EXPLOITING NODE MOBILITY FOR ENERGY OPTIMIZATION
IN WIRELESS SENSOR NETWORKS
By
Fatme Mohammad El-Moukaddem

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILOSOPHY
Computer Science
2012

ABSTRACT

EXPLOITING NODE MOBILITY FOR ENERGY OPTIMIZATION
IN WIRELESS SENSOR NETWORKS
By
Fatme Mohammad El-Moukaddem
Wireless Sensor Networks (WSNs) have become increasingly available for data-intensive
applications such as micro-climate monitoring, precision agriculture, and audio/video surveillance. A key challenge faced by data-intensive WSNs is to transmit the sheer amount of data
generated within an application’s lifetime to the base station despite the fact that sensor
nodes have limited power supplies such as batteries or small solar panels. The availability
of numerous low-cost robotic units (e.g. Robomote and Khepera) has made it possible to
construct sensor networks consisting of mobile sensor nodes. It has been shown that the controlled mobility offered by mobile sensors can be exploited to improve the energy efficiency
of a network.
In this thesis, we propose schemes that use mobile sensor nodes to reduce the energy
consumption of data-intensive WSNs. Our approaches differ from previous work in two main
aspects. First, our approaches do not require complex motion planning of mobile nodes, and
hence can be implemented on a number of low-cost mobile sensor platforms. Second, we
integrate the energy consumption due to both mobility and wireless communications into a
holistic optimization framework.
We consider three problems arising from the limited energy in the sensor nodes. In
the first problem, the network consists of mostly static nodes and contains only a few mobile nodes. In the second and third problems, we assume essentially that all nodes in the

WSN are mobile. We first study a new problem called max-data mobile relay configuration
(MMRC) that finds the positions of a set of mobile sensors, referred to as relays, that maximize the total amount of data gathered by the network during its lifetime. We show that the
MMRC problem is surprisingly complex even for a trivial network topology due to the joint
consideration of the energy consumption of both wireless communication and mechanical
locomotion. We present optimal MMRC algorithms and practical distributed implementations for several important network topologies and applications. Second, we consider the
problem of minimizing the total energy consumption of a network. We design an iterative
algorithm that improves a given configuration by relocating nodes to new positions. We show
that this algorithm converges to the optimal configuration for the given transmission routes.
Moreover, we propose an efficient distributed implementation that does not require explicit
synchronization. Finally, we consider the problem of maximizing the lifetime of the network. We propose an approach that exploits the mobility of the nodes to balance the energy
consumption throughout the network. We develop efficient algorithms for single and multiple round approaches. For all three problems, we evaluate the efficiency of our algorithms
through simulations. Our simulation results based on realistic energy models obtained from
existing mobile and static sensor platforms show that our approaches significantly improve
the network’s performance and outperform existing approaches.

ACKNOWLEDGMENTS

I would like express my gartitude to my advisors Dr. Eric Torng and Dr. Guoliang
Xing for their constant guidance and useful critiques. I would also like to extend my thanks
to Dr. Li Xiao, Dr. Abdol-Hossein Esfahanian, and Dr. Haydar Radha for serving on
my committee. Finally, I wish to thank my family for their support and encouragement
throughout my study.

iv

TABLE OF CONTENTS

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vii

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Background . . . . . . .
2.1 Mobile Base Stations
2.2 Data Mules . . . . .
2.3 Mobile Relays . . . .
2.4 Discussion . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

1

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

7
7
9
10
11

3 System Model . . . . . . . . . .
3.1 Network Model . . . . . . . .
3.2 Energy Consumption Models .
3.2.1 Mobility . . . . . . . .
3.2.2 Communication . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

14
14
15
15
16

4 Maximization of Data Gathering Capacity Using Mobile Nodes
4.1 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Max-Data Mobile Relay Configuration . . . . . . . . . . . .
4.2 1-MMRC: Single Relay per Link . . . . . . . . . . . . . . . . . . . .
4.2.1 Base Case of 1-MMRC . . . . . . . . . . . . . . . . . . . . .
4.2.2 The Assignment Problem . . . . . . . . . . . . . . . . . . .
4.2.3 1-MMRC in a Line, Star, and Tree with Data Reduction . .
4.2.4 1-MMRC in a Tree without Data Reduction . . . . . . . . .
4.2.5 Distributed Algorithm . . . . . . . . . . . . . . . . . . . . .
4.3 2-MMRC: Two Relays per Link . . . . . . . . . . . . . . . . . . . .
4.3.1 Base Case . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.2 Distributed Algorithm . . . . . . . . . . . . . . . . . . . . .
4.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.2 1-MMRC in a Line . . . . . . . . . . . . . . . . . . . . . . .
4.4.3 1-MMRC in a Star . . . . . . . . . . . . . . . . . . . . . . .
4.4.4 Fair 1-MMRC in Trees with Data Reduction . . . . . . . . .
4.4.5 Fair 1-MMRC in Trees without data reduction . . . . . . . .
4.4.6 Two Relays Per Link . . . . . . . . . . . . . . . . . . . . . .
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

18
20
22
24
24
32
33
34
35
38
39
41
42
42
44
46
48
50
51
53

5 Total Energy Minimization Using Mobile Nodes
5.1 Problem Definition . . . . . . . . . . . . . . . . .
5.1.1 An Illustrative Example . . . . . . . . . .
5.1.2 Problem Formulation . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

55
56
56
58

v

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

5.2

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

59
59
62
64
65
66
66
68
71
72
76
76
78
80
85
87

6 Maximizing Network Lifetime Using Node Rotation
6.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Problem Definition . . . . . . . . . . . . . . . . . . . .
6.3 Node Rotation Algorithms . . . . . . . . . . . . . . . .
6.3.1 Algorithm NR1 . . . . . . . . . . . . . . . . . .
6.3.2 Centralized Algorithm CNR(r) . . . . . . . . .
6.3.3 Distributed Algorithm DNR . . . . . . . . . . .
6.4 Upper Bounds on Lifetime Improvement Ratio . . . . .
6.5 Simulation Results . . . . . . . . . . . . . . . . . . . .
6.5.1 Single Rotation Round . . . . . . . . . . . . . .
6.5.2 Multiple Rotation Rounds . . . . . . . . . . . .
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

89
90
92
94
94
96
96
99
101
102
103
108

5.3

5.4
5.5
5.6

5.7

Centralized Solution . . . . . . . . . . .
5.2.1 Energy Optimization Framework
5.2.2 Base Case . . . . . . . . . . . . .
5.2.3 Static Tree Construction . . . . .
5.2.4 Node Insertion . . . . . . . . . .
Tree Optimization . . . . . . . . . . . .
5.3.1 Extended Base Case . . . . . . .
5.3.2 Optimization Algorithm . . . . .
Efficiency and Optimality . . . . . . . .
Distributed Algorithms . . . . . . . . . .
Simulations . . . . . . . . . . . . . . . .
5.6.1 Setup . . . . . . . . . . . . . . .
5.6.2 Performance Metrics . . . . . . .
5.6.3 Centralized Algorithm . . . . . .
5.6.4 Distributed Algorithm . . . . . .
Summary . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.2 Open problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
A Proofs of Claims and Theorems . . . . . . . . . . . . . . . . . . . . . . . . . 114
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

vi

LIST OF TABLES

Table 3.1

Parameter values for energy consumption during communication . .

17

Table 4.1

Average Improvement Ratios for Base Case Problems . . . . . . . .

51

Table 5.1

Energy consumption comparison . . . . . . . . . . . . . . . . . . . .

57

Table 6.1

Upper Bounds on Lifetime Improvement Ratios in Balanced Trees . 101

vii

LIST OF FIGURES

Figure 4.1

Two static tree topologies; each link is labeled with its capacity. . .

23

Figure 4.2

Example scenario illustrating the base case. Using relay r at position
r∗ increases the amount of data that can be sent. . . . . . . . . . . .

25

Figure 4.3

The highlighted portion of gr (α) is part of M (α) . . . . . . . . . . .

28

Figure 4.4

The highlighted portion of hb (α) is part of M (α) . . . . . . . . . . .

28

Figure 4.5

The set of point R ∪ B form a unimodal set . . . . . . . . . . . . . .

28

Figure 4.6

Points of f (α) connect pieces of R and B and form a unimodal continuous function . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

28

The function M is unimodal. For interpretation of the references
to color in this and all other figures, the reader is referred to the
electronic version of this dissertation. . . . . . . . . . . . . . . . . .

28

Figure 4.8

Local algorithm executed by each static node . . . . . . . . . . . . .

36

Figure 4.9

Local algorithm executed by each mobile node . . . . . . . . . . . .

37

Figure 4.10

Using r1 and r2 at r1 and r2 increases the amount of data that can
be transmitted from s1 to s2 . . . . . . . . . . . . . . . . . . . . . . .

39

Improvement in data gathering capacity of our algorithms in line
topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

45

Improvement in data gathering capacity of our algorithms in star
topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

47

Improvement in data gathering capacity of our algorithms in tree
topologies with data reduction at the nodes . . . . . . . . . . . . . .

48

Improvement in data gathering capacity of our algorithms in tree
topologies without data reduction at the nodes . . . . . . . . . . . .

49

Figure 4.7

Figure 4.11

Figure 4.12

Figure 4.13

Figure 4.14

viii

Figure 4.15

Improvement in data gathering capacity of the distributed implementations for 1-MMRC and 2-MMRC for trees with data reduction . .

53

Reduction in energy consumption due to mobile relay. As the data
chunk size increases, the optimal position converges to the midpoint
of s1 s3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

56

Figure 5.2

Optimal routing tree for 0 ≤ m ≤ 15MB. . . . . . . . . . . . . . . .

60

Figure 5.3

Optimal configuration for 15MB≤ m ≤ 25MB: same topology but
nodes relocate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

Optimal configuration for 25MB≤ m ≤ 60 MB. A new topology is
used. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

Optimal configuration for 100MB≤ m ≤ 150MB: a new topology
with more nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

Example of optimal configurations as a function of amount of data
to be transferred. In each part, source nodes s1 and s2 must send m
bits of data to the sink r. We consider m up to 150MB. . . . . . . .

60

Algorithm to compute the optimal position of a relay node that receives data from a single node and transmits the data to a single
node. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

64

Figure 5.1

Figure 5.4

Figure 5.5

Figure 5.6

Figure 5.7

Figure 5.8

Centralized algorithm to compute the optimal positions in a given tree 69

Figure 5.9

Convergence of iterative approach to the optimal solution. Each line
shows the configuration obtained after 2 iterations. The optimal configuration is reached after 6 iterations. . . . . . . . . . . . . . . . . .

70

Figure 5.10

Local algorithm executed by tree nodes . . . . . . . . . . . . . . . .

74

Figure 5.11

Graph of the average static energy consumption ratio of TREE+INS+FO
as a function of data chunk size for our three tree construction strategies PB, HB, and GG . . . . . . . . . . . . . . . . . . . . . . . . . . 79

Figure 5.12

Graph of the average reduction ratio of optimization INS+FO as a
function of data chunk size for our three tree construction strategies
PB, HB, and GG . . . . . . . . . . . . . . . . . . . . . . . . . . . .

81

Graph of the average reduction ratio of optimization FO as a function
of data chunk size for our three tree construction strategies PB, HB,
and GG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

81

Figure 5.13

ix

Figure 5.14

Figure 5.15

Figure 5.16

Figure 6.1

Figure 6.2

Figure 6.3

Graph of the average reduction ratio of optimization INS as a function
of data chunk size for our three tree construction strategies PB, HB,
and GG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

82

Graph of the average reduction ratio of the centralized and distributed
GG+INS+FO optimizations as a function of the number of sources,
a data chunk of 75MB and different values of k. . . . . . . . . . . .

84

Graph of the average reduction ratio of the centralized optimization
(INS+FO) and three distributed optimizations (FO, INS, INS+FO)
as a function of data chunk size for the greedy geographic tree GG. .

86

Algorithm NR1 for computing the optimal lifetime through a single
round of rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . .

95

Algorithm executed at the beginning of each round in a centralized
setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

97

Local algorithm executed at the beginning of each round in a distributed setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

98

Figure 6.4

Balanced tree topology of degree d + 1 and lowest level h . . . . . . 100

Figure 6.5

Average lifetime improvement ratios of CNR(r) and DNR as a function of r plotted as a fraction of L(I) . . . . . . . . . . . . . . . . . 104

Figure 6.6

Average number of relocations of CNR(r) and DNR as a function of
r plotted as a fraction of L(I). . . . . . . . . . . . . . . . . . . . . . 104

Figure 6.7

CCDF of lifetime improvement ratio of CNR and DNR with r =
L(I)/2 and h = 1, 2, and 4. . . . . . . . . . . . . . . . . . . . . . . . 106

Figure 6.8

CCDF of the lifetime comparison ratios of DNR versus LEACH and
multihop LEACH. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

Figure 6.9

Average improvement ratio of our distributed approach DNR with respect to energy aware LEACH and multi-hop LEACH as the number
of nodes in the network increases . . . . . . . . . . . . . . . . . . . . 107

Figure A.1

gr (α) is unimodal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

Figure A.2

Subdivision of s1 s2 into 3 intervals. . . . . . . . . . . . . . . . . . . 119

Figure A.3

In I1 , hα is on the line s1 sα . In I2 , s1 hα increases with sα . In I3 , hα
coincides with r. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
x

Figure A.4

Any point c between a and b such that m(a) > m(b) satisfies the
condition m(c) > m(b) . . . . . . . . . . . . . . . . . . . . . . . . . 120

xi

CHAPTER 1
Introduction
Recent years have seen the deployment of WSNs in a variety of data-intensive applications
including micro-climate and habitat monitoring [51], precision agriculture, and audio/video
surveillance [37]. It is shown in [16] that a moderate-size WSN can gather up to 1 Gb/year
from a biological habitat. Due to the limited storage capacity of sensor nodes, most data
must be transmitted to the base station for archiving and analysis. However, sensor nodes
must operate on limited power supplies such as batteries or small solar panels. Therefore, a
key challenge faced by data-intensive WSNs is to minimize the energy consumption of sensor
nodes such that the sheer amount of data generated within the lifetime of the application
can be transmitted to the base station.
Several approaches have been introduced to reduce the energy consumption of nodes and
prolong the lifetime of the network. In general, these approaches can be classified into four
main groups: duty cycling, power control, data reduction and controlled mobility.
Duty cycling approaches reduce the energy consumption during a period of time by
reducing the duty cycle of nodes, which is the fraction of time the node is turned on during

1

its lifetime. Nodes follow a sleep/wake up schedule. When in sleep mode, nodes enter in a
low power consumption stage since one of the main energy consumption components (the
radio) is turned off. When nodes wake up, they resume their data reception and transmission
activities. These approaches take advantage of redundancy of sensor nodes in the deployment
area; a single node can perform the same job of multiple nodes in a more energy efficient way.
However, connectivity and coverage [4, 47, 29] issues arise when duty cycling is used. Nodes
need to collaboratively set their sleep/wake up schedules to ensure data can successfully
reach the sink. Different power management protocols have been proposed; they include ondemand [44, 66], scheduled rendez-vous [56, 34, 27] and asynchronous [55, 39, 68] strategies.
In power control approaches, the main idea is that nodes reduce their energy consumption
by reducing their transmission power to the minimum levels needed. In [5, 25], the authors
propose MAC protocols for power control in which nodes determine the minimum transmit
levels needed based on several criteria such as channel gains and noise and interference levels
at the receiver. In [32, 31, 6, 48], the authors propose different transmission power assignment
schemes to reduce the total energy consumption while maintaining connectivity.
In data reduction approaches the goal is to reduce the amount of data that gets sent
to the sink. This is usually done by reducing the amount of data that is generated by the
sensors. In these cases, the energy consumed by the sensing component is reduced. In other
cases, the strategy is to reduce the number of transmissions needed to relay data to the
sink, and this reduces the energy consumed by the radio component. Different techniques
have been proposed. The first is data aggregation [67, 30, 65]. This consists of in network
processing of the data and generating a smaller size result to send to the sink as opposed
to sending all the data gathered to the sink where processing takes place. This technique
2

is application specific; for example in an environment monitoring application, we might be
interested in average temperatures so data aggregation can be used. The second technique
is data compression. The data sensed is encoded into smaller size data at the sensors and
then decoded at the sink [11, 60, 53]. The third technique is data prediction. This technique
uses the data sensed to build a model describing data evolution. It can be done either at
the sink or at the sources to predict (within some error bound) new data instead of sensing
it. The last technique is efficient data acquisition which uses efficient sampling strategies.
For example, adaptive sampling [10, 35] exploits temporal and/or spatial correlation of data
to reduce the amount of data sensed. These approaches regularly measure the error margin
between the actual data and the predicted data and dynamically adjust the frequency of
actual data generation.
The last scheme for improving energy efficiency is using mobility. Recent work showed
that the energy cost of WSNs can be significantly reduced by utilizing the mobility of nodes.
Several different approaches have been proposed. A robotic unit may move around the
network and collect data from static nodes through one-hop or multi-hop transmissions
[15, 36, 63, 26, 64]. The mobile node may serve as the base station or a “data mule” that
transports data between static nodes and the base station [24, 46, 23]. Mobile nodes may
also be used as relays [62] that forward the data from source nodes to the base station.
Several movement strategies for mobile relays have been studied in [62, 17].
Although the effectiveness of mobility in energy conservation is demonstrated by previous
studies, the following two key issues have not been addressed by any mobility technique.
First, the movement cost of mobile nodes is not accounted for in the total network energy
consumption. Instead, mobile nodes are often assumed to have replenishable energy supplies.
3

For instance, a mobile node may periodically recharge its battery at a fixed charging dock
[26]. However, energy replenishment is not always feasible due to the constraints of the
physical environment. Second, complex motion planning of mobile nodes is often assumed in
existing solutions which introduces significant design complexity and manufacturing costs. In
[50, 18, 26, 64], mobile nodes need to compute optimal motion paths and continuously change
their orientation and/or speed of movement. Such capabilities are usually not supported by
existing low-cost mobile sensor platforms. For instance, Robomote [12] nodes are designed
using 8-bit CPUs and small batteries that only last for about 25 minutes in full motion.
Complex motion planning is not practical for such platforms due to the limited computational
power and short battery lifetime.
We explore the use of low-cost mobile sensor nodes to improve the energy efficiency of
a WSN. Different from mobile base station or data mules, mobile nodes do not transport
data; instead, they move to different locations and then remain stationary to forward data
along the paths from sources to the base station. As a result, the communication delays can
be significantly reduced compared with using mobile stations or data mules. We propose
different approaches to reduce the total energy consumption of data-intensive WSNs, to
maximize the efficiency of the network in gathering data and to maximize the network
lifetime. Our approaches are motivated by the current state of mobile sensor platform
technology. On the one hand, numerous low-cost mobile sensor prototypes such as Robomote
[12], Khepera [1], and FIRA [28] are now available. Their manufacturing cost is comparable
to that of typical static sensor platforms. As a result, they can be massively deployed in
a network. Our approaches take advantage of this capability by assuming that we have
a large number of mobile sensor nodes. On the other hand, due to low manufacturing
4

cost, existing mobile sensor platforms are typically powered by batteries and only capable
of limited mobility. Consistent with this constraint, our approaches only require simple
motions of mobile relays, i.e., one-shot relocation to designated positions after deployment.
Compared with our approach, existing mobility approaches (such as mobile base station and
data mule) typically assume a small number of powerful mobile nodes.
We note that all our formulations are based on realistic energy consumption models of
both node movement and wireless communications that we developed using empirical results
obtained from two widely used radios (CC2420 [3], CC1000 [2]) which is in contrast to existing
mobility approaches that only account for the communication energy consumption. We make
the following contributions.
1. We leverage mobility of sensor nodes to improve the performance of wireless sensor
networks. We consider different criteria for measuring the performance of the network
such as the data gathering capacity, the total energy consumption and the network’s
lifetime.
2. For each performance criterion, we present a new problem formulation. We develop
optimal algorithms as well as fast heuristics to reconfigure the positions of the mobile
nodes in the network such that the performance is greatly improved. We also propose
practical distributed implementations for all our algorithms. We show that these algorithms have a low overhead as they converge to the final configuration after only a
few iterations and they require little synchronization between nodes.
3. We conduct extensive simulations based on the realistic energy models we adopted. We
show that our algorithms outperform existing non mobility solutions and significantly

5

improve the network’s energy efficiency. Moreover, our distributed implementations
are shown to have fast convergence and low messaging overhead.
The rest of this thesis is organized as follows. Chapter 2 reviews existing mobility approaches. In Chapter 3, we introduce the network model and develop models for energy
consumption of different operations such as communication and mobility. In Chapter 4, we
consider networks consisting of mostly static nodes and containing a small number of mobile
nodes. We define the problem of maximizing the efficiency of a network by maximizing the
total amount of data that the sink can gather throughout the lifetime of the network. We
propose optimal algorithms as well as heuristics that produce close to optimal results. In
Chapter 5, we define the problem of minimizing the total energy consumption in networks
consisting entirely of mobile nodes and we propose iterative algorithms for the single-flow
and multiple flows problems. We also show the advantage of our approach through extensive simulation results. In Chapter 6, we consider the problem of maximizing the network
lifetime. We propose algorithms to balance the energy consumption across the network and
show through simulations that the lifetime is substantially increased. We conclude the thesis
in Chapter 7.

6

CHAPTER 2
Background
Several approaches have been used to reduce the energy consumption in battery operated
wireless sensor networks. They include duty cycling strategies, data driven strategies and
mobility. In this chapter, we review mobility strategies for reducing energy consumption.
These strategies can be divided further into three classes: mobile base stations, data mules
and mobile relays. In the following sections, we discuss each of these classes.

2.1

Mobile Base Stations

One of the main energy efficiency challenges of static wireless sensor networks is that
nodes closer to the base station deplete their energy at a faster rate than other nodes since
they have to transmit more data. One remedy is to use a mobile base station. When the
base station relocates to a new position, it reduces the traffic load of the nodes close to its
old position and increases the load at nodes close to its new position. In general, mobile base
station approaches aim at balancing the energy consumption in the network by constantly
rotating which nodes are close to the base station.

7

In some work, all the nodes are always performing multiple hop transmissions to the base
station regardless of its position [15, 36, 63]. In [15], the authors divide the (known) lifetime
of the network into equal periods of time called rounds. Base stations are relocated at the
beginning of every round using a linear program with the objective of either minimizing the
total energy spent at the nodes or minimizing the energy spent at any node in the network.
In [36], the authors claim that the best mobility strategy for the base station is to move
along the periphery of the network (assuming nodes are placed in a circular area) because
this strategy minimizes the average load on any node. The authors also propose a different
mobility strategy that takes into consideration the routing paths. This heuristic shortens
the trajectory of the base station to a smaller circle inside the network area and combines
straight and round routing paths from the sensors towards the base station. In [63], the
authors also use a linear programming formulation for finding the mobility paths of the sink.
Unlike [15, 36], their objective is to maximize the system lifetime defined as the time until
the first node dies because of energy depletion. In both [15, 63], the base station can only
move to a predefined and finite set of locations.
In other work, nodes only transmit to the base station when it is close to them (or a
neighbor). The goal is to compute a mobility path to collect data before any node experiences
a buffer overflow [50, 18, 26]. These approaches incur high latencies due to the low to
moderate speed, e.g. 0.1-1 m/s [12, 50], of mobile base stations. In [50], each node in the
network has an overflow period based on its data generation rate and buffer that is known in
advance. The traveling time between any two nodes is also known. The authors prove that
the problem of scheduling visits to static nodes such that no node’s buffer overflows is NPComplete. They propose several heuristics such as earliest deadline first, earliest deadline
8

first with k-lookaheads, greedy based on the weighted sum of two criteria (earliest deadline
and closeness to the sink) as well as known heuristics for solving vehicle routing problems
with time windows. In [18], the authors present an algorithm for finding a route for the
mobile base station that visits all the nodes in the network, some more frequently than
others, such that no node’s buffer overflows between visits. The main idea of the algorithm
is to divide nodes into bins according to their geographic location and their overflow time.
In each bin, a route that visits each node once is computed using a solution for the Traveling
Salesperson Problem. These routes are then concatenated to form a schedule for visiting
all the nodes in the network. Once this schedule is found, the speed of the base station is
adjusted to make the route feasible. In [26], the authors propose a solution that achieves
a balance between the latency of relaying data to the base station and the total energy
consumption at static nodes. Their approach consists of dividing static nodes into clusters.
In each cluster, data is transmitted to the head node. The mobile base station then follows
a shorter path that only communicates with the cluster heads.

2.2

Data Mules

Data mules are similar to mobile base stations [24, 46, 23] in that the mobile node
mechanically carries the data. Data mules pick up data from the sensors and transport it to
the sink. Similar to mobile base stations, data mules introduce large delays. First, sensors
have to wait for a mule to pass by before starting their transmission. Second, data mules
take a long time to then deliver the data to the sink due to their limited speed.
In [38], the data mule visits all the sources to collect data, transports data over some
distance, and then transmits it to the static base station through the network. The goal is to

9

find a movement path that minimizes both communication and mobility energy consumption.
The data mule has a sequence of tasks to perform. The plane is divided into regions such
that each request falls in a different region. Each region is divided using a two dimensional
grid resulting in potential points the mule can visit. The data mule then picks a point from
each region to perform its corresponding task using Dijkstra’s shortest path algorithm.
Similar to [26], the approach in [64] achieves a balance between energy savings and
data collection delay. Additionally, it takes into consideration the deadline requirements for
relaying data to the base station. The goal is to find a set of static nodes in the network
at which the mobile nodes meet to pick up the data and transport it to the base station.
The authors propose an optimal solution when the mobile node path is constrained to the
routing tree and a utility based greedy algorithm when the path is unconstrained.

2.3

Mobile Relays

In the third approach, the network consists of mobile relay nodes along with static base
station and data sources. Relay nodes do not transport data; instead, they move to different
locations to decrease the communication costs. Goldenberg et al. [17] showed that an
iterative mobility algorithm where each relay node moves to the midpoint of its neighbors
converges on the optimal solution for a single routing path. However, they do not take
into consideration the cost of moving the relay nodes. In [52], the authors observe that
the optimal positions for nodes participating in a single flow transmission are the evenly
spaced positions along a straight line between the source and the sink. They propose a
mobility algorithm in which each node iteratively moves to midpoint of its neighbors. In this
approach, mobile nodes decide to move only when moving is beneficial, (i.e., mobility costs

10

are covered by the savings in transmission costs). It is shown that this strategy converges
to the optimal positions. However, the only target point considered is the midpoint of the
neighbors whereas other points may be better when we consider the energy consumed by
movement of the node.

2.4

Discussion

All three problems that we consider aim at efficiently using the limited energy available
in the network by exploiting the mobility of nodes. They all fall in the class of mobile
relays. However, they differ in the network setting, the specific objective they each optimize,
and subsequently in the approach we propose to reach that objective. In the first problem,
MMRC, the network is mostly static and contains only a small number of mobile nodes. A
fraction of the static nodes generates data; the rest are used as relays. We study how to
exploit this small number of mobile nodes to maximize the total amount of energy gathered
during the lifetime of the network. We propose inserting the mobile nodes into the routing
lines to help the weak links in the networks. The second problem we propose considers a
similar setup as MMRC in that only a fraction of the nodes generates data but differs in that
it considers networks that entirely consist of mobile nodes. The objective is to minimize the
total amount of energy consumed by all nodes in the network when relaying data from the
source nodes to the sink. We propose an approach that first constructs a routing tree that
includes both nearby and distant nodes. Then, our approach collectively moves all the mobile
relay nodes in the tree closer together while maintaining the connectivity of the network. The
third problem, maximizing the network lifetime, also considers networks entirely composed
of mobile nodes. Additionally, it considers scenarios in which all the nodes generate data

11

that need to be relayed to the sink. We propose to rotate the nodes within the network one
or multiple times to balance the energy consumption across different nodes.
A general advantage of our approaches is that we keep a relatively stable routing structure; namely, we reduce the disruptions to the network topology caused by mobility. In
MMRC, only a few nodes are inserted into the routing topology and each insertion affects
exactly two neighbors. In our solution to the energy minimization problem, once the routing
tree is constructed, its exact structure is maintained throughout the moving process. In
our final approach, node rotation, the routing topology is maintained with respect to the
locations of nodes.
Compared to the approaches of mobile base stations and data mules, all our approaches
consider the energy consumption of both mobility and transmission. Additionally, with the
exception of the multiple rotation solution to the network lifetime problem, they all require
much simpler motion planning. Specifically, after deployment, each mobile relay relocates
only once and then remains stationary for data forwarding. Such a simple mobility strategy
can be implemented on a number of existing mobile sensor platforms [12, 1, 50].
Our proposed solution to the total energy minimization problem is similar to the mobile relay schemes in [17] and [52]. But one key difference between our approach and those
schemes is that we consider all possible locations as possible target locations for a mobile
node instead of just its current position and the midpoint of its neighbors. Our approaches
(specifically for reducing the total energy consumption and maximizing data gathering capacity) accurately compute the optimal target position of each mobile relay node. Moreover,
we allow nodes to behave differently. If it is beneficial for one node to move, then this node
moves independently of other nodes. We observe that any local improvement is a global
12

improvement. In each transmission round, different subsets of nodes might move to new
locations while the rest remain in place. Our approaches, in this aspect, are flexible to
accommodate for additional mobility constraints on individual nodes.

13

CHAPTER 3
System Model
In this chapter, we introduce the model abstracting the network behavior and properties.
We also define the energy consumption models of several components of sensor nodes.

3.1

Network Model

We consider wireless sensor networks in which nodes are randomly deployed in a two
dimensional field. We consider all links to be bidirectional. Furthermore, we assume that all
nodes know their locations either by GPS units mounted on them or by a localization service
in the network so any node can determine the euclidean distance separating it from any other
node. We model the network as an undirected graph G in which vertices represent network
nodes. Each vertex v is placed in the 2-dimensional Euclidean plane. An edge uv exists in
G if u and v are within communication range from each other so it is possible to establish a
communication channel between both nodes. These communication channels are determined
as a connected subset of the existing edges by the routing algorithm that we adopt in each
application. We assume that all communication channels have low interference and low data

14

loss rate as the communication power level is set such that the signal reception is strong and
reliable as explained in Section 3.2.2.

3.2

Energy Consumption Models

Sensor nodes consume energy during various operations such as communication, computation, movement and sensing. However, communication and mobility energy consumption
are the two most critical sources of energy consumption and battery depletion. In this work,
we develop models for the energy consumed by nodes during transmissions and receptions,
and we adopt energy models for wheeled sensor nodes to model energy consumed by nodes
during movement, all of which facilitate our holistic objective in all the problems that we
consider. This advances significantly upon previous work that considers only the energy
consumed during transmissions. We do not consider the energy consumed in idle listening
states as it can be significantly reduced through existing sleep scheduling schemes [59]. Alternatively, we can view our results as the limits of what can be achieved assuming a perfect
sleep scheduling algorithm.

3.2.1

Mobility

First, we define the mobility energy consumption models. We consider wheeled sensor
nodes with differential drives such as Khepera [1], Robomote [12] and FIRA [28]. This type
of node usually has two wheels, each controlled by an independent engine. The direction
and speed of the node are determined by the angular velocity and direction of each wheel.
We adopt the distance proportional energy consumption model which is appropriate for this

15

kind of node [58]. The energy EM (d) consumed by moving a distance d is modeled as:

EM (d) = kd

The value of the parameter k depends on the speed of the node. In general, there is an
optimal speed at which k is lowest. The parameter k increases as the difference between the
node velocity and the optimal velocity increases. In [58], the authors discuss in details the
variation of the energy consumption with respect to the speed of the mote. When the node
is running at optimal speed, k = 2 J/m [58]. In our simulations, we use k = 1, 2, 3 and 4 to
model different speeds and different terrains.

3.2.2

Communication

We develop our transmission energy consumption model by analyzing empirical results
obtained from two radios CC2420 [3] and CC1000 [2] that are widely used on sensor network
platforms. For CC2420, the authors of [45] studied the power needed for transmitting packets
reliably (e.g., above 95% packet reception ratio) over different distances. Let ET (d, m) be
the energy consumed to transmit m bits reliably over distance d and ER (m) the energy
consumed by receiving m bits. We have

ET (d, m) = m(a + bd2 )
ER (m) = ma

where a , a and b are constants depending on the environment. We assume that a node only
receives data when it has enough energy to transmit that data to the next node. In this case,

16

a node receiving m bits and transmitting it over a distance d consumes m(a + a + bd2 ) =
m(a + bd2 )J.
We now discuss the instantiation of the above model for the CC2420 and CC1000 radio
platforms. We used measurements in [45] on CC2420 obtained in an outdoor environment
under the default parameters to determine the power level needed for different distances to
get a packet reception ratio higher than 95%. The same measurements were collected for the
CC1000 platform. Using these values along with current and voltage requirements from [3]
and [2], we computed the transmission energy consumption model parameters as summarized
in table 3.1. These values are consistent with the theoretical analysis in [20]. We observe
that a > a . On the other hand, there is no dependence on distance d for reception whereas
there is a dependence on d for transmission.
To capture a wider range of motes and environment conditions, we varied the values
for parameters a , a , and b in our simulations over an interval containing the values we
empirically obtained (Table 3.1).
Table 3.1: Parameter values for energy consumption during communication
a’ J/bit
a” J/bit b Jm−2 /bit
CC2420 0.6 ×10−7 1.4 ×10−7 4.0 ×10−10
CC1000 0.3 ×10−7 2.6 ×10−7 2.0 ×10−10

We also note that although the mobility parameter k is roughly 1010 times larger than
the transmission parameter b, the relay node does not move much whereas large amounts of
data are transmitted. For large enough data chunk sizes, the savings in energy transmission
costs compensate for the energy expended to move the nodes resulting in a decrease in total
energy consumed.

17

CHAPTER 4
Maximization of Data Gathering
Capacity Using Mobile Nodes
In this chapter, we propose to use low-cost disposable mobile nodes to increase the ability
of a network to collect data and transmit it to the sink. We consider the following scenario.
We assume that we have a network consisting of static nodes that generate or relay data to
the sink. We also assume that we have a small number of mobile nodes that were added to
the network. We propose an approach that uses these mobile nodes to alleviate bottleneck
transmission links and thus increase the network’s data gathering capacity.
We first present the formulation of a new problem called Max-Data Mobile Relay Configuration (MMRC) for hybrid WSNs composed of both static and mobile nodes. The objective
of MMRC is to configure the positions of mobile relays to maximize the data gathering capacity of WSNs, which is defined as the total amount of data that can be transmitted to the
BS during the lifetime of a network topology. We define k-MMRC as the variant in which
at most k mobile relay nodes may join one link to improve its data gathering capacity. Our

18

formulation is based on realistic energy consumption models of both wireless communications and node movement, which is in contrast to existing mobility approaches that only
account for the communication energy consumption. We then develop an optimal algorithm
for the base case of the 1-MMRC problem with one mobile and two static nodes. Despite
the trivial network configuration, we show that the base case of 1-MMRC has a surprisingly
high complexity. In particular, a direct analytical solution requires finding the roots of a
degree six bivariate polynomial. To the best of our knowledge, this result is the first to
reveal the complexity of data gathering capacity maximization when both communication
and mobility energy consumption are jointly considered. Based on the algorithm to the
base case of 1-MMRC, we develop optimal algorithms for several important network topologies including lines, stars, and trees with and without data reduction techniques. We also
present a practical distributed implementation for these algorithms. Moreover, we consider
the 2-MMRC variant of the problem and develop efficient algorithms for the base case and
general topologies. Finally, we conduct extensive simulations and show that our algorithms
can increase the amount of data gathered during the system lifetime by a factor of 2 or more.
We also show that our algorithms for 1-MMRC achieve similar data gathering capacities as
our 2-MMRC algorithms.
The rest of this chapter is organized as follows. In Section 4.1, we formally define the
MMRC problem. Section 4.2 presents several optimal algorithms for 1-MMRC along with
a practical distributed implementations. In Section 4.3, we propose a solution for the 2MMRC variant. Section 4.4 describes our simulation results and Section 4.5 summarizes the
chapter.

19

4.1

Problem Definition

The fundamental problem we address in this chapter is maximizing the amount of data
that can be gathered by the base station of hybrid WSNs composed of mostly static nodes
and a small number of mobile nodes. A key challenge of utilizing low cost mobile relay nodes
is that they often are more unreliable than static nodes. For example, mobile relay nodes
often experience mechanical failures when moving in rough terrain. To provide a balance
between maximizing data gathering capacity and mitigating the impact of mobile sensor
nodes on network reliability, we make the following assumptions. We first assume that a
routing topology connecting the static nodes to the sink has already been established. We
also assume that each mobile relay can help at most one link. These assumptions allow
mobile relays to improve the system’s data gathering capacity but minimize the dependence
of the communication on mobile nodes.
We define the data gathering capacity of a network to be the total amount of data
that can be transmitted from the sources to the sink during the lifetime of the current
routing topology. Consistent with several other papers [49, 9, 43] the lifetime of the current
topology ends when the first node dies. We use this definition because the death of any node
entails a reconfiguration of the underlying routing topology of static nodes and a consequent
reconfiguration of the mobile nodes. However, we do assume that data that is in transit
from a source to the sink and still has a viable path to the sink will be sent along the
remaining viable links to the sink. For example, suppose a node u expends all its energy to
transmit data to an intermediate node v. Since u has expended all its energy, the system
lifetime has ended. However, we assume that node v and other nodes on the path to the

20

sink can still send this data to the sink assuming they have enough energy. We use this
data gathering capacity metric as an alternative to two other important but conflicting
metrics, lifetime and throughput, since it directly measures the network’s ability to perform
its target functionality of transmitting data from the sources to the sink. In one variant
with a star routing topology where each source node is directly connected to the sink node,
we simply measure the raw data gathering capacity without regard to system lifetime as
no reconfiguration will be performed. That is, we simply measure how much data can be
transmitted from the sources to the sink until the single sink node dies or all the source
nodes die.
We define the k-MMRC problem as follows. Given a network N consisting of a large
number of static nodes and a small number of mobile nodes, and the routing topology
connecting the static nodes to the sink, the goal is to find the optimal positions of the
mobile nodes to join the transmission routes such that the total data gathering capacity of
N is maximized and at most k relay nodes join each transmission link. In this definition, we
limit the number of mobile relays that can help a single link to the parameter k for two main
reasons. First, since the number of mobile nodes in the network is relatively small, we want
to allow a fair number of links to be helped. Second, as mobile nodes may be unreliable,
we want to reduce the dependence of a single link on a large number of mobile nodes. We
consider two variants of the problem, with k = 1 and 2. Our results show that it is enough
to solve the 1-MMRC variant as it results in comparable improvements in data gathering
capacity to the 2-MMRC variant.
Different from previous work that only accounts for the energy required for transmission,
we take into account the energy required to transmit and receive messages plus the energy
21

required to move the mobile relay nodes into position. We use the energy models defined
in chapter 3. These model parameters fit the quadratic path loss model for distances up
to 35m. For larger transmission distances, the energy consumption increased and diverged
from the quadratic model and could be modeled by higher path loss coefficients (4 and 6) as
discussed in [61]. To cope with higher path loss coefficients, the only modification we need
is a new optimal point location algorithm to replace our algorithm in Section 4.2.1.

4.1.1

Max-Data Mobile Relay Configuration

We consider four variants of our problem that cover three different topologies and several
different metrics.
(1) MMRC in a line: The static nodes form a line between the source and the sink. We
want to maximize the total amount of data that can be relayed from the source to the sink
before any node runs out of energy. This is equivalent to maximizing the bottleneck capacity
of any transmission link on the line.
(2) MMRC in a star: The star transmission topology consists of n static sources and
one static sink. Each of the n source nodes transmits data directly to the sink. We again
want to maximize the total amount of data that can be sent from all n sources to the sink.
The star topology is widely used in networks running clustering protocols and is the primary
topology supported by the 802.15.4 standard [22]. For the star topology, we measure the
total amount of data gathered until either the sink node dies or all the source nodes die since
we will not reconfigure the topology when any source node dies.
(3) Fair MMRC in a tree with data reduction: The static nodes form a tree with n source
nodes sending data through the tree to a single sink. When an event occurs, each source

22

node sends one data unit to its parent. To ensure fairness, we require that each source’s data
must reach the sink. We assume perfect data aggregation where any internal node is able
reduce the amount of data it sends to its parent to one data unit per event. Many natural
sensor network applications use such an N-to-1 aggregation model. For instance, a user may
periodically query the maximum or average value of temperature readings from all sensors
deployed in a large biological habitat. Moreover, many in-network information processing
techniques such as data fusion [57] can aggregate noisy measurements or local decisions
from multiple sensors to improve the accuracy of sensing. In both applications described
above, N sensor readings can be aggregated into one, which significantly reduces the energy
consumption of the network. The net result is that we must maximize the bottleneck capacity
of any link in the tree.

Figure 4.1: Two static tree topologies; each link is labeled with its capacity.

Consider the two examples in Figure 4.1. In the left tree, the data for 3 events can be
sent to the sink before source c2 runs out of energy; in the right tree, the data for 5 events
can be sent to the sink before node c3 runs out of energy. In both cases, the number of
events reachable to the sink is limited by the edge with the smallest capacity.
(4) Fair MMRC in a tree without data reduction: this problem is identical to the previous
problem except no data reduction is possible. That is, if an internal node receives one data
23

unit from 3 children, it must send 3 data units to its parent. In both trees in Figure 4.1,
after 2 events are processed, c3 has only enough energy to send one data unit to the sink.
With our fairness constraint, no more events can be processed as a third event would require
sending 2 data units to the sink.

4.2

1-MMRC: Single Relay per Link

The MMRC problem of maximizing system data gathering capacity is much more complex
than the seemingly similar problem of minimizing the total energy required to transmit a
fixed amount of data in a network [14] because we must account for the energies consumed
by each individual node relative to its initial energy rather than worrying only about the
total energy consumed by all nodes together. To the best of our knowledge, we are the
first to optimize data gathering capacity taking into account the energy consumed by both
movement and communication. In this section, we consider the special case in which a
link may be helped by at most one relay. We first present an optimal algorithm for the
simplest possible “base case” of the 1-MMRC problem: a topology with only one source, one
sink, and one mobile relay. Next, we discuss the assignment problem, a weighted matching
problem that several of our application problems can be reduced to. We then present optimal
algorithms for our four variants of 1-MMRC.

4.2.1

Base Case of 1-MMRC

Our base case consists of computing the optimal position r∗ to move relay node r to
that will maximize the total number of bits that can be sent from static source node s1 to
static sink node s2 using r as a relay at position r∗ given initial energies e1 , e2 , and er .

24

Figure 4.2: Example scenario illustrating the base case. Using relay r at position r∗ increases
the amount of data that can be sent.
An example scenario is given in Figure 4.2. For cases where it may be better to not use r
because r has low energy, we assume that s1 sends data directly to s2 . Note that we focus
on maximizing the amount of data gathered by s2 during the lifetime of the current network
topology. Therefore, we do not allow s1 to send data to r until r dies and then have s1
send data directly to s2 because this requires a topology reconfiguration. In the rest of this
section, we study the problem of finding r∗ when it is helpful to use r.
Optimal Algorithm
We now show that a direct analytical solution to the base case requires finding the roots
of a degree 6 bivariate multi-modal polynomial. For any candidate point r , we define two
key values: m(s1 , r ) which is the amount that s1 can send to r located at r , and m(r , s2 )
which is the amount that r can send to s2 when r is located at r taking into account the
energy r loses by moving to r . It follows that the maximum amount of data s1 can send
to s2 using r as a relay at r , denoted m(r ), is the minimum of m(s1 , r ) and m(r , s2 ).
Optimal positions r∗ have one of the following properties:
(O1) m(s1 , r∗ ) = m(r∗ , s2 ) or
(O2) ∀r , m(s1 , r∗ ) ≥ m(s1 , r ) and m(s1 , r ) < m(r , s2 ) or
(O3) ∀r , m(r∗ , s2 ) ≥ m(r , s2 ) and m(s1 , r ) > m(r , s2 )

25

Let’s assume condition (O1) holds where r∗ is located at (x, y) and (x1 , y1 ), (xr , yr ), and
(x2 , y2 ) are the initial coordinates of nodes s1 , r and s2 , respectively. Let d(u, v) be the
distance between two nodes u and v. Finding point r∗ is equivalent to solving the following
equation.
m(s1 , r∗ ) = m(r∗ , s2 )
⇔

(4.1)

e1
er − kr d(r, r∗ )
=
a1 + b1 d(s1 , r∗ )2
ar + br d(r∗ , s2 )2

(4.2)

⇔ A(x, y)B(x, y) + C(x, y) = 0
⇒ B(x, y)2 A(x, y)2 − C(x, y)2 = 0

where

(4.3)
(4.4)

A(x, y) =kr a1 + b1 (x − x1 )2 + b1 (y − y1)2
B(x, y) = (x − xr )2 + (y − yr )2
C(x, y) =e1 ar + br (x − x2 )2 + br (y − y2 )2
− er a1 + b1 (x − x1 )2 + b1 (y − y1 )2 .

Because A(x, y) and B(x, y)2 have degree 2, equation (4.4) has degree 6; this equation is
also a nonconvex (multi-modal) polynomial. Thus, solving the base case using analytical
methods is numerically complex.
We present an efficient algorithm that converges on an optimal solution by solving the
special case of the problem in which we are given the direction that r will move. Let α
be any point on the line segment between s1 and s2 including s1 and s2 , let posr (α) be
the optimal position for r when it moves towards α, and let M (α) be the resulting amount
of data that can be transmitted from s1 to s2 when r moves to posr (α). We develop an
algorithm that computes posr (α) and M (α) for any α between s1 and s2 , and we prove
26

that M (α) is unimodal on this range. Thus, we can converge on the global optimum by
performing a golden ratio search for α between s1 and s2 .
We now discuss the algorithm for computing posr (α) and M (α). We first consider optimality conditions (O2) and (O3) by finding the unique points r1 and r2 on the line to α that
maximize m(s1 , r ) and m(r , s2 ), respectively. Point r1 is returned if m(r1 , s2 ) > m(s1 , r1 )
as condition (O2) applies. Otherwise, if m(r2 , s2 ) < m(s1 , r2 ), then we return r2 as condition
(O3) applies. Otherwise, it must be the case that condition (O1) applies. When restricting
the movement to a given line (with slope sα ), the distance traveled becomes

1 + s2 |x − xr |
α

and equation (4.2) becomes

er − kr 1 + s2 |x − xr |
e1
α
=
2 + b (y − y1)2
2 + b (y − y2)2
a1 + b1 (x − x1 )
ar + br (x − x2 )
r
1

Identifying such points is equivalent to solving a degree 3 polynomial. Thus, there are at
most 3 possible points, all of which can be computed in O(1) time. We return the point
that maximizes the data sent. In summary, we find the optimal position along a given line
by examining at most five well-defined points.
This base case algorithm only converges on the optimal r∗ for energy models with path
loss coefficient 2 or smaller. For other energy models with higher path loss coefficients, we
need to use different techniques to find the optimal relocation point. We do not have efficient
algorithms that do this for arbitrary energy models. We now prove that M (α) is unimodal.

Theorem 1. M (α) is unimodal for α ∈ s1 s2 .
Proof. We define the function gr (α) to be the value of the maximum message that can be

27

Figure 4.3: The highlighted portion of
gr (α) is part of M (α)

Figure 4.4: The highlighted portion of
hb (α) is part of M (α)

Figure 4.5: The set of point R∪B form
a unimodal set

Figure 4.6: Points of f (α) connect
pieces of R and B and form a unimodal
continuous function

Figure 4.7: The function M is unimodal. For interpretation of the references to color in this
and all other figures, the reader is referred to the electronic version of this dissertation.

28

transmitted by r when it moves to posr (α) and gb (α) the largest value that s1 can transmit
to posr (α). Similarly, we define posb (α) to be the point along rsα that maximizes the value
that s1 can transmit, we denote this value by the function hb (α). We also associate with
posb (α) the function hr (α), denoting the maximum value that r can transmit to s2 when
it moves to posb (α). Lastly, we define the function f (α) to be the value at the point on
segment rsα such that both s1 and r can transmit the same amount of data to/from (when
there are more than one point with this property, then f (α) denotes the highest value).
We observe that

M (α) = max min{gr (α), gb (α)},
(4.5)
min{hr (α), hb (α)}, f (α)

We base the proof that M (α) is unimodal on the following claims (which we prove in appendix
A):
Claim 1. gr (α) is unimodal
Claim 2. hb (α) is decreasing
Claim 3. f (α) is unimodal
Claim 4. gb (α) is continuous
Claim 5. hr (α) is continuous
Since, gr (α) maximizes the message that r can transmit, then

f (α) ≤ gr (α)
(4.6)
hr (α) ≤ gr (α)

29

Similarly, from the perspective of s1

f (α) ≤ hb (α)
(4.7)
gb (α) ≤ hb (α)
We consider two cases:
1. gr (α) < gb (α)
2. hb (α) < hr (α)
When the first condition holds, (4.6) and (4.7) imply

hr (α) ≤ gr (α) ≤ gb (α) ≤ hb (α)
M = max min{gr (α), gb (α)}, min{hr (α), hb (α)}, f (α)
= max gr (α), hr (α), f (α)
= gr (α)

So in this case, M (α) = gr (α) as shown in Figure 4.3. Similarly, when the second condition
holds, M (α) = hb (α) as shown in Figure 4.4.
Since gr (α) is unimodal and hb (α) is decreasing, the function min{gr (α), hb (α) is unimodal. Let R be the set points of gr (α) ∩ M (α) and B the set of points of hb (α) ∩ M (α).
Since points of R satisfy the first condition and points of B satisfy the second condition,
then
R ∪ B ⊆ min{gr (α), hb (α)}
and hence, R ∪ B form a unimodal set (Figure 4.5).
30

The remaining points of M (α) are points satisfying M (α) = f (α). We know using claims
4 and 5 that the functions gb (α) and hr (α) are continuous. So the endpoint of each interval
in R and B is also a point on f (α). On the other hand, f (α) forms a unimodal set (by claim
3). Therefore, the function M (α) is continuous and unimodal as Figure 4.6 illustrates.
Load Balancing Heuristic
In this section, we propose a simple heuristic for solving the base case of 1-MMRC. This
heuristic is faster than the optimal algorithm proposed above and achieves comparable data
gathering capacities. This heuristic increases the data gathering capacity of the base case by
balancing the energy consumption loads between the transmitting node s1 and the mobile
relay r proportionally according to their available energy. The main idea is to consider only
points on the line (s1 s2 ) as target positions for r, find an approximate position r on (s1 s2 )
and then refine r into the final position r . We compute our solution as follows. First, we
set r as the midpoint of s1 s2 . We then compute the amount of energy em , consumed by r
when it moves from its original position to r . This amount corresponds to an estimate of
the amount of energy consumed when r moves to its final position r . We use e = er − em
as an estimate of the amount of energy available at r after it moves to r . Finally, we use
e1 and e to compute the position r that maximizes the data gathering capacity of s1 , r,
and s2 . Let L be the length of the segment (s1 s2 ) and d the distance between s1 and r ,
the optimal d satisfies:

e1
e
2 = a + b (L − d)2 ⇔
a1 + b 1 d
r
r
(e1 br − r b1)d2 − 2r1 br Ld + e1 (ar + br L2 ) − e a1 = 0

31

(4.8)

We solve the quadratic equation (4.8) to obtain the optimal value d and then compute r as
d
r = s1 + (s2 − s1 ).
L

4.2.2

The Assignment Problem

Because of our assumption that each transmission link can be helped by at most one relay
node, our objective is to find a matching between mobile relay nodes and transmission links
in the underlying transmission topology. Thus, all of our centralized algorithms employ a
matching or assignment algorithm as a key step. It is possible that some mobile relay nodes
may not be used in the matching, but this is unlikely in practice because the number of relay
nodes will typically be much smaller than the number of static nodes. We also must find the
optimal position of each mobile relay node.
With our optimal base case algorithm, we now know where each relay node r should move
if we know which transmission link r should help; that is, we need to match each mobile relay
with a transmission link such that the application objective is optimized. This is essentially
the assignment problem from combinatorial optimization where we typically have n tasks
and n people, each task needs to be assigned to a single person, and the cost or efficiency
of the i-th person performing the j-th task is given by cij . The goal is to assign people to
tasks in order to maximize efficiency or minimize costs. The assignment problem is often
formalized as a 0-1 integer program where the 0-1 variable xij takes on value 1 if person
i is assigned to task j and 0 otherwise. Fortunately, we can relax this integer program to
a linear program where xij can take on fractional values because the constraint matrix of
the linear program is totally unimodular which means that an integral optimal solution is
guaranteed. The linear programming formulation is the following:

32




















n

n

max

cij xij
i=1 j=1

n

xij = 1
j=1
 n





xij = 1




 i=1





 x ∈ {0, 1},
ij

(i = 1, 2, . . . , n)
(j = 1, 2, . . . , n)
(i = 1, , . . . , n,

j = 1, . . . , n)

The first constraint guarantees that each person is assigned to one task, and the second constraint guarantees that each task is assigned to one person. The objective function attempts to maximize the sum of task efficiencies and can be solved in O(n3 ) time
[8]. We use this formulation for 1-MMRC in a star. We can create the maximum bottleneck assignment problem by changing the objective function from max
max min1≤j≤n

n c x .
i=1 ij ij

n
i=1

n c x
j=1 ij ij

to

This formulation can be solved in O(n2.5 ) time [8]; we use it for

1-MMRC in a line and for 1-MMRC in a tree with data reduction.

4.2.3

1-MMRC in a Line, Star, and Tree with Data Reduction

Consider an arbitrary 1-MMRC in a line instance. Let r be the number of mobile relay
nodes, k be the number of static nodes on the transmission path, and n = r + k be the total
number of nodes. Let P = (s1 , . . . , sk ) be the path of static nodes where s1 is the source
node and sk is the sink node.
The basic idea in transforming this 1-MMRC instance to a maximum bottleneck assignment problem instance is the following. The relay nodes take the place of people. The
transmission links lj = (sj , sj+1 ) take the role of jobs. To allow for the case where no relay
nodes are used, we add k −1 trivial “people” dj to represent the case when node sj transmits
directly to sj+1 and r trivial jobs outi to represent the case where relay node ri is not used.

33

For actual relays and transmission links, we use the base case algorithm in Section 4.2.1 to
compute cij , the size of the largest data chunk that can be transmitted from sj to ri to
sj+1 . We can alternatively use the approximation of Section 4.2.1 to obtain a faster close to
optimal value. If relay node ri cannot increase the size of data that sj can transmit, then we
set cij to 0. Let cj be the size of data if the transmission is done directly from sj to sj+1 .
The efficiency matrix values would then be:
r1

...

rr

l1
l2

c11
c12
...
lk−1 c1k−1

. . . cr1
. . . cr2
...
. . . crk−1

out1 ∞
...
outr ∞

...
...
...

∞
∞

d1 . . .

dk−1

c1 . . . 0
0 ... 0
...
0 . . . ck−1
∞ ... ∞
...
∞ ... ∞

More formally,
A(lj , ri ) = cij

A(outj , ri ) = ∞

A(lj , dj ) = cj

A(outj , di ) = ∞

A(lj , di ) = 0 if i = j
We use the exact same formulation for 1-MMRC in a Tree with Data Reduction. For
1-MMRC in a Star, we change the objective function to max

4.2.4

n
i=1

n c x .
j=1 ij ij

1-MMRC in a Tree without Data Reduction

The assignment based solution we described in Section 4.2.2 can not be applied to this
1-MMRC variant since the LP formulation does not capture the correct bottleneck edge
which is no longer the edge with the smallest capacity; adding more constraints to solve
this issue breaks the unimodularity property of the constraint matrix and subsequently the

34

integer program can not be solved using linear programming techniques. Instead, we solve
a decision version of this problem using unweighted bipartite matching algorithms. For the
decision version, we add an input M and the yes/no question is whether or not there exists
an assignment of mobile relay nodes to transmission links such that every source node is able
to send M events to the sink. We then converge on an optimal solution by using standard
doubling and binary search techniques. We make an initial guess M and keep doubling this
guess until it is infeasible. We then do a binary search between the last feasible value and
the first infeasible value. If the initial M was not feasible, we binary search between 0 and
M.
To solve the decision version, we first compute for each node the amount of data that it
needs to send so that every M data units initiating at a source and passing through it can
reach the sink by doing a bottom-up traversal of the tree. For each source, we set its value
to M . For each internal node, we set its value to be the sum of the values of its children. Let
Mi be the value that node si needs to transmit. Second, we eliminate any node that can send
Mi bits without help. Third, we create a bipartite graph Gb as follows. For each remaining
static node si , we create an edge between si and relay node rj if rj can help si reach its
Mi value. Finally, we search for a bipartite matching in Gb . If there is a bipartite matching
such that all nodes in Gb are matched with some relay node, we return yes. Otherwise, we
return no. Bipartite matchings can be found in O(n2.5 ) time [21].

4.2.5

Distributed Algorithm

Before describing our distributed implementation, we propose three centralized greedy
algorithms that provide insight to the distributed approach; these greedy strategies differ

35

procedure staticRun
Initialize empty priority queue
offers ← ∅;
committed ← false;
repeat
Listen to mobile nodes in range and collect offers presented
repeat
receive(mobile, type, data);
if type = MOBILE IN RANGE then
send(mobile, META DATA, info);
else if type = OFFER then
offers.add(data);
end if
until timeout
Process offers and pick best
if offers = ∅ then
bestOffer ← offers.dequeue();
send(bestOffer.sender, ACCEPT OFFER);
committed ← true;
end if
while offers = ∅ do
offer ← candidates.dequeue();
send(offer.sender, REJECT OFFER);
end while
until committed
end procedure

Figure 4.8: Local algorithm executed by each static node
only in their link priority strategies which determine which link to process first. The three
algorithms proceed as follows. First, for each link l between two static nodes si and sj , we
compute the maximum amount of data c(l) that si can transmit to sj . Second, for each
mobile node r and for each link l, we compute the total amount of data c(r, l) that can
be transmitted from si to r to sj using the optimal algorithm for the base case (section
4.2.1). If si can send more data directly to sj than going through r, we set c(r, l) to zero.
We now choose the link with the highest priority using three different priority schemes:
(1) the highest value c(r, l), (2) the largest improvement c(r, l) − c(l) and (3) the smallest
(bottleneck) direct value c(l). In case of a tie for the third metric, we pick the link with the
36

procedure mobileRun
Initialize empty priority queue
candidates ← ∅;
committed ← false;
repeat
broadcast(MOBILE IN RANGE);
Collect responses from static nodes in range
repeat
receive(sender, META DATA, sender.info);
Decision data include: data that can be sent directly,
through mobile node, position of mobile node, priority
sender.decisionData ← computeAmounts(sender.info)
candidates.add(sender);
until timeout
Process candidates queue
while candidates = ∅ do
candidate ← candidates.dequeue();
send(candidate, OFFER, data(candidate, self));
response ← receive(candidate, responseType);
if responseType = ACCEPT OFFER then
committed ← true;
break;
end if
end while
until committed
moveTo(candidate.optimalPosition);
end procedure

Figure 4.9: Local algorithm executed by each mobile node
highest c(r, l). In all three cases, we match the selected link to the corresponding mobile
relay r and then repeat the greedy selection process until all the mobile nodes are matched
to links or cannot help the remaining links.
We now propose a distributed approach that mimics our greedy algorithms. For MMRC
in a line or tree, we use the bottleneck link priority scheme. For MMRC in a star, we
use the largest improvement link priority scheme. The algorithm begins with each mobile
relay node broadcasting its existence (a MOBILE IN RANGE message) to nodes within its
neighborhood. We define the neighborhood of a mobile node r as the set of nodes that can
37

be reached within h hops from r for a given parameter h. Each static node that receives
such a message responds by sending its location, its remaining energy, and the location of its
destination static node (a META DATA message) to each mobile relay node that sends it a
broadcast message. Each mobile relay node r then calculates for each responding static node
s how much that s can transmit now and how much r can improve s’s transmission capacity.
It then sends a message (of type OFFER) to the static relay node with the highest priority
saying how much it can help it and waits for a response from that static node. Each static
node that receives an offer accepts the best offer it receives, sending an acceptance message
to the one relay node it accepts and rejection messages to the other relay nodes. Rejected
relay nodes then send messages to the next node on their priority list. If a static node that
is already committed receives a late offer from a relay node, it automatically rejects the
offer even if that relay node can provide more help. We show the pseudocode for the local
algorithm performed by the static and mobile nodes in Figures 4.8 and 4.9 respectively.

4.3

2-MMRC: Two Relays per Link

The previous algorithms are optimal or close to optimal under the constraint that a single
mobile relay node can help at most a single link. In some cases, the network contains a large
number of mobile relays, or long transmission links. These cases may benefit from further
decomposing the transmission links into multiple smaller links, to exploit the availability of
relay nodes and distribute the energy consumption over several nodes. In this section, we
relax the above constraint and study the benefit of allowing two relays to help a single static
link. We first consider the modified base case when two relay nodes are available in the
proximity of a link. Then we extend our distributed algorithm of Section 4.2.5 to take into

38

account two relays joining a link.

Figure 4.10: Using r1 and r2 at r1 and r2 increases the amount of data that can be transmitted from s1 to s2 .

4.3.1

Base Case

We consider the base case of 2-MMRC as shown in Figure 4.10. Similar to the base case
of Section 4.2.1, we would like to find the optimal positions of r1 and r2 such that the amount
of data that can be sent from s1 to s2 through r1 and r2 is maximized. A direct analytical
solution is equivalent to solving the following nonlinear program which is a complex and
challenging task.

Maximize

xr1 ,yr1 ,xr2 ,yr2

m

subject to
e1
a1 + b1 (p1 − xr1 )2 + (q1 − yr1 )2

≥m

er1 − kr1 (xr1 − pr1 )2 + (yr1 − qr1 )2
ar1 + br1 (xr1 − xr2 )2 + (yr1 − yr2 )2

≥m

er2 − kr2 (xr2 − pr2 )2 + (yr2 − qr2 )2
≥m
ar2 + br2 (xr2 − p2 )2 + (yr2 − q2 )2

39

We observe that in optimal solutions, the final locations for r1 and r2 are close to the line
connecting s1 and s2 even when the mobility parameter k is high. This is because when two
mobile relays join the transmission line, the corresponding transmission distances are much
smaller than the original direct transmission distance between s1 and s2 . Since the energy
consumption due to transmissions is quadratic in the transmission distance, the savings in
transmission energy are high and make up for a relatively long movement for the mobile
nodes.
Consistent with this observation, we extend the load balancing heuristic of 1-MMRC
to solve the base case of 2-MMRC. Again, our new load balancing heuristic only considers
the line (s1 s2 ) for target positions of r1 and r2 . This constraint reduces the complexity
of the computation without noticeable loss in the performance as we show in Section 5.6.
The target positions for r1 and r2 are computed as follows. First, we estimate the amount
of energy that r1 and r2 consume to get to their target positions on s1 s2 by picking two
estimate destination points on that line, r1 and r2 respectively (shown in Figure 4.10). These
points are computed as the evenly spaced points along (s1 s2 ) as follows:

2
1
r1 = s1 + s2
3
3

1
2
r2 = s1 + s2
3
3

Then, we use these positions to estimate the remaining energies er1 and er2 at r1 and r2
when they relocate to r1 and r2 . Finally, we compute the final destination positions r1 and
r2 as the points at which nodes s1 , r1 and r2 consume their remaining energy at the same
rate. We solve for the values d1 (distance between s1 and r1 ) and d2 (distance between r1

40

and r2 ) by finding the values satisfying the following equation:
er1
er2
e1
2 = a + b d2 = a + b (L − d − d )2
a1 + b 1 d 1
r2
r2
1
2
r1
r1 2

(4.9)

where L is the length of the segment s1 s2 . Equation (4.9) is equivalent to

d2 =
2
(L − d1 − d2 )2 =

er1 a1 − e1 ar1
e1 br1
er2 a1 − e1 ar2
e1 br2

+

er1 b1 2
d
e1 br1 1

(4.10)

+

er2 b1 2
d
e1 br2 1

(4.11)

(4.10) and (4.11) imply that d1 satisfies

((β2 − β1 − 1)2 − 4β1 )d4 + (4Lβ1 − 4L + 4Lβ2 )d3
1
1
+ (−4L2 β1 − 4α1 + 2(α2 − α1 − L2 )(β2 − β1 − 1) + 4L2 )d2
1

(4.12)

+ (4Lα2 + 4Lα1 − 4L3 )d1
+ (α2 − α1 − L2 )2 − 4L2 α1 = 0
which is a quartic polynomial whose roots can be found efficiently using Ferrari’s method
[54] for quartic equations. Once d1 is computed, d2 and consequently r1 and r2 can be found
in a straightforward manner.

4.3.2

Distributed Algorithm

The transformation into an assignment problem described in Section 4.2.3 does not apply
in this case because it allows for a mobile relay node to be used multiple times, each time
coupled with a different relay. Adding a constraint to the linear program to ensure this does

41

not happen breaks the unimodularity property of the linear program and consequently the
guarantee of an integral solution.
Based on the distributed implementation of Section 4.2.5, we propose a greedy algorithm
to match mobile relays to links as follows. The main idea is to consider the critical links
first. Each such link considers the surrounding mobile nodes that did not join the tree, both
individually and in pairs. When r is considered by itself by a link si sj , the optimal position
for the single relay base case (Section 4.2.1) is used to compute the corresponding data
gathering capacity of the subpath si , r, sj . Likewise, when two mobile relays r1 and r2
are considered, the solution described in Section 4.3.1 is used to compute the data gathering
capacity of si , r1 , r2 , sj . Node si then picks the relay node or relay pair that results in
the highest priority. We note that we allow two relay nodes to join a single link only if the
expected priority is greater than the highest priority of a single relay for that link by a factor
f . This is to ensure the improvement is significant when two relay nodes are used for one
link. This process is repeated until all the mobile relays are assigned to some link or the
available mobile relays cannot benefit the links by joining the transmission lines.

4.4
4.4.1

Simulation Results
Setup

We evaluated our algorithms using simulations. For a given simulated network, we varied
the number of mobile nodes by 5 between 5 and 30. For the star and tree topologies, we
varied the number of source nodes between 5 and 30 by 5. For each number of source
nodes and number of mobile relay nodes, we generated 20 random networks (120 for the

42

line, 720 for the tree and star) consisting of 100 nodes in a 150m by 150m area. To ensure
high quality transmissions based on empirical data using Tmotes [45], we set the maximum
communication distance to 35m. We set the initial battery energy of nodes uniformly at
random between half charged and fully charged. We modeled different platforms and different
settings by having 16 possible choices for node energy consumption parameters (4 values for
a and b and 4 values for k). To build the communication trees between the sources and
sink, we used greedy geographic forwarding in which each node forwards the packet to the
neighbor closest to the sink. We note that our solutions only require the establishment of a
routing topology before mobile relay configuration, and hence can work with other routing
algorithms as well.
For all networks that we generated, we evaluated our optimal algorithm as well as seven
different greedy strategies. Six of our greedy strategies differed in two main ways: point
selection and link priority. We consider two different point selection strategies. The first is
to move the relay node to the midpoint of the two static nodes defining the link that will be
helped; we label such greedy strategies with the suffix -MP. The other option is to compute
the optimal position using our optimal base case algorithm; these strategies are labeled with
the suffix -OPT. We consider the three link priority schemes we described in the distributed
implementation section: the highest value, the largest improvement, and the bottleneck value
with ties broken by largest improvement. The net result is six greedy strategies. We also
consider a seventh greedy algorithm that uses matching to minimize the total travel distance
of all mobile relay nodes. Finally, we evaluated our distributed implementations, one for each
of the three link priority strategies. To model the asynchronicity of our protocol, we vary the
fraction of mobile relay nodes that participate in the protocol at various percentages between
43

25% and 100%. We also vary the neighborhood parameter h between 1, 2, 3 and unlimited.
We did not implement existing algorithms because their fundamental assumptions do not
match ours making meaningful comparisons difficult if not impossible.
We also generated a new set of topologies, consisting of 100 base cases to assess our
optimal algorithm and load balancing heuristics for solving the base case. Each network in
this set consists of a single static link and two mobile relays. When assessing the base case
for 1-MMRC, only one of the existing mobile relays (the more beneficial one) is allowed to
join the link whereas in 2-MMRC solutions, both relay nodes are expected to join.
For a network N and an algorithm A, let A(N ) be the data gathering capacity of the
resulting network N computed by algorithm A. Let B denote the baseline algorithm that
uses no mobile relay nodes; thus, the resulting network N is the original network of static
nodes. We measure the performance of algorithm A on network N by its improvement ratio
which is A(N )/B(N ). When reporting results for an algorithm A, we report the average of
A(N )/B(N ) for all N that have the same number of mobile relay nodes. For line topologies,
this will be an average of 20 networks; for tree and star topologies, this will be an average of
120 networks (20 for each of 6 values of the number of source nodes). We also report results
for where we average over all networks for the given application. We now present a summary
of our simulation results for each application.

4.4.2

1-MMRC in a Line

We highlight some of our key results for MMRC in a Line in Figure 4.11. The average
improvement ratio for our optimal algorithm over all 120 networks is 232%, and the improvement ratio increases from 180% to 285% as we increase the number of mobile relay nodes.

44

Improvement in Data Gathering
Capacity

Optimal
Bottleneck-OPT
Bottleneck-MP

Distributed (h=1)
Distributed (h=3)

300%

250%

200%

150%

100%
5

10

15

20

25

30

Number of Mobile Nodes

Figure 4.11: Improvement in data gathering capacity of our algorithms in line topologies
The best greedy strategy, Bottleneck-OPT, also has an average improvement ratio of 232%.
As we can see from Figure 4.11, its improvement ratio is very close to that of the optimal
algorithm for all numbers of mobile relay nodes. This implies that we can use a simple greedy
strategy to perform the matching rather than using more sophisticated weighted matching
algorithms. The remaining greedy strategies perform signficantly worse with average improvement ratios ranging from 165% to 210%. We also observe the importance of optimal
point selection; the Bottleneck-MP greedy strategy achieves an average improvement ratio
of 175% which is much worse than that of Bottleneck-OPT. In fact, for each link priority
scheme, the midpoint point selection strategy has a significantly lower improvement ratio
than its optimal point selection counterpart.
Our distributed implementation of the Bottleneck-OPT greedy strategy has an average
improvement ratio of 211% when the neighborhood parameter h = 1. This ratio is a bit lower
than both the optimal algorithm and Bottleneck-OPT. There are two possible explanations
45

for the slight decrease in performance. The first is that a relay node may not be matched
to the correct link because it cannot reach the source node of the link with its transmission
(e.g. it is farther than 35m away from the source node of the link); the second is that a relay
node may not be matched to the correct link because its offer of help arrives after the source
node of the link has already accepted the help of another relay node. Our results indicate
that the 35m transmission distance limitation is the more important factor, particularly
as the number of relay nodes increases. When h increases to 3, the improvement ratio
increases by 12% to an average of 223%. We also observe that it is enough to set h to
3 to get the maximum improvement for the distributed approach. Moreover, we see little
difference between the distributed implementation with 33% of the relay nodes participating
and 100% of the relay nodes participating for any number of relay nodes; this also holds
for all values between 33% and 100%. We also see that the distributed implementation’s
average improvement ratio approaches that of the optimal algorithm as the number of relay
nodes increases. In this case, each transmission link is likely to receive an offer from a good
if not optimal relay node.

4.4.3

1-MMRC in a Star

We highlight some of our key results for 1-MMRC in a Star in Figure 4.12. The average improvement ratio for our optimal algorithm over all 720 networks is 155%, and the
improvement ratio increases from 130% to 170% as we increase the number of mobile relay nodes. The best greedy strategy, Improvement-OPT, has an average improvement ratio
of 152% which is nearly identical to the optimal algorithm for all numbers of relay nodes.
The remaining greedy strategies perform significantly worse with improvement ratios ap-

46

Optimal
Improvement-OPT
Improvement-MP

Distributed (h=1)
Distributed (h=2)

Improvement in Data
Gathering Capacity

170%
160%
150%
140%
130%
120%
110%
100%
5

10

15

20

25

30

Number of Mobile Nodes

Figure 4.12: Improvement in data gathering capacity of our algorithms in star topologies
proximately 20% lower. We again observe the importance of optimal point selection; all
the greedy strategies that use optimal point selection have improvement ratios roughly 17%
better than their midpoint point selection counterparts.
Our distributed implementation of the Improvement-OPT greedy strategy has an improvement ratio of 133% for h = 1 and 152% for h = 2. We note that setting h to 2 allows
each mobile node to consider all the links in the network and consequently the improvement
ratios are very close to the optimal ratios. We again observe that the main cause of the
performance loss is the 35m transmission distance limit as our distributed algorithms perform essentially identically once at least 33% of the relay nodes participate in the protocol.
We again see the distributed implementation’s performance increase significantly with the
number of mobile relay nodes.

47

Optimal
Bottleneck-OPT
Bottleneck-MP

Distributed (h=1)
Distributed (h=2)
Distributed (h=3)

Improvement in Data
Gathering Capacity

300%

250%

200%

150%

100%
5

10

15

20

25

Number of Mobile Nodes

30

Figure 4.13: Improvement in data gathering capacity of our algorithms in tree topologies
with data reduction at the nodes

4.4.4

Fair 1-MMRC in Trees with Data Reduction

We highlight some of our key results for 1-MMRC in trees with data reduction in Figure
4.13. The average improvement ratio for our optimal algorithm over all 720 networks is 270%,
and the improvement ratio increases from 215% to 300% as the number of mobile relays
increases. The best greedy strategy, Bottleneck-OPT, has an average improvement ratio of
268% which is nearly identical to the optimal algorithm in all cases. The remaining greedy
strategies perform significantly worse with average improvement ratios of at most 240%.
We again observe the importance of optimal point selection; the greedy strategies that use
optimal point selection have improvement ratios roughly 65% higher than their midpoint
point selection counterparts. For Bottleneck-OPT and Bottleneck-MP, the divergence in
performance increases as the number of mobile relays increases.
We now evaluate the performance of our distributed implementation for different values

48

Improvement in Data Gathering
Capacity

Optimal
Improvement-OPT
Improvement-MP

Distributed (h=1)
Distributed (h=2)

300%

250%

200%

150%

100%
5

10

15

20

25

30

Number of Mobile Nodes

Figure 4.14: Improvement in data gathering capacity of our algorithms in tree topologies
without data reduction at the nodes
of h. We obtain improvement ratios of 226%, 242% and 252% for h = 1, 2 and 3 respectively.
We also obtain the same improvement ratios for all h ≥ 3. We again observe that the main
cause of the performance loss is the 35m transmission distance limit since the improvement
ratios increase as h increases. Similar to the previous results, our distributed algorithms
perform essentially identically once at least 33% of the relay nodes participate in the protocol.
We again see the distributed implementation’s performance approaches that of the optimal
algorihtm as the number of mobile nodes increases. We also study how many communication
rounds are required before all the mobile nodes commit to some edge or remove themselves
from consideration. In all cases, the number of rounds grows very slowly with the number of
mobile nodes. On average, only 4 to 6 rounds are needed. Even when the number of relays
was high (≈ 30), the number of rounds never exceeded 8. These results also hold for MMRC
in trees without data reduction.

49

4.4.5

Fair 1-MMRC in Trees without data reduction

We highlight some of our key results for 1-MMRC in trees without data reduction in
Figure 4.14. The average improvement ratio for our optimal algorithm over all 720 networks
is 233%, and the improvement ratio increases from 216% to 250% as the number of mobile
relay increases. The best greedy strategy, Bottleneck-OPT, has an average improvement ratio
of 220% and it approaches that of the optimal algorithm as we increase the number of mobile
relay nodes. Improvement-OPT performs almost as well as Bottleneck-OPT, particularly for
many mobile relay nodes. The remaining greedy strategies perform significantly worse with
improvement ratios of at most 175%. We again observe the importance of optimal point
selection; the greedy strategies that use optimal point selection have improvement ratios
roughly 40% higher than their midpoint point selection counterparts. For Bottleneck-OPT
and Bottleneck-MP, the divergence in performance increases as the number of mobile nodes
increases.
Our distributed implementation based on Bottleneck-OPT again performs well with average improvement ratios of 200% and 208% for h = 1 and 2 respectively. We note that it is
enough to set h to 2 to get similar improvement ratios as the unlimited neighborhood case.
When h ≥ 2, the improvement ratios obtained are within 5% from Bottleneck-OPT when
the number of mobile nodes exceeds 5. We also show that the average number of messages
generated per mobile relay node grows slowly and linearly with number of mobile nodes.
For example, when a static node receives offers from 33% of the mobile relay nodes in its
neighborhood, the average number of messages sent by each relay node is only 2.1. Even
with large number of mobiles nodes (30) and high communication ratio (100%), the number

50

of messages generated by each mobile node is only 4.
Finally, in our simulations for all four applications, the approach that matches mobile
relays based on minimizing the total distance traveled produces noticeably inferior results
to the optimal approach and, in most cases, the other greedy approaches. This shows that
optimizing only to minimize movement without regard to transmission is not helpful. In
particular, it is helpful to spend a little more energy on movement if this will significantly
decrease energy consumed by transmissions.

4.4.6

Two Relays Per Link

We now evaluate the benefit of allowing two relay nodes to join a link. We first compare
our base case algorithms for 1-MMRC and 2-MMRC. Then, we compare our distributed
algorithms for general topologies.
Base Case
To evaluate our algorithms for the base case, we ran the following simulations on the
set of base cases we generated. For each base case topology, we computed the optimal data
gathering capacity using our optimal algorithm for 1-MMRC and exhaustive search for 2MMRC and using our load balancing heuristics for both variants. The average improvement
ratios are shown in Table 4.1.
Table 4.1: Average Improvement Ratios for Base
Optimal
Im- Heuristic
Improvement Ratio provement Ratio
Single Relay 460.7%
460.5%
Double Relay 591.2%
585.6%

Case Problems
Heuristic Ratio
/ Optimal Ratio
99.3%
98.9%

We observe that in both 1-MMRC and 2-MMRC base cases, our approximation algo-

51

rithms result in solutions very close to optimal, within 99% for both cases. However, in
2-MMRC base cases, the improvement in data gathering capacity is larger, reaching an average of 591% compared to 460% in 1-MMRC base cases. So, when considering base case
topologies, 2-MMRC outperforms 1-MMRC.
Fair 2-MMRC in Trees with Data Reduction
We now evaluate the performance of 2-MMRC in general topologies. For this purpose,
we compare the total data gathering capacity achieved by 1-MMRC and 2-MMRC in tree
topologies with data reduction. For 1-MMRC, we set h to 3. For 2-MMRC, we varied h
between 1 and 3 and set the improvement threshold f to 1.25. We highlight the results in
Figure 4.15. We make the following observations. First, consistent with the previous results,
the performance increases as h increases from 1 to 3. We note that it is enough to set h to
3 to get very close performance to the unrestricted case where each mobile relay considers
the whole network. Second, for h = 3, 2-MMRC slightly outperforms 1-MMRC with the
difference being at most 4% which is much less than the divergence observed for base cases.
This is because when the number of mobile nodes in the network is limited, using at most
one relay node per links allows more links in the network to be helped and consequently
more bottlenecks to be improved. In 2-MMRC solutions, only a small number of relay nodes
are used in pairs (≤ 3 pairs), most links are helped by a single relay node. This is due to two
factors: the parameter f and the geographic distribution of mobile nodes. Aside from a few
cases, most relay nodes are scattered in the network and are not close enough to each other
to collectively increase the data gathering capacity of a link. We note that we get similar
results for topologies without data compression at the nodes. We conclude that it is more

52

1-MMRC (h=3)
2-MMRC (h=1)

2-MMRC (h=2)
2-MMRC (h=3)

Improvement in Data
Gathering Capacity

300%
275%
250%
225%
200%
175%
150%
125%
100%
5

10

15

20

25

30

Number of Moble Nodes
Figure 4.15: Improvement in data gathering capacity of the distributed implementations for
1-MMRC and 2-MMRC for trees with data reduction
beneficial and efficient to compute the data gathering capacity using the 1-MMRC variant
of MMRC.

4.5

Summary

In this chapter, we study a new problem, MMRC, maximizing the data gathering capacity
of hybrid wireless sensor networks consisting of both mobile and static nodes. Our solutions
consist of inserting the mobile nodes along the transmission lines to improve the capacity of
weak links. We presented optimal solutions to four variants of MMRC when a single relay
is allowed per link, and approximate solutions when two relay nodes are allowed per link.
Unlike most previous work that exploits controlled mobility, we consider both the energy
consumed during the transmission process as well as the energy consumed by mechanical
locomotion. In most variants, our optimal algorithm improved system lifetime by a factor of 2

53

or more. For the star topology, it increased system lifetime by a factor of 1.5. Our distributed
protocols were almost as effective at improving data gathering capacity, particularly as the
number of mobile relay nodes increases. Our simulations also showed that these protocols
quickly converge on a final solution with little messaging overhead. Moreover, we show using
simulations, that although using multiple relay nodes on a single link significantly improves
the data gathering capacity of that link, it is better to use each relay node by itself on a
single link to increase the number of links that can be helped and achieve a globally higher
data gather capacity.

54

CHAPTER 5
Total Energy Minimization Using
Mobile Nodes
In the previous scheme, the network contained a small number of mobile nodes relative
to the network size. In some scenarios, the number of mobile nodes is relatively large and
our previous solutions may not fully exploit this fact. In this chapter, we consider dataintensive networks that consist entirely of mobile nodes. We exploit the mobility of all
the nodes in such networks to efficiently use the available energy. We consider the total
energy used in the network as an indirect measure for the network lifetime. Specifically,
we integrate the energy consumption due to both mobility and wireless transmissions into
a holistic optimization framework and consider the problem of minimizing the total energy
consumption of all participating nodes. We propose an approach in which mobile nodes join
the transmission routes then all participating nodes collectively move closer to each other
while maintaining connectivity. We show that our algorithms significantly reduce the energy
consumption and in particular converge to the optimal configuration when the routing tree

55

is given.

5.1
5.1.1

Problem Definition
An Illustrative Example

Before formally defining the problem of total energy minimization, we first describe the
main idea of our approach using a simple example. Suppose we have three nodes s1 , s2 , s3
located at positions x1 , x2 , x3 , respectively (Fig. 5.1), such that s2 is a mobile relay node.
The objective is to minimize the total energy consumption due to both movement and
transmissions. Data storage node s1 needs to transmit a data chunk to sink s3 through relay
node s2 . One solution is to have s1 transmit the data from x1 to node s2 at position x2 and
node s2 relays it to sink s3 at position x3 ; that is, node s2 does not move. Another solution,
which takes advantage of s2 ’s mobility, is to move s2 to the midpoint of the segment x1 x3 ,
which is suggested in [17]. This will reduce the transmission energy by reducing the distances
separating the nodes. However, moving relay node s2 also consumes energy. We assume the
following parameters for the energy models: k = 2, a = 0.6 × 10−7 , b = 4 × 10−10 .

Figure 5.1: Reduction in energy consumption due to mobile relay. As the data chunk size
increases, the optimal position converges to the midpoint of s1 s3 .
In this example, for a given data chunk mi , the optimal solution is to move s2 to xi (a
2
56

position that we can compute precisely). This will minimize the total energy consumption
due to both transmission and mobility. For small messages, s2 moves very little if at all. As
the size of the data increases, relay node s2 moves closer to the midpoint. In this example,
it is beneficial to move when the message size exceeds 4 MB. We illustrate in Table 5.1
the energy savings achieved using our optimal approach and the other two approaches for
the relevant range of data sizes. For large enough data chunks (≈ 13 MB), one relay node
can reduce total energy consumption by 10% compared to the other two approaches. As
the data chunk size increases further, the energy savings decrease, and the optimal position
converges to the midpoint when the data size exceeds 43 MB. In general, the reduction in
energy consumption is higher when there are multiple mobile relay nodes.
Table 5.1: Energy consumption comparison
Data Size
(MB)
5.00
11.00
12.00
13.00
14.00
15.00
16.00
17.00
18.00
40.00

Costs at
Costs at
Costs at
Reduction
Original Pos. Midpoints Optimal Pos.
42.78
70.71
42.04
1.73%
94.12
101.93
88.39
6.09%
102.68
107.13
94.71
7.75%
111.23
112.33
100.87
9.32%
119.79
117.53
106.89
9.06%
128.35
122.74
112.80
8.09%
136.90
127.94
118.62
7.28%
145.46
133.14
124.37
6.58%
154.01
138.34
130.06
5.98%
342.26
252.77
247.58
2.05%

The above example illustrates two interesting results. The optimal position of a mobile
relay is not the midpoint between the source and sink when both mobility and transmissions
costs are taken into consideration. This is in contrast to the conclusion of several previous
studies [17, 62] which only account for transmission costs. Second, the optimal position of a
mobile relay depends on not only the network topology (e.g., the initial positions of nodes)
but also the amount of data to be transmitted. Moreover, as the data chunk size increases,

57

the optimal position converges to the midpoint of s1 and s3 . These results are particularly
important for minimizing the energy cost of data-intensive WSNs as the traffic load of such
networks varies significantly with the sampling rates of nodes and network density.

5.1.2

Problem Formulation

In our definitions, we assume that all movements are completed before any transmissions
begin. We also assume there are no obstacles that affect mobility or transmissions. In this
case, as we show in Section 5.2.2, the distance moved by a mobile relay is no more than the
distance between its starting position and its corresponding position in the evenly spaced
configuration which often leads to a short delay in mobile relay relocation. We focus on the
case where all nodes are in a 2-dimensional plane

2,

but the results apply to

3

and other

metric spaces.
Our problem can be described as follows. Given a network containing one or more static
source nodes that store data gathered by other nodes, a number of mobile relay nodes and
a static sink, we want to find a directed routing tree from the sources to the sink as well as
the optimal positions of the mobile nodes in the tree in order to minimize the total energy
consumed by transmitting data from the source(s) to the sink and the energy consumed by
relocating the mobile relays. The source nodes in our problem formulation serve as storage
points which cache the data gathered by other nodes and periodically transmit to the sink, in
response to user queries. Such a network architecture is consistent with the design of storagecentric sensor networks [42]. Our problem formulation also considers the initial positions of
nodes and the amount of data that needs to be transmitted from each storage node to the
sink. The formal definition of the problem is given below.

58

Definition 1. (Optimal Mobile Relay Configuration):
Input Instance: S, a list of n nodes (s1 , . . . , sn ) in the network; O, a list of n locations
(o1 , . . . , on ) where oi is the initial position of node si for 1 ≤ i ≤ n; Ssources , a subset of
S representing the source nodes; r, a node in S, representing the single sink; Msources =
{Mi | si ∈ Ssources }, a set of data chunk sizes for all sources in Ssources ;
We define mi , which we compute later, to be the weight of node si which is equal to the
total number of bits to be transmitted by node si . We define a configuration E, U as a
pair of two sets: E, a set of directed arcs (si , sj ) that represent the directed tree in which all
sources are leaves and the sink is the root and U , a list of locations (u1 , . . . , un ) where ui is
the transmission position for node si for 1 ≤ i ≤ n. The cost of a configuration E, U is
given by:
ami + b ui − uj 2 mi + k oi − ui

c( E, U ) =
(si ,sj )∈E

Output: E, U , an optimal configuration that minimizes the cost c( E, U ).

5.2
5.2.1

Centralized Solution
Energy Optimization Framework

The Optimal Mobile Relay Configuration (OMRC) problem is challenging because of the
dependence of the solution on multiple factors such as the routing tree topology and the
amount of data transferred through each link. For example, when transferring little data,
the optimal configuration is to use only some relay nodes at their original positions. As
the amount of data transferred increases, three changes occur: the topology may change by
adding new relay nodes, the topology may change by changing which edges are used, and

59

Figure 5.2: Optimal routing tree for
0 ≤ m ≤ 15MB.

Figure 5.3: Optimal configuration for
15MB≤ m ≤ 25MB: same topology
but nodes relocate.

Figure 5.4: Optimal configuration for
25MB≤ m ≤ 60 MB. A new topology
is used.

Figure 5.5: Optimal configuration for
100MB≤ m ≤ 150MB: a new topology
with more nodes.

Figure 5.6: Example of optimal configurations as a function of amount of data to be transferred. In each part, source nodes s1 and s2 must send m bits of data to the sink r. We
consider m up to 150MB.

60

the relay nodes may move closer together. In many cases, we may have restrictions such as
no mobility for certain relay nodes or we must use a fixed routing tree. These constraints
affect the optimal configuration.
We illustrate how the optimal configuration depends on the amount of data to transfer
using the example from Fig. 5.2. When there is very little data to transfer, the optimal
routing tree Ta depicted in Fig. 5.2 uses only some of the relay nodes in their original
positions. When the amount of data to transfer from s1 and s2 increases to 15 MB, the relay
nodes in tree Ta move to their corresponding positions in tree Tb of Fig. 5.3 but the topology
does not change. When the amount of data to transfer from s1 and s2 is between 25 and 60
MB, the optimal routing tree has a different topology as shown in Fig. 5.4. For even larger
messages, new trees with even more nodes included are optimal. For example, when the
amount of data to be transferred is between 100 and 150 MB, the optimal tree is depicted
in Fig. 5.5. No existing tree construction strategy handles all these cases. For example, the
minimum spanning tree that includes all network nodes has two fundamental problems. It
will typically include unneeded nodes, and it typically creates non-optimal topologies as it
focuses only on the current location of nodes as opposed to where nodes may move to.
We now present a centralized approach to solve OMRC that breaks the problem into three
distinct steps: initial tree construction, node insertions, and tree optimization. For each step,
we present an algorithm to solve the corresponding subproblem. Our algorithm for initial
tree construction is optimal for the static environment where nodes cannot move. However,
we can effectively apply the later algorithms if we must start with a different topology.
Our greedy heuristic for improving the routing tree topology by adding nodes exploits the
mobility of the newly added nodes. Our tree optimization algorithm improves the routing
61

tree by relocating its nodes without changing its topology. This iterative algorithm converges
on the optimal position for each node given the constraint that the routing tree topology is
fixed. Our node insertion and tree optimization algorithms use the LocalP os algorithm we
propose in Fig. 5.7 that optimally solves the simplest case (see Section 5.2.2) of the mobile
relay configuration problem where there is a single source, a single sink, and a single relay
node. Our approach is not guaranteed to produce an optimal configuration because we do
not necessarily find the optimal topology, but our simulation results show that it performs
well.

5.2.2

Base Case

Before presenting our algorithm for OMRC, we revisit the example of Section 5.1.1 as it
represents the simplest possible base case of the problem in which the network consists of
one source si−1 , one mobile relay node si and one sink si+1 . In this section, we calculate
the optimal position for the relay node. We use the following notation. In

2,

let the

original position of a node sj be oj = (pj , qj ), and let uj = (xj , yj ) its final position in
configuration U . According to our energy models, the total transmission and movement
energy cost incurred by the mobile relay node si is

ci (U ) = k ui − oi + am + b ui+1 − ui 2 m

We also define

Ci (U ) = ci (U ) + am + b ui − ui−1 2 m

62

This corresponds to the transmission cost of node si−1 plus the total cost of node si , which
is the total cost of the final configuration in this example. We need to compute a position ui
for si that minimizes Ci (U ) assuming that ui−1 = oi−1 and ui+1 = oi+1 ; that is, node si ’s
neighbors remain at the same positions in the final configuration U . We calculate position
ui = (xi , yi ) for node si by finding the values for xi and yi where the partial derivatives of
the cost function Ci (U ) with respect to xi and yi become zero. Position ui will be toward
the midpoint of positions ui−1 and ui+1 . The partial derivatives

δCi (U ) δCi (U )
,
at xi and
δxi
δyi

yi , respectively are defined as follows.

δCi (U )
= − 2bm(xi+1 − xi ) + 2bm(xi − xi−1 )
δxi
+k

(xi − pi )
(xi − pi )2 + (yi − qi )2

δCi (U )
= − 2bm(yi+1 − yi ) + 2bm(yi − yi−1 )
δyi
+k

Setting

(yi − qi )
(xi − pi )2 + (yi − qi )2

δCi (U )
δC (U )
= 0, i
= 0, we get the following two cases. Suppose si needs to
δxi
δyi

move left. This means pi is to the right of the midpoint of nodes si−1 and si+1 . Let
k
4bm

1
. The optimal position is then xi = (xi−1 +xi+1 )+Yi .
2
(y
+ yi+1 − 2qi )2
1 + i−1
(xi−1 + xi+1 − 2pi )2
If si needs to move right, then pi is to the left of the midpoint of nodes si−1 and si+1 . The
Yi =

1

1
optimal position is then xi = (xi−1 + xi+1 ) − Yi . The corresponding yi in both cases is
2
(xi−1 + xi+1 − 2pi )
(x − pi ) + qi .
(yi−1 + yi+1 − 2qi ) i
We note that in some cases it might not be beneficial to move, so the optimal position

63

for the relay node is its original position. The algorithm to compute the optimal position of
a relay node given its neighbors is shown in Fig. 5.7.
function localPos(oi , ui , ui−1 , ui+1 )
Consider case si moves right
valid ← F ALSE;
xi ← 1 (xi−1 + xi+1 ) − Yi ;
2
if xi > pi then
valid ← T RU E;
else
Consider case si moves left
xi ← 1 (xi−1 + xi+1 ) + Yi
2
if xi < pi then
valid ← T RU E;
end if
end if
Record if new position is different from previous one
if valid then
i+1 −2p )
yi ← (xi−1 +xi+1 −2qii) (xi − pi ) + qi ;
(yi−1 +y
ui = (xi , yi );
if ui − ui > threshold then
return (ui , TRUE);
end if
end if
not beneficial to move, stay at original position
return (oi , FALSE);
end function

Figure 5.7: Algorithm to compute the optimal position of a relay node that receives data
from a single node and transmits the data to a single node.

5.2.3

Static Tree Construction

Different applications may apply different constraints on the routing tree. When only
optimizing energy consumption, a shortest path strategy (as discussed below) yields an
optimal routing tree given no mobility of nodes. However, in some applications, we do not
have the freedom of selecting the routes. Instead, they are predetermined according to some
other factors (such as delay, capacity, etc). In other less stringent cases, we may be able to

64

update the given routes provided we keep the main structure of the tree. Depending on the
route constraints dictated by the application, we start our solution at different phases of the
algorithm. In the unrestricted case, we start at the first step of constructing the tree. When
the given tree must be loosely preserved, we start with the relay insertion step. Finally, with
fixed routes, we apply directly our tree optimization algorithm. Our simulations (Section
5.6) show that our approach outperforms existing approaches for all these cases.
We construct the tree for our starting configuration using a shortest path strategy. We
first define a weight function w specific to our communication energy model. For each pair of
nodes si and sj in the network, we define the weight of edge si sj as: w(si , sj ) = a+b oi −oj 2
where oi and oj are the original positions of nodes si and sj and a and b are the energy
parameters discussed in Section 3. We observe that using this weight function, the optimal
tree in a static environment coincides with the shortest path tree rooted at the sink. So
we apply Dijkstra’s shortest path algorithm starting at the sink to all the source nodes to
obtain our initial topology.

5.2.4

Node Insertion

We improve the routing tree by greedily adding nodes to the routing tree exploiting the
mobility of the inserted nodes. For each node sout that is not in the tree and each tree edge
si sj , we compute the reduction (or increase) in the total cost along with the optimal position
of sout if sout joins the tree such that data is routed from si to sout to sj instead of directly
from si to sj using the LocalPos algorithm described in Fig. 5.7. We repeatedly insert the
outside node with the highest reduction value modifying the topology to include the selected
node at its optimal position, though the node will not actually move until the completion

65

of the tree optimization phase. After each node insertion occurs, we compute the reduction
in total cost and optimal position for each remaining outside node for the two newly added
edges (and remove this information for the edge that no longer exists in the tree). At the
end of this step, the topology of the routing tree is fixed and its mobile nodes can start the
tree optimization phase to relocate to their optimal positions.

5.3

Tree Optimization

In this section, we consider the subproblem of finding the optimal positions of relay nodes
for a routing tree given that the topology is fixed. We assume the topology is a directed
tree in which the leaves are sources and the root is the sink. We also assume that separate
messages cannot be compressed or merged; that is, if two distinct messages of lengths m1
and m2 use the same link (si , sj ) on the path from a source to a sink, the total number of
bits that must traverse link (si , sj ) is m1 + m2 .
First, we extend the base case solution of Section 5.2.2 to handle multiple flows passing
through a mobile relay node. Then, we propose an iterative algorithm that uses the solution
for this base case to compute the new positions of the relay nodes in the routing tree. We
also show that this algorithm converges to the optimal solution for the given tree given the
topology is fixed.

5.3.1

Extended Base Case

Before we describe our optimal algorithm for this problem, we extend the solution to
the base case presented in Section 5.2.2 to the more general multiple flow traffic pattern.
The network now consists of multiple sources, one relay node and one sink such that data is

66

transmitted from each source to the relay node and then to the sink. We modify our solution
as follows. Let si be the mobile relay node, S(si ) the set of source nodes transmitting to si
and sd the sink collecting nodes from si . The cost incurred by si in this configuration U is:
i

ci (U ) = k ui − oi + ami + bmi ud − ui 2

where mi is the total amount of data that si transmits to sd . Similar to the single source
i
base case, we define

aml + b ui − ul 2 ml

Ci (U ) = ci (U ) +
sl ∈S(si )

This corresponds to the transmission cost of all nodes sl that send messages to node si plus
the total cost of node si . In this case, this also corresponds to the total cost of configuration
ml . We then follow the

U which we wish to minimize. First, we compute mi as
sl ∈S(si )

same routine of computing the points at which both partial derivatives
become zero. We obtain the following positions:

xi = p i +

y i = qi +

2
2
−Bx ( Bx + By ± k)
2
2
A Bx + By
2
2
−By ( Bx + By ± k)
2
2
A Bx + By

67

δCi (U )
δCi (U )
and
δxi
δyi

where

A = mi +

ml
sl ∈S(si )

Bx = mi xd +

ml xl + Api
sl ∈S(si )

By = mi yd +

ml yl + Aqi
sl ∈S(si )

The details of the derivation are provided in proof of Theorem 4 of appendix A. We note
that these values correspond to two candidate points moving in each direction (left/right).
The optimal position is the valid value yielding the minimum cost.

5.3.2

Optimization Algorithm

We propose a simple iterative approach to compute the optimal position ui for each node
j

j

j

si . We define the following notations. Let ui = (xi , yi ) be the position of node si after the
j

j

jth iteration of our algorithm for j ≥ 0 and U j = (u1 , . . . , un ) the computed configuration
of nodes s1 through sn after j iterations. We define u0 = oi . Note that the mobile relay
i
nodes do not move until the final positions are computed.
Our algorithm starts by an odd/even labeling step followed by a weighting step. To obtain
consistent labels for nodes, we start the labeling process from the root using a breadth first
traversal of the tree. The root gets labeled as even. Each of its children gets labeled as
odd. Each subsequent child is then given the opposite label of its parent. We define mi ,
the weight of a node si , to be the sum of message lengths over all paths passing through si .
This computation starts from the sources or leaves of our routing tree. Initially, we know

68

mi = Mi for each source leaf node si . For each intermediate node si , we compute its weight
as the sum of the weights of its children.
procedure OptimalPositions(U 0 )
converged ← false;
j ← 0;
repeat
anymove ← false;
j ← j + 1;
Start an even iteration followed by an odd iteration
for idx = 2 to 3 do
for i = idx to n by 2 do
(uj , moved) ← localPos(oi , S(si ), sd );
i
i
anymove ← anymove OR moved
end for
end for
converged ← NOT anymove
until converged
end procedure

Figure 5.8: Centralized algorithm to compute the optimal positions in a given tree

Once each node gets a weight and a label, we start our iterative scheme. In odd iterations
j

j, the algorithm computes a position ui for each odd-labeled node si that minimizes Ci (U j )
j

j−1

j

j−1

assuming that ui−1 = ui−1 and ui+1 = ui+1 ; that is, node si ’s even numbered neighboring
nodes remain in place in configuration U j . In even-numbered iterations, the controller does
the same for even-labeled nodes. The algorithm behaves this way because the optimization of
j

ui requires a fixed location for the child nodes and the parent of si . By alternating between
optimizing for odd and even labeled nodes, the algorithm guarantees that the node si is
always making progress towards the optimal position ui . Our iterative algorithm is shown
in Fig. 5.8
Fig. 5.9 shows an example of an optimal configuration for a simple tree with one source
node. Nodes start at configuration U 0 . In the first iteration, odd nodes (s3 and s5 ) moved to
their new positions (u1 , u1 ) computed based on the current location of their (even) neighbors
3 5
69

Figure 5.9: Convergence of iterative approach to the optimal solution. Each line shows
the configuration obtained after 2 iterations. The optimal configuration is reached after 6
iterations.
(u0 , u0 , u0 ). In the second iteration, only even nodes (s2 and s4 ) moved to their new positions
2 4 6
(u2 , u2 ) computed based on the current location of their (odd) neighbors (u1 , u1 , u1 ). Since
2 4
1 3 5
s3 and s5 did not move, their position at the end of this iteration remains the same, so
u1 = u2 and u1 = u2 . In this example, nodes did two more sets of iterations, and finally
3
3
5
5
converged to the optimal solution shown by configuration U 6 .
Even though configurations change with every iteration, nodes only move after the final
positions have been computed. So each node follows a straight line to its final destination.
As the data size increases, nodes in the optimal configuration get more evenly spaced. In
fact, in any given configuration, the maximum distance traveled by a node is bounded by the
distance between its starting position and its final position in the evenly spaced configuration.
The above example shows another property of our algorithm. When a node si moves
and its neighbors (si−1 and si+1 ) remain in place, it moves in the direction of the midpoint
of si−1 si+1 . This results in a reduction in the length of one of the transmission links. The
other may increase in length but will never exceed the new length of the first link. This

70

remains valid for multiple children case. So in any configuration U i+1 , the length of the
largest link is at most the length of the largest link in the previous configuration U i . So if
we start with a route along good quality links, this quality will be preserved in the optimal
configuration (and throughout intermediate configurations).

5.4

Efficiency and Optimality

We first consider efficiency. Our initial tree construction algorithm is essentially a single
source shortest path algorithm. Using Dijkstra’s algorithm, the time complexity is O(n2 )
where n is the number of nodes. Our second algorithm needs to compute the reduction in cost
for each pair of node and tree edge, so the time complexity is O(n2 ). Our tree optimization
algorithm runs until the change in position for each node falls below a predefined threshold.
The value of this threshold represents a tradeoff between precision and cost. As the threshold
decreases, more iterations are needed for convergence. Upon termination, no node can move
by itself to improve the overall cost (within the threshold bound). We have not completed
a rate of convergence analysis for this algorithm. However, in our simulations, we reach
our error threshold within 8 to 10 iterations. Since each iteration involves only half the
j

nodes and each computation of ui can be performed in constant time, the time complexity
of our algorithm is O(λn), where λ is the number of iterations to reach convergence. Given
that λ ≤ 10 in our simulations, our observed time complexity is O(n). The resulting time
complexity for the full approach is O(n2 ).
With respect to optimality, our resulting configuration is not necessarily optimal because
we do not necessarily find the optimal topology. However, two of our algorithms, the initial
tree construction algorithm and the tree optimization algorithm, are optimal for their re71

spective subproblems. That is, our initial tree construction algorithm is optimal in a static
environment where nodes cannot move so that only the original positions of the nodes are
considered. Likewise, for our tree optimization algorithm, we prove that the final configuration where no node can move by itself to improve the overall cost (within the threshold
bound) is globally optimal; that is, no simultaneous relocation of multiple nodes can improve
the overall cost. We present the proof of optimality in Theorem 5 of appendix A. The key
intuition is that for a configuration in which no relay node can move and improve the cost
by itself, the directional derivative [7] in any direction at that configuration is positive; this
is a sufficient condition for the optimality of that configuration because the cost function is
convex.

5.5

Distributed Algorithms

Our solutions to the three subproblems assume a centralized scheme in which one node
has full knowledge of the network including which nodes are on the transmission paths to each
source, the original physical position oi of each node si , and the total message length m to be
sent from each source. Whereas the centralized algorithm computes the optimal static tree
and the optimal position of each node in the restructured tree, it incurs prohibitively high
overhead in large-scale networks. We now present a distributed and decentralized version of
each of our algorithms.
We modify the first phase, the tree construction phase, to use a fully distributed routing
algorithm. We pick greedy geographic routing since it does not require global knowledge of
the network although any algorithm with such property can be used.
After a routing tree is constructed, the tree restructuring phase begins. Network nodes

72

outside the tree broadcast their availability (as NODE IN RANGE message) to tree nodes
within their communication range and wait for responses for a period of time Tw . Similarly,
tree nodes enter a listening phase Tu . During that period, tree nodes receive messages of
different types (NODE IN RANGE, OFFER, . . .). Each tree node that receives one or more
NODE IN RANGE message responds to the sender by giving it its location information and
its parent’s location information. Each non-tree node so that receives location information
from a tree node si during Tw computes the reduction in cost if it joins the tree as parent
of si and adds si to a list of candidates. At the end of Tw , the non-tree node selects from
the candidate list the node that results in the largest reduction and sends it an offer. It
also sends the tree node with the second largest reduction a POTENTIAL OFFER message.
At the end of Tu , each tree node vt that collected one or more offers and potential offers
operates as follows. If vt ’s best potential offer exceeds its best offer by a certain threshold B
and vt has not already waited R rounds, vt waits rather than accepting its best offer in the
hopes that its best potential offer will become an actual offer in another round. By waiting,
it sends everyone a REJECT OFFER, restarts the listening phase, and records that it has
waited another round. Otherwise, vt accepts its best offer by responding to its sender p
with an ACCEPT OFFER message and to the remaining nodes with a REJECT OFFER
message. It then updates its parent in the tree to p, resets Tu and starts the listening phase
again.
A non-tree node p that receives an ACCEPT OFFER message moves to the corresponding
local optimal location and joins the tree. It becomes a tree node and enters the listening
phase. On the other hand, if p does not receive an ACCEPT OFFER, p repeats the process
by broadcasting its availability again and resetting Tw . We note that values in p’s candidate
73

procedure treeRun
Phase I: Run routing algorithm to discover parent and children
(parent, children) ← DistributedRouting;
Phase II: Start tree restructuring phase
offers ← ∅; potentialoffers ← ∅; wait ← 0;
repeat
Listen to incoming offers or changes in structure
repeat
receive(sender, type, data);
if type = MOBILE IN RANGE then
send(sender, META DATA, info);
else if type = OFFER then
offers.add(data);
else if type = POTENTIAL OFFER then
potentialloffers.add(data);
else if type = UPDATE STRUCTURE then
children.add(data.newchild);
children.remove(data.oldchild);
end if
until timeout
Process offers and pick best
if offers = ∅ then
bestOffer ← offers.dequeue();
bestPotentialOffer ← potentialoffers.dequeue();
if bestPotentialOffer > bestOffer*B and wait < R then
send(bestOffer.sender, REJECT OFFER);
wait++;
else
send(bestOffer.sender, ACCEPT OFFER);
parent ← bestOffer.sender;
end if
end if
while offers = ∅ do
offer ← candidates.dequeue();
send(offer.sender, REJECT OFFER);
end while
until timeout
Phase III: Iterate moving to optimal local positions
converged ← f alse;
while not converged do
(u, converged) ← LocalPos(o, parent, children);
Exchange location info with parent and children
send(parent, NEW LOCATION, u);
for all child ∈ children do
send(child, NEW LOCATION, u);
end for
receive(parent, NEW LOCATION, parent.u);
for all child ∈ children do
receive(child, NEW LOCATION, child.u);
end for
end while
end procedure

Figure 5.10: Local algorithm executed by tree nodes

74

list cannot be reused to extend offers to old tree nodes since those tree nodes could have
a new parent at this point in time. When the second phase ends, any remaining non-tree
nodes stop processing whereas tree nodes enter the tree optimization phase. Fig. 5.10 shows
the algorithm executed by each tree node.
Giving tree nodes the ability to wait before accepting an offer increases the chances of
using mobile relay nodes to their full potential. For example, consider a scenario where
several mobile relay nodes can greatly improve the capacities of several tree links but are all
closest to one specific link. They will all send offers to the same tree node while the rest of
the tree nodes in their proximity will receive modest offers from more distant mobile nodes.
If the tree nodes cannot wait, they will be forced to accept a modest offer and the mobile
nodes will either remain unused or they will help more distant tree nodes where their impact
is reduced since they use up more energy to get to their new location.
The centralized tree optimization algorithm can be transformed into a distributed alj

gorithm in a natural way. The key observation is that computing each ui for node si only
depends on the current position of si ’s neighbors in the tree (children and parent), nodes that
si normally communicates with for data transfers. Thus, si can perform this computation.
The distributed implementation proceeds as follows. First, there is a setup process where
the sender s1 sends a discover message that ends with the receiver sn ; the two purposes of
this message are (1) to assign a label of odd or even to each node si and (2) for each node
si to learn the current positions of its neighbors. A node si sends its current position to
node sj when acknowledging receipt of the discover message. Second, there is a distributed
process by which the nodes compute their transmission positions. We make each iteration
of the basic algorithm a “round”, though there does not need to be explicit synchronization.
75

In odd rounds, each odd node computes its locally optimal position and transmits this new
position to its neighbors. In even rounds, each even node does the same. A node begins
its next round when it receives updated positions from all its neighbors. The final step is
to have the nodes move to their computed transmission positions, send messages to their
neighbors saying they are in position, and finally perform the transmission. To ensure the
second process does not take too long, we limit the number of rounds to 8; that is, each
node computes an updated position four times. Simulation results show that this is enough
to obtain costs close to optimal (see Section 5.6).

5.6
5.6.1

Simulations
Setup

We carried out simulations on 100 randomly generated initial topologies, each of which
has 100 nodes placed uniformly at random within a 150m by 150m area. We used these
initial topologies to generate two subsequent sets of complete topologies with established
sources and sink. We used the first set to study the effectiveness of our algorithms as the
amount of data transferred to the sink varies and the second set to study the effectiveness
of our algorithms for different numbers of sources. In the first set, we selected sources and
sinks uniformly at random from these 100 nodes. We varied the number of sources from 4 to
12, by increments of 2, and used each number of sources for 20 initial topologies. For each
resulting topology, we created many separate input instances by varying the data chunk size
from 1MB to 150MB where the data chunk size for an input instance is the common amount
of data to be transferred from each source to the sink. In the second set, for each initial

76

topology, we generated 10 different complete topologies by starting with 2 randomly selected
sources, and adding two new sources to the previous set at each step.
We used the following settings to model the transmission and mobility costs of our nodes.
For transmission, we use a = 0.6 × 10−7 and b = 4 × 10−10 as the standard setting which
is consistent with the empirical measurements on CC2420 motes [45]. For mobility, we used
different settings in each of our two sets. In the first set, we used k = 2 as the standard
setting because it models several platforms such as Robomote [12, 1]. In the second set,
we set k to be 1, 2 and 4 since we additionally use that set to study the effect of different
mobility costs on the energy reduction. Furthermore, we set the maximum communication
distance of a node to be 30m, which was shown to result in a high packet reception ratio
for the CC2420 radio [45]. We ran simulations using different values for the convergence
threshold. We obtained similar gains for values less than or equal to 0.01. In the following
simulations, we set the threshold to 0.01.
Our algorithmic framework starts with an initial routing tree. In the centralized setting,
we construct this initial routing tree using the following three widely used routing algorithms:
power based routing, hop based routing, and greedy geographic routing. Power based routing
computes a shortest path from the sink to each source with each edge weight being the square
of the distance between the two corresponding nodes plus some constant value to represent
the energy consumed a + bd2 to transmit each byte of data over that edge. Hop based
routing minimizes the number of hops between each source and the sink and is the base of
several widely used algorithms in wireless networks (e.g. AODV [40]). Given our maximum
communication range of 30m, we do not have any links with poor quality which is a common
concern with hop based routing. Greedy geographic routing is a greedy strategy in which
77

each node forwards messages to the reachable node (within the communication range of
the node) that is closest to the sink. The first two tree construction approaches require
global knowledge of the network whereas the last one is fully localized. For the distributed
setting, we construct the initial routing tree using greedy geographic routing because it
is fully localized. Of the 100 initial topologies, the distributed routing algorithm resulted
in a disconnected path between the sources and the sink in only four networks given our
maximum communication distance of 30m.

5.6.2

Performance Metrics

We study variants of our strategy where we use only one optimization, inserting nodes or
optimizing a given tree, to determine the benefit of both optimizations. Specifically, we use
TREE to represent the variant where we only construct an initial tree and do no optimizations, TREE+FO to represent the variant where we optimize the initial tree, TREE+INS to
represent the variant where we insert nodes into the initial tree, and TREE+INS+FO to represent the variant where we insert nodes into the initial tree and then optimize the final tree.
The three possibilities for TREE are PB, HB, and GG which represent the Power Based, Hop
Based, and Greedy Geographic tree construction algorithms, respectively. For each input instance I, we let T REE(I) denote the energy consumed by the initial tree constructed by our
three tree construction algorithms PB, HB, and GG, and we let T REE +OP T (I) denote the
energy consumed by the final optimized tree where TREE can be PB, HB, or GG and OPT
can be INS, FO or INS+FO. The reduction ratio achieved by optimization OPT on input I
for tree construction algorithm TREE is (T REE(I) − T REE + OP T (I))/T REE(I). We
measure the performance of optimization OPT on initial tree strategy TREE by computing

78

Power Based
Hop Based

Geographic Greedy

Comparison to power based
original configuration

120%
110%
100%
90%
80%
70%
60%
50%
0

25

50

75

100

125

150

Data Chunk Size (MB)

Figure 5.11: Graph of the average static energy consumption ratio of TREE+INS+FO as a
function of data chunk size for our three tree construction strategies PB, HB, and GG
the average reduction ratio achieved by OPT over all input instances I of set 1 that have the
same data chunk size. Moreover, for each input instance I and each algorithm TREE+OPT,
we define the static energy ratio (T REE + OP T (I))/P B(I) where P B(I) is the cost of the
power based tree which is the optimal cost for the static version of this problem where no
nodes can move. The static energy ratio measures the benefit of our algorithms which exploit
mobility of nodes versus the static optimal configuration. We measure the overall performance of algorithm TREE+OPT by computing the average static energy ratio achieved by
TREE+OPT over all input instances I of set 1 that have the same data chunk size. Finally, we measure the performance of optimization INS+FO on initial tree strategy TREE
by computing the average reduction ratio achieved by INS+FO over all input instances I of
set 2 that have the same number of sources.

79

5.6.3

Centralized Algorithm

We first show the benefit of exploiting the mobility of relay nodes by computing the
average static energy consumption ratio of TREE+INS+FO for all data chunk sizes for each
of our three tree building strategies PB, HB, and GG as shown in Fig. 5.11. For all three
initial tree strategies, we see that the average static energy consumption ratio drops quickly
as the data chunk size increases. For HB and GG, the average static energy consumption
ratio starts out higher than 100% because P B(I), the optimal tree for the static case, is
roughly 37% lower than HB(I) and GG(I) for any of our input instances. Even given this
initial disadvantage of a poor starting tree from an energy consumption perspective, we see
that the average static energy consumption ratios of HB+INS+FO and GG+INS+FO drop
below 100% for data chunk sizes of 12 MB and 15 MB, respectively. As the data chunk
size increases further, both HB+INS+FO and GG+INS+FO achieve average static energy
consumption ratios of 75% and 60% for data chunk sizes of 60 MB and 150MB, respectively.
The results for PB+INS+FO are even better because we start with the optimal tree for the
static case. Thus, the average static energy consumption ratio for PB+INS+FO is always
below 100% and reaches 55% for 150 MB.
We now evaluate the benefit achieved by our optimizations FO, INS, and INS+FO for
each of our tree building strategies PB, HB, and GG. We note that in this set of simulations,
we used our centralized improvement schemes with the distributed tree building approach
GG. The purpose is to test the limits of our optimizations given a non-optimal starting tree.
A fully distributed setup is studied later in this section.
We start with optimization INS+FO. Fig. 5.12 plots the average reduction ratio for

80

Power Based
Hop Based

Geographic Greedy

Average Reduction Ratio

70%
60%
50%
40%
30%
20%
10%
0%
0

25

50

75

100

125

150

Data Chunk Size (MB)
Figure 5.12: Graph of the average reduction ratio of optimization
INS+FO as a function of data chunk size for our three tree construction strategies PB, HB, and GG
Power Based
Hop Based

Greedy Geographic

Average Reduction Ratio

70%
60%
50%
40%
30%
20%
10%
0%
0

25

50

75

100

125

Data Chunk Size (MB)

150

Figure 5.13: Graph of the average reduction ratio of optimization
FO as a function of data chunk size for our three tree construction
strategies PB, HB, and GG

81

Power Based
Hop Based

Greedy Geographic

Average Reduction Ratio

70%
60%
50%
40%
30%
20%
10%
0%
0

25

50

75

100

125

Data Chunk Size (MB)

150

Figure 5.14: Graph of the average reduction ratio of optimization INS as a function of data
chunk size for our three tree construction strategies PB, HB, and GG
optimization INS+FO for PB, HB, and GG. In all three cases, we see the same basic trend;
the average reduction ratio increases as data chunk size increases. For both HB and GG, the
average reduction ratio starts at roughly 25% for small data chunk sizes and exceeds 60%
for large data chunk sizes; for PB the average reduction ratio starts near 0% and exceeds
43% for large data chunk sizes. The difference in average reduction ratio, in particular for
small data chunk sizes, is due to the quality of the initial tree. For PB, the initial tree is
good so there is little that our optimization INS+FO can do to improve energy consumption
for small data chunk sizes. For HB and GG, the initial tree can be very poor, so INS+FO
can provide immediate improvement to the tree to significantly reduce energy consumption
by an average of 25% for data chunk sizes of only 1MB. We note that although INS+FO
achieves higher reduction ratios for HB and GG than for PB, the total energy consumed by
PB+INS+FO is lower than the total energy consumed by HB+INS+FO or GG+INS+FO.

82

We next consider optimization FO alone. Fig. 5.13 plots the average reduction ratio for
optimization FO for PB, HB, and GG. In all three cases, the average reduction ratio starts
at 0% for small data chunk sizes and increases to roughly 18% for HB and GG and 33%
for PB for large data chunk sizes. It is interesting to note that FO is most effective for PB
whereas INS+FO achieved significantly greater reduction ratios for GG and HB for all data
chunk sizes.
Finally, we consider optimization INS alone. Fig. 5.14 plots the average reduction ratio
for optimization INS for PB, HB, and GG. In all three cases, we see the average reduction
ratio of INS alone is comparable to that of INS+FO (within 5%-8% for data chunk sizes of
at least 15MB). For very small data chunk sizes, the average reduction ratio is constant until
a certain threshold is exceeded and then rises significantly.
We now evaluate our approach as we vary the number of sources. We used the greedy
geographic tree GG as our initial tree and INS+FO as our optimization algorithm. Fig.
5.15 shows the average reduction ratio as a function of the number of sources. We observe
that this ratio remains almost constant for different values of k as the difference in ratios for
different number of sources does not exceed 3.5%. Fig. 5.15 also shows the effect of mobility
costs on the reduction in energy consumption costs in general. As mobility costs decrease,
it becomes more effective for mobile nodes to move over longer distances and reduce the
communication consumption further so the reduction in total costs increases as k decreases.
Given our simulation results, we draw the following five conclusions. First, we achieve the
best results when we use the power based tree PB as our initial tree. Second, if we use the
power based tree, either optimization alone is very effective and both optimizations together
achieve the best results. Third, if we start with either the hop based tree HB or the greedy
83

Average Reduction Ratio

Centralized (k = 1)
Centralized (k = 2)
Centralized (k = 4)

Distributed (k = 1)
Distributed (k = 2)
Distributed (k = 4)

65%
60%
55%
50%
45%
40%
35%
30%
25%
2

4

6

8 10 12 14 16 18 20

Number of Sources

Figure 5.15: Graph of the average reduction ratio of the centralized and distributed
GG+INS+FO optimizations as a function of the number of sources, a data chunk of 75MB
and different values of k.
geographic tree GG, the most effective optimization is the node insertion optimization INS
which achieves nearly as good an average reduction ratio as INS+FO. Fourth, if we start
with the hop based or greedy geographic tree, we can achieve a static energy ratio that is
close to that achieved by starting with the power based tree if we apply both optimizations.
In particular, the node insertion optimization INS helps alleviate the initial disadvantage
by adding a lot of new nodes into the tree. We briefly explain the reason for all of these
conclusions. The key observation is that the hop based and greedy geographic trees HB and
GG tend to create initial trees with relatively long edges and relatively few nodes whereas
the power based tree PB tends to create trees with lots of nodes and relatively short edges
because of the quadratic cost metric. As a result, for HB and GG, optimization FO alone
which rearranges nodes is relatively ineffective as it can only balance the relatively long
edges. On the other hand, optimization INS alone can insert new nodes into the tree and

84

thus create a new tree with significantly shorter edges on average given HB or GG as the
initial tree. Because PB starts with many more nodes and shorter edges, PB does not benefit
as much from node insertion INS as HB and GG do, and PB benefits a lot more from node
rearrangement FO than HB and GG do. Fifth, the improvement ratios that we obtain are
almost independent of the number of sources in the network.
In all our simulation results, the standard deviation varied between 4% and 6.5%. We
identified six outlier topologies which deviated from the mean by more than 10%. In these
topologies, the sources were either very close to the sink so there was little room for improvement or very far from the sink so the improvement was much greater than the average
case.

5.6.4

Distributed Algorithm

We now evaluate how well our distributed implementation works. Our initial tree is the
greedy geographic tree GG. We consider four optimizations: the centralized implementation
of INS+FO, the distributed implementation of just FO, the distributed implementation of
just INS, and the distributed implementation of INS followed by the distributed implementation of FO. For the distributed implementation of INS, we set parameter B to 10% (a
potential offer must be 10% better than the best actual offer to cause a node to wait). Fig.
5.16 shows the average reduction ratio of each of these optimizations.
The average reduction ratio for distributed INS+FO starts at 20% for small data chunk
sizes, reaches 30% for data chunk sizes around 20MB, and exceeds 40% for data chunk sizes
larger than 75MB. The gap between the average reduction ratio for centralized INS+FO
and distributed INS+FO starts at roughly 5% for small data chunk sizes and increases to

85

Distributed - GG+FO

Distributed - GG+INS+FO
60%

Distributed - GG+INS

Average Reduction Ratio

GG+INS+FO

50%
40%
30%
20%
10%
0%
0

25

50

75

100

125

150

Data Chunk Size (MB)
Figure 5.16: Graph of the average reduction ratio of the centralized optimization (INS+FO)
and three distributed optimizations (FO, INS, INS+FO) as a function of data chunk size for
the greedy geographic tree GG.
roughly 15% for large data chunk sizes. This gap is due to the lack of global information
when performing the insertion step. Expensive links in the tree that do not have nearby
relay nodes are not able to communicate with further but available relay nodes whose help is
only offered to cheaper but nearby links. This problem is exacerbated as the data chunk size
increases. We varied values for B between 10% and 50% and for R between 1 and 3. For all
combinations of B and R that we tested, we obtained similar results to those of Fig. 5.16.
As in the centralized case, distributed INS is more effective than distributed FO. However,
doing both distributed optimizations does result in roughly a 10% improvement compared
to only doing the distributed INS optimization for most data chunk sizes.
Similar to the centralized implementation, we observe a slow reduction in the improvement ratio as the number of sources increases for k = 2 and 4 (Fig. 5.15). For cheaper

86

mobility cost (k = 1), the difference in improvement ratios increases at a faster rate and
reaches 9% as the number of sources increases from 2 to 20. This is because when mobility
is cheaper, in an optimal setting, nodes can move over longer distances to help expensive
links. However, as we mentioned earlier, in a distributed setting, mobile nodes are not aware
of those distant expensive edges. Moreover, as the number of sources increases, the number
of mobile nodes available to help decreases. Both factors combined make the distributed
implementation slightly less effective for a high number of sources.

5.7

Summary

In this chapter, we proposed a holistic approach to minimize the total energy consumed
by both mobility of relays and wireless transmissions. When we model both sources of
energy consumption, the optimal position of a node that receives data from one or multiple
neighbors and transmits it to a single parent is not the midpoint of its neighbors; instead,
it converges to this position as the amount of data transmitted goes to infinity. Ideally,
we start with the optimal initial routing tree in a static environment where no nodes can
move. However, our approach can work with less optimal initial configurations including
one generated using only local information such as greedy geographic routing. Our approach
improves the initial configuration using two iterative schemes. The first inserts new nodes
into the tree. The second computes the optimal positions of relay nodes in the tree given a
fixed topology. This algorithm is appropriate for a variety of data-intensive wireless sensor
networks. It allows some nodes to move while others do not because any local improvement
for a given mobile relay is a global improvement. This allows us to potentially extend our
approach to handle additional constraints on individual nodes such as low energy levels or

87

mobility restrictions due to application requirements.

88

CHAPTER 6
Maximizing Network Lifetime Using
Node Rotation
In this chapter, we directly address the problem of extending the network lifetime. Similar to the previous setting, we consider networks that consist entirely of mobile nodes.
Additionally, all the nodes in the network generate data that need to be relayed to the sink.
Most sensor networks have a many-to-one traffic pattern and consequently a tree topology.
One of the main reasons limiting the lifetime of such networks is that nodes that are closer
to the sink usually transmit a large amount of data and consequently drain their battery at
a much faster rate than nodes further from the sink. As a result, the lifetime of the network
is substantially reduced even though the network still contains nodes rich in energy. We propose an approach that exploits the mobility of the nodes to balance the energy consumption
throughout the network. The main idea is to redistribute the nodes in the network such that
nodes at low traffic (i.e. low consumption) positions switch places with nodes at high traffic
(i.e. high consumption) positions.

89

The rest of this chapter is organized as follows. We first describe the motivation behind
our approach in Section 6.1. Then, we define the problem of maximizing the network lifetime
using node rotations in Section 6.2. In Section 6.3, we present our solutions for the problem.
In Section 6.4, we derive some upper bounds on the effectiveness of any node rotation
algorithm for several classes of input instances. Section 6.5 describes our simulation results
and Section 6.6 summarizes the chapter.

6.1

Motivation

Recently, the controlled mobility of sensors has been exploited to improve the energy
efficiency of WSNs. For instance, by relocating mobile sensors, the communication topology
of a network can be dynamically configured to reduce power consumption. Moreover, mobile
sensors can physically carry large chunks of data to reduce energy consumption in wireless
transmissions [50]. Such approaches become increasingly attractive due to the emergence of
numerous low-cost mobile sensor prototypes such as Robomote [12], Khepera [1], and FIRA
[28].
However, many applications have constraints which make existing approaches infeasible.
We identify three key constraints. The first is that the location of the nodes and the communication topology of the network may not be mutable. For example, in a surveillance
or environment monitoring application, the exact placement of sensor nodes may not be
adjusted without compromising the operation of the system. The second is that all nodes
have equal, typically limited, capabilities. This rules out approaches that require a few nodes
with extra capabilities and the ability to perform complex motion planning. The third is
that nodes face differential power consumption where some nodes consume significantly more

90

power than other nodes. For example, nodes closer to the sink in a given routing topology
often have to transmit more data and thus consume more power than nodes farther from
the sink in the given topology.
To address these three constraints that limit existing techniques, we propose a new approach that we call mobile node rotation which is inspired by the huddling and rotation
behavior of emperor penguins that help them breed in the fierce arctic winter. Penguins
on the outside of the huddle face temperatures as low as −45 ◦ C and strong winds while
those on the inside of the huddle enjoy warm ambient temperatures as high as 37 ◦ C and
significant wind protection. Emperor penguins rotate positions to share the burden of being
on the outside [69]. In mobile node rotation, we propose to rotate the physical positions of
mobile sensors to share the burden of any high power consumption location. We observe
that mobile node rotation is particularly suitable for mobile sensor platforms with limited
mobility as we can impose mobility constraints on individual nodes. For example, we can
model the constraints of the NIMS sensors [41] that are only capable of moving along fixed
cables by only allowing such sensors to exchange with other sensors on the same set of cables. Likewise, mobile node rotation does not require powerful nodes capable of performing
complex motion planning calculations or developing new mobility-aware routing topologies
since all movements are to known positions and the topology does not change.
We first present a new problem, Max-lifetime Node Rotation (MaxLife), that models
maximizing the lifetime of a WSN using rounds of mobile node rotation. MaxLife can incorporate any energy consumption model for both wireless communication and node movement.
We then efficiently solve the one round MaxLife problem by reducing it to the assignment
problem and the general multiple round MaxLife problem by performing localized rotations
91

of node pairs. We prove upper bounds on the lifetime improvement ratio of mobile node
rotation approaches. Finally, we conduct extensive simulations based on energy models obtained from existing mobile sensor platforms. We show that our algorithms can significantly
increase the network lifetime. With just one rotation round, we can almost double the network lifetime. With multiple rotation rounds, we can increase network lifetime by more than
a factor of seven.

6.2

Problem Definition

We consider WSNs consisting of many wireless mobile sensor nodes and a single static
sink. The sensor nodes gather data from their surroundings and transmit the data through
one or multiple hops to the sink forming a directed routing tree. We divide time into
intervals. In each interval, each sensor node transmits the data it gathered as well as the
data it received from its children to its parent along the routing tree. The goal is to maximize
the lifetime of the WSN, i.e. the number of time intervals until the first node dies. We use
this definition of lifetime assuming that all nodes are needed in their exact positions in order
to not compromise the operation of the system.
One of the main reasons limiting the lifetime of such networks is differential power consumption; nodes closer to the sink usually transmit a large amount of data and consequently
consume more power than nodes further away from the sink. As a result, the lifetime of the
WSN is substantially reduced even though the WSN still contains nodes rich in energy. Mobile base stations mitigate differential power consumption by having a powerful mobile sink
move around the WSN. We propose a new solution, mobile node rotation, that uses multiple
low-cost mobile nodes that rotate or swap positions and roles allowing nodes to share the

92

burden of high consumption locations and the benefits of low consumption locations. We
formally define the problem as follows.
Definition 2 (One-Round Max-Lifetime Node Rotation (1-MaxLife)).
Input Instance:
• S = (s1 , . . . , sn ), a list of sensor nodes
• u, the network sink, and pu , its position
• P = (p1 , . . . , pn ), a list of positions such that node si starts at position pi
• E = (e1 , . . . , en ), a list of initial energies for nodes in S
• T , a directed routing tree represented as a set of non-zero values tij for every arc
(pi , pj ) in the tree corresponding to the amount of energy consumed when transmitting
one data unit from pi to pj
• K, a set of values kij for every pair of positions pi and pj , corresponding to the amount
of energy consumed by a node when it moves between pi and pj
• Λ = (λ1 , . . . , λn ), the amount of data gathered at each position per time interval
Output Instance: A matching M of nodes in S to locations in P , and two durations r1 and
r2 such that nodes transmit data from their original positions for r1 time intervals, relocate
to their new position according to M , then generate and transmit data for r2 time intervals
such that the total duration (r1 + r2 ) is maximized and no node’s energy goes to 0 before
r1 + r2 .
We define the Max-Lifetime Node Rotation (MaxLife) problem to be the general
version where nodes can switch positions any number of times.

93

6.3

Node Rotation Algorithms

In this section, we first present our centralized node rotation algorithm NR1 to the 1MaxLife problem. For comparison purposes, we present CNR, an extended version of NR1
that applies to the general MaxLife problem. Finally, we present a practical distributed
algorithm DNR for the general MaxLife problem that uses only local information and reduces
the number of node movements. All algorithms begin by having nodes compute the load
lj at each position pj in P which is the total energy consumed in transmitting all the data
gathered in one time interval from the subtree rooted at pj to its parent. More formally,
lj = tjq

λi where T (pj ) is the subtree rooted at pj and pq is the position of the parent
i∈T (pj )

of pj in the tree.

6.3.1

Algorithm NR1

NR1 transforms the input instance into an instance of the assignment problem [8] (outlined in Section 4.2.2). We first assume that we know the optimal length of the first time
interval r1 for which nodes remain at their initial positions. To compute the optimal matching M of sensor nodes to positions, we transform the instance I into an instance of the
maximum bottleneck assignment problem I for the given r1 as follows. Each mobile node si
corresponds to a person and each location pj in P corresponds to a task. The efficiency cij
of a person si performing task pj corresponds to the total lifetime of si after transmitting
for a period r1 from its original position pi , then moving to pj where it transmits until its
energy is depleted; that is, cij = r1 +

e−li r1 −kij
.
lj

The optimal solution for the maximum

bottleneck assignment instance I corresponds to an optimal matching M for the given r1 .
We now need to compute the best duration r1 such that the total duration of trans94

procedure NR1(S, P, F, E)
Initialize empty priority queue
√
GOLDEN RATIO ← 3+2 5 ;
rlo ← 0;
rhi ← StaticLifetime(S, P, F, E);
Run golden ratio search for best r
while rhi - rlo ≥ error do
ra ← rlo + GOLDEN RATIO*(rhi - rlo);
rb ← rlo - ra + rhi;
(La, Π) ← optimalAssignmentValue(S, P, E, F, ra);
(Lb, Π) ← optimalAssignmentValue(S, P, E, F, rb);
if Lb ≤ La then
rhi ← rb;
else
rlo ← ra;
end if
end while
return (ra, La, Π);
end procedure

procedure optimalAssignmentValue(S, P, E, F, r)
Build efficiency matrix
for all si ∈ S do
for all pj ∈ P do
if si .location = pj then
cij ← ei /lj
else
if ei − li r − kdij > 0 then
cij ← (1 − li /lj )r + (ei − kdij )/lj
else
cij ← 0
end if
end if
end for
end for
Find optimal matching
(L, Π) ← SolveBottleneckAssignment(c)
return (L, Π)
end procedure

Figure 6.1: Algorithm NR1 for computing the optimal lifetime through a single round of
rotations

95

missions (r1 + r2) is maximized. We observe that as r1 increases, r2 decreases since there
is less energy for the second round. So we can express r2 as a function L2 (r1 ). We can
then define the total network lifetime as a function L(r1 ) = r1 + L2 (r1 ). We use golden
ratio search to find the best r1 . When L(r1 ) is unimodal, the golden ratio search yields
an optimal r1 that maximizes L(r1 ). To start the golden ratio search algorithm, we first
compute L(I) = minn ej /lj which is an upper bound on r1 . Our algorithm NR1 runs in
j=1
O(log L(I)n2.5 ) time because the golden ratio search has O(log L(I)) time complexity and
each assignment problem has O(n2.5 ) time complexity. Algorithm NR1 is shown in Fig. 6.1.

6.3.2

Centralized Algorithm CNR(r)

CNR(r) extends NR1 as follows. First, a node is selected to be the controller. The
controller collects energy and location information from all the other nodes once. CNR(r)
then proceeds in rounds, each of which has a common duration r. At the beginning of each
round, the controller computes matchings of nodes to positions for the following round using
duration r as each round length until the first node dies. We note that the controller has
enough information to estimate each node’s available energy for each future round. The
controller then broadcasts the number of rounds and all the matchings to the other nodes.
Finally, the nodes carry out the rounds synchronizing as needed to initiate each rotation.
We present CNR(r) solely for benchmarking purposes. We show CNR in Fig 6.2.

6.3.3

Distributed Algorithm DNR

Our practical multiple rotation round algorithm is DNR(r, h, lcr , f ) which requires only
local information. We typically drop the four parameters and refer to this algorithm as
DNR. The main idea is that only critical nodes, i.e. nodes at locations with a high power
96

procedure CNRRound(r)
if controller then
Compute expected energy at all sensors
Lr ← compute remaining lifetime
(L, Π) ← optimalAssignmentValue(S, P, E, F, r)
if Lr > r then
Broadcast new positions
Run for duration r
Relocate to new position according to Π
else
Run for duration Lr
end if
else
Receive new position
Run for duration r
Relocate to new position according to Π
end if
end procedure

Figure 6.2: Algorithm executed at the beginning of each round in a centralized setup
consumption rate, relocate to lower consumption positions. Each critical node tries to find
a low power consumption rate node within its neighborhood to swap positions with. As
with CNR(r), all rounds will have duration r. In each round, each node si with li > lcr
is a critical node. Critical nodes collect the position, load, and current energy level from
descendants that are at most h hops away and have not committed to switch with other
critical nodes. For each candidate node c, critical node s computes L1(c), the minimum
of s’s and c’s expected lifetime without swapping, and L2(c), the minimum of s’s and c’s
expected lifetime if they swap positions. Critical node s selects candidate c∗ with maximum
value L2(c). The swap is performed if and only if L2(c)/L1(c) ≥ f . This is to eliminate
switches that increase the network lifetime by negligible amounts. At this point c∗ commits
to switch with s. Figure 6.3 shows the algorithm executed by each node in each round.

97

procedure DNRRound(r, h, lcr , f )
if isCriticalNode then
Compute current lifetime
Compute minimum target lifetime
Send REQUEST INFO to children
end if
status = uncommitted
Collect information from descendants for a period of time
repeat
Receive data from sender
if data.type = STATUS CHECK and data.target = self then
sender status to sender
else if data.type = SELECT and data.target = self then
status = committed
Perform switch with sender
break
else if data.type = INFO and isCriticalNode then
If a critical node, record best candidate based on info received
Compute expected lifetime if switched with data.candidate
Update current bestCandidate
if data.hops > 0 then
data.hops = data.hops - 1
Forward data to parent
end if
else if data.type = REQUEST INFO then
Send INFO to parent
if data.hops > 0 then
data.hops = data.hops - 1
Forward data to children
end if
end if
until timeout
Select best candidate, and switch
if isCriticalNode then
Send(bestCandidate, STATUS CHECK)
Receive(bestCandidate, candidateStatus)
if ( thencandidateStatus = uncommitted)
Send(bestCandidate, SELECT)
Perform switch with bestCandidate
end if
end if
Run for duration r
end procedure

Figure 6.3: Local algorithm executed at the beginning of each round in a distributed setup

98

6.4

Upper Bounds on Lifetime Improvement Ratio

We now prove some upper bounds on lifetime improvement ratios for any node rotation
algorithms. We first consider the 1-MaxLife problem. We then consider the general MaxLife
problem. We use the following notation in our analysis. For any node rotation algorithm A
and input instance I, let L(A, I) denote the lifetime achieved using algorithm A on I, L(I) the
lifetime without node rotation, and RA (I) = minI L(A, I)/L(I) the lifetime improvement
ratio (LIR) of A on I. Finally, let EV (I) = maxn ei / minn ej be the initial energy
i=1
j=1
variance of I.
Theorem 2. For any one round node rotation algorithm A and any input instance I RA (I) ≤
1 + EV (I).
Proof. Let si denote the bottleneck node in I that determines L(I); that is, L(I) = ei /li .
Clearly, r1 ≤ L(I). We now observe that r2 ≤ EV (I) · L(I) because the node that moves
to position pi can have at most energy EV (I) · ei . Thus, r1 + r2 ≤ (1 + EV (I))L(I) and
RA (I) ≤ 1 + EV (I).
This leads to two corollaries, one for the special case where all nodes have the same initial
energy and one for multiple round rotation algorithms.
Corollary 1. For any one round node rotation algorithm A and any input instance I with
EV (I) = 1, RA (I) ≤ 2.
Corollary 2. For any j round node rotation algorithm A and any input instance I, RA (I) ≤
1 + (j − 1)EV (I).
We note that although our one round solution NR1 is optimal for 1-MaxLife only when
L(r1 ) is unimodular, our simulations (Section 5.6) show that NR1’s LIR is usually very close

99

⊗ sink

Nodes
Level in level
0

h

dj

1

1

Size of
subtree

d

j=0

h−i

i

dj

di
j=0

h−h

h

dj = 1

dh
j=0

Figure 6.4: Balanced tree topology of degree d + 1 and lowest level h
to the upper bound of 2 for input instances I with EV (I) = 1. Moreover, for input instances
I where EV (I) > 1, NR1’s LIR is often better than 2.
We now prove upper bounds on the RA (I) for inputs I corresponding to balanced trees
of degree d + 1 shown in Fig. 6.4.
Theorem 3. For any node rotation algorithm A and any input I where T represents a
balanced tree of degree d + 1 and tij = t for all non-zero tij , h is the lowest level of the tree
where the root is at level 0, and for 1 ≤ i ≤ n, ei = e and λi = 1, then

RA (I) ≤

(dh+1 − 1)2
(h(d − 1) + d − 2)dh+1 + 1

Proof. To prove this upper bound, we ignore the energy consumed by movement. The best
solution is to then have all nodes equally share the transmission of all data. We first compute
L(I) which is constrained by the bottleneck node at level 0 that transmits data from the
entire tree to the sink. The total number of nodes n =

100

h dj
j=0

as there are dj nodes at

level j. It follows that
L(I) =

e
d−1 e
= h+1
nt
d
−1t

We next compute the lifetime L∗ of the WSN if we are able to perfectly share the transmission
of all data among all n sensor nodes. A node at level i is the root of a subtree of size
Di =

h−i j
j=0 d .

Thus, the total amount of data transmitted each time interval is

h

D=

di D i =

i=0

h
i=0

di

h−i

dj =

j=0

dh+1 (h(d − 1) + d − 2) + 1
(d − 1)2

This implies
L∗ =

ne
(dh+1 )(d − 1)
e
=
Dt
(h(d − 1) + d − 2)dh+1 + 1 t

Dividing L∗ by L(I) gives us the result.
Table 6.1 displays some of the upper bounds for different values of d and h. The improvement ratio starts at a factor of 1.8 for a tree with 3 nodes and rapidly increases with
both the degree and the level.
d/h
2
3
4
5

1
1.80
2.29
2.78
3.27

2
3
2.88 4.59
4.97 11.27
7.74 23.08
11.17 41.53

4
5
7.45 12.36
26.77 66.08
72.99 240.82
164.37 679.26

Table 6.1: Upper Bounds on Lifetime Improvement Ratios in Balanced Trees

6.5

Simulation Results

In this section, we evaluate the performance of our NR1 and DNR algorithms through
simulations. For comparison purposes, we also evaluate the performance of the baseline

101

CNR algorithm. We generated 100 networks each consisting of 100 nodes placed uniformly
at random in a 150m by 150m area with the sink node chosen uniformly at random. We
set the maximum communication distance to 35m, which was shown in [45] to lead to a
high packet reception ratio for TelosB motes in outdoor environments. For each network, we
constructed the routing tree from the sources to the sink using greedy geographic routing in
which each node forwards its data to the neighbor that is closest to the sink. We note that
our algorithms are applicable to network topologies generated by any routing algorithms.
We set the tij and the kij values based on the energy models in [45, 13] since they are based
on realistic platforms; any other energy model for communication or mobility could be used
without any algorithmic change. We usually set each node’s ei to the same value typically
ranging from half full to full, though in some simulations different nodes have different ei
values. For our distributed algorithm DNR, we set the local improvement factor f to 1.25
and the critical consumption rate threshold lcr to 10% of the range of ratios in the network.
We describe our choices for r and h below. We assess the performance of algorithms using
several criteria. The main criteria is lifetime improvement ratio. We also assess the number
of rounds required, the number of nodes that move per round, and the number of movements
each node makes over the network lifetime.

6.5.1

Single Rotation Round

We first evaluate the performance of NR1 for the 1-MaxLife problem in trees. For all
100 inputs, 1.91 ≤ RN R1 (I) ≤ 1.99, and the average value of RN R1 (I) = 1.95. We observe
that the results are very close to the theoretical upper bound of 2 from Corollary 1. When
we varied the starting energy level of the nodes between half full and completely full, the

102

average lifetime improvement ratio of NR1 increased to 2.3. We also observe that most of
the nodes change their positions; on average 86 nodes relocate to new positions and 14 nodes
remain at their original location. This is not too expensive since only a single rotation is
performed which means the network’s activity is interrupted only once.

6.5.2

Multiple Rotation Rounds

Round Duration
We now evaluate the performance of our DNR algorithm with h = 2 for the general
MaxLife problem using our CNR(r) algorithm as a baseline. We first study the effect of
varying r on the performance of both algorithms. Figure 6.5 shows the average lifetime
improvement ratios for both algorithms as we increase r denoted as a fraction of the static
lifetime L(I). We see that DNR outperforms CNR(r) for all values of r but especially smaller
r. For r ≤ 3L(I)/5, RDN R (I) ≥ 2.5 + RCN R (I). At r = L(I)/2, the difference in lifetime
improvement ratio is 2.5. One notable feature of DNR’s performance is that the lifetime
improvement ratio is relatively flat for L(I)/5 ≤ r ≤ 3L(I)/5; it takes on a maximum value
of 7.4 at r = L(I)/2 and decreases to 7.1 at r = 3L(I)/5. When r is too large, the LIR of
both algorithms drops because nodes stay at high consumption positions for too long.
For most of the remaining simulations, we set r = L(I)/2. For DNR, this almost maximizes lifetime improvement ratio while maximizing r which minimizes the number of disruptions to the network. For our comparison algorithm CNR(r), this choice maximizes the
lifetime improvement ratio.

103

CNR

DNR

CNR

Number of nodes relocating

Lifetime Improvement Ratio

800%
700%
600%
500%
400%
300%
200%
100%
0%
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

DNR

100
90
80
70
60
50
40
30
20
10
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Round Length

Round Length

Figure 6.5: Average lifetime improvement
ratios of CNR(r) and DNR as a function of
r plotted as a fraction of L(I)

Figure 6.6: Average number of relocations of
CNR(r) and DNR as a function of r plotted
as a fraction of L(I).

Node Relocations
We next compare the two algorithms with respect to how much the sensor nodes move.
We first consider the number of relocations per round. As we can see from Fig. 6.6, DNR
outperforms CNR again. In particular, with CNR, more than 90% of the nodes relocate in
each round whereas with DNR, the average number of relocating nodes stays below 16%.
These nodes are usually different in each round as 96% of the nodes move between 0 and 3
times overall and the remaining 4% move up to 6 times during their lifetime. For DNR, most
of the energy available at each node is consumed by communication rather than movement:
85% of the nodes spend at least 80% of their energy on communication and no node spends
more than 35% of its energy on movement. We see observe that node relocations increase
only slightly for DNR as round duration r increases because more nodes become critical in
each round due to more time at high consumption locations.

104

Distributed Approach and Hop Distance Parameter h
We now analyze the effect of the hop distance parameter h on the performance of DNR.
We plot the complementary cumulative distribution function (CCDF) of the lifetime improvement ratio for DNR with h set to 1, 2, and 4 in Fig. 6.7. The CCDF gives us the
probability that the lifetime improvement ratio exceeds a given threshold. For comparison,
we also plot the CCDF of CNR. From this data, we see that setting h = 2 is sufficient to
achieve excellent performance as the CCDF for h = 2 is almost identical to that of h = 4.
In both cases, 93% of the topologies have lifetime improvement ratios of at least 400% and
more than 50% of topologies have lifetime improvement ratios over 700%. With h = 1,
DNR is much less effective; the number of nodes taking turns transmitting from the high
consumption position may be too low. In the remaining simulations, we set h = 2 for DNR.
Lifetime Improvement Ratio Increase per Round
We now assess how much effect each round has on the lifetime improvement ratio. Specifically, if we stop node rotations after round n, what will the lifetime be? For both CNR
and DNR, we obtain a lifetime improvement ratio that is essentially linear in the number of
rounds with each round increasing the lifetime improvement ratio by between 40 and 50%.
This analysis shows that both algorithms are effective in increasing the lifetime improvement
ratio but that DNR is more effective in minimizing distance moved and maintaining a reserve
of energy rich nodes for later rounds. This is why DNR outperforms CNR which moves 93%
of the nodes in each round.

105

DNR h=2
DNR h=4

vs Energy Aware LEACH
vs Multi-Hop LEACH
Lifetime Improvement Ratio

Lifetime Improvement Ratio

CNR
DNR h=1
1400%
1200%
1000%
800%
600%
400%
200%
0%
0

25

50

75

600%
500%
400%
300%
200%
100%
0%
0

100

Figure 6.7: CCDF of lifetime improvement
ratio of CNR and DNR with r = L(I)/2 and
h = 1, 2, and 4.

25

50

75

100

Figure 6.8: CCDF of the lifetime comparison
ratios of DNR versus LEACH and multihop
LEACH.

Comparison with Existing Approaches
In this section, we compare DNR to existing approaches. We consider only approaches
that do not change the positions of where nodes are placed. This rules out existing mobile
relay approaches and leaves us with non-mobility approaches like LEACH [19] that rotate
the roles of different nodes by periodically changing the topology of the network but not
modifying any node positions. Specifically, we compare DNR with LEACH [19] as it serves
as the base for several other clustering algorithms that seek to increase network lifetime
and (2) multihop LEACH [33], an improved variation that uses multihop transmissions
between cluster heads. Both LEACH approaches assume that data is compressed before
being transmitted while DNR does not. To compare all approaches in a similar setting, we
run them all without data compression.
We compare DNR to LEACH and multihop LEACH on an input instance I by comput-

106

Lifetime Improvement Ratio

vs Energy Aware LEACH
vs Multi-Hop LEACH
400%
350%
300%
250%
200%
150%
100%
50%
0%
60

80

100 120 140 160 180 200

Number of Nodes

Figure 6.9: Average improvement ratio of our distributed approach DNR with respect to
energy aware LEACH and multi-hop LEACH as the number of nodes in the network increases
ing the lifetime comparison ratio RDN R (I)/RA (I) where A is either LEACH or multihop
LEACH. Figure 6.8 shows the complementary cumulative distribution function of both lifetime comparison ratios. First, we note that DNR outperforms both LEACH variations for
every topology, attaining lifetimes between 2 and 5.75 times better than LEACH and between 1.4 and 3.7 times better than multihop LEACH. Additionally, we observe that DNR
needs many fewer rounds than both LEACH variations. On average, DNR needs 8 rounds of
rotations whereas the average number of rounds is 1800 for LEACH and 2000 for multihop
LEACH. We also note that the round duration r used for the LEACH approaches was 20%
of the r used by DNR as using the same r resulted in much lower lifetime improvements
ratios for LEACH.
We now compare the performance of all three approaches as the density of the network
varies. We varied the number of nodes between 60 and 200 in increments of 10. We generated
a new set of topologies in which we varied the number of nodes between 60 and 200 in incre-

107

ments of 10. For each size, we randomly generated 75 networks. We observe in Fig. 6.9 that
the average improvement ratio of our distributed approach with respect to both LEACH
variations decreases slowly as the density of the network increases. Compared to energy
aware LEACH, our distributed approach yields lifetimes between 3.3 and 4.0 times better.
Compared to multihop LEACH, our approach maintains an average improvement ratio between 2 and 2.3 times better. We also observe that as the density of the network increases,
the number of rounds performed increases significantly for both LEACH approaches (from
1600 to 2200 rounds in energy aware LEACH and from 1600 to 3000 rounds in multihop
LEACH) whereas DNR requires on average only two more rounds.

6.6

Summary

In this chapter, we consider the problem of maximizing the lifetime of mobile WSNs. We
exploit the mobility of nodes to mitigate differential power consumption by having nodes take
turns in high power consumption positions without modifying the existing communication
topology. We present efficient algorithms for both the single round and the general multiple
round MaxLife problem. This approach is very different than other schemes such as data
mules in that all nodes expend relatively little energy on movement and move only a few
times during the network lifetime. Our simulations show that our algorithms can improve
average lifetime by more than a factor of 7 and that our algorithms outperform existing
non-mobility approaches for mitigating differential power consumption to prolong network
lifetime.

108

CHAPTER 7
Conclusion
In this thesis, we leverage node mobility to improve the energy efficiency of wireless sensor
networks. We propose several approaches for optimizing different performance metrics. In
this chapter, we first summarize our proposed approaches. Then, we discuss some possible
open problems.

7.1

Summary

We propose to use low-cost disposable mobile nodes to reduce the energy consumption
in wireless sensor networks and increase the networks lifetime. Our proposed approaches are
holistic as they take into consideration the energy consumed by both mobility of relays and
wireless transmissions and receptions. Most previous work ignored the energy consumed by
moving mobile relays. Unlike most previous work, for each problem we consider, we carefully
compute the optimal target location of each node. Relocating to the optimal position showed
its superiority over existing approaches that use an approximate location.
We first consider a new problem, MMRC, maximizing the data gathering capacity of

109

hybrid wireless sensor networks consisting of mainly static nodes and a few mobile nodes. We
study how to exploit a small number of mobile nodes to improve the data gathering capacity
of the network. We examine two cases in particular, when one and then two mobile nodes are
allowed to join a single link. We present optimal and approximate solutions to four variants
of MMRC. In most variants (tree topologies), our algorithms improved the system lifetime
by a factor of 2 or more. For the star topology, it increased system lifetime by a factor
of 1.5. Our distributed protocols are almost as effective at improving the data gathering
capacity, particularly as the number of mobile relay nodes increases. Our simulations also
show that these protocols quickly converge on a final solution with little messaging overhead.
Moreover, our simulations show that it is enough to consider a single mobile relay per link
to get the highest improvements.
Second, we study the problem of minimizing the total energy consumption in a wireless
network. For this problem, we consider networks entirely consisting of mobile nodes and
utilize the mobility of all nodes to reduce the energy consumption. We design a framework for
minimizing the total energy consumption in a sensor network consisting in both constructing
low consumption transmission routes and repositioning the nodes to their optimal positions.
We first identify a tree construction strategy that results in an optimal configuration in a
static environment. Then we develop an algorithm that uses this optimal configuration and
iteratively refines it by inserting new relay nodes into the tree. Moreover, we develop a
second iterative algorithm that improves the given configuration by relocating nodes to new
positions. We show that this algorithm converges to the optimal configuration for the given
transmission routes. Our simulations show that our algorithm can reduce the total energy
consumption by an average of 45%.
110

Finally, we consider the problem of maximizing the network lifetime. Again we exploit
the mobility of all nodes to balance the energy consumption across the network. We propose
to relocate nodes rich in energy to locations with higher consumption rates and conversely
relocate nodes low in energy to locations with low consumption rates. We address this
problem for single and multiple rounds of rotations. We show that a single rotation improves
the lifetime by a factor of almost 2. Our simulations show that our algorithms for multiple
rounds of rotations can improve the average lifetime by more than a factor of 7 and that they
outperform existing non-mobility approaches for mitigating differential power consumption
to prolong network lifetime in both overhead and lifetimes achieved.

7.2

Open problems

In our solutions for maximizing the data gathering capacity, we assumed that the routing
tree connecting the static nodes to the sink is predetermined and we proposed an insertion
approach to improve the given tree. Likewise, in our solutions for the lifetime maximization
problem, we maintained the same tree throughout the process of node rotation. These
solutions significantly improve the performance of the given tree. However, in some cases,
the initial routing structure may not be defined or the application may not dictate any route
constraints. These cases may benefit from joint optimization of the routing structure and
the nodes’ locations to fully exploit mobility of nodes.
Throughout this study, we adopted one particular definition of network lifetime which is
the time until the first node dies. This definition coincides with the lifetime of a configuration
as the death of a single node entails the need for a new configuration. However, some
applications may continue to function properly even after a small fraction of the nodes fails.

111

It is interesting to extend our solutions to such applications where the lifetime is defined
by several criteria such as the fraction of the network alive and some minimum coverage
requirements.

112

APPENDICES

113

APPENDIX A
Proofs of Claims and Theorems

Claim 1. gr (α) is unimodal
Proof. Recall that the function gr (α) is defined as the value of the maximum message that
can be transmitted by r when it moves to posr (α). We define and mα (t) as the maximum
r
message r can transmit to s2 when it moves to point t along rα. We also define γ to be the
angle between rα and rs2 and δ to be the distance traveled by r to get to point t as shown
in Figure A.1. Since both γ and δ uniquely determine the point t, we redefine mα (t) as a
r
function of δ as follows:
Mγ (δ) =
=

er − kδ
a + b((ts2 )2
er − kδ
a + b(δ 2 + d2 − 2δd2 cos(γ))
2

We define δl to be the maximum distance r can travel before it consumes all its energy by

114

r
γ
d1

δ

d2
t

s1

α

s2

Figure A.1: gr (α) is unimodal
et
moving. We have δl = . We compute the derivative of Mγ with respect to δ. We get:
k
(er − kδ)b(2δ
k
2 − 2δd cos(γ)) − (a + b ∗ (δ 2 + d2
a + b(δ 2 + d2
2
2
2 − 2e bδ − k(a + bd2 ) + 2e bd cos(γ)
kbδ
r 2
r
2
2 + d2 − 2d δ cos(γ)))2
(a + b(δ
2
2

Mγ (δ) = −
=

− 2d2 cos(γ))
− 2δd2 cos(γ)))2

So Mγ (δ) has the same sign as E = kbδ 2 − 2er bδ − k(a + bd2 ) + 2er bd2 cos(γ).
2
This is a second degree polynomial, we study its variation as follows. Let ∆ be its determinant.
∆ = (−2er b)2 − 4(kb)(−k(a + bd2 ) + 2er bd2 cos(γ))
2
= 4e2 b2 + 4k 2 b(a + bd2 ) − 8kb2 er d2 cos(γ)
r
2
∆>0
e2 b + k 2 (a + bd2 )
2
⇔ cos(γ) < r
2kbd2 er
e 2 + k 2 d2
2 + ka
⇔ cos(γ) < r
2kd2 er
2bd2 er
ka
e2 + k 2 d2
2 > 1 and
But r
> 0 ∀ er , k, d2 > 0
2kd2 er
2bd2 er
So ∆ > 0 ⇔ cos(γ) < 1 + v + , for some v + > 0 and consequently
∆ > 0 ∀ er , k, d2 , a, b > 0, 0 < γ < π

115

(A.1)

The roots of E are:
√
er
∆
=
−
k
2kb
√
er
∆
δ+ =
+
> δl
k
2kb

δ−

(A.2)

In this case, the function Mγ varies as follows with δ:
δ−

−∞
δ
−)
(δ − δ
−
(δ − δ + )
−
− )(δ − δ + )
kb(δ − δ
+
Mγ

δ+
+
−
−

+∞
+
+
+

So the maximum of Mγ is at δ − if δ >= 0 and at 0 otherwise. We now study how the
γ
maximum at each δ − varies has γ varies between 0 and π. We define this function gr as:
γ
gr (γ)

er − kδ − (γ)
=
a + b(δ − (γ)2 + d2 − 2d2 δ − (γ) cos(γ))
2
=

k 2 e2 b2 + k 2 ab + k 2 b2 d2 − 2v1
r
2
v2

where

v1 = kber d2 cos(γ)
v2 = e2 b2 − er b e2 b2 + k 2 ab + k 2 b2 d2 − 2v1 + k 2 ab + k 2 b2 d2 − 2v1
r
r
2
2
+ d2 cos(γ)k e2 b2 + k 2 ab + k 2 b2 d2 − 2v1
r
2
γ

We compute the derivative gr (γ) which corresponds to:
γ
gr (γ) =

2d2 T
k B2

116

with
T = − sin(γ)(e2 b2 + k 2 ab + k 2 b2 d2 − 2v)(er b −
r
2

e2 b2 + k 2 ab + k 2 b2 d2 − 2v)
r
2
γ

γ

So gr (γ) has the same sign as T . We analyze the sign of gr (γ) for 0 ≤ γ ≤ π as follows.
We observe that T is the product of three factors:
1. sin(γ) > 0,

∀ 0≤γ≤π

er b + k 2 a + k 2 bd2
2
2. e2 b2 + k 2 ab + k 2 b2 d2 − 2v > 0 ⇔ cos(γ) ≤
r
2
2kber d2
er b + k 2 a + k 2 bd2
2 < 1 ∀ e ≥ 0, b ≥ 0, k ≥ 0, a ≥ 0
but we can show that
r
2kber d2
so e2 b2 + k 2 ab + k 2 b2 d2 − 2v > 0 ∀ 0 ≤ γ ≤ π
r
2
3. er b −

e2 b2 + k 2 ab + k 2 b2 d2 − 2v : this becomes 0 when cos(γ) =
r
2

k(a + bd2 )
2 . We
d2 ber

distinguish two cases:
• Case 1:
2
k(a + bd2 )
2 < 1 ⇔ e2 > k(a + bd2 )
d2 ber
db

for γ0 = arccos(

k(a + bd2 ) 2 2
2 ), e b + k 2 ab + k 2 b2 d2 − 2v = 0
r
2
d2 ber

e2 b2 + k 2 ab + k 2 b2 d2 − 2v > 0 ∀ 0 < γ < γ0
r
2
e2 b2 + k 2 ab + k 2 b2 d2 − 2v < 0 ∀ γ0 < γ < π
r
2
• Case 2:
2
k(a + bd2 )
2 > 1 ⇔ e < k(a + bd2 )
r
d2 ber
sb

⇒ cos(γ) <

k(a + bd2 )
2
d2 ber

∀ 0≤γ≤π

e2 b2 + k 2 ab + k 2 b2 d2 − 2v < 0 ∀ 0 < γ < π
r
2

117

γ

The following tables represent the variation of function gr (γ) for 0 < γ < π:
Case 1
δ
-1
sin(γ)
e2 b2 + k 2 ab + k 2 b2 d2 − 2v
r
2
2 b2 + k 2 ab + k 2 b2 d2 − 2v
er b − er
2

0

∞

γ0
−
+
−
+

−

γ

gr (γ)
γ
gr (γ)

−
+
+
+

+

Case 2
δ
-1
sin(γ)
2 b2 + k 2 ab + k 2 b2 d2 − 2v
er
2
er b − e2 b2 + k 2 ab + k 2 b2 d2 − 2v
r
2
γ

∞

0
−
+
−
+
−

gr (γ)
γ
gr (γ)
γ

As γ increases, α decreases. So when gr (γ) is increasing, gr (α) is decreasing and conversely,
γ

when when gr (γ) is decreasing, gr (α) is increasing. Therefore the function gr (α) is unimodal.

Claim 2. hb (α) is decreasing between s1 and s2 .
Proof. Given a direction rα, hb (α) is the largest message that s1 can transmit to r when it
moves along rα. This corresponds to r moving to the closest point on rα to s1 , which is the
intersection of the perpendicular line to rα taken from s1 . We call this point hα . We now
show that the distance between s1 and rα increases as α increases. We consider two cases:
when x1 ≤ xr ≤ x2 and when xr ≤ x1 ≤ x2 .

118

r

s1

H1

s2

H2

I1

I2

I3

Figure A.2: Subdivision of s1 s2 into 3 intervals.
Case I: x1 ≤ xr ≤ x2
Let H1 be the intersection of s1 s2 and the perpendicular to it from r, and H2 the intersection
between s1 s2 and the perpendicular to s1 r from r. We divide the segment s1 s2 into three
intervals: I1 = [s1 , H1 ], I2 = [H1 , H2 ], I3 = [H2 , s2 ] as shown by Figure A.2.
For any point α ∈ I1 , hα falls outside the triangle s1 rs2 . Since we only consider points
inside the triangle, the point maximizing hb (α) is α. The distance s1 α increases as α moves
towards H1 . So hb (α) is decreasing in I1 .

(hIII )
α
r
h2

h1
s1

s2
hI
α

Figure A.3: In I1 , hα is on the line s1 sα . In I2 , s1 hα increases with sα . In I3 , hα coincides
with r.

Consider two points sα1 ∈ I2 and sα2 ∈ I2 such that xα1 < xα2 as shown in Figure A.3.
119

Let s1 hα1 be the perpendicular segment from s1 to rsα1 . Let C be the circle of center s1
and radius s1 hα1 . The line rhα1 is tangent to C from r. Since α1 < α2 , the line rhα2 does
not intersect C so s1 hα2 > radius(C) and s1 hα2 > s1 hα1 . So hb (α) is decreasing in interval
I2 .
For any point α ∈ I3 , hα falls outside the triangle s1 rs2 . In this case, the point maximizing
hb (α) is s2 . The distance s1 r remains constant as α moves from H2 towards s3 .

Case II: xr ≤ x1 ≤ x2
The same argument as case I, interval I2 applies in this case and the optimal transmission
distance for s1 increases with α. So hb (α) is decreasing between s1 s2 .

Claim 3. f (α) is unimodal
C1 (b)

r
CM (a)
b

C1 (a)

b′
a

c

′

s1

s2
αa

αc

αb

Figure A.4: Any point c between a and b such that m(a) > m(b) satisfies the condition
m(c) > m(b)

Proof. Consider directions along three points αa < αc < αb such that M (αa ) = f (αa ) >
f (αb ) = M (αb ). We consider the case when αa ∈ I1 but the remaining cases (for αa ∈ I2
and αa ∈ I3 ) follow similar geometric arguments.
Let a be the point of segment rαa where f (αa ) reaches its maximum and b the point along

120

segment rαb where f (αb ) reaches its maximum. We define C1 (a) to be the circle of center s1
and radius s1 a, C1 (b) the circle of center s1 and radius s1 b, and CM (a) the circle of center
r and radius ra.
We now show that if M (αc ) = f (αc ) for some point c on rαc then f (αc ) > f (αb ) and
consequently M (αa ) > M (αc ) > M (αb )
First, we show the position of point b relative to point a. We define the point b to be the
intersection of circle CM (a) with segment rαb . At point b , r traveled the same distance as
at point a but b s2 < as2 , so
m2 (b ) > m2 (a)

(A.3)

But since b is the optimal point on the segment rαb , f (αb ) > min{m1 (b ), m2 (b )}. On the
other hand, f (αa ) > f (αb ), therefore
f (αa ) > min{m1 (b ), m2 (b )}
This along with (A.3) ⇒
f (αa ) > m1 (b )
So m1 (b ) = min{m1 (b ), m2 (b )}. But since f (αb ) > m1 (b ) then m1 (b) > m1 (b ). So the
point b must be within the circle of center s1 and radius s1 b as shown in Figure A.4.
We now show that the function f never increases after it starts decreasing. Let c be the
intersection of segment rαc with circle CM (a). Point c is on the circle CM (a) and is closer
to s2 than a is, so:
m2 (c ) > m2 (a) > m2 (b)

121

Besides, c is on the inside of circle C1 (b) so:
m1 (c ) > m1 (b)
⇒ min{m1 (c ), m2 (c )} > m1 (b) = m2 (b)
But

f (αc ) > min{m1 (c ), m2 (c )}

since c is the optimal point on rαc

⇒ f (αc ) > f (αb )

Claim 4. gb (α) is continuous
Proof. We prove that the function gb (α) is continuous by showing it is differentiable for any
α in s1 s2 . First, we define α as the angle s1 rα and θ as the angle s1 rs2 . We note that θ is
ˆ
a constant. We then redefine gb (α) in terms of α as:
ˆ
gb (ˆ ) = e1 /(a + bu(ˆ )2 )
α
α
with u(ˆ ) being the transmission distance for s1 to r at its new position (i.e. the distance
α
s1 t in Figure A.1). We note that gb (α) and gb (ˆ ) are equal and gb (α) is continuous and
α
differentiable whenever gb (ˆ ) is. Consider triangle s1 rt, we have
α
u(ˆ )2 = δ 2 (γ) + d2 − 2δ(γ)d1 cos(ˆ )
α
α
1

122

So
gb (ˆ ) = e1 /(a + b(δ 2 (γ) + d2 − 2δ(γ)d1 cos(ˆ )))
α
α
1
= e1 /(a + b(δ 2 (θ − α) + d2 − 2δ(θ − α)d1 cos(ˆ )))
ˆ
ˆ
α
1
Using the value of δ(γ) in (A.2), and simplifying the result, we get:
α
gb (ˆ ) = e1 k 2 / 2k 2 a + 2e2 b − 2er b(e2 b + k 2 a + k 2 bd2 − 2kber d2 cos(ˆ ))+
α
r
r
2
k 2 bd2 − 2kber d2 cos(ˆ ) + d2 k 2 b − 2d1 cos(θ − α)kber +
α
ˆ
2
1
2d1 cos(θ − α)k b(e2 b + k 2 a + k 2 bd2 − 2kber d2 cos(ˆ ))
ˆ
α
r
2
and its derivative with respect to α is:
ˆ
√
er
∆
( −
)ber d2 sin(ˆ )
α
ber d2 sin(ˆ )d1 cos(θ − α)
α
ˆ
k
kb √
√
gb (ˆ ) = −e1 b − 2
α
+2
∆
∆
√
er
∆
+2( −
)d sin(θ − α)
ˆ
k
kb 1
/
√

√
2
er
∆ 2
2 − 2( er − ∆ )d cos(θ − α))
a + b(( −
) + d1
ˆ
1
k
kb
k
kb
The derivative gb (ˆ ) exists when the following two conditions hold:
α
1.

√

∆ = 0 which is true according to (A.1) in the proof of claim (1)
√
√
√
∆ 2
er
2 − 2( er − ∆ )d cos(θ − α)) = 0: but a > 0 and ( er − ∆ )2 +
2. a + b(( −
) + d1
ˆ
k
kb
k
kb 1
k
kb
√
er
∆
d2 − 2( −
)d cos(θ − α) = u > 0
ˆ
1
k
kb 1

So gb (ˆ ) is differentiable and consequently continuous and the same applies to gb (α).
α

123

Claim 5. hr (α) is continuous
Proof. We prove that the function hr (α) is continuous by showing it is differentiable. As
previously defined, α is the angle s1 rα, δ the distance traveled by r, and u the distance
ˆ
between s1 and r at its new position. We additionally define w as the distance between the
new position of r and s2 . We consider the three intervals I1 , I2 , I3 of Figure A.2.
Case I: α ∈ I1
We define β to be the angle between s1 r and rH1 , and h to be the distance rH1 .
δ=

h
cos(β − α)
ˆ

w2 = δ 2 + d2 − 2δd2 cos(γ)
2
=

h
h2
2
ˆ
2 + d2 − 2 cos(β − α) d2 cos(θ − α)
ˆ
cos(β − α)
ˆ

hr (α) = hr (ˆ ) =
α

er − kδ
a + bw2

h
cos(β − α)
ˆ
=
2
h
h
2
a + b(
ˆ
2 + d2 − 2 cos(β − α) d2 cos(θ − α))
ˆ
cos(β − α)
ˆ
er − k

And the derivative of hr (ˆ ) is:
α
kh sin(β − α)
ˆ

hr (ˆ ) =
α
cos(β − α)2 a + b(
ˆ

h2
cos(β

− α )2
ˆ

+ d2 − 2
2

h
d cos(θ − α))
ˆ
cos(β − α) 2
ˆ

kh
)
cos(β − α)
ˆ
−
∗
2
h2
h
2−2
a + b(
+ d2
d cos(θ − α))
ˆ
cos(β − α) 2
ˆ
cos(β − α)2
ˆ
b(er −

(−2

h2 sin(β − α)
ˆ
hd2 cos(θ − α) sin(β − α)
ˆ
ˆ
hd sin(θ − α)
ˆ
+2 2
)
3 +2
2
cos(β − α)
ˆ
cos(β − α)
ˆ
cos(β − α)
ˆ

The derivative exists when the following conditions hold:
1. cos(β − α) = 0
ˆ
124

2. a + b(

h2
h
2
ˆ
2 + d2 − 2 cos(β − α) d2 cos(θ − α)) = 0
ˆ
cos(β − α)
ˆ

But since s1 r and rH1 are not parallel, β < π and cos(β − α) > 0, ∀ 0 ≤ α < β. So the
ˆ
ˆ
2
first condition is satisfied. On the other hand,
a + b(

h2
h
+ d2 − 2
d cos(θ − α)) = a + bw2 > 0
ˆ
2
cos(β − α) 2
ˆ
cos(β − α)2
ˆ

So the second condition is also satisfied and consequently the derivative hr (ˆ ) exists in this
α
case and hr (ˆ ) is continuous and so is hr (α).
α

Case II: α ∈ I2
In this case,
δ = d1 cos(ˆ )
α
u = d1 sin(ˆ )
α
w2 = δ 2 + d2 − 2δd2 cos(γ)
2
= d2 cos(ˆ )2 + d2 − 2d1 d2 cos(ˆ ) cos(θ − α)
α
α
ˆ
1
2
hr (ˆ ) =
α
=

er − kδ
a + bw2
a + b(d2 cos(ˆ )2
α
1

er − kd1 cos(ˆ )
α
2 − 2d d cos(ˆ ) cos(θ − α))
+ d2
α
ˆ
1 2

The derivative of hr (ˆ ) is:
α
hr (ˆ ) =
α

kd1 sin(ˆ )
α
−
a + b(d2 cos(ˆ )2 + d2 − 2d1 d2 cos(ˆ ) cos(θ − α))
α
α
ˆ
1
2

(er − kd1 cos(ˆ ))b(−2d2 cos(ˆ ) sin(ˆ ) + 2d1 d2 sin(ˆ ) cos(θ − α) − 2d1 d2 cos(ˆ ) sin(θ − α))
α
α
α
α
ˆ
α
ˆ
1
a + b(d2 cos(ˆ )2 + d2 − 2d1 d2 cos(ˆ ) cos(θ − α))
α
α
ˆ
1
2

2

The derivative exists when a + b(d2 cos(ˆ )2 + d2 − 2d1 d2 cos(ˆ ) cos(θ − α)) = 0. But this is
α
α
ˆ
1
2

125

equal to a + bw2 > 0. So hr is differentiable and continuous.

Case III: α ∈ I3
In this case, posb (α) = r and the function hr (α) is constant and continuous.
As a result of all three cases, hr (α) is continuous in each interval. Also, the value of hr at
point H1 is the same in both intervals I1 and I2 and the value at point H2 is the same in I2
and I3 . So hr is continuous over the interval I1 ∪ I2 ∪ I3 , ie, for α between s1 and s2 .

Theorem 4. The optimal position for node si when it moves by itself is given by
xi = p i +

y i = qi +

2
2
−Bx ( Bx + By ± k)
2
2
A Bx + By
2
2
−By ( Bx + By ± k)
2
2
A Bx + By

where

A = 2b(

ml + mi )

sl ∈S(si )

Bx = 2b(

ml + mi )pi − 2b(

sl ∈S(si )

By = 2b(

ml xl + mi xd )

sl ∈S(si )

ml + mi )qi − 2b(

sl ∈S(si )

ml yl + mi yd )
sl ∈S(si )

126

Proof. We compute the derivatives

δCi
δCi
and
as follows.
dxi
dyi

aml + bml ui − ul 2

Ci =
sl ∈S(si )

+ ami + bmi ui − ud 2 + k ui − oi
aml + bml (xi − xl )2 + bml (yi − yl )2

=
sl ∈S(si )

+ ami + bmi (xi − xd )2 + bmi (yi − yd )2
+ k (xi − pi )2 + (yi − qi )2
δCi
=2b
ml (xi − xl ) + 2bmi (xi − xd )
dxi
sl ∈S(si )
k(xi − pi )

+

(xi − pi )2 + (yi − qi )2
ml + mi )xi − 2b(

=2b(
sl ∈S(si )

sl ∈S(si )

k(xi − pi )

+

(xi − pi )2 + (yi − qi )2
ml + mi )xi − 2b(

=2b(
sl ∈S(si )

− 2b(

ml xl + mi xd )

sl ∈S(si )

ml + mi )pi + 2b(
sl ∈S(si )

+

ml xl + mi xd )

ml + mi )pi

sl ∈S(si )

k(xi − pi )
(xi − pi )2 + (yi − qi )2

δCi
=2b(
ml + mi )(xi − pi )
dxi
sl ∈S(si )
ml + mi )pi − 2b(

+ 2b(
sl ∈S(si )

+

sl ∈S(si )

k(xi − pi )
(xi − pi )2 + (yi − qi )2

127

ml xl + mi xd )

In addition to A, Bx , By , we define Xi and Yi as follows:
X i = xi − p i
so we can rewrite

δCi
as
dxi

Similarly, we compute

Yi = y i − qi

δCi
= AXi + Bx +
dxi

kXi
2
Xi + Yi2

δCi
and rewrite it as
dyi
δCi
= AYi + By +
dyi

kYi
2
Xi + Yi2

δCi
We first solve for the value for Yi at which
= 0:
dxi
δCi
=0
dxi
kXi

AXi + Bx +
kXi
2
Xi + Yi2

2
Xi + Yi2

=0

= −(AXi + Bx )
kXi
AXi + Bx

(A.4)

(kXi )2
(AXi + Bx )2

(A.5)

2
Xi + Yi2 = −
2
Xi + Yi2 =

2
Yi2 = −Xi +

(kXi )2
(AXi + Bx )2

2
Xi (k 2 − (AXi + Bx )2 )
=
(AXi + Bx )2

Yi = ±

Xi k 2 − (AXi + Bx )2
AXi + Bx

(A.6)

Since k > 0, (A.4) implies that
−

Xi
>0
AXi + Bx

128

(A.7)

We now replace Yi in
which

δCi
by its value obtained above, and we solve for the value of Xi at
dyi

δCi
= 0. We obtain:
dyi

δCi
=0
dyi
kY

AYi + By +
Yi (A +

k
2
Xi + Yi2

Yi = −

Yi2 =

2
Xi + Yi2

=0

) + By = 0

By
A+

(A +

k
2 +Y 2
Xi
i
2
By
k
)
2 +Y 2
Xi
i

2
Xi (k 2 − (AXi + Bx )2 )
=
(AXi + Bx )2
A2 +

2
By
k2
2 +Y
Xi
i

+

2Ak
2
Xi +Yi

2
Using the value of Xi + Yi2 obtained in (A.5), we get:

2
Xi (k 2 − (AXi + Bx )2 )
=
(AXi + Bx )2

2
Xi (

2
By

A2 +

(AXi +Bx )2
2
Xi

2
By
k2
− 1) = 2 2
A Xi +(AXi +Bx )2
(AXi + Bx )2
2
Xi

129

+

2A
Xi
| AX +B |
x
i

+ 2A|

AXi +Bx
|
Xi

(A.7) implies that

Xi
Xi
=−
AXi + Bx
AXi + Bx
2
Xi (

2
By
k2
− 1) = 2 2
A Xi +(AXi +Bx )2
(AXi + Bx )2
2
Xi

2
Xi (

− 2A

AXi +Bx
Xi

2
By
k2
− 1) = 2 2
A Xi +(AXi +Bx )2
(AXi + Bx )2
AX +B
− 2AXi i 2 x
2
Xi

Xi

2
By
k2
−1= 2 2
(AXi + Bx )2
A Xi + (AXi + Bx )2 − 2AXi (AXi + Bx )
2
By
k2
−1= 2
(AXi + Bx )2
Bx

(AXi + Bx )2 =
(AXi + Bx −

kBx
2
2
Bx + By

Xi =

2
k 2 Bx
2
2
Bx + By

)(AXi + Bx −

kBx
2
2
Bx + By

2
2
−Bx ( Bx + By ± k)
2
2
A Bx + By

and the corresponding Yi is:

Yi =

2
2
−By ( Bx + By ± k)
2
2
A Bx + By

and hence

xi = p i +

yi = qi +

2
2
−Bx ( Bx + By ± k)
2
2
A Bx + By
2
2
−By ( Bx + By ± k)
2
2
A Bx + By

130

)=0

Theorem 5. If no node can improve the current configuration U by moving by itself, then
U is an optimal configuration
Proof. The key intuition is that for a configuration in which no relay node can move and
improve the cost by itself, the directional derivative in any direction at that configuration
is positive; this is a sufficient condition for the optimality of that configuration because the
cost function is convex.
We define the following notations. Let x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) denote a final
configuration where xi and yi represent the final positions of relay node si along the x-axis
and the y-axis respectively. We note that xi and yi are fixed for the source nodes and the sink
since these nodes do not move. The cost of this final configuration c(x, y) can be represented
as Tx (x) + Ty (y) + M (x, y) where
mi (xi − xj )2

Tx (x) = b
ij∈E

mi (yi − yj )2

Ty (y) = b

(A.8)

ij∈E

(xi − pi )2 + (yi − qi )2

M (x, y) = k
i∈N

Let α = (α1 , . . . , αn ) and β = (β1 , . . . , βn ) denote the final configuration computed by our
centralized algorithm. Let (g, h) be a possible direction in 2 : g = (g1 , . . . , gn ) represents the
possible movement of each node along the x-axis and h = (h1 , . . . , hn ) the possible movement
along the y-axis. We also define Gi = (0, . . . , 0, gi , 0, . . . , 0) and Hi = (0, . . . , 0, hi , 0, . . . , 0)
to be the restriction of (g, h) where only node si moves. We want to show that the directional

131

derivative c at (α, β) is positive in any direction (h, g). By definition,
c(α + tg, β + th) − c(α, β)
t
t↓0

c ((α, β), (h, g)) = lim

By the properties of our algorithm, no node can reduce the global cost by moving by itself
away from position (α, β). More specifically, if node si moves from position (αi , βi ) to
(αi + tgi , βi + thi ), the global cost increases. Let ∆i be this difference.
∆i ≥0

(A.9)

∆i =c(α + tGi , β + tHi ) − c(α, β)
∆i =Tx (α + tGi ) − Tx (α)
+ Ty (β + tHi ) − Ty (β)
+ M (α + tGi , β + tHi ) − M (α, β)
We replace Tx , Ty , M in ∆i and c(α + tg, β + th) by their values according to (A.8). We
obtain the following:
c(α + tg,β + th) − c(α, β) =
n

n

∆i − 2b
i=1

2
mi t2 (gi + h2 )
i

i=1

mi t2 (gi − gj )2 + b

+b
ij∈E
n

We rewrite −2b
i=1

mi t2 (hi − hj )2
ij∈E

2
mi t2 (gi +h2 )+b
i

mi t2 (gi −gj )2 +b
ij∈E

mi t2 (hi −hj )2 in the previous
ij∈E

132

equation as t2 V for some value V independent of t. We get
n

lim (c(α + tg, β + th) − c(α, β)) /t = lim
t↓0

t↓0 i=1
n

= lim

t↓0 i=1

∆i /t − tV
∆i /t

(A.10)

We now examine the limit in (A.10) closely. By using values Tx , Ty and M again, we get:
n

lim

t↓0 i=1

mi (gi − gj )(αi − αj )

∆i /t = 2b
ij∈E

mi (hi − hj )(βi − βj )

+ 2b
ij∈E


( (αi + tgi − pi )2 + (βi + thi − qi )2

+ lim k
t↓0

ij∈E


( (αi − pi )2 + (βi − qi )2  /t

−k

(A.11)

ij∈E

Using the difference of two squares identity, we can rewrite the limit term in (A.11) as
(αi − pi )gi + (βi − qi )hi

k
ij∈E

(αi − pi )2 + (βi − qi )2

(A.12)

(A.9), (A.11), (A.12) ⇒ c ((α, β), (h, g)) ≥ 0 so (α, β) is a global minimizer and represents
an optimal configuration.

133

BIBLIOGRAPHY

134

BIBLIOGRAPHY

[1] http://www.k-team.com/robots/khepera/index.html.
[2] Cc1000
single
chip
very
low
http://focus.ti.com/lit/ds/symlink/cc1000.pdf.

power

rf

transceiver.

[3] Cc2420 datasheet.
http://inst.eecs.berkeley.edu/cs150/Documents/CC2420.pdf.
[4] S. Adlakha and M. Srivastava. Critical density thresholds for coverage in wireless sensor networks. In Wireless Communications and Networking, 2003. WCNC 2003. 2003
IEEE, volume 3, pages 1615–1620, march 2003.
[5] S. Agarwal, R.H. Katz, S.V. Krishnamurthy, and S.K. Dao. Distributed power control
in ad-hoc wireless networks. In Personal, Indoor and Mobile Radio Communications,
2001 12th IEEE International Symposium on, volume 2, 2001.
[6] D.M. Blough, M. Leoncini, G. Resta, and P. Santi. The k-neighbors approach to interference bounded and symmetric topology control in ad hoc networks. Mobile Computing,
IEEE Transactions on, 5(9):1267–1282, 2006.
[7] Jonathan M. Borwein and Adrian S Lewis. Convex Analysis and NonLinear Optimization. Theory and Exmaples. Springer, 2000.
[8] Rainer E. Burkard. Selected topics on assignment problems. Discrete Applied Mathematics, pages 257–302, 2002.
[9] Jae-Hwan Chang and Leandros Tassiulas. Energy conserving routing in wireless ad-hoc
networks. In INFOCOM, pages 22–31, 2000.
[10] Supriyo Chatterjea and Paul J. M. Havinga. An adaptive and autonomous sensor
sampling frequency control scheme for energy-efficient data acquisition in wireless sensor
networks. In DCOSS, pages 60–78, 2008.
[11] Jim Chou, Dragan Petrovic, and Kannan Ramchandran. A distributed and adaptive
signal processing approach to reducing energy consumption in sensor networks. In
INFOCOM, 2003.
[12] Karthik Dantu, Mohammad Rahimi, Hardik Shah, Sandeep Babel, Amit Dhariwal, and
Gaurav S. Sukhatme. Robomote: enabling mobility in sensor networks. In IPSN, pages
404–409, 2005.

135

[13] Fatme El-Moukaddem, Eric Torng, and Guoliang Xing. Maximizing data gathering
capacity of wireless sensor networks using mobile relays. In IEEE MASS, pages 312–
321, 2010.
[14] Fatme El-Moukaddem, Eric Torng, Guoliang Xing, and Sandeep Kulkarni. Mobile relay
configuration in data-intensive wireless sensor networks. In IEEE MASS, pages 80–89,
2009.
[15] S. R. Gandham, M. Dawande, R. Prakash, and S. Venkatesan. Energy efficient schemes
for wireless sensor networks with multiple mobile base stations. In Globecom, 2003.
[16] Deepak Ganesan, Ben Greenstein, Denis Perelyubskiy, Deborah Estrin, and John Heidemann. An evaluation of multi-resolution storage for sensor networks. In SenSys,
2003.
[17] David Kiyoshi Goldenberg, Jie Lin, and A. Stephen Morse. Towards mobility as a
network control primitive. In MobiHoc, pages 163–174, 2004.
[18] Y. Gu, D. Bozdag, and E. Ekici. Mobile element based differentiated message delivery
in wireless sensor networks. In WoWMoM, pages 83–92, 2006.
[19] M.J. Handy, M. Haase, and D. Timmermann. Low energy adaptive clustering hierarchy
with deterministic cluster-head selection. In International Workshop on Mobile and
Wireless Communications Network, pages 368–372, 2002.
[20] Wendi Rabiner Heinzelman, Anantha Chandrakasan, and Hari Balakrishnan. Energyefficient communication protocol for wireless microsensor networks. In HICSS, 2000.
[21] John E. Hopcroft and Richard M. Karp. An n5/2 algorithm for maximum matchings
in bipartite graphs. SIAM J. Comput., 2(4):225–231, 1973.
[22] IEEE. Wireless medium access control (mac) and physical layer (phy) specifications for
low-rate wireless personal area networks (lr-wpans). In IEEE Standard 15.4, 2003.
[23] Sushant Jain, Rahul Shah, Waylon Brunette, Gaetano Borriello, and Sumit Roy.
Exploiting mobility for energy efficient data collection in wireless sensor networks.
MONET, 11(3):327–339, 2006.
[24] David Jea, Arun A Somasundara, and Mani B Srivastava. Multiple controlled mobile
elements (data mules) for data collection in sensor networks. In DCOSS, pages 244–257,
2005.
[25] Eun-Sun Jung and Nitin H. Vaidya. A power control mac protocol for ad hoc networks.
In Proceedings of the 8th annual international conference on Mobile computing and
networking, MobiCom ’02, pages 36–47, 2002.
136

[26] Aman Kansal, David D. Jea, Deborah Estrin, and Mani B. Srivastava. Controllably
mobile infrastructure for low energy embedded networks. IEEE Transactions on Mobile
Computing, 5(8):958–973, 2006.
[27] Abtin Keshavarzian, Huang Lee, and Lakshmi Venkatraman. Wakeup scheduling in
wireless sensor networks. In MobiHoc, pages 322–333, 2006.
[28] J.-H. Kim, D.-H. Kim, Y.-J. Kim, and K.-T. Seow. Soccer Robotics. Springer, 2004.
[29] Santosh Kumar, Ten-Hwang Lai, and J´zsef Balogh. On k-coverage in a mostly sleeping
o
sensor network. In MOBICOM, pages 144–158, 2004.
[30] Euisin Lee, Soochang Park, Fucai Yu, and Sang-Ha Kim. On selection of energy-efficient
data aggregation node in wireless sensor networks. IEICE Transactions, 93-B(9):2436–
2439, 2010.
[31] Li Li and J.Y. Halpern. A minimum-energy path-preserving topology-control algorithm.
Wireless Communications, IEEE Transactions on, 3(3):910–921, may 2004.
[32] N. Li, J.C. Hou, and L. Sha. Design and analysis of an mst-based topology control
algorithm. In INFOCOM 2003. Twenty-Second Annual Joint Conference of the IEEE
Computer and Communications. IEEE Societies, volume 3, pages 1702–1712, 2003.
[33] Yuhua Liu, Zhenrong Luo, Kaihua Xu, and Lilong Chen. A reliable clustering algorithm
base on leach protocol in wireless mobile sensor networks. In International Conference
on Mechanical and Electrical Technology (ICMET), pages 692–696, 2010.
[34] Gang Lu, Narayanan Sadagopan, Bhaskar Krishnamachari, and Ashish Goel. Delay
efficient sleep scheduling in wireless sensor networks. In INFOCOM, pages 2470–2481,
2005.
[35] Hong Luo, Jinge Wang, Yan Sun, Huadong Ma, and Xiang-Yang Li. Adaptive sampling
and diversity reception in multi-hop wireless audio sensor networks. In ICDCS, pages
378–387, 2010.
[36] J. Luo and J-P. Hubaux. Joint mobility and routing for lifetime elongation in wireless
sensor networks. In INFOCOM, pages 1735–1746, 2005.
[37] Liqian Luo, Qing Cao, Chengdu Huang, Tarek F. Abdelzaher, John A. Stankovic, and
Michael Ward. Enviromic: Towards cooperative storage and retrieval in audio sensor
networks. In ICDCS, page 34, 2007.
[38] Chia-Ching Ooi and Christian Schindelhauer. Minimal energy path planning for wireless
robots. In ROBOCOMM, page 2, 2007.

137

[39] Vamsi Paruchuri, Shivakumar Basavaraju, Arjan Durresi, Rajgopal Kannan, and
S. Sitharama Iyengar. Random asynchronous wakeup protocol for sensor networks.
In BROADNETS, pages 710–717, 2004.
[40] Charles E. Perkins and Elizabeth M. Royer. Ad-hoc on-demand distance vector routing.
In WMCSA ’99: Proceedings of the Second IEEE Workshop on Mobile Computer Systems and Applications, page 90, Washington, DC, USA, 1999. IEEE Computer Society.
[41] R. Pon, M.A. Batalin, J. Gordon, A. Kansal, D. Liu, M. Rahimi, L. Shirachi, Y. Yu,
M. Hansen, W.J. Kaiser, M. Srivastava, G. Sukhatme, and D. Estrin. Networked infomechanical systems: a mobile embedded networked sensor platform. In Fourth International Symposium on Information Processing in Sensor Networks. IPSN, pages
376–381, april 2005.
[42] Sylvia Ratnasamy, Brad Karp, Scott Shenker, Deborah Estrin, Ramesh Govindan,
Li Yin, and Fang Yu. Data-centric storage in sensornets with ght, a geographic hash
table. MONET, 8(4):427–442, 2003.
[43] Paolo Santi, Douglas M. Blough, and Feodor S. Vainstein. A probabilistic analysis for
the range assignment problem in ad hoc networks. In MobiHoc, pages 212–220, 2001.
[44] Curt Schurgers, Vlasios Tsiatsis, and Mani B. Srivastava. Stem: Topology management
for energy efficient sensor networks. In IN IEEE AEROSPACE CONFERENCE, pages
78–89, 2002.
[45] Mo Sha, Guoliang Xing, Gang Zhou, Shucheng Liu, and Xiaorui Wang. C-mac: Modeldriven concurrent medium access control for wireless sensor networks. In INFOCOM,
pages 1845–1853, 2009.
[46] R. Shah, S. Roy, S. Jain, and W. Brunette. Data mules: Modeling a three-tier architecture for sparse sensor networks. In IEEE SNPA Workshop, 2003.
[47] S. Shakkottai, R. Srikant, and N. Shroff. Unreliable sensor grids: coverage, connectivity
and diameter. In INFOCOM 2003. Twenty-Second Annual Joint Conference of the IEEE
Computer and Communications. IEEE Societies, volume 2, pages 1073–1083, march-3
april 2003.
[48] Yao Shen, Yunze Cai, Xinyan Li, and Xiaoming Xu. The restricted shortest-path-based
topology control algorithm in wireless multihop networks. Communications Letters,
IEEE, 11(12):937–939, december 2007.
[49] Suresh Singh, Mike Woo, and C. S. Raghavendra. Power-aware routing in mobile ad
hoc networks. In MOBICOM, pages 181–190, 1998.
[50] Arun A. Somasundara, Aditya Ramamoorthy, and Mani B. Srivastava. Mobile element
138

scheduling with dynamic deadlines. IEEE Transactions on Mobile Computing, 6(4):395–
410, 2007.
[51] Robert Szewczyk, Alan Mainwaring, Joseph Polastre, John Anderson, and David Culler.
An analysis of a large scale habitat monitoring application. In SenSys, pages 214–226,
2004.
[52] Chiping Tang and Philip K. McKinley. Energy optimization under informed mobility.
IEEE Trans. Parallel Distrib. Syst., 17(9):947–962, 2006.
[53] B¨lent Tavli and Onur Ceylan. Joint routing and multi level data compression for
u
lifetime optimization in wireless sensor networks. In ISCIS, pages 692–696, 2009.
[54] J.-P. Tignol. Galois’ theory of algebraic equations. John Wiley & Sons, Inc., New York,
NY, USA, 1987.
[55] Yu-Chee Tseng, Chih-Shun Hsu, and Ten-Yueng Hsieh. Power-saving protocols for ieee
802.11-based multi-hop ad hoc networks. In INFOCOM, 2002.
[56] Tijs van Dam and Koen Langendoen. An adaptive energy-efficient mac protocol for
wireless sensor networks. In SenSys, pages 171–180, 2003.
[57] P Varshney. Distributed Detection and Data Fusion. Springer-Verlag, New York, NY,
1996.
[58] Guiling Wang, Mary Jane Irwin, Piotr Berman, Haoying Fu, and Thomas F. La Porta.
Optimizing sensor movement planning for energy efficiency. In ISLPED, pages 215–220,
2005.
[59] Lan Wang and Yang Xiao. A survey of energy-efficient scheduling mechanisms in sensor
networks. Mob. Netw. Appl., 11(5), 2006.
[60] Pu Wang, Rui Dai, and Ian F. Akyildiz. Collaborative data compression using clustered
source coding for wireless multimedia sensor networks. In INFOCOM, pages 2106–2114,
2010.
[61] Qin Wang, Mark Hempstead, and Woodward Yang. A realistic power consumption
model for wireless sensor network devices. In IEEE SECON, Reston, VA, September
2006.
[62] Wei Wang, Vikram Srinivasan, and Kee-Chaing Chua. Using mobile relays to prolong
the lifetime of wireless sensor networks. In MobiCom, 2005.
[63] Z. Maria Wang, Stefano Basagni, Emanuel Melachrinoudis, and Chiara Petrioli. Exploiting sink mobility for maximizing sensor networks lifetime. In HICSS, 2005.
139

[64] Guoliang Xing, Tian Wang, Weijia Jia, and Minming Li. Rendezvous design algorithms
for wireless sensor networks with a mobile base station. In MobiHoc, pages 231–240,
2008.
[65] Hongli Xu, Liusheng Huang, Yindong Zhang, He Huang, Shenglong Jiang, and Gang
Liu. Energy-efficient cooperative data aggregation for wireless sensor networks. J.
Parallel Distrib. Comput., 70(9):953–961, 2010.
[66] Xue Yang and Nitin H. Vaidya. A wakeup scheme for sensor networks: Achieving
balance between energy saving and end-to-end delay. In IEEE Real-Time and Embedded
Technology and Applications Symposium, pages 19–26, 2004.
[67] Ossama Younis and Sonia Fahmy. Distributed clustering in ad-hoc sensor networks: A
hybrid, energy-efficient approach. In INFOCOM, 2004.
[68] Rong Zheng, Jennifer C. Hou, and Lui Sha. Asynchronous wakeup for ad hoc networks.
In MobiHoc, pages 35–45, 2003.
[69] Daniel Zitterbart, Barbara Wienecke, James Butler, and Ben Fabry. Coordinated movements prevent jamming in an emperor penguin huddle. PLoS one, 6(6):e20260, 2011.

140