Computer Networks: Habib M. Ammari

Computer Networks 57 (2013) 19131935
Contents lists available at SciVerse ScienceDirect
Computer Networks
journal homepage: www.elsevier.com/locate/comnet
On the energy-delay trade-off in geographic forwarding

in always-on wireless sensor networks: A multi-objective
optimization problem
Habib M. Ammari
Wireless Sensor and Mobile Ad-hoc Networks (WiSeMAN) Research Lab, Department of Computer and Information Science, College of Engineering and
Computer Science, University of Michigan-Dearborn, Dearborn, Michigan 48128, United States
a r t i c l e
i n f o
Article history:
Received 5 May 2012
Received in revised form 26 February 2013
Accepted 22 March 2013
Available online 2 April 2013
Keywords:
Wireless sensor networks
Data forwarding
Slicing
Optimization
Trade-off
Energy
Delay
Proxy forwarders
a b s t r a c t
The design and development of multi-hop wireless sensor networks are guided by the specic requirements of their corresponding sensing applications. These requirements can be
associated with certain well-dened qualitative and/or quantitative performance metrics,
which are application-dependent. The main function of this type of network is to monitor a
eld of interest using the sensing capability of the sensors, collect the corresponding
sensed data, and forward it to a data gathering point, also known as sink. Thus, the longevity of wireless sensor networks requires that the load of data forwarding be balanced
among all the sensor nodes so they deplete their battery power (or energy) slowly and uniformly. However, some sensing applications are time-critical in nature. Hence, they should
satisfy strict delay constraints so the sink can receive the sensed data originated from the
sensors within a specied time bound. Thus, to account for all of these various sensing
applications, appropriate data forwarding protocols should be designed to achieve some
or all of the following three major goals, namely minimum energy consumption, uniform
battery power depletion, and minimum delay. To this end, it is necessary to jointly consider
these three goals by formulating a multi-objective optimization problem and solving it. In
this paper, we propose a data forwarding protocol that trades off these three goals via slicing the communication range of the sensors into concentric circular bands. In particular, we
discuss an approach, called weighted scale-uniform-unit sum, which is used by the source
sensors to solve this multi-objective optimization problem. Our proposed data forwarding
protocol, called Trade-off Energy with Delay (TED), makes use of our solution to this multiobjective optimization problem in order to nd a best trade-off of minimum energy consumption, uniform battery power depletion, and minimum delay. Then, we present and
discuss several numerical results to show the effectiveness of TED. Moreover, we show
how to relax several widely used assumptions in order to enhance the practicality of our
TED protocol, and extend it to real-world network scenarios. Finally, we evaluate the performance of TED through extensive simulations. We nd that TED is near optimal with
respect to the energy delay metric. This simulation study is an essential step to gain more
insight into TED before implementing it using a sensor test-bed.
2013 Elsevier B.V. All rights reserved.
1. Introduction
Tel.: +1 313 593 5239.

E-mail address: hammari@umd.umich.edu
1389-1286/$ - see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.comnet.2013.03.009
Recent advances in sensor technology and wireless

communications have enabled the design and development of inexpensive, large-scale wireless sensor networks,
1914
H.M. Ammari / Computer Networks 57 (2013) 19131935
which are suitable for various civilian applications, such as

health environments monitoring; natural applications,
such as seism monitoring; and military applications, such
as battleelds surveillance, to name a few. A wireless sensor network (WSN) is a collection of tiny, low-powered
sensors that communicate with each other through multi-hop wireless links, and collaborate together to accomplish a common task. This type of network suffers from
severe limitations of the sensors with respect to their battery power, computation, communication, and storage
capabilities. It is worth mentioning that battery power
(or energy) is the most crucial resource in WSNs. In fact,
when the sensors are deployed in hostile environments,
such battleelds, it is sometimes difcult or even impossible to recharge or replenish their battery. It is well known
that the main function of WSNs is to monitor a eld of
interest using the sensing capability of the sensors, collect
the corresponding sensed data, and forward it to a central
data gathering point, called the sink. Thus, it is necessary to
design energy-efcient data forwarding protocols for
WSNs, which are an essential component and critical
determinant of the effectiveness of the network design.
These protocols should guarantee uniform energy depletion
of the sensors in the network. This helps the sensors operate for longer periods of time, thus extending the network
operational lifetime [31]. However, ensuring the longevity
of WSNs becomes a challenging issue for sensing applications with strict source-to-sink delay (or simply delay)
constraints [32], [33]. These delay constraints must be
satised at the sink so it can make decisions in a timely
fashion based on the collected sensed data (or simply data)
regarding the observed phenomenon in the eld of interest. A comprehensive survey on WSNs can be found in [2].
It is clear that the above-mentioned goals, namely minimum energy consumption, minimum delay, and uniform
energy depletion, are conicting goals. This conict can
be explained by the following three interpretations. First,
the minimization of the energy consumption requires
transmitting the data over short distances. Indeed, the energy (Etx) spent in data transmission over a distance d between any pair of consecutive forwarders, is proportional
to d, i.e., Etx / da, with 2 6 a 6 4 being the path-loss exponent. Second, the minimization of the delay requires minimizing the number of intermediate forwarders between a
source and the sink. This can be achieved by maximizing
the distance between any pair of consecutive forwarders.
It is worth noting that Haenggi [14] and Haenggi and Puccinelli [15] took an extreme position by arguing that longhop routing is a very competitive strategy compared to
short-hop routing. However, this sacrices the very scarce
energy resource of the sensors. Haenggi [14] provided
twelve reasons explaining the advantages of long-range
over short-range forwarding. We believe that a more balanced approach should be used to account for delay and
energy uniformity. Third, usually, the search space of next
candidate forwarders is a cone centered at the current sensor holding the data to be forwarded. A small cone yields
an unbalanced distribution of the data forwarding load
among the sensors. In fact, this causes a non-uniform
depletion of the available energy of the sensors. Indeed,
the candidate forwarders located in a small cone would
suffer heavy depletion of their energy as they will be frequently selected as forwarders. In contrast, a large cone ensures a more balanced data forwarding load among the
sensors and hence helps achieve uniform energy depletion
of the sensors. Therefore, it is necessary to nd a trade-off
of these three goals, which are jointly considered.
1.1. Major contributions
Our major contributions in this paper are fourfold and
can be summarized as follows:
First, we propose an approach based on slicing the communication range of the sensors in order to trade-off
three conicting goals of sensing applications. More
precisely, our approach aims to decompose the communication range of the sensors into concentric circular
bands and classify them with a goal to satisfy the specic requirements of sensing applications in terms of
energy consumption, delay, and energy depletion. For
tractability, we assume that the communication ranges
of the sensors are modeled by a disk. In addition, we
suppose that all the sensors have the same radius of
their communication range.
Second, we formulate a trade-off of these three conicting goals as a multi-objective optimization problem,
which is solved using a weighted scale-uniform-unit
sum (WES) approach [19]. Then, we propose a data forwarding protocol for WSNs, which exploits a solution to
this multi-objective optimization problem to nd an
optimum trade-off of three conicting goals, namely
minimum energy consumption, minimum delay, and
uniform energy depletion. To account for the third goal,
we propose an approach to characterize the uniform
energy depletion of the sensors based on the size of
the cone that includes a subset of candidate forwarders.
Although there are other methods, such as multiobjective optimization genetic algorithm (MOGA) [12],
we nd that the WES approach offers more exibility
to nd solutions to an optimization problem with several weighted objective functions. We introduce these
weighting coefcients to reect the relative importance
of the individual objective functions and address
the problem of their different units and order of magnitude. Our theoretical results show that an optimum
trade-off of the three goals exists. Moreover, this optimum trade-off depends on these weighting coefcients.
Third, we relax several widely used assumptions in the
design of WSNs and which we adopted in our study. Our
ultimate goal is to enhance the practicality and effectiveness of our proposed TED protocol.
Fourth, we evaluate the performance of TED through
extensive simulations, and compare it with existing
ones. We nd that the performance of TED is near optimal with respect to the energy delay metric. This simulation study seems to be an essential step to gain more
insight into TED before implementing it on a sensor
test-bed. To the best of our knowledge, although the
design of energy-efcient data forwarding protocols
for WSNs has received much attention, there is no
previous work that jointly considers minimum energy

consumption, minimum delay, and uniform energy
depletion to nd their best trade-off.
The remainder of this paper is organized as follows. In
Section 2, we review existing data forwarding protocols
for WSNs. In Section 3, we present some denitions and
assumptions that are used in our study. In Section 4, we
introduce the concepts of slicing the communication range
of the sensors and proxy forwarders. Also, we characterize
the uniform energy depletion of the sensors. In Section 5,
we present an approach to trade-off between energy consumption, delay, and energy depletion in data forwarding
based on the needs of the sensing applications. Moreover,
we discuss our proposed data forwarding protocol TED,
which trades off energy with delay, and show how to relax
the assumptions stated in Section 3. In Section 6, we evaluate the performance of our protocol TED. Finally, in Section 7, we conclude the paper.
2. Related work
In this section, we review a sample of energy-efcient
data forwarding protocols for WSNs. In addition, we
describe previous protocols that trade off energy with
other metrics, such as delay and robustness.
2.1. Energy-efcient data forwarding
Kim et al. [20] proposed a Scalable Energy-efcient
Asynchronous Dissemination (SEAD) protocol for WSNs,
which is based on dissemination trees that are built to disseminate data to mobile sinks. Every mobile sink is supported by a special node, called access node, which acts
as the relay between the mobile sink and source sensors.
Boukerche et al. [7] proposed a novel protocol, called
energy-aware data-centric (EAD), which builds a virtual
backbone composed of active sensors that are responsible
for in-network data processing and trafc relaying. EAD attempts to construct a broadcast tree that approximates an
optimal spanning tree with a maximum number of leaves.
This helps reduce the size of the backbone formed by active
sensors. Xing et al. [49] proposed a greedy geographic routing protocol that allows sensing-covered networks to
achieve a lower routing path length compared to other
existing protocols. Luo and Hubaux [30] discussed an
energy-efcient routing protocol for WSNs that exploits
base station mobility and multi-hop routing. Different
mobility strategies of a mobile station were studied in
order to identify the optimum one in terms of balanced
load distribution. Intanagonwiwat et al. [18] proposed a
data-centric paradigm for sensor query dissemination
and processing in static WSNs, called directed diffusion
(DD), which uses attribute-based naming to match data
to sensors. The DD paradigm provides robust multi-path
delivery and achieves energy savings when intermediate
nodes aggregate responses to queries. Ye et al. [53]
proposed an efcient data delivery protocol to multiple
mobile sinks using a two-tier data dissemination model
(TTDD). When a data source detects a stimulus, it builds
a data dissemination grid structure over the sensor eld
1915
and sets up the forwarding information at sensors closest

to grid points. Liu et al. [29] developed energy efcient protocols to disseminate information sensed at a source node
to any other node that is interested in the information. Precisely, they proposed two data dissemination protocols.
While the rst one is based on the quorum scheme, the
second is based on the home agent scheme.
2.2. Trading-off energy with other metrics
Yang and Vaidya [51] proposed a wakeup scheme, called
Pipelined Tone Wakeup (PTW), which achieves a balance
between energy savings and end-to-end delay. The PTW
scheme is based on an asynchronous wakeup pipeline that
overlaps the wakeup procedures with the packet transmissions. It uses wakeup tones that allow a large value of duty
cycle ratio without causing a large wakeup delay at each
hop. Miller et al. [35] studied the trade-off between energy,
latency, and reliability. They presented a Probability-Based
Broadcast Forwarding (PBBF) scheme that minimizes energy usage and optimizes latency and reliability. Zorzi and
Rao [59] proposed a transmission technique for WSNs called
geographic random forwarding (GeRaF), where relay nodes
are decided only after the transmission has started. The
packet duplication problem is solved using a scheme for
contention among receivers. Also, Zorzi and Rao [60] gave
a detailed description of a MAC protocol and an evaluation
of the latency and energy performance. Bandyppadhyay
and Coyle [6] proposed a transmission scheduling scheme
using a collision-free protocol for gathering sensor data.
Moreover, they studied many trade-offs between energy
usage, sensor density, and temporal and spatial sampling
rates. Sohrabi et al. [43] proposed a sequential assignment
routing (SAR) protocol that is used by the sensors to select
a path among multiple ones to the sink node. The SAR protocol selects a path based on the energy resources and the
priority level of a packet. Lindsey et al. [26] presented a
scheme, called Power-Efcient Gathering in Sensor Information Systems (PEGASIS), where each node can receive
from and send to close neighbors. The data gathered by
nodes in each round has to be collected and transmitted to
the base station by only one designated node in order to reduce energy consumption and extend the network lifetime.
PEGASIS, which outperforms LEACH protocol [16], considered energy delay as the optimization metric per round
of data gathering in WSNs [27,28]. Using well-selected
examples, Krishnamachari et al. [21] showed that when
robustness and energy-efciency are the main concerns,
single-path routing outperforms multipath routing under
the assumption of perfectly reliable source and destination
sensors. Choi and Das [10] proposed a data gathering
scheme that trades off coverage and data reporting latency
while enhancing energy conservation. Hynh and Hong
[17] proposed an energy delay-aware routing protocol
for wireless sensor networks using cluster-based and
chain-based approaches. Each communication round consists of a cluster and chain formation phase, and a data
transmission phase. The construction of the network conguration, which is dened by the cluster of sensors and their
cluster-heads, is accomplished by a base station. Soltan et al.
[44] proposed a mobility-aware multi-hop routing scheme
1916
for a hierarchical WSN using mobile relays with a goal to

optimize the network lifetime, delay, and local storage size.
They proposed to solve an optimization problem whose
objective function is to maximize the network lifetime under local storage, delay, and maintenance cost constraints.
2.2.1. Our approach
Unlike most existing approaches described earlier in Sections 2.1 and 2.2, our proposed approach for data forwarding
in WSNs considers three conicting performance metrics
simultaneously, namely minimum energy consumption,
minimum delay, and uniform energy depletion. Specically,
our approach aims to minimize the total energy consumption of the sensors, minimize the delay, and ensure that all
the sensors deplete their energy as uniformly as possible
during data forwarding from the source sensors to the sink.
Given the conicting nature of these three metrics as stated
earlier, our approach focuses rst on nding the best tradeoff of these metrics by solving a multi-objective optimization
problem that is formulated based on these metrics. Then, it
exploits the solution found to this optimization problem in
the design of a data forwarding protocol, called TED.
3. Denitions and assumptions
In this section, we state the assumptions made about
the network and the energy model being considered.
Relaxations of these assumptions and their impact will
be further discussed in Section 5. First, we dene the notions of neighbor set, delay, and largest enclosed disk.
Denition 1 (Neighbor set). A neighbor set of a sensor si,
denoted by NS(si), is a set of sensors located within sis
communication range.
Denition 2 (Delay). The delay, denoted by D(s0, sm), is
dened as the time elapsed between the departure of the
data from a source s0 and its arrival to the sink sm. This
delay is given by
Assumption 2 (Communication disk model). The communication ranges of the sensors follow the unit disk model, i.e.,
they are modeled by disks with the same radius R. The
communication disk of sensor si is centered at its location
fi and denoted by CD(fi, R).
Assumption 3 (Dense network model ). The sensors are
densely deployed in a planar eld. In fact, the limited
energy of the sensors and the difculty of replacing and/
or recharging batteries on the sensors in hostile environments require that the sensors be deployed with high density (up to 20 sensors per m3 [42]) in order to extend the
network operational lifetime.
Assumption 4 (Energy consumption model ). The energy
consumption of the sensors is dominated by data transmission and reception. Let si and si be two neighbors. According to the energy consumption model specied by
Heinzelman et al. [16], when data is sent from a transmitter to a receiver, there is energy consumption incurred at
both ends (i.e., transmitter and receiver). While at the
receiver end, the energy consumption is due to only one
component, called the transceiver, the energy consumption
at the transmitter end depends on two components,
namely the transceiver and the transmitter amplier. The
energy consumed by the latter component depends on
the size of the data packet being sent, the distance between
the transmitter and the receiver, and another constant,
called transmitter amplier and denoted by e. The value of
this constant depends on whether the free space or the
multi-path model is being considered. Formally, according
to [16], the energy spent in data transmission is given by
Etx si ; sj aEelec eda si ; sj

whereas the energy spent in data reception is computed as
Erx si aEelec
Ds0 ; sm qd td pdNf s0 ; sm
where qd is the average queuing delay per intermediate
forwarder, td is the average transmission delay, pd is the
average propagation delay, and Nf(s0, sm) is the number of
intermediate forwarders between s0 and sm.
Given that the size of the eld is in the order of a few
miles, the average propagation delay pd is negligible. Thus,
the delay D(s0, sm) is proportional to Nf(s0, sm), i.e.
Ds0 ; sm cN f s0 ; sm / Nf s0 ; sm
where d is the Euclidean distance function, a is the data

size in bits, Eelec is the electronics energy, e 2 {efs, emp} is
the transmitter amplier in the free-space (efs) or the multi-path (mps) model, and 2 6 a 6 4 is the path-loss exponent. When the identities of sender and receiver are not
important, we simply write Etx = a(Eelec + eda) and Erx = aEelec, where d stands for the transmission distance used
by a sender. The values used for the energy consumption
model are given in Table 1.
where c = qd + td. A similar result can be found in [20].

Denition 3 (Largest enclosed disk). The largest enclosed
disk of a closed convex area A is a disk that lays inside A and
whose diameter is equal to the minimum distance
between any pair of points on A0 s boundary.h
Assumption 1 (Static and homogeneous sensor deployment). The sensors are randomly and uniformly distributed in a planar eld. Furthermore, they are static and
homogeneous.
Table 1
Parameters setting.
Eelec
efs
emp
50 nJ/bit
d(s0, sm)
3500 m
a=2
dmin = 70.71 m
nccb = 5
10 pJ/bit/m2
R
350 m
a=3
dmin = 156.68 m
nccb = 3
0.013 pJ/bit/m2
c
0.001
a=4
dmin = 44.29 m
nccb = 8
256 bits
k
0.001
Assumption 5 (Always-on sensor model). The sensors are

always active, i.e., each sensor is on all the time.
Assumption
6 (Location
and
energy
awareness
model). The sensors and the single sink are location-aware
using, for instance, a localization technique [8] and advertise them only once at the beginning of the monitoring
task. Also, each sensor advertises its remaining energy to
its neighbors periodically.
Assumption 7 (Transmission power model ). Each sensor
has transmit-power control. Thus, it can adjust its transmission distance. That is, it can transmit its data over a distance that is less than or equal to the radius R of its
communication disk.
4. A slicing approach
This section shows how to slice the communication
range of the sensors into concentric circular bands (CCBs).
Also, it characterizes the uniform battery power depletion
of the sensors.
4.1. Slicing of communication range
The idea of slicing the communication range of the sensors stems from the simple fact that any sensor has higher
preference to some of its neighbors than to others. This notion of preference becomes apparent in the next-forwarder
selection process when a sensor has to decide to which sensor it wants to forward its data so it reaches the sink while
meeting some energy and/or delay constraints. The slicing
approach is based on an approximation of the minimum
transmission distance dmin in data transmission [6]. For this
paper to be self-contained, both of the result regarding dmin
(Lemma 1) and the proof are given below. As it will be discussed, the communication range slicing approach helps a
sensor si classify its neighbors based on which of the
above-mentioned metrics si wishes to optimize.
Lemma 1. The minimum transmission distance that can be
used in data transmission can be approximated by
dmin

1
Eelec a
Proof. According to the energy model [16] discussed earlier,

Erx = aEelec and Etx(d) = aEelec + aeda, where d is the transmission distance used by a sender. By Assumption 4, Erx represents the minimum energy that can be spent by any
sensor. It is clear that the energy consumption due to data
reception is less than that due to data transmission. Moreover, the value aEelec depends only on the size of the data to
be sent. Thus, if the data packet size is xed, the quantity
aEelec is constant as the electronic energy is a constant given
by Eelec. Thus, it can be considered as the minimum energy
consumption that would be spent during the operation of
1917
the sensor. Now, if we consider the second part aeda(si, sj) in

the formula of the energy consumption in data transmission
Etx(d), it cannot be less than aEelec. In other words, we have a
eda P aEelec. Consequently, we conclude that the minimum
a
transmission distance dmin should satisfy aedmin aEelec .
Thus, the minimum transmission distance dmin that can be
used by a sensor in data transmission is equal to
dmin

1
Eelec a
Notice that the minimum transmission distance

means that when a sensor si adjusts its transmission distance to forward its data toward the sink, it cannot transmit
that data over a distance shorter than dmin. In other words,
the transmitted data would be received by at least the sensors located at a distance dmin from the sensor si. Moreover,
we proved that dmin corresponds to the minimum energy
consumption due to data transmission and reception when
data is forwarded from a source to the sink [6]. Also, it is
worth noting that the minimum transmission distance dmin
is totally different from the minimum distance between
neighboring sensors. While the former is quantied using
energy parameters, the latter is a geometric parameter that
can be quantied based, for instance, on the sensor density.
In fact, the minimum distance between any pair of neighboring sensors can be very small if the sensor density if very
high.
In order to achieve a better balance between minimum
energy consumption, minimum delay, and uniform energy
depletion, we propose to slice the communication range
l m
R
CD(fi, R) of a sensor si into nccb dmin
CCBs, each of which
is centered at si and has a width equal to dmin. As it can be
seen, a set of CCBs can be divided into three categories
(Fig. 1). The inner CCBs favor minimizing energy consumption over minimizing delay and uniform energy depletion;
the middle CCBs give the same degree of interest to the
three performance metrics; the outer CCBs favor minimizing delay and uniform energy depletion over minimizing
sensors energy consumption. As it is discussed in Section
5.1, the path loss exponent has an impact on the classication of the CCBs of the communication range of the
sensors. More precisely, we nd that the second CCB favors
minimum energy consumption over the other two metrics
in the case of a = 2. However, in the cases where a = 3 and
a = 4, it is the rst CCB that yields the minimum energy
consumption. That is the reason why Fig. 1 shows that both
Fig. 1. Slicing of the communication range of sensors (Case a = 2).
1918
of the rst and second CCBs are associated with minimizing the energy consumption during data forwarding to the
sink. The outmost CCB favors delay and uniform energy
depletion over minimum energy consumption. The CCBs
located in the middle help achieve a trade-off of the three
metrics. It all depends on the values of the weighting
coefcients w1, w2, and w3. These coefcients reect the
importance of each of the three metrics, and will be
discussed in Section 5.2.
Notice that our proposed approach may nd a longer
data forwarding path from a source to the sink that
involves several proxy forwarders. A longer data forwarding path is due to the fact that it is not always possible to
forward data through the shortest path between a source
and the sink. The main reason is that the selection of proxy
forwarders is based on the remaining energy of the
sensors. If we keep selecting the shortest path all the time,
the sensors located on this path would deplete their energy
very fast and die before expected. However, the presence of
more proxy forwarders does not always mean more energy
consumption. According to our energy model, which was
proposed by Heinzelman et al. [16], the energy consumption due to data transmission depends on the distance
between the sender and receiver. Thus, less energy
consumption requires sending data over short distances.
That is, we need to have more proxy forwarders. In [5], we
proved that there is an upper bound on the number of
forwarders on the data forwarding path from a source to
the sink. In fact, we computed the optimum number of
forwarders and proved that it depends on the path loss
exponent a and the ratio of the Euclidean distance between
a source and the sink, and dmin. Given that the width of CCB
is dmin, our protocol TED is able to improve the network
lifetime. h
4.2. Selection of candidate proxy forwarders
We consider a set of sensors and a single sink sm that are
deployed in a planar eld. Next, we dene the notion of
candidate proxy forwarder.
Denition 4 (Candidate proxy forwarder ). A candidate
proxy forwarder set of a sensor si, denoted by CPF(si, sm, k, b),
is a subset of sensors from the neighbor set NS(si) of si that
belong to the kth CCB and located within a zone determined
by a wedge with an angle b centered at si (Fig. 2).
The size of CPF(si, sm, k, b) depends on the values of k
and b, where 1 6 k 6 nccb and 0 < b < p. We design our
trade-off protocol between energy and delay based on
the following rule: When a source selects a specic concentric circular band, say k, from which it will designate
a sensor as a forwarder, any future proxy forwarder will
have to use the same value of k. We use this design decision for the sake of ease of the analysis of our proposed
protocol based on this trade-off. Also, this decision would
help us nd some theoretical results in terms of upper
bounds on some specic metrics. We will further relax
this decision by allowing the sensors to select their own
values of k before forwarding data originated from a
source to the sink.
Fig. 2. Impact of k on the size of the subset CPF(si, sm, k, b).
4.3. Uniform energy depletion characterization

Now, we propose to characterize the concept of uniform
energy depletion. Let si be a sensor that holds the data to be
forwarded to the sink. Clearly, if si can choose its proxy forwarder from a large subset of sensors located in its communication range, it is clear that all of those sensors will
be equally likely to participate in the data forwarding process. This ensures that those sensors deplete their energy
uniformly. To the contrary, if si is able to choose its proxy
forwarder from a small subset of sensors, it is obvious that
those sensors deplete their energy quickly before other
sensors in the communication range of si. Indeed, each of
those sensors would be selected very often to act as proxy
forwarder for si. Next, we characterize this concept of uniform energy depletion using our slicing approach through
the following property, which implies that a larger number
of neighbors in the given CCB is guaranteed.
Property 1 (Characterization of uniform energy depletion). We say that uniform energy depletion is achieved if
each sensor guarantees that a large number of its neighbors
located in a given CCB are equally likely to act as candidate
proxy forwarders of the data being sent to the sink.
Our proposed data forwarding protocol TED considers
all the sensors as source sensors (or data generators). Given that the sensors are randomly distributed in a eld
of interest, TED denitely balance the energy consumption
of the sensors in the entire network. As it can be seen from
Fig. 2, the size of the subset CPF(si, sm, k, b) increases with k.
On the one hand, high values of k yield a large subset of
sensors to participate in the selection process of a proxy
forwarder. This leads to a uniform depletion of the battery
power of sensors. On the other hand, when the size of
CPF(si, sm, k, b) is small, the same sensors will be frequently
selected as proxy forwarders. This would causes the
remaining energy of those sensors to drain faster compared to the other sensors in the network, which leads to
a non-uniform energy depletion. Thus, uniform energy
depletion of the sensors can be characterized by a large
size of the subset CPF(si, sm, k, b), which implies the use of
higher values of k.
1919
It is worth noting that higher values of k mean that the

proxy forwarders would be selected from CCBs that are
closer to the boundary of the communication range of
the sensors. This will introduce an additional computational cost due to the selection of a proxy forwarder from
a large subset CPF(si, sm, k, b) compared to choosing a proxy
forwarder from a CCB that is closer to the center of the
communication range of the sensor. Thus, the origin of this
additional computational cost is due to the size of the set
of candidate proxy forwarders. Indeed, a sensor si would
select a candidate proxy forwarder that has the highest
remaining energy as it is stated in Section 5.3.3. To this
end, si has to sort its neighboring sensors based on their
remaining energy. However, the cost of this sorting step depends on the size of the set of candidate proxy forwarders.
Clearly, choosing proxy forwarders from CCBs closer to the
boundary of the communication ranges of the sensors requires higher sorting cost compared to choosing from CCBs
farther from the boundary of the communication ranges of
the sensors.
5. Trading-off energy with delay
In this section, we discuss our approach that trades off
minimum energy consumption, minimum delay, and uniform energy depletion in data forwarding in WSNs. Also,
we compute upper bounds on these three metrics and
the optimal values of k corresponding to their optimum
trade-offs.
5.1.2. Data forwarding along non-direct paths

In general, shortest paths between the senders and the
sink are not always available given that our data forwarding protocol is based on the remaining energy of the sensors. Precisely, the sensors that lie on the shortest path
between a sender and the sink could be used as proxy forwarders a few times before their remaining energy get
smaller than that of other sensors. The following analysis
focuses on non-direct paths between the sources and the
sink based on the angle h = \si+1, si, sm (Fig. 3). In Lemma
3, we assume that the progress that is made toward the
sink, which is denoted by w(k, h), is constant. This assumption is to simplify the analysis so the formulas that we obtain for the expected total energy consumption, delay, and
number of candidate proxy forwarders, look like the ones
given in Eqs. (5)(7). Otherwise, we obtain three summations that depend on the value of the angle h. Note that
the maximum value of h is 2b, where the value of b determines the size of the candidate proxy forwarder set as
shown in Fig. 2. Thus, in practice, the value of h could be
chosen between 0 and 2b. Lemma 3 is a generalization of
Lemma 2.
Lemma 3. Let w(k, h) be a constant progress that is made
toward the sink sm at each forwarding action, i.e., h is constant
(Fig. 3). The expected total energy consumption, delay, and
number of candidate proxy forwarders considered when a
data packet is forwarded along a non-direct path from a
source s0 to the sink sm with respect to the kth CCB are given
by
Eexp s0 ; sm ; k; h

a a
a 2Eelec ek dmin ds0 ; sm
wk; h
Dexp s0 ; sm ; k; h
cds0 ; sm
wk; h
5.1. Simple analytical bounds

5.1.1. Data forwarding along shortest paths
Lemma 2 computes upper bounds on the expected
number of candidate proxy forwarders, energy consumption, and delay. We omit its proof since it is verbatim.
Lemma 2. Let k be the sensor spatial density (i.e., the number
of sensors per unit area) and c = qd + td. The expected total
number of candidate proxy forwarders, energy consumption,
and delay associated with the kth CCB in forwarding a data
packet from a source s0 to the sink sm along the shortest path
[s0, sm] are computed as
kb2k 1ds0 ; sm dmin

2k

2Eelec
a1 a1
Eexp s0 ; sm ; k a
ek dmin ds0 ; sm
kdmin
jCPF exp s0 ; sm ; k; bj
Dexp s0 ; sm ; k
cds0 ; sm
kdmin
2
3
4
respectively, and their respective upper bounds are given by
kb2R dmin ds0 ; sm dmin

jCPF exp s0 ; sm ; bj 6

2R
2Eelec
Eexp s0 ; sm 6 a
eRa1 ds0 ; sm
R
cds0 ; sm
Dexp s0 ; sm 6
dmin
jCPF exp s0 ; sm ; k; hj
kb2k 1dmin ds0 ; sm

2wk; h
respectively,
where

wk; h d s0 ; sp1 kdmin cosh.
1 6 k 6 nccb
and
Proof. By
Pythagorean
Theorem,
we
have

wk; h d s0 ; sp1 kdmin cosh, where sp1 is the orthogonal
projection of s1 on the segment [s0, sm] and hmax 2b. Also,
0 ;sm
. Thus, we obtain
we have N f ds
wk;h

a a
Eexp s0 ; sm ; k; h a 2Eelec ek dmin Nf

a a
a 2Eelec ek dmin ds0 ; sm
wk; h
Similarly, the expected delay is given by
Dexp s0 ; sm ; k; h cN f
cds0 ; sm
wk; h
Using the same reasoning as in Lemma 2, we obtain

2
jCPF exp s0 ; sm ; k; hj
kb2k 1dmin ds0 ; sm

2wk; h
1920
Fig. 3. Non-shortest path between s0 and sm.
5.1.3. Numerical results

Fig. 4 shows the impact of k, h, and a on Eexp(s0, sm, k, h)
Eq. (5). Notice that the energy consumption increases with
k and reaches its minimum at k = 2 as discussed earlier.
This result can be proved as follows. Let Eexp(s0, sm, k) = a(A1(k) + A2(k))d(s0, sm),
where
A1 k
2Eelec
kdmin
MinimizeFx
Subjectto x 2 X
where ci
and
Fx F 1 x; . . . ; F n xT
where Fi(x) is an objective function, for 1 6 i 6 n. WES is a
simple approach that introduces a weighting coefcient wici for each Fi(x), where wi is a weight selected by a network
designer to reect the relative importance of Fi(x) and ci is a
coefcient that not only scales Fi(x) but also helps produce
a one-dimensional function F(x). A survey on similar approaches for solving multi-objective optimization problems can be found in [34]. Using WES, a multi-objective
optimization problem can be formulated as follows:
i1
l max F max
:16i6n
i
and X is a set of admissible solutions. It is assumed that

F max
q0 : 1 6 i 6 n.
i
5.2.2. Solving the trade-off problem using WES
Notice that the function CPFexp(s0, sm, k) Eq. (2) is concave
for any h, and, in particular, for h = 0. In fact
@ 2 jCPF exp s0 ; sm ; kj
5.2. Multi-objective optimization approach
5.2.1. Overview of the WES approach

Assume we want to minimize the following multiobjective function:
l
F max
i
F max
maxfF i x : 8x 2 Xg
i
n
X
wi 1
wi P 0; where
@2k
In this section, we solve our multi-objective optimization problem for trading off the above-mentioned three
metrics, namely minimum energy consumption, minimum
delay, and uniform energy depletion. First, we discuss a
weighted scale-uniform-unit sum (WES) approach [19] that
we use in solving this multi-objective optimization
problem.
wi ci F i x
i1
a1 a1
A2 k ek dmin . Thus, A2(k) becomes a dominant factor

(i.e., A2(k) P A1(k)) for some k that makes Eexp(s0,sm, k) grow
proportionally to k. In fact, A2(k) P A1(k) implies ka P 2,
which yields k P 2 given that 2 6 a 6 4. Moreover, more
energy consumption would be incurred as we deviate from
the shortest path [s0, sm] between the source s0 to the sink
sm, i.e., as h increases.
Fig. 5 shows that Dexp(s0, sm, k, h) Eq. (6) decreases with k
since fewer proxy forwarders are needed. Also, a has an
impact on the delay. Indeed, Dexp(s0, sm, k, h) is inversely
proportional to dmin, which in turn depends on a. Similarly,
more delay would be incurred as we deviate from the
shortest path [s0, sm].
n
X
< 0; for 1 6 k 6 nccb
Thus, we consider its opposite function given by
PF exp s0 ; sm ; k jCPF exp s0 ; sm ; kj
Now, PFexp(s0, sm, k) is a convex function (Fig. 6). Thus, we

can formulate our unconstrained multi-objective optimization problem using the WES approach. Consequently,
uniform energy depletion requires minimizing PFexp(s0, sm, k). Notice that Eexp (s0, sm, k), Dexp(s0, sm, k), and PFexp(s0, sm, k)
(Eqs. 3, 4 and 8, respectively) are conicting objective functions. In fact, to minimize Eexp(s0, sm, k), we need to minimize
k,
whereas
minimizing
Dexp(s0, sm, k)
and
PFexp(s0, sm, k) requires maximizing k. Also, Eexp(s0, sm, k),
Dexp(s0, sm, k), and PFexp(s0, sm, k) are strictly convex [13] on
the interval [1, nccb] (i.e., 1 6 k 6 nccb) given that
@ 2 Eexp s0 ; sm ; k
@2k
@ Dexp s0 ; sm ; k
>0
>0
@2k
@ PF exp s0 ; sm ; k
>0
@2k
2
In addition, because the feasible set {1, . . . , nccb} is a convex set, the WES approach yields correct solutions [9].
1921
Fig. 4. Impact of CCB id (k), angle h, and path loss exponent (a) on the energy consumption.
Let M(k) = (Eexp(s0, sm, k), Dexp(s0, sm, k), PFexp(s0, sm, k))T
be our multi-objective function, which we wish to
minimize. Notice that Eexp(s0, sm, k) reaches its
maximum
Emax
exp

2Eelec
a
eRa1 ds0 ; sm at k nccb for
R
a 2 f2; 3; 4g
Also, Dexp(s0, sm, k) and PFexp(s0, sm, k) reach their respecmax
tive maximum Dmax
exp and PF exp at k = 1. Both maximum values are computed as follows:
Dmax
exp
cds0 ; sm
dmin
Subject to 1 6 k 6 nccb
where
8
Emax
Emax PF s ;s ;k
exp Dexp s0 ;sm ;k
>
>
w3 exp PFexpmax0 m
> w1 Eexp s0 ; sm ; k w2
Dmax
>
exp
exp
>
n
o
>
>
>
max
max
max
max
>
if
E
max
E
;
D
;
PF
>
exp
exp
exp
exp
>
>
>
>
>
Dmax E s0 ;sm ;k
Dmax PF s ;s ;k
>
< w1 exp exp
w
D
s
;
s
;
k
w3 exp PFexpmax0 m
max
2
exp
0
m
Eexp
exp
n
o
Mk
max
max
max
max
>
>
if
D
max
E
;D
;
PF
>
exp
exp
exp
exp
>
>
>
>
>
PF max
PF max
exp Eexp s0 ;sm ;k
exp Dexp s0 ;sm ;k
>
>
w
w3 PF exp s0 ;sm ; k
w
max
max
1
2
>
Eexp
Dexp
>
>
n
o
>
>
max
max
max
>
:
if PF exp max Eexp ; Dexp ; PF max
exp
0 6 w1 ; w2 ; w3 6 1 with w1 w2 w3 1
Emax
exp maxfEexp s0 ; sm ; k : 1 6 k 6 nccb g
and
PF max
exp
MO Minimize Mk
kpdmin ds0 ; sm
6
Using the WES approach, where the weights w1, w2, and
w3 indicate the relative importance of Eexp(s0, sm, k), Dexp
(s0, sm, k), and PFexp(s0, sm, k), respectively, our unconstrained
multi-objective optimization problem can be written as
follows:
Dmax
exp maxfDexp s0 ; sm ; k : 1 6 k 6 nccb g
PF max
exp maxfPF exp s0 ; sm ; k : 1 6 k 6 nccb g
Let us study the non-linear multi-objective function M
(k), which depends on the maximum values Emax
Dmax
exp ;
exp ,
and PF max
of
their
corresponding
objective
functions.
Thus,
exp
we consider the following three cases depending on the
max
values Emax
Dmax
exp ;
exp , and PF exp .
1922
Fig. 5. Impact of CCB id (k), angle h, and path loss exponent (a) on the delay.
n
o
max
max
max
Case 1: Emax
exp max Eexp ; Dexp ; PF exp

Let k1 be a solution to the (MO) problem stated above.
@Mk

0 implies k1
k
v
u
u
w2 cEmax
w3 pkEmax
2Eelec
a
exp
exp
t
a
a
a2
a 1edmin w1 a 1aedmin Dmax
6w1 a 1aedmin PF max
exp
exp
@ 2 M k1

Notice that
@2 k1 > 0. Thus, k1 corresponds to the minimum of M k1 . Furthermore, varying the weights w1, w2,
and w3 from 0 to 1, where w1 + w2 + w3 = 1, generates

the corresponding minimum solutions of M k1 .
n
o
max
max
max
Case 2: Dmax

Let k2 be a solution to the (MO) problem dened
earlier.
@Mk

0 implies k2
k
v
u
u
w2 c Emax
w3 pk Emax
2Eelec
a
exp
exp
t
a
a
a2
6w
a
1aedmin PF max
1
exp
exp
n
o
max
max
max
Case 3: PF max

Let k3 be a solution to the abovementioned (MO)
problem.
@Mk

0 implies k3
k
v
u
u
w2 c Emax
w3 pk Emax
2Eelec
a
exp
exp
t
a
a
a2
6w1 a 1aedmin PF max
exp
exp
Notice that our multi-objective optimization problem is

solved exactly once at the beginning of sensor deployment to accomplish their monitoring task. Also, WES

generates a unique, optimum solution k1 k2 k3 to
the multi-objective optimization problem (MO) regardless of the outcome of the comparison between the maxmax
max
imum values Emax
exp ; Dexp , and PF exp . Although there are
other methods for solving multi-objective optimization
problems, such as multi-objective optimization genetic
algorithm (MOGA) [11], WES ts well our purpose. Indeed, the TED protocol is designed for sensing applications with different requirements in terms of energy
and delay. That is, real-world applications may give priority to one or more metrics rather than others and vice
versa. The exibility of WES enables a network designer
to emphasize the importance of a specic objective function over the others by choosing a suitable weight vector. Particularly, the different objective functions may
be given the same preference by using equal weights.
1923
Fig. 6. Impact of CCB id,, path loss exponent a, angle h on PF exp s0 ; sm ; k, which is convex.
Moreover, we obtain a Pareto-optimum solution [9].5.2.3.

Numerical results
In Fig. 7, we plot M(k) with respect to various weight
vectors (w1, w2, w3), where w1 + w2 + w3 = 1, while varying
the path-loss exponent 2 6 a 6 4. As expected, the minimum of M(k) depends on the weights assigned by a network designer to the individual objective functions.
When a = 2, the minimum of M(k) is obtained at k = 2 for
high values of w1(w1 2 {0.8, 0.7}), meaning that the network designer wants to minimize the energy consumption
of sensors, which could be achieved for k = 2 as was proved
theoretically. However, when w1 has comparable values to
those of w2 and w3, such as in the following combinations
((w1, w2, w3) 2 {(0.5, 0.3, 0.2), (0.4, 0.4, 0.2)}), M(k) reaches
its minimum at k = 3 as the network designer wants to
achieve some balance of the three objective functions.
For the following combinations, ((w1, w2, w3) 2 {(0.8, 0.1,
0.1), (0.7, 0.2, 0.1), (0.6, 0.2, 0.2), (0.5, 0.3, 0.2)}), when a = 3,
the best trade-off corresponds to k = 1, which favor minimizing energy consumption over minimizing delay and
guaranteeing uniform energy depletion, and k = 2 for lower
values of w1. However, when a = 4, M(k) reaches its minimum at k = 2 and k = 3, depending on the weights w1, w2,
and w3. It is clear that as w1 decreases, the value of k
corresponding to the minimum of M(k) increases, meaning

that inner CCBs are more preferred.
Fig. 8 plots (k), which favors minimizing delay over the
other two metrics using appropriate values of w1, w2, and
w3. When a = 2, the minimum of M(k) is obtained at
k = 3, k = 4, and k = 5. As it can be observed, the value of k
increases with w2. Recall that the minimum delay is
reached for the outer CCBs. When a = 3, the best trade-off
is obtained at k = 2 and k = 3. However, when a = 4, M(k)
reaches its minimum at k = 4 and k = 5. Also. the value of
k corresponding to the minimum of M(k) increases as w1
increases. This means that the outer CCBs are more preferred to minimize the delay objective function.
The plot of M(k) in Fig. 9 uses the weights
w3 > w1 P w2, which favor uniform battery power depletion over the other two metrics. As it can be seen from
Fig. 9, the minimum of M(k) occurs at 3 6 k 6 5 for
a = 2. However, when a = 3,M(k) attains its minimum at
k = 2 (middle CCB). When a = 4, the minimum of M(k) is
reached at 3 6 k 6 5. Regardless of the value of a, we notice that the value of k corresponding to the minimum of
M(k) decreases as w1 decreases. It is clear that the outer
CCBs are more preferred for the uniform energy depletion
objective function.
1924
Fig. 7. Trade-off between energy consumption, delay, and uniform energy depletion (w1 > w2, w3).
Fig. 10 shows that the best trade-off of the three objective functions with the same weight corresponds to k = 4
for a = 2, k = 2 for a = 3, and k = 3 for a = 4. It is worth noting that minimizing delay and guaranteeing uniform energy depletion are not conicting metrics since both of
them require maximizing k. Thus, the weights w2 and w3
can be viewed as a combined weight against the weight w1.
5.3. TED detailed description
Our proposed data forwarding protocol TED is given in
Fig. 11. It has three phases: Communication range slicing,
concentric circular band selection, and proxy forwarder
selection. Next, we describe each of those phases in details.
5.3.1. Communication range slicing
This phase is run by each sensor only once at the beginning of the sensing task. Each sensor creates a table with
nccb entries, each including a subset of neighbors located
in the corresponding CCB. Notice that the sensors that
are located at the boundaries of two consecutive CCBs
are assigned to the inner one.
5.3.2. Concentric circular band selection

When a source s0 wishes to transit its data to the sink
sm, it computes the id of the CCB to be used. This id is denoted by k, where 1 6 k 6 nccb. The value of k corresponds
to the optimum solution of the multi-objective optimization problem dened earlier. It is the responsibility of a
source to choose the appropriate weights associated with
each of the three metrics. The selection of the values of
these weights is guided by the requirements of the
underlying sensing application in terms of energy and delay. In other words, a source may wish to favor one of the
metrics over the others or nd a best trade-off between
them. Once a source s0 has selected the corresponding
weights, it solves the multi-objective optimization
problem to nd the value of k that meets the sensing application needs.
5.3.3. Proxy forwarder selection
A source s0 selects its proxy forwarder sPF1 among the
subset CPF(s0, sm, k), such that
Erem sPF1 maxfErem sj : sj 2 CPFs0 ; sm ; k; hg
1925
where Erem(sj) is the remaining energy of sensor sj. In other

words, the protocol takes into consideration the remaining
energy of the sensors, and prefers the ones with high available energy. This helps avoid the sensors with low remaining energy whose selection would possibly lead to network
disconnections. When a proxy forwarder receives data, it
runs the same protocol to select its next proxy forwarder.
This process continues until the sink receives the data originated from a source.
5.3.4. Is k xed for all proxy forwarders or not?
The
approach generates a unique, optimum solu WES

tion k k1 k2 k3 to the multi-objective optimization
problem (MO). Furthermore, the value of k depends only on
the weighting coefcients w1, w2, and w3, which are associated with each of the three metrics. The question that we
want to address is whether the same value of k computed
by a source is used by all proxy forwarders. In order to answer this question, we consider the following two scenarios for computing the value of k. In the rst one, the TED
protocol requires that all proxy forwarders on the data forwarding path between s0 and sm use the same value of k to
identify their proxy forwarders. In other words, the formu-
lation of the multi-objective optimization problem and its

solving is done only once by a source s0 and any proxy forwarder should use the value of k found by s0. In the second
scenario, however, each of the senders (i.e., sources and
proxy forwarders) computes its own value of k based on
its own values of w1, w2, and w3. The second scenario
seems more practical for the following reason. As we get
closer to the sink, sensors may get more active in data forwarding toward the sink. In particular, sensors located
around the sink act as relay of all data originated from all
sources. Specically, the data trafc model would have
an impact on the energy consumption. In fact, if each sensor is required to transmit data to the sink, sensors nearer
the sink would consume more energy than all other sensors in the network. Thus, for these sensors, it would be
more important for them to minimize their energy consumption as much as possible by sending/forwarding data
to the sink over short distances. This implies that the value
of w1 is higher than those of w2 and w3. Therefore, in our
design and simulation of the TED protocol, we consider
the second scenario in which the value of k is not the same
for a source and all subsequent proxy forwarders when forwarding data to the sink.
1926
5.4. Relaxation of several key assumptions

The assumptions that we have made in Section 3,
namely sensor homogeneity (Assumption 1), communication disk model (Assumption 2), highly dense WSNs
(Assumption 3), energy consumption dominated by energy
spent in data transmission and reception (Assumption 4),
and always-on sensor model (Assumption 5), form the basis for most of the existing data forwarding and topology
control protocols for WSNs. In this section, we discuss
how to relax each of these assumptions and assess their
impact on the TED protocol. Our goal is to promote the
applicability of TED to real-world sensing applications.
5.4.1. Relaxing the sensor homogeneity model

In real-world scenarios, our assumption (Assumption 1)
may not be true. Indeed, sensors may have different capabilities, particularly in terms of communication ranges.
Several studies [11,45,52] showed that WSNs with heterogeneous sensors possessing unequal energy levels and different
sensing,
processing,
and
communication
capabilities, have increased reliability and lifetime. In the
second scenario (i.e., each sensor computes its own value
of k), there is no change to TED. However, for the rst scenario (i.e., same value of k for all source and proxy forwarders), it may happen that the number of CCBs of some proxy
forwarder si, say nccb(si), is less than k (i.e., solution to the
multi-objective optimization problem (MO)). In this case,
the sensor si selects its proxy forwarder from its k0 th CCB,
where k0 < nccb (si) < k. Given that k0 < k, the data forwarding
process will consume less energy than expected. Also, given that the proxy forwarder si selects its last CCB, a minimum delay and a uniform energy depletion will be
ensured. Thus, we obtain the best trade-off with respect
to the three goals.
5.4.2. Relaxing the communication disk model

The circular radio model (Assumption 2), which is used
to simplify the analysis, may be unrealistic. Indeed, empirical studies have found that the communication range of
radios is highly probabilistic and irregular [50,55,57]. For
tractability, we consider convex communication ranges
that are not necessarily disks. Using the notion of largest
enclosed disk (Denition 3), each sensor would be able to
run the TED protocol by slicing the largest enclosed disk
of its communication range. As in Section 5.4.1, there is
1927
Fig. 10. Trade-off between energy consumption, delay, and uniform energy depletion (w3 = w1 = w2).
no change to TED protocol in the second scenario. However, for the rst scenario, the number nccb (si) of CCBs of
a sensor si may be less than the value of k computed by a
source. Similarly, sensor si would consider its last CCB from
which it would selects its proxy forwarder. Using the same
argument as in Section 5.4.1, we can see that we obtain the
best trade-off between all the three individual objective
functions. Furthermore, the relation of the radio sending
power and the radio communication range has been reported irregular and also vary with time [27]. In order to
account for the dynamic nature of the communication
range of the sensors, it would be useful to use the algorithm ATPC [25] jointly with TED. In fact, the slicing approach depends on the radius of the communication
range of the sensors. Given that the latter is irregular and
time-varying, TED would be more practical if the concept
of the largest enclosed disk is used based on the prediction
of the communication range produced by ATPC algorithm.
This helps improve the effectiveness of our protocol TED
and its generalization to real-world scenarios.
We should point out the fact that in reality and as demonstrated in [58], the radio pattern is not likely to be con-
vex. We assumed convexity of the communication range of

the sensors only to make the problem more tractable. Our
proposed approach based on the largest enclosed disk
(Denition 3) still works. However, this enclosed disk
could be so small that the TED protocol always favors minimizing the total energy consumption of the sensors. As a
consequence, our TED protocol may not be suitable for
time-critical applications, where delay is the metric to be
considered. In other words, time-critical sensed data may
not reach the sink on time as it would be forwarded
through several proxy forwarders, thus, introducing more
delay. This is a limitation of our proposed work with regard
to the above type of sensing applications. We leave the
more general problem of non-convexity of the radio pattern of the sensors as a future work.
5.4.3. Relaxing the Dense network model
In general, one of the main characteristics of WSNs is
their high density compared to ad hoc wireless networks
[2]. Thus, our assumption (Assumption 3) is realistic.
Now, what would happen if TED is used by a sparse
WSN? In this case, a sensor may not nd any sensor
1928
Fig. 11. The TED data forwarding protocol.
located in the selected kth CCB. Although some proxy forwarder may be located in another CCB, say the ith CCB,
where i k, it may not be always possible to nd the best
trade-off required by the underlying sensing application.

Thus, the TED protocol would be able to trade-off energy
with delay with respect to the specic needs of a sensing
application only when the network has a sensor density

such that each CCB includes some sensors that would be
selected to act as proxy forwarders. This is, indeed, a valid
assumption since WSNs are densely deployed in general to
provide high-quality monitoring.
5.4.4. Relaxing the energy consumption model
The slicing approach of the communication range of
sensors depends on the minimum transmission distance
dmin. Recall that dmin is computed based on our assumption
(Assumption 4) that considers only the energy spent in
data transmission and reception. Nevertheless, sensors
spend energy in computation and sensing. However, both
types of energy are negligible compared to the one spent
in communication. Also, neither the energy spent in computation nor the one spent in sensing is quantiable.
Hence, if either type of energy has some known formula,
the value of dmin would be modied accordingly. Precisely,
this value would be smaller than the one that we have
computed in Lemma 1. Hence, the number of CCBs would
be higher, which would have an impact on TED
performance.
5.4.5. Relaxing the always-on sensors model
Most of the existing geographic routing and data forwarding protocols assume that the radios of all the sensors
are turned on all the time (Assumption 5). In particular, all
the sensors in the network are awake during the forwarding activity. However, in real-world scenarios, the sensors
may switch between on and off states in order to save their
limited energy. It is not even practical to keep a sensor
awake all the time while it is in reality active for only some
short periods of time. Indeed, keeping the sensors always
on may cause failures that have severe impact on the performance of the WSN. Precisely, the network could be partitioned into at least two non-communicating subnetworks. Hence, the existence of the whole network
may become meaningless. Therefore, it is important to
duty-cycle the sensors so that they deplete their energy resources uniformly and slowly [50,54]. Unfortunately, dutycycling may create a problem for routing the current message to the next hop while it is asleep. The challenge is how
to duty-cycle the sensors that are running TED while guaranteeing good routing performance [36,59,60]. The handling of highly dynamic WSNs that experience timevarying connectivity due to sensor duty-cycling is left as
one of our future work.
5.5. Application contexts: how to choose the weighting
coefcients?
In this paper, we do not specify the values of the
weights and we have attempted to nd an optimal solution
given the weights. The real problem is to precisely nd, for
a given set of cost functions and for a given set of preferences of the decision maker, what weights are best for each
network setup. It is clear that the choice of the values of
those weighting coefcients w1, w2, and w3 is application-dependent. In other words, the network designer
should be able to determine the appropriate values to meet
the underlying sensing application requirements. For in-
1929
stance, for time-critical applications, which have hard

deadlines, the coefcient w3 should have the highest value.
In the case of sensing application, where time is not really
a constraint, the coefcient w1 should have the highest value compared to those of w2 and w3. As it is discussed earlier, there are two scenarios for computing the value of k,
i.e., CCBs id. In the rst scenario, the value of k is computed
exactly once by a source si based on the values of those
weighting coefcients w1, w2, and w3. All other proxy forwarders of the data being originated from si have to use
the same value of k. For instance, the source computes
the values of these weighting coefcients w1, w2, and w3
based on the nature of the collected data. That is, whether
this data is time-critical or not. In the second scenario, the
selection of the values of w1, w2, and w3 should be done locally by each proxy forwarder. Also, this computation process may account for the distribution of the trafc that
varies depending on the location of the sensors with respect to the sink. Indeed, this trafc is expected to be higher at the nodes closer to the sink. Clearly, those sensors
nearer the sink are heavily used in forwarding data to the
sink on behalf of all other sensors in the network. This is
a non-trivial problem that we plan to solve to enhance
the practicality and applicability of our proposed TED
protocol.
6. Simulation results
Our approach to trading-off minimum energy consumption, minimum delay, and uniform energy depletion
is unique in several ways as described below. Thus, it is
impossible to make a fair quantitative comparison between TED and other existing approaches, such as the ones
given in [27,28,31,35,40,51,56,59,60] and reviewed in Section 2.2. First, TED allows a network designer to optimize
the above-mentioned three metrics according to the specic needs of the underlying sensing application. This is
done by using weights that specify the interest in each of
these three metrics. Second, TED works at the network
layer and does not assume any sleep-wakeup scheduling
protocol, where sensors are duty-cycled (i.e., turned on or
off) to save their energy. TED applies to always-on and
many-to-one WSNs, where sensors are always onto collect
data about a specic phenomenon in a eld of interest, and
send them to a single sink for further analysis and processing. Third, TED does not assume any aggregation of the
data originated from sources toward the sink. In other
words, all data should be received by the sink without
undergoing any fusion or aggregation at any intermediate
sensor. On the other hand, PEGASIS (Power-Efcient GAthering in Sensor Information Systems) [27,28] is a simple
and elegant data aggregation protocol. It forms a chain
among sensors so that each sensor receives from and
transmits to a close neighbor. Also, only one designated
sensor sends the combined data to the sink in each round.
Precisely, all sensors take turns to directly transmit the
combined data to the sink. PBBF (Probability-Based Broadcast Forwarding) [35] works at the MAC layer and assume
a sleep-wakeup mechanism for all the sensors. Indeed,
PBBF benets from the redundancy in broadcast
1930
communication and forwards packets using a probabilitybased approach with two parameters that need to be used
by a sleep-wakeup scheduling protocol. The rst parameter is a probability p that a sensor rebroadcasts a packet
in the current active time while not all its neighbors are
guaranteed to be awake to receive the broadcast. The second parameter is a probability q that a sensor would remain awake after the active time when it would be
asleep. Furthermore, the goal of PBBF applies to many-tomany WSNs and ensures that a sensor receives at least
one copy of each broadcast packet with high probability,
while reducing the latency due to sleeping. Also, PTW
(Pipelined Tone Wakeup) [51] helps the sensors forward
their data to their nal destinations based on a wakeup
scheme that helps achieve a balance between energy saving and end-to-end delay. Also, GeRaF [59,60] assumes a
sleep-wakeup scheduling protocol, where each sensor
keeps forwarding data until at least one of its neighboring
sensor is awake and able to receive them.
6.1. Simulation settings
In this section, we present the performance results of
the TED protocol for the free-space model (a = 2) based
on simulation programs written in the C programming language. We consider a square sensor eld of side length
equal to 1000 m, where the sensors are randomly and uniformly distributed. Furthermore, we assume that every
sensor continuously generates constant bit rate (CBR) data
of 1024 bits/s (i.e., 4 data packets of size 256 bits/s). Moreover, the radius of the communication range of the sensors
is equal to 250 m. Also, we assume that the total number of
deployed sensors in the default case is 1000, which corresponds to a sensor spatial density equal to k = 0.001 sensor
per m2. In addition, we assume that all sensors have the
same amount of initial energy that is equal to 1 J.
6.2. Impact of selection space size
In this experiment, we consider one metric, namely
number of communication rounds, and show the impact of
the size of the selection space of proxy forwarders on
TED. We assume that all sensors use the same value of
the angle b. As it is discussed earlier, the values of k and
b dene the size of the space of the kth CCB from which
proxy forwarders are selected. According to Fig. 12, the
number of communication rounds increases with the value
of b. The existence of more communication rounds means
that there are more communication paths between the
sensors and the sink so that data originated from the former would reach the latter. In the absence of a balance of
the load of data forwarding among all the sensors, some
of them are more heavily used as proxy forwarders than
others. As a result, they die quickly. Thus, holes (i.e., void
regions) in the network appear and the network consequently disconnects, which prohibits data from reaching
the sink. This situation appears if the space from which
the sensors select their proxy forwarders is small. In this
case, the same neighbors are selected frequently to forward data to the sink, thus, depleting their energy faster
than others. Thus, uniform energy depletion would guar-
antee that most of sensors would evenly participate in data

forwarding to the sink and deplete their energy as slowly
and uniformly as possible. However, we should mention
that the energy sink-hole problem [23,24,35,37,47,48] cannot be avoided while using a static sink.
6.3. Using the energy delay metric
In this experiment, we consider another metric, called
energy delay, which was introduced by Lindsey et al.
[28] to evaluate the performance of their PEGASIS protocol.
Since it is impossible to compare our TED protocol with
any existing approach, we consider two instances of TED,
namely short-range forwarding (SR) and long-range forwarding (LR). Using SR, the sensors forward data over short
distances. With LR, the sensors forward data over long distances. SR performs the best in terms of energy consumption, and, hence, provides us with lower bound on energy.
But, it yields a highest delay. LR performs the best in terms
of delay, and, hence, provides us with lower bound on delay. But, it consumes the maximum energy. On the other
hand, TED helps nd a balance between energy with delay.
Thus, we expect that TED performs like SR when sensors
care more about energy consumption (w1 w2, w3), and
performs like LR when sensors care more about delay
and uniform battery depletion (w2, w3 w1). Given that
our goal is to nd a trade-off between energy and delay,
we consider the energy delay metric to compare TED
against SR and LR.
Fig. 13 shows that LR has the lowest energy delay,
whereas SR yields the highest energy delay. As it is expected, Fig. 13 shows that energy delay for the three
schemes increases with the sensor spatial density. Fig. 14
shows that energy delay depends on the location of the
sink. It is worth mentioning that the optimum location of
the sink in terms of energy-efcient data gathering corresponds to the center of the sensor eld [30]. Fig. 14 shows
that the best performance of the three protocols is obtained only when the sink is located at the center of the
sensor eld. Thus, the obtained results are conforming to
Fig. 12. Uniform energy depletion.
the claim given in [30]. Fig. 15 shows the impact of sensed

data packet size on energy delay. Recall that the energy a
sensor spends in data transmission (Etx) depends on the
size of the data forwarded to the sink. As it can be observed, TED has energy delay close to the one produced
by LR. Given the orders of magnitude of energy and delay,
we can claim that energy delay reaches its minimum for
the smallest value of delay. Our intuition matches the results of the analysis of the function Eexp(s0, sm, k) Dexp(s0, sm, k). Indeed, it is easy to check that this function reaches
its minimum at k = 5, which corresponds to LR as it always
selects proxy forwarders from the last CCB. Thus, LR is optimal with respect to the energy delay metric, and hence
TED is near optimal. Although LR has the best energy delay, it is not suitable for energy-constrained WSNs.
On the other hand, SR is not appropriate for time-critical
sensing applications.
As shown in Fig. 16, SR yields better network lifetime
than TED. However, according to Fig. 17, we nd that
TED outperforms LR in terms of network lifetime. This is
due to the nature of forwarding schemes of SR and LR.
Compared to SR and LR, TED is more balanced in terms
of energy and delay. Thus, TED is a best candidate data forwarding protocol for WSNs, where both energy and delay
are equally important. Also, TED offers more exibility to
WSN designers to meet the specic energy and delay
requirements of sensing applications.
1931
Fig. 14. Impact of location of the sink.
6.4. Impact of variability of k

In this experiment, we focus on the variability of k
whose value may not be the same for all the proxy forwarders. We have considered two cases: In the rst case,
the source sensor computes the value of k and each proxy
forwarder involved in the forwarding data toward the sink
has to use the same value of k. In the second case, the computation of the value of k is done locally. In fact, each proxy
forwarder computes its proper value of k to identify the
CCB from which it would select its next proxy forwarder.
Fig. 18 shows the simulation results for both cases.
Although both results are very similar, we nd that the
Fig. 15. Impact of sensed data packet size.
variability of k yields better performance. It is worth mentioning that the sensors nearer the static sink represent
hot-spot trafc points as they are heavily used in forwarding data to the sink. This situation creates a problem
known as the energy sink-hole problem, which could possibly isolate the sink, thus, disconnecting the network. These
sensors care more about minimizing their energy consumption by forwarding the data to the sink over short distances even when the sink is within their communication
range. Thus, their selection scheme enables them to extend
their individual lifetime, which will prolong the network
operational lifetime.
6.5. Impact of sensor heterogeneity
Fig. 13. Impact of sensor density.
In this experiment, we evaluate the performance of our

protocol TED in the presence of heterogeneous sensors. We
assume that the sensors may not necessarily have the same
amount of initial energy and the same radii of their communication ranges. The initial energy of the sensors is randomly picked from the interval [1 J, 2 J]. Similarly, the radii
1932
Fig. 16. TED compared to SR.
Fig. 18. Impact of variability of the value of k.
Fig. 19. Impact of sensor heterogeneity.

Fig. 17. TED compared to LR.
of their communication ranges are randomly selected from

the interval [250 m, 350 m]. Given that the sensors may not
have the same radii of their communication range, they
may not have the same number of CCBs when they slice
their communication ranges. Thus, a value of k that is initially computed by the source sensor may exceed the total
number of CCBs a proxy forwarder would have in its communication range. In that case, a proxy forwarder simply
chooses the last CCB from which it selects its next proxy
forwarder. Therefore, the value of k would not be the same
for all the proxy forwarders, i.e., k is dynamic. It is as if each
sensor has its proper value of k. Fig. 19 shows the results
for homogeneous sensors and heterogeneous sensors network setup. Not surprisingly, we nd that the performance
of TED using heterogeneous sensors outperforms that of
the network with homogeneous sensors. Here, we nd that
sensor heterogeneity helps extend the lifetime of the sensors, thus prolonging the network lifetime. Indeed, Yarvis
et al. [52] presented several results showing the potential
benet and impact of energy and link heterogeneity on
WSNs.
Now, we consider sensor with irregular but convex

communication ranges. Fig. 19 shows that the results using
heterogeneous sensors and those using sensors with irregular but convex communication ranges are quite similar.
However, the performance of TED using each of those types
of sensors outperforms that of the network using homogeneous sensors.
7. Conclusion
7.1. Summary
While energy is a crucial resource in the design of WSNs,
delay is an important factor, especially for real-time sensing
applications, where the data is time-critical. Since minimizing energy consumption, minimizing delay, and guaranteeing uniform energy depletion are conicting goals, we
formulated their trade-off as a multi-objective optimization
problem based on the idea of slicing the sensors communication range into concentric circular bands (CCBs). Then, we
solved it using a weighted scale-uniform-unit sum (WES)
approach to nd an optimum trade-off of the three goals.

Also, we proposed a data forwarding protocol, called TED,
which exploits the solution to the multi-objective optimization problem in order to identify particular CCBs and select
appropriate proxy forwarders. Moreover, we derived upper
bounds on the energy consumption, delay, and the size of
the subset of candidate proxy forwarders. We found that a
trade-off of the above-mentioned three goals exists and depends on their corresponding weighting coefcients, which
in turn depend on the specic needs of the underlying sensing application. We presented extensive numerical results
showing the impact of different parameters on each of the
three metrics as well as the inuence of the weighting coefcients on their optimum trade-off. Moreover, we found
that relaxing the assumptions regarding homogeneous sensor network and disk radio models can only yield the best
trade-off of the three objective functions based on their
associated weighting coefcients. In addition, we compared
TED with SR and LR, and found that the performance of TED
lies between those of SR and LR. Surprisingly, we found that
LR has a little greater energy delay compared to TED. Despite this result, LR is not suitable for energy-constrained
WSNs, while TED helps nd the best trade-off between energy and delay for various sensing applications.
7.2. Future work
Our future work is sixfold: First, we intend to extend
TED by considering irregular (but, not necessarily convex)
radio communication ranges [46] and heterogeneous sensors and using more powerful solutions. Second, in order
to deal with the energy sink-hole problem [23,24,35,37,48]
in static WSNs, we plan to extend TED to mobile WSNs
[22,39,45] in which the sink is mobile [41]. Indeed, sink
mobility helps the neighbors of the sink change over time,
thus, leading to a load balance among all the sensors in
the network. Third, the assumption that all the sensors
are always on during forwarding is not realistic. We plan
to design a duty-cycling framework on top of which TED
would run, where the sensors are turned on or off to save
energy while achieving the desired quality of surveillance
[50,54]. Moreover, we will consider m-covered WSNs [3],
where each point in a eld of interest is covered (or sensed)
by at least m sensors, and investigate the use of TED. Furthermore, we plan to extend TED to m-covered mission-oriented mobile WSNs [4]. Fourth, we focus on extending TED
to three-dimensional WSNs, such as underwater WSNs [1],
[38] which require a design in a three-dimensional space
instead of a two-dimensional space. Fifth, we plan to investigate the use of ATPC algorithm [25] to account for the
irregular and time-varying communication range of the
sensors. Finally, our main concern is to implement TED on
a real sensor test-bed for the free space and multi-path
models to assess its performance in real-world sensing
applications.
References
[1] I.F. Akyildiz, D. Pompili, T. Melodia, Underwater acoustic sensor
networks: research challenges, Ad Hoc Networks 3 (2005) 257279.
1933
[2] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless

sensor networks: a survey, Computer Networks 38 (2002) 393422.
[3] H.M. Ammari, S.K. Das, Centralized and clustered k-coverage
protocols for wireless sensor networks, IEEE TC 61 (1) (2012) 118
133.
[4] H.M. Ammari, On the problem of k-coverage in mission-oriented
mobile wireless sensor networks, Elsevier Computer Networks
(Elsevier COMNET) 56 (7) (2012) 19351950.
[5] H.M. Ammari, S.K. Das, Forwarding via checkpoints: geographic
routing on always-on sensors, Elsevier Journal of Parallel and
Distributed Computing 70 (7) (2010) 719731.
[6] S. Bandyopadhyay, E. Coyle, Spatio-temporal sampling rates and
energy efciency in wireless sensor networks, Proceedings of IEEE
INFOCOM (2004) 17291740.
[7] A. Boukerche, X. Cheng, J. Linus, Energy-aware data-centric routing in
microsensor networks, Proceedings of ACM MSWiM (2003) 4249.
[8] N. Bulusu, J. Heidemann, D. Estrin, GPS-less low cost outdoor
localization for very small devices, IEEE Personal Communications
Magazine 7 (5) (2000) 2834.
[9] V. Chankong, Y. Haimes, Multiobjective decision making: theory and
methodology, North Holland Series in System Science and
Engineering (1983).
[10] W. Choi, S. Das, A novel framework for energy-conserving data
gathering in wireless sensor networks, Proceedings of IEEE
INFOCOM (2005) 19851996.
[11] X. Du, F. Lin, Improving routing in sensor networks with
heterogeneous sensors, Proceedings of IEEE VTC (2005) 25282532.
[12] C. Fonseca, P. Fleming, Genetic algorithms for multiobjective
optimization:
formulation,
discussion
and
generalization,
Proceedings of the International Conference on Genetic Algorithms
(1993) 416423.
[13] I. Gradshteyn, I. Ryzhik, Tables of Integrals, Series, and Product, sixth
ed., Academic Press, San Diego, CA, 2000.
[14] M. Haenggi, Twelve reasons not to route over many short hops, IEEE
VTC (2004) 31303134.
[15] M. Haenggi, D. Puccinelli, Routing in ad hoc networks: a case
for long hops, IEEE Communication Magazine 43 (10) (2005)
93101.
[16] W. Heinzelman, A. Chandrakasan, H. Balakrishnan, An applicationspecic protocol architecture for wireless microsensor networks,
IEEE TWC 1 (4) (2002) 660670.
[17] T.T. Huynh, C.S. Hong, An energy delay efcient routing scheme for
wireless sensor networks, Proceedings of MMNS, LNCS 3754 (2005)
1122.
[18] C. Intanagonwiwat, R. Govindan, D. Estrin, Directed diffusion: a
scalable and robust communication paradigm for sensor networks,
Proceedings of ACM MobiCom (2000) 5667.
[19] R.L. Keeney, H. Raiffa, Decision Making with Multiple Objectives,
Cambridge University Press, 1993.
[20] H. Kim, T. Abdelzaher, W. Kwon, Minimum-energy asynchronous
dissemination to mobile sinks in wireless sensor networks,
Proceedings of ACM SenSys (2003) 193204.
[21] B. Krishnamachari, Y. Mourtada, S. Wicker, The energy-robustness
tradeoff for routing in wireless sensor networks, Proceedings of IEEE
ICC (2003) 18331837.
[22] L. Li, J.Y. Halpern, Minimum-energy mobile wireless networks
revisited, Proceedings of IEEE ICC (2001) 278283.
[23] J. Li, P. Mohapatra, An analytical model for the energy hole problem
in many-to-one sensor networks, Proceedings of IEEE VTC (2005)
27212725.
[24] J. Lian, K. Naik, G. Agnew, Data capacity improvement of wireless
sensor networks using non-uniform sensor distribution,
International Journal of Distributed Sensor Networks 2 (2)
(2006) 121145.
[25] S. Lin, J. Zhang, G. Zhou, L. Gu, J.A. Stankovic, T. He, ATPC: adaptive
transmission power control for wireless sensor networks,
Proceedings of ACM SenSys (2006) 223236.
[26] S. Lindsey, C.S. Raghavendra, PEGASIS: power-efcient gathering in
sensor information systems, Proceedings of IEEE Aerospace
Conference 3 (2002).
[27] S. Lindsey, C.S. Raghavendra, K.M. Sivalingam, Data gathering in
sensor networks using the energy delay metric, Proceedings of
IPDPS (2001) 20012008.
[28] S. Lindsey, C. Raghavendra, K. Sivalingam, Data gathering algorithms
in sensor networks using energy metrics, IEEE TPDS 13 (9) (2002)
924935.
[29] D. Liu, X. Hu, X. Jia, Energy efcient information dissemination
protocols by negotiation for wireless sensor networks, Computer
Communications 29 (issue 11) (2006) 21362149.
1934
[30] J. Luo, J.-P. Hubaux, Joint mobility and routing for lifetime elongation
in wireless sensor networks, Proceedings of IEEE Infocom (2005)
17351746.
[31] D. Luo, X. Zhu, X. Wu, G. Chen, Maximizing lifetime for the shortest
path aggregation tree in wireless sensor networks, Proceedings of
IEEE INFOCOM (2011) 15661574.
[32] A. Manjeshwar, D.P. Agrawal, APTEEN: a hybrid protocol for efcient
routing and comprehensive information retrieval in wireless sensor
networks, Proceedings of IPDPS (2002) 195202.
[33] A. Manjeshwar, D.P. Agrawal, TEEN: a routing protocol for enhanced
efciency in wireless sensor networks, Proceedings of IPDPS (2001)
20092015.
[34] R.T. Marler, J.S. Arora, Survey of multi-objective optimization
methods for engineering, Structural and Multidisciplinary
Optimization 26 (6) (2004) 369395.
[35] M. Miller, C. Sengul, I. Gupta, Exploring the energy-latency trade-off
for broadcasts in energy-saving sensor networks, Proceedings of
IEEE ICDCS (2005) 1726.
[36] S. Nath, P.B. Gibbons, Communicating via reies: geographic
routing on duty-cycled sensors, Proceedings of IPSN (2007) 440449.
[37] S. Olariu, I. Stojmenovic, Design guidelines for maximizing lifetime
and avoiding energy holes in sensor networks with uniform
distribution and uniform reporting, Proceedings of IEEE INFOCOM
(2006) 112.
[38] D. Pompili, T. Melodia, I.F. Akyildiz, Routing algorithms for delayinsensitive and delay-sensitive applications in underwater sensor
networks, Proceedings of ACM MobiCom (2006) 298309.
[39] V. Rodoplu, T.H. Meng, Minimum energy mobile wireless networks,
IEEE Journal on Selected Areas in Communications 17 (8) (1999)
13331344.
[40] R. Sarkar, X. Yin, J. Gao, F. Luo, X.D. Gu, Greedy routing with
guaranteed delivery using ricci ows, Proceedings of IPSN (2009)
121132.
[41] R.C. Shah, S. Roy, S. Jain, W. Brunette, Data MULEs: modeling a threetier architecture for sparse sensor networks, Proceedings of SNPA
(2003) 3041.
[42] E. Shih, S. Cho, N. Ickes, R. Min, A. Sinha, A. Wang, A. Chandrakasan,
Physical layer driven protocol and algorithm design for energyefcient wireless sensor networks, Proceedings of ACM MobiCom
(2001) 272287.
[43] K. Sohrabi, J. Gao, V. Ailawadhi, G. Pottie, Protocols for selforganization of a wireless sensor network, IEEE Personal
Communications 7 (5) (2000) 1627.
[44] M. Soltan, M. Maleki, M. Pedram, Lifetime-aware hierarchical
wireless sensor network architecture with mobile overlays,
Proceedings of IEEE Radio and Wireless Symposium (2007) 325328.
[45] W. Wang, V. Srinivasan, K.-C. Chua, Using mobile relays to prolong
the lifetime of wireless sensor networks, Proceedings of ACM
MobiCom (2005) 270283.
[46] A. Woo, T. Tony, D. Culler, Taming the underlying challenges of
reliable multi-hop routing in sensor networks, Proceedings of ACM
SenSys (2003) 1427.
[47] X. Wu, G. Chen, S.K. Das, On the energy hole problem of nonuniform
node distribution in wireless sensor networks, Proceedings of IEEE
MASS (2006) 180187.
[48] X. Wu, G. Chen, S.K. Das, Avoiding energy holes in wireless sensor
networks with nonuniform node distribution, IEEE TPDS 19 (5)
(2008).
[49] G. Xing, C. Lu, R. Pless, Q. Huang, On greedy geographic routing
algorithms in sensing-covered networks, Proceedings of ACM
MobiHoc (2004) 3142.
[50] G. Xing, X. Wang, Y. Zhang, C. Lu, R. Pless, C. Gill, Integrated coverage
and connectivity conguration for energy conservation in sensor
networks, ACM TOSN 1 (1) (2005) 3672.
[51] X. Yang, N. Vaidya, A wakeup scheme for sensor networks: achieving
balance between energy saving and end-to-end delay, Proceedings
of IEEE RTAS (2004) 1926.
[52] M. Yarvis, N. Kushalnagar, H. Singh, A. Rangarajan, Y. Liu, S. Singh,
Exploiting heterogeneity in sensor networks, Proceedings of IEEE
INFOCOM (2005) 878890.
[53] F. Ye, H. Luo, J. Cheng, S. Lu, L. Zhang, A two-tier data dissemination
model for large-scale wireless sensor networks, Proceedings of ACM
MobiCom (2002) 148159.
[54] H. Zhang, J. Hou, Maintaining sensing coverage and connectivity in
large sensor networks, Ad Hoc & Sensor Wireless Networks 1 (1-2)
(2005) 89124.
[55] J. Zhao, R. Govindan, Understanding packet delivery performance in
dense wireless sensor networks, Proceedings of ACM SenSys (2003)
113.
[56] W. Zeng, R. Sarkar, F. Luo, X.D. Gu, J. Gao, Resilient routing for sensor
networks using hyperbolic embedding of universal covering space,
Proceedings of IEEE INFOCOM (2010) 16941702.
[57] G. Zhou, T. He, S. Krishnamurthy, J. Stankovic, Impact of radio
irregularity on wireless sensor networks, Proceedings of MobiSys
(2004) 125138.
[58] G. Zhou, T. He, S. Krishnamurthy, J. Stankovic, Models and solutions
for radio irregularity in wireless sensor networks, ACM TOSN 2 (2)
(2006) 221262.
[59] M. Zorzi, R. Rao, Geographic random forwarding (GeRaF) for ad hoc
and sensor networks: Multi-hop performance, IEEE TMC 2 (4) (2003)
337348.
[60] M. Zorzi, R. Rao, Geographic random forwarding (GeRaF) for ad hoc
and sensor networks: energy and latency performance, IEEE TMC 2
(4) (2003) 349365.
Habib M. Ammari is an Associate Professor in

the Department of Computer and Information
Science, College of Engineering and Computer
Science, University of Michigan-Dearborn,
and the Founding Director of Wireless Sensor
and Mobile Ad-hoc Networks (WiSeMAN)
Research Lab at the University of MichiganDearborn since September 2011. Prior to that,
he was on the faculty of the Department of
Computer Science, Hofstra University from
September 2008 - August 2011. He obtained
his second Ph.D. degree in Computer Science
and Engineering from the University of Texas at Arlington in May 2008,
and his second Masters degree in Computer Science from Southern
Methodist University in December 2004. Also, he obtained his rst Ph.D.
(Highly Honorable with Praise) and Masters degrees in Computer Science
from the Faculty of Sciences of Tunis in December 1996 and July 1992,
respectively.
His main research interests lie in the areas of wireless sensor and mobile
ad hoc networking, multi-hop mobile wireless Internet architectures and
protocols, and cyber physical systems. In particular, he is interested in
coverage, connectivity, energy-efcient data routing and information
dissemination, fault tolerance, and security in wireless sensor networks,
and the interconnection between wireless sensor networks, mobile ad
hoc networks, and the global IP Internet. He has a strong publication
record in top-quality journals, such as IEEE TPDS, IEEE TC, ACM TAAS,
Elsevier COMNET, Elsevier PMC, Elsevier JPDC, and high-quality conferences, such as IEEE SECON, IEEE ICDCS, EWSN, and ICDCN. He published
his rst Springer book, titled Challenges and Opportunities of Connected
k-Covered Wireless Sensor Networks: From Sensor Deployment to Data
Gathering, in August 2009.
He received several prestigious awards, including the Certicate of
Appreciation Award at the 17th ACM MobiCom in September 2011, the
Outstanding Leadership Award at the 20th IEEE ICCCN in August 2011,
the Best Symposium Award at the 7th IEEE IWCMC in July 2011, the
Lawrence A. Stessin Prize for Outstanding Scholarly Publication from
Hofstra University in May 2010, the Faculty Research and Development
Grant Award from Hofstra College of Liberal Arts and Sciences in May
2009, the Best Paper Award at the 5th EWSN in February 2008, and the
Best Paper Award at the Google Ph.D. Forum-6th IEEE PerCom in March
2008. Also, he was an ACM Student Research Competition (ACM SRC)
nominee at ACM MobiCom 2005. He is the recipient of the Nortel Outstanding CSE Doctoral Dissertation Award in February 2009, and the John
Steven Schuchman Award for 20062007 Outstanding Research by a PhD
Student in February 2008. He received a 3-year US National Science
Foundation (NSF) Research Grant Award (400 K) in June 2009, and the US
National Science Foundation Faculty Early Career Development(CAREER)
Award(450 K) in January 2011. He is the University of Michigan-Dearborns second NSF CAREER Award recipient, and the Hofstras rst ever
NSF CAREER Award recipient.
He serves as Associate Editor of several journals, including ACM TOSN,
IEEE TC, Wiley WCMC, AHSWN, Wiley IJCS, and NPA. He is on the Editorial
board of the International Journal of Mobile Communications and the
International Journal on Advances in Networks and Services. Also, he is on
the Editorial Review Board of the International Journal of Distributed
Systems and Technologies. He serves as Track Co-Chair of IEEE ICCCN
2012-SEP, Demo Co-Chair of IEEE WoWMoM 2012, Program Co-Chair of

CPNS 2012, Program Vice-Chair of MUSIC 2012, Workshop Chair of
WiMAN 2012, Program Vice-Chair of IEEE/IFIP EUC 2011, Workshop Chair
of WiMAN 2011, and Symposium Co-Chair of IWCMC 2011 Wireless
Sensor Networks Symposium. Also, he served as Workshop Co-Chair of
WiMAN 2010, Symposium Co-Chair of IWCMC 2010 Wireless Sensor
Networks Symposium, Program Co-Chair of IQ2S 2009, and Workshop CoChair of WiMAN 2009. He has served as Publicity (Co) Chair of numerous
conferences, symposia, and workshops, including ACM MobiCom 2011.
Also, he has served as a reviewer for several international journals,
1935
including IEEE TMC, IEEE TPDS, IEEE SMC, ACM TOSN, IEEE TVT, IEEE
TWireless, IEEE CL, Springers MONE, Springers WINET, Elseviers ADHOC,
Elseviers COMNET, AHSWN, IJSNet, Elseviers IPL, Elseviers COMCOM,
Elseviers JPDC, Elseviers INS, IJCA, and Elseviers DKE, and as a Technical
Program Committee member of numerous IEEE and ACM conferences,
symposia, and workshops, including IEEE Infocom, IEEE ICDCS, IEEE
DCOSS, IEEE INSS, IEEE PerCom, IEEE GlobeCom, IEEE ICC, SSS, IEEE MASS,
IEEE MSN, IEEE LCN, IEEE VTC, IEEE ICCCN, EWSN, ICDCN, and AdHocNets.

Computer Networks: Habib M. Ammari

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Computer Networks: Habib M. Ammari

Enviado por

Direitos autorais:

Formatos disponíveis

Computer Networks 57 (2013) 19131935

Contents lists available at SciVerse ScienceDirect

On the energy-delay trade-off in geographic forwarding

Tel.: +1 313 593 5239.

Recent advances in sensor technology and wireless

H.M. Ammari / Computer Networks 57 (2013) 19131935

which are suitable for various civilian applications, such as

H.M. Ammari / Computer Networks 57 (2013) 19131935

previous work that jointly considers minimum energy

and sets up the forwarding information at sensors closest

H.M. Ammari / Computer Networks 57 (2013) 19131935

for a hierarchical WSN using mobile relays with a goal to

Etx si ; sj aEelec eda si ; sj

where d is the Euclidean distance function, a is the data

where c = qd + td. A similar result can be found in [20].

H.M. Ammari / Computer Networks 57 (2013) 19131935

Assumption 5 (Always-on sensor model). The sensors are

Proof. According to the energy model [16] discussed earlier,

the sensor. Now, if we consider the second part aeda(si, sj) in

Notice that the minimum transmission distance

Fig. 1. Slicing of the communication range of sensors (Case a = 2).

H.M. Ammari / Computer Networks 57 (2013) 19131935

Fig. 2. Impact of k on the size of the subset CPF(si, sm, k, b).

4.3. Uniform energy depletion characterization

H.M. Ammari / Computer Networks 57 (2013) 19131935

It is worth noting that higher values of k mean that the

5.1.2. Data forwarding along non-direct paths

5.1. Simple analytical bounds

kb2k 1ds0 ; sm dmin

respectively, and their respective upper bounds are given by

kb2R dmin ds0 ; sm dmin

kb2k 1dmin ds0 ; sm

Using the same reasoning as in Lemma 2, we obtain

kb2k 1dmin ds0 ; sm

H.M. Ammari / Computer Networks 57 (2013) 19131935

Fig. 3. Non-shortest path between s0 and sm.

5.1.3. Numerical results

and X is a set of admissible solutions. It is assumed that

5.2. Multi-objective optimization approach

5.2.1. Overview of the WES approach

A2 k ek dmin . Thus, A2(k) becomes a dominant factor

< 0; for 1 6 k 6 nccb

Thus, we consider its opposite function given by

PF exp s0 ; sm ; k jCPF exp s0 ; sm ; kj

Now, PFexp(s0, sm, k) is a convex function (Fig. 6). Thus, we

H.M. Ammari / Computer Networks 57 (2013) 19131935

H.M. Ammari / Computer Networks 57 (2013) 19131935

Notice that our multi-objective optimization problem is

H.M. Ammari / Computer Networks 57 (2013) 19131935

Moreover, we obtain a Pareto-optimum solution [9].5.2.3.

corresponding to the minimum of M(k) increases, meaning

H.M. Ammari / Computer Networks 57 (2013) 19131935

5.3.2. Concentric circular band selection

Erem sPF1 maxfErem sj : sj 2 CPFs0 ; sm ; k; hg

H.M. Ammari / Computer Networks 57 (2013) 19131935

where Erem(sj) is the remaining energy of sensor sj. In other

lation of the multi-objective optimization problem and its

H.M. Ammari / Computer Networks 57 (2013) 19131935

5.4. Relaxation of several key assumptions

5.4.1. Relaxing the sensor homogeneity model

5.4.2. Relaxing the communication disk model

H.M. Ammari / Computer Networks 57 (2013) 19131935

vex. We assumed convexity of the communication range of