2004 and 2006 respec-
Conference on Genetic and Evolutionary Computation (London, tively. He is currently pursuing his Ph.D. in the
England, July 07–11, 2007) GECCO’07, ACM Press, New York, NY, area of energy efficient protocols for wireless
2007, p. 2260. sensor networks. His other research interests
[35] G. Lee, M. Lee, W. Seok, J. Kong, O. Byeon, A base station centralized include optical networks and bioinformatics.
simple clustering protocol for sensor networks, in: IFIP International
Federation for Information Processing, LNCS, vol. 4096, 2006, pp.
682–691.
[36] R. Madan, S. Cui, S. Lall, A. Goldsmith, Cross-layer design for lifetime
maximization in interference-limited wireless sensor networks, in:
Proceedings of 24th IEEE Conference on Computer Communications
(IEEE INFOCOM 2005), vol. 3, 2005, pp. 1964–1975.
[37] S.D. Muruganathan, D.C.F. Ma, R.I. Bhasin, O. Abraham, A. Fapojuwo,
Shamsul Wazed completed his BCS[H] and
Centralized energy-efficient routing protocol for wireless sensor
networks, IEEE Radio Communications 43 (3) (2005) S8–S13. M.Sc. from the University of Windsor in 2004
[38] J. Pan, Y.T. Hou, L. Cai, Y. Shi, S.X. Shen, Topology control for wireless and 2007 respectively. He is currently work-
sensor networks, in: Proceedings of the Ninth Annual International ing in a software development company in
Conference on Mobile Computing and Networking, 2003, pp. 286– the USA.
299.
[39] Y. Pan, X. Liu, Energy-efficient lifetime maximization and sleeping
scheduling supporting data fusion and QoS in Multi-Sensor Net,
Signal Processing 87 (12) (2007) 2949–2964.
[40] Q. Qiu, Q. Wu, D. Burns, D. Holzhauer, Lifetime aware resource
management for sensor network using distributed genetic
algorithm, in: Proceedings of the 2006 International Symposium
on Low Power Electronics and Design (Tegernsee, Bavaria, Germany,
676 A. Bari et al. / Ad Hoc Networks 7 (2009) 665–676

Subir Bandyopadhyay received the B.Sc, the


Arunita Jaekel obtained her B.Engg. degree in
B.Tech., the M.Tech. and the Ph.D. degree from
electronics and telecommunications engi-
Calcutta University and the M. Math degree
neering from Jadavpur University, India. She
from the University of Waterloo. He has
received her M.Sc. and Ph.D. degree in elec-
taught at the Calcutta University and the
trical engineering from University of Windsor,
Indian Statistical Institute in Calcutta, India
Canada. She has been with the School of
and the University of Lethbridge in Alberta,
Computer Science at University of Windsor
Canada. Since 1984 he has been at the School
since 1995, where she is currently an associ-
of Computer Science at the University of
ate professor. Her research interests include
Windsor in Ontario, Canada, where currently
optical networks, survivable topology design
he is a professor. His research interests
and wireless sensor networks.
include Optical Networks and Sensor
Networks.

You might also like