Self-Organisation in LTE Networks An Inv PDF

SELF-ORGANISATION IN LTE NETWORKS:
AN INVESTIGATION
by
Andrew Thompson
Submitted for the degree of

Doctor of Philosophy
Department of Computer Science
School of Mathematical and Computer Sciences

Heriot-Watt University
February 2014
The copyright in this thesis is owned by the author. Any quotation from the report or
use of any of the information contained in it must acknowledge this report as the source
of the quotation or information.

Abstract
Mobile telecommunications networks based on Long Term Evolution (LTE) technol-

ogy promise faster throughput to their users. LTE networks are however suscepti-
ble to a phenomenon known as inter-cell interference which can greatly reduce the
throughput of the network causing unacceptable degradation of performance for cell
edge users.
A number of approaches to mitigating or minimising inter-cell interference have
been presented in the literature such as randomisation, cancellation and coordina-
tion. The possibility of coordination between network nodes in an LTE network is
made possible through the introduction of the X2 network link.
This thesis explores approaches to reducing the effect of inter-cell interference on
the throughput of LTE networks by using the X2 link to coordinate the scheduling
of radio resources. Three approaches to the reduction of inter-cell interference were
developed. Localised organisation is a centralised scheme in which a scheduler is
optimised by a Genetic Algorithm (GA) to reduce interference. Networked organ-
isation makes use of the X2 communications link to enable the network nodes to
exchange scheduling information in a way that lowers the level of interference across
the whole network. Finally a more distributed and de-centralised approach is taken
in which each of the network nodes optimises its own scheduling in coordination
with its neighbours.
An LTE network simulator was built to allow for experimental comparison be-
tween these techniques and a number of existing approaches and to serve as a test
bed for future algorithm development. These approaches were found to significantly
improve the throughput of the cell edge users who were most affected by interefer-
ence. In particular the networked aspect of these approaches yielded the best initial
results showing clear improvement over the existing state of the art. The distributed
approach shows significant promise given further development.
Acknowledgements
This research, funded by an EPSRC CASE studentship in collaboration with Mo-

torola, was only possible with the help and support of others.
First and foremost I would like to thank my supervisor David Corne for his
guidance throughout the process and for his initial suggestion that a PhD might
be an interesting project; you were right! Thanks is also due to Chris Murphy of
Motorola whose help in understanding both the wider context, and technical details
of LTE were invaluable.
I would also like to thank my family and friends for their continued love and
support. Id particularly like to thank my parents who taught me to value knowledge,
exploration and inquisitiveness, without which such research would be impossible.
Finally, a special thanks to my wife Cara, for whose support, encouragement and
patience throughout this process I am deeply grateful.
i
Contents
1 Introduction 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Overview of LTE Networks . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Radio Access Network Architecture . . . . . . . . . . . . . . . 3
1.2.2 UMTS Radio Access Network Architecture . . . . . . . . . . . 6
1.2.3 LTE Radio Access Network Architecture . . . . . . . . . . . . 7
1.2.4 Transmission Scheduling in LTE . . . . . . . . . . . . . . . . . 9
1.3 Self-Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Outline of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 Literature Review 16
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Optimisation issues in LTE affecting Interference . . . . . . . . . . . 16
2.2.1 Base Station Placement . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Antenna Set-up . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Interference in LTE Networks . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Downlink Transmission in LTE Networks . . . . . . . . . . . . 20
2.3.1.1 Downlink Resources . . . . . . . . . . . . . . . . . . 21
2.3.1.2 Signal Quality . . . . . . . . . . . . . . . . . . . . . 23
2.3.1.3 Scheduling of Resources . . . . . . . . . . . . . . . . 25
2.3.1.4 CQI and Link Adaptation . . . . . . . . . . . . . . . 26
2.3.1.5 Existing Schedulers . . . . . . . . . . . . . . . . . . . 26
2.3.2 Interference . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
ii
2.4 Inter-Cell Interference Mitigation Techniques . . . . . . . . . . . . . . 30
2.4.1 Inter-cell Interference Randomisation . . . . . . . . . . . . . . 31
2.4.2 Inter-cell Interference Cancellation . . . . . . . . . . . . . . . 32
2.4.3 Inter-cell Interference Coordination . . . . . . . . . . . . . . . 33
2.4.3.1 Static Inter-Cell Interference Coordination Schemes . 33
2.4.3.2 Semi-Static Inter-Cell Interference Coordination Schemes 34
2.4.3.3 De-centralised Inter-Cell Interference Coordination
Schemes . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5 Evolutionary Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5.1 Single-Objective Genetic Algorithms . . . . . . . . . . . . . . 37
2.5.2 Multi-Objective Genetic Algorithms . . . . . . . . . . . . . . . 42
2.5.3 Evolutionary Algorithms in Telecommunication . . . . . . . . 46
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3 Localised Organisation at the Level of a Single eNodeB 49

3.1 Adaptive Softer Frequency Reuse . . . . . . . . . . . . . . . . . . . . 50
3.2 LTE Network Simulator . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2.1 Overview of Simulator Execution . . . . . . . . . . . . . . . . 53
3.2.2 Datarate Model . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.3 Channel Estimation Model . . . . . . . . . . . . . . . . . . . . 55
3.2.4 Traffic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2.5 Performance Metrics . . . . . . . . . . . . . . . . . . . . . . . 58
3.2.6 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3 Baseline LTE Schedulers . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3.1 Random Scheduler . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3.2 Round Robin Scheduler . . . . . . . . . . . . . . . . . . . . . 60
3.3.3 Maximum Rate Scheduler . . . . . . . . . . . . . . . . . . . . 60
3.3.4 Proportionate Fair Scheduler . . . . . . . . . . . . . . . . . . . 61
3.3.5 Soft Frequency Reuse Scheduler . . . . . . . . . . . . . . . . . 61
3.3.6 Softer Frequency Reuse Scheduler . . . . . . . . . . . . . . . . 62
3.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
iii
3.4.1 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . 64
3.4.2 MO-GA Parameters . . . . . . . . . . . . . . . . . . . . . . . 65
3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.5.1 Overall Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.5.2 Light Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5.3 Mixed Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.5.4 Heavy Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.5.5 Full Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4 Networked Organisation for eNodeB Co-ordination 85

4.1 Overview of the X2 Protocol . . . . . . . . . . . . . . . . . . . . . . . 86
4.2 Adaptive Softer Frequency Reuse with X2 . . . . . . . . . . . . . . . 87
4.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.4.2 Light Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.4.3 Mixed Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.4.4 Heavy Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.4.5 Full Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5 De-centralised Organisation 106

5.0.1 Central and Distributed Organisation . . . . . . . . . . . . . . 107
5.0.2 A De-centralised ICIC Algorithm . . . . . . . . . . . . . . . . 108
5.1 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2.2 Light Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
iv
5.2.3 Mixed Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.2.4 Heavy Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.2.5 Full Traffic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6 Conclusion 129
6.1 Main Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.1.1 Local Organisation . . . . . . . . . . . . . . . . . . . . . . . . 129
6.1.2 Networked Organisation . . . . . . . . . . . . . . . . . . . . . 130
6.1.3 De-centralised Organisation . . . . . . . . . . . . . . . . . . . 131
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.2.1 Improvements to the Network Simulator . . . . . . . . . . . . 133
6.2.1.1 Improved Traffic Model . . . . . . . . . . . . . . . . 133
6.2.1.2 Changes in Network Environment . . . . . . . . . . . 134
6.2.2 Further Investigation into ICIC Schemes . . . . . . . . . . . . 135
6.2.3 Alternative Optimisation Techniques . . . . . . . . . . . . . . 136
6.2.3.1 Cellular GAs . . . . . . . . . . . . . . . . . . . . . . 136
6.2.3.2 Embodied Evolution . . . . . . . . . . . . . . . . . . 137
6.3 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
Bibliography 139
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
v
List of Tables
1.1 The relationship between channel bandwidth, number of sub-carriers

and the number of available Resource Blocks (RBs) . . . . . . . . . . 11
2.1 The relationship between the operators channel bandwidth and the
number of available RBs [3] . . . . . . . . . . . . . . . . . . . . . . . 22
3.1 The GA parameters for the Adaptive Softer Frequency Reuse (aSerFR)
scheduler. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2 The probabilities associated with the traffic classes. . . . . . . . . . . 56
3.3 The proportion of User Equipments (UEs) created in each traffic class. 57
3.4 The possible number of UEs in each network environment and the
resulting average number of UEs per cell . . . . . . . . . . . . . . . . 63
3.5 The possible traffic levels of the UEs in each network environment . . 63
3.6 The possible speeds of the UEs in each network environment . . . . . 63
3.7 The parameters used for the simulator in experiment 1. . . . . . . . . 64
3.8 The parameters used for the Multi-Objective Genetic Algorithm (MO-
GA) in experiment 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.1 The parameters used for the MO-GA in experiment 2. . . . . . . . . 88

4.2 The parameters used for the simulator in experiment 2. . . . . . . . . 88
5.1 The parameters used for the De-centralised Softer Frequency Reuse
(dSerFR). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
vi
List of Figures
1.1 A conceptual representation of two wireless telecommunications net-

works, each of which is connected to both the public telephone net-
work and the internet. . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 A simple cellular network arranged in 3 clusters, each with 7 cells. . . 4
1.3 A simple cellular network demonstrating frequency reuse . . . . . . . 5
1.4 A simple cellular network demonstrating both frequency reuse and
sectorisation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 UMTS radio access network architecture . . . . . . . . . . . . . . . . 7
1.6 LTE radio access network architecture . . . . . . . . . . . . . . . . . 8
1.7 The time structure of the LTE physical resource. . . . . . . . . . . . 9
1.8 The frequency structure of the LTE physical resource. . . . . . . . . . 10
2.1 The importance of base station positioning for the movement of UEs
between cells. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 The time domain structure of LTE downlink physical resource. . . . 21
2.3 The frequency structure of the LTE downlink physical resource. . . . 22
2.4 The effect of distance on the propagation loss experienced by a UE
using the propagation model outlined in [6]. . . . . . . . . . . . . . . 24
2.5 Shadowing causes signal strength to diminish due to environmental
factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.6 Multi-path fading causes multiple copies of the transmitted signal to
be received by the UE due to interactions with the environment. . . . 25
2.7 Network cells without overlapping coverage. The UE moving from
cell A to cell B will lose its connection in the process. . . . . . . . . 28
vii
2.8 Network cells with overlapping coverage allowing the UE to move
from one cell to the other without losing connection. . . . . . . . . . 28
2.9 Mobile 1.B and 2.B will experience interference since they both oc-
cupy an area of overlapping coverage and are both being scheduled
on the same group of RBs. . . . . . . . . . . . . . . . . . . . . . . . . 29
2.10 Relationship between Signal to Interference plus Noise Ratio (SINR)
and UE datarate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.11 Frequency allocation in a static SFR scheme. . . . . . . . . . . . . . . 34
2.12 An example one-dimensional fitness landscape. . . . . . . . . . . . . . 37
2.13 Overview of the main steps taken by a GA during each iteration. . . . 38
2.14 Structure of a GA chromosome. . . . . . . . . . . . . . . . . . . . . . 38
2.15 Mutation on a GA chromosome. . . . . . . . . . . . . . . . . . . . . . 39
2.16 One point crossover operating on two GA chromosomes. . . . . . . . 39
2.17 Two point crossover operating on two GA chromosomes. . . . . . . . 40
2.18 The operation of uniform crossover on a GA chromosome. . . . . . . 40
2.19 Main steps taken by a Multi-Objective Genetic Algorithm . . . . . . 43
2.20 An example of a Pareto front. . . . . . . . . . . . . . . . . . . . . . . 44
2.21 Overview of NSGA-II algorithm. . . . . . . . . . . . . . . . . . . . . . 45
2.22 Overview of the MO-GA used in this research. . . . . . . . . . . . . . 46
3.1 The effect of UE speed on the variation of signal strength. . . . . . . 56

3.2 The 19 cell network used in experiment 1. . . . . . . . . . . . . . . . 65
3.3 The convergence of 5th percentile UE throughput for each algorithm. 65
3.4 The convergence of average UE throughput for each algorithm. . . . 66
3.5 Boxplot showing the average UE throughput of each algorithm across
all the network environments and traffic levels. . . . . . . . . . . . . . 68
3.6 Boxplot showing the 5th percentile UE throughput of each algorithm
across all the network environments and traffic levels. . . . . . . . . . 69
3.7 Boxplot showing the average throughput of each algorithm across all
the network environments under light traffic. . . . . . . . . . . . . . . 70
viii
3.8 Boxplot showing the 5th percentile throughput of each algorithm
across all the network environments under light traffic. . . . . . . . . 71
3.9 The 5th Percentile UE throughput vs average UE throughput in light
traffic across nine network environments. Across the nine sub-plots
the number of UEs increases from left to right while the UE speed
increases from top to bottom. . . . . . . . . . . . . . . . . . . . . . . 72
the network environments under mixed traffic. . . . . . . . . . . . . . 73
across all the network environments under mixed traffic. . . . . . . . 74
3.12 The 5th Percentile UE throughput vs average UE throughput in
mixed traffic environments. Across the nine sub-plots the number
of UEs increases from left to right while the UE speed increases from
top to bottom. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
the network environments under heavy traffic. . . . . . . . . . . . . . 76
across all the network environments under heavy traffic. . . . . . . . . 77
3.15 The 5th Percentile UE throughput vs average UE throughput in heavy
traffic environments. Across the nine sub-plots the number of UEs
increases from left to right while the UE speed increases from top to
bottom. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
the network environments under full traffic. . . . . . . . . . . . . . . 80
across all the network environments under full traffic. . . . . . . . . . 80
3.18 The 5th Percentile UE throughput vs average UE throughput in full
traffic environments. Across the nine sub-plots the number of UEs
increases from left to right while the UE speed increases from top to
bottom. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
ix
all all traffic levels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
across all all traffic levels. . . . . . . . . . . . . . . . . . . . . . . . . 91
all light traffic network environments. . . . . . . . . . . . . . . . . . . 92
4.4 Boxplot showing the average 5th percentile UE throughput of each
algorithm across all light traffic network environments. . . . . . . . . 93
all mixed traffic network environments. . . . . . . . . . . . . . . . . 95
algorithm across all mixed traffic network environments. . . . . . . . 96
mixed traffic across nine network environments. Across the nine sub-
plots the number of UEs increases from left to right while the UE
speed increases from top to bottom. . . . . . . . . . . . . . . . . . . . 97
all heavy traffic network environments. . . . . . . . . . . . . . . . . . 98
algorithm across all heavy traffic network environments. . . . . . . . 99
all full traffic network environments. . . . . . . . . . . . . . . . . . . 101
x
algorithm across all full traffic network environments. . . . . . . . . 102
5.1 The convergence of the average UE throughput values in the simulator

over time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.2 The convergence of the 5th percentile UE throughput values in the
simulator over time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
all traffic levels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
across all traffic levels. . . . . . . . . . . . . . . . . . . . . . . . . . . 114
all light traffic network environments. . . . . . . . . . . . . . . . . . . 115
algorithm across all light traffic network environments. . . . . . . . . 116
all mixed traffic network environments. . . . . . . . . . . . . . . . . . 118
algorithm across all mixed traffic network environments. . . . . . . . 119
mixed traffic across nine network environments. Across the nine sub-
plots the number of UEs increases from left to right while the UE
speed increases from top to bottom. . . . . . . . . . . . . . . . . . . . 120
xi
all heavy traffic network environments. . . . . . . . . . . . . . . . . . 121
algorithm across all heavy traffic network environments. . . . . . . . 122
all full traffic network environments. . . . . . . . . . . . . . . . . . . 124
algorithm across all full traffic network environments. . . . . . . . . 125
xii
Acronyms
3GPP 3rd Generation Partnership Project.
ARQ Automatic Repeat reQuest.
aSerFR Adaptive Softer Frequency Reuse.
aSerFR+X2 Adaptive Softer Frequency Reuse + X2.
CDMA Code Division Multiple Access.
CQI Channel Quality Indicator.
dSerFR De-centralised Softer Frequency Reuse.
EA Evolutionary Algorithm.
eNB Evolved NodeB.
GA Genetic Algorithm.
ICIC Inter-Cell Interference Coordination.
IDMA Interleave Division Multiple Access.
LTE Long Term Evolution.
MCS Modulation and Coding Scheme.
MO-GA Multi-Objective Genetic Algorithm.
MOEA Multi-Objective Evolutionary Algorithm.
xiii
NB NodeB.
NGMN Next Generation Mobile Networks Alliance.
OFMD Orthogonal Frequency Division Multiple.
OFMDA Orthogonal Frequency Division Multiple Access.
PF Proportional Fair.
QoS Quality of Service.
RAN Radio Access Network.
RB Resource Block.
RE Resource Element.
RNC Radio Network Controller.
SerFR Softer Frequency Reuse.
SFR Soft Frequency Reuse.
SINR Signal to Interfernce plus Noise Ratio.
SON Self-Organising/Optimizing Network.
UE User Equipment.
UMTS Universal Mobile Telecommunications System.
xiv
Declaration Statement
There will be two blank pages here so I can add the pre-printed declaration form.
1
2
Chapter 1
Introduction
1.1 Introduction
This thesis describes an investigation into self-organisation in Long Term Evolution

(LTE) networks. LTE is a wireless networking standard designed to enable high
speed data connections for mobile phones and other devices. LTE networks typically
consist of a number of mobile terminals (known as User Equipments (UEs)) each
connected to a wireless access point (known as an Evolved NodeB (eNB)). These
eNBs are connected both to other eNBs and crucially to the operators core network
which enables communication with other networks such as the internet.
The group responsible for standardising LTE networks, known as the 3rd Gen-
eration Partnership Project (3GPP), have identified self-organisation as a key way
of enabling network operators to better manage the increasing complexity of their
networks. It is envisaged that by incorporating self-organisation into the network
nodes, such as the eNBs, a number of common network planning and optimisa-
tion tasks can be automated, allowing network operators to reduce their operating
expenditure by reducing the manual involvement in such tasks [2]. For example,
currently a large number of network parameters must be set and tuned over time
by human intervention. It is envisioned that the use of self-organisation will enable
the network itself to determine the optimal values for such parameters without the
need for human intervention.
This chapter provides a high level introduction to LTE networks and self-organisation
1
Chapter 1: Introduction
in order to provide the necessary information to understand the wider context of

this thesis. The specific contributions made by this thesis are also highlighted.
1.2 Overview of LTE Networks
LTE networks are cellular wireless networks based on an existing technology known
as Universal Mobile Telecommunications System (UMTS). UMTS is the current
wireless network standard and has been widely adopted by the telecommunications
industry. LTE is an evolutionary redesign of UMTS designed primarily to signifi-
cantly increase the peak data rate (the maximum amount of information that can
be transmitted), increase the performance of cell edge users (users that are located
at the edge of a cell typically suffer from reduced data rates due to interference) and
simplify the network architecture [1].
Figure 1.1: A conceptual representation of two wireless telecommunications net-

works, each of which is connected to both the public telephone network and the
internet.
Conceptually wireless telecommunication networks can be viewed as consisting

of two main elements: the radio access network and the core network, see Figure 1.1.
The core network is the part of an operators network infrastructure that provides
systems for routing user calls and data sessions to their destinations either within
2
the operators network or to external networks. As such it maintains connections

to both the internet, so that users can access standard web services, and to the
public telephone network so that incoming and outgoing phone calls can be correctly
routed. The core network is also responsible for many other functions such as
authenticating users on their service, billing users for calls made and data sessions
used, and providing services such as voice-mail.
The radio access network is that part of the network that wirelessly connects a
users device with the operators core network. It is responsible for setting up and
maintaining a wireless connection to any user devices and forwarding any calls or
data sessions back to the core network for appropriate handling. The radio access
network is typically implemented using a cellular network architecture as this allows
a single radio access network to cover a significant area.
One of the main reasons self-organisation is being considered for use in LTE
networks is due to changes in the design of the radio access network. To understand
the significance of these changes it is first important to understand how radio access
networks and more specifically cellular networks operate.
This section presents an overview of the cellular network architecture adopted by
the 3GPP for use in the radio access network within UMTS and LTE and highlights
some of the main differences in UMTS and LTE which have led to the possibility of
implementing Self-Organising/Optimizing Network (SON) techniques.
1.2.1 Radio Access Network Architecture
Radio access networks are used to provide a wireless link between users devices
and the network operators core network and typically take the form of a cellular
network.
Cellular networks aim to provide connectivity for users (known as UEs) over a
geographic area by using radio waves to transmit and receive information. The total
area to be covered by the network is divided up into much smaller areas known as
cells, each of which is served by a single base station.
Each base station is equipped with a number of antennae with which it commu-
3
nicates to the UEs within its cell and is connected to the operators core network
allowing it to route calls and data sessions made by the UEs in its area to their
appropriate destination in other parts of the network. These cells can have a range
of sizes depending on their locality and setup. For example a cell radius of 5km
would be common in rural areas where the population is less dense, while in an
urban environment cells would typically have a radius of 1km or less. Typically
these cells are represented as hexagons as in Figure 1.2.
Figure 1.2: A simple cellular network arranged in 3 clusters, each with 7 cells.
The cellular architecture allows for a greater coverage area to be served by the
network. The fact that any wireless transmission inevitably loses strength as it
travels through space means that any base station will have an inherently limited
range. Thus by dividing the desired coverage area into a number of smaller cells,
each of which has a limited range, a much larger area can be covered [84, 94]. The
cellular system has the additional benefit of increasing the ease with which the
network can be expanded by the operator. To increase the size of the coverage area
additional cells can be created by adding more base stations.
Another compelling reason for the use of a cellular system architecture is that
it allows for the reuse of the available spectrum. Network operators are typically
granted a license to transmit on a specific and limited range of frequencies. The
frequency spectrum is a finite resource and thus there are a limited number of radio
4
channels available to wireless systems [56]. This means that the total bandwidth of
the system, and thus the amount of data that it can transmit, is dictated by the
amount of available spectrum. Since the capacity of the network is constrained by
the available spectrum, re-using that spectrum can greatly increase the capacity of
the network.
Traditionally a cellular system allows the available spectrum to be divided into
a number of bands, each of which can then be allocated to a cell. These frequency
bands can then be reused in other cells provided that there is sufficient distance
to prevent any interference. By dividing a large network into a number of smaller
clusters the spectrum can be reused across the network in a predictable way. For
example Figure 1.3 shows a small network composed of 3 clusters, each with 7 cells.
The available spectrum has been divided into 7 bands, labelled F 1 to F 7, which
have then been allocated to each cluster in a similar fashion.
Figure 1.3: A simple cellular network demonstrating frequency reuse
Each cell can be further subdivided into sectors, typically three or four, each of
which is served by a distinct directional antenna on the base station. Dividing an
ideal cell into N sectors should increase the capacity by N , however there are costs
involved such as the increased cost of equipment and signalling overhead due to the
need for handovers. A handover procedure occurs when a UE moves from one cell
or sector to another. In order to correctly route incoming calls and data packets
5
the network must be able to detect when a UE has moved from one cell or sector to
another. Handovers typically reduce the throughput experienced by a UE and put
additional pressure on the base stations resources. Sectorising cells can also reduce
both the co-channel interference within the cell and the interference caused by the
transmissions of neighbouring cells [25] and can lead to increased frequency reuse
as shown in Figure 1.4, allowing several frequencies to be used within a single cell.
Within UMTS networks this has been shown to significantly increase the capacity
of cells [62].
Figure 1.4: A simple cellular network demonstrating both frequency reuse and sec-
torisation.
1.2.2 UMTS Radio Access Network Architecture
The radio access network utilised by UMTS takes the form of a cellular network
and introduces additional network infrastructure as shown in Figure 1.5. As with
traditional cellular networks each cell is served by a base station, known as a NodeB
(NB) in UMTS terminology, which is responsible for sending and receiving the actual
wireless transmissions to and from the UEs in its cell. However a new network node,
known as the Radio Network Controller (RNC), is introduced.
Each NB in a UMTS network is controlled by a single RNC, but a single RNC
will control many NBs. The exact number of NBs controlled an individual RNC
6
Figure 1.5: UMTS radio access network architecture
varies across deployments, but several hundred is not uncommon.

The RNC is responsible for a number of important functions including setting up
the various connections and resources required to initiate an incoming or outgoing
call or data session, the handling of Quality of Service (QoS) requirements for on-
going connections and the management and allocation of the radio resources within
each of the cells that it controls. The RNC is also responsible for implementing the
Automatic Repeat reQuest (ARQ) protocol, which handles the re-transmission of
erroneous or missing data, on all connections.
All handovers are handled by the RNC since NBs cannot communicate directly
to other NBs. If a UE moves from a cell controlled by one RNC to a cell controlled
by another RNC the RNCs will establish a connection in order to coordinate the
transfer of any ongoing connections.
Most of the processing power within the UMTS radio access network is located
in the RNC to the extent that the NB is little more than a modem that connects
the UEs to the RNC.
1.2.3 LTE Radio Access Network Architecture
The radio access network in LTE is an enhanced version of that found in UMTS
and as such shares many similarities.
The most significant difference, in the context of this research, is the removal of
the RNC node, see Figure 1.6. The functionality and roles which were previously
fulfilled by the RNC have now been devolved into the NBs, which have been renamed
to Evolved NodeBs (eNBs) to reflect their more autonomous status. The eNBs are
7
now directly responsible for assigning and scheduling radio resources, dealing with
handovers to other eNBs and routing calls and data connections to their destination.
Figure 1.6: LTE radio access network architecture
Another important change in LTE is the addition of the X2 interface. Each eNB
is directly connected to the core network and can also set up and maintain direct
connections to other eNBs in the network using the X2 protocol [18]. This allows the
eNBs to coordinate directly with each other regarding decisions such as handovers
and scheduling rather than requiring the involvement of the RNC. It is this change
in architecture which has led to an increasing interest in self-organisation since each
eNB can be thought of as an independent agent in a distributed system.
Another change in the radio access network in LTE is the adoption of Orthogo-
nal Frequency Division Multiple Access (OFMDA) as the transmission scheme. The
transmission scheme determines how the data to be transmitted is actually encoded
and transmitted using the radio wave as the transmission medium. The particular
benefit of OFMDA, in the context of this thesis, is the fact that it essentially elimi-
nates intra-cell interference. This means that there is very little interference within
a single cell from the transmissions sent from that eNB. The immediate benefit of
this is that the whole available spectrum can be re-used in each cell or even each
sector within each cell. Since the amount of bandwidth within a system is directly
proportional to the available spectrum this means that LTE networks have signif-
icantly more bandwidth available to them, thus potentially being able to provide
high data rates and throughput.
8
1.2.4 Transmission Scheduling in LTE
Much of the detail of this research is concerned with the scheduling of transmissions
within LTE networks and as such this section provides a general overview of how
transmissions take place in LTE networks. The details of wireless transmission in
LTE networks is contained in a number of standards documents authored by 3GPP,
primarily [15], but with other details contained in [16, 17, 10, 3]. Their relevant
content is summarised here since it provides important context to this research.
The resources that must be scheduled in an LTE system are the physical resources
that are used to transmit the data to and from the UEs. These resources have both
a time dimension and a frequency dimension.
The time domain is sub-divided in a number of different ways as seen in Figure
1.7. The largest division is known as a radio frame and has a duration of 10ms.
Within each radio frame are ten subframes each with a duration of 1ms and each
being divided into two further divisions known as slots with a duration of 0.5ms
each. Finally, each slot can also be divided into 7 Orthogonal Frequency Division
Multiple (OFMD) symbols. These symbols are essentially the physical wave-forms
that are used to encode and transmit the actual data to be sent across the wireless
channel.
Figure 1.7: The time structure of the LTE physical resource.
In the frequency domain the physical resource is divided into a large number of
sub-carriers. The exact number of sub-carriers in an LTE system depends mainly
on transmission bandwidth (see Table 1.1), but can vary between 72 and 1200.
The basic scheduling unit in LTE is known as a Resource Block (RB). In the
time domain a RB spans 0.5ms (or one slot), meaning that there are two RBs per
9
sub-frame in the time domain. In the frequency domain a RB spans 12 sub-carriers.

The actual number of RBs varies based on the network operators configuration but
can vary between 6 and 100, see Table 1.1. Figure 1.8 shows the typical breakdown
of the frequency domain in LTE.
Figure 1.8: The frequency structure of the LTE physical resource.
The amount of data that can be transmitted in an RB is dependent on the

modulation and coding scheme used. Modulation and coding schemes are essentially
a set of parameters, such as modulation type and code rate, that determine how the
data to be transmitted is encoded on the physical medium. Different modulation and
coding schemes are better suited to different radio conditions. For example when
the radio conditions are good, that is there is little interference, a more efficient
modulation and coding scheme can be used, one that maximises the amount of data
transmitted at the expense of error checking codes and redundant data. However
10
in more adverse radio conditions a modulation and coding scheme that sacrifices
the total amount of data for additional error checking codes and redundancy will
perform better. The most optimistic modulation and coding scheme can transmit 6
bits per OFMD symbol.
Channel Bandwidth (MHz) 1.4 3 5 10 15 20

Maximum number of sub-carriers 72 180 300 600 900 1200
Maximum number of RBs 6 15 25 50 75 100
Table 1.1: The relationship between channel bandwidth, number of sub-carriers and
the number of available RBs
During each slot an eNB may schedule an RB to any of the UEs within its cell. A
scheduler typically tries to allocate each RB to a UE such that the maximum amount
of data can be transmitted to that UE. There are however a number of reasons why
this is not always possible. The quality of the wireless channel between the eNB
and UE is constantly changing due to interference, movement and background noise.
The quality of the wireless channel dictates how much data can be transmitted to
the UE based on the modulation and coding scheme.
The scheduling of RBs can be thought of as an optimisation problem in the sense
that there are typically two competing requirements: maximising the throughput of
the eNB while also ensuring that all UEs are able to maintain adequate data rates.
For example if an eNB was simply to maximise its throughput it could schedule
as many RBs as possible to those UEs with the best wireless channel conditions,
this would maximise the amount of data transmitted in each RB since an optimistic
modulation and coding scheme could be used. But this would result in those UEs
experiencing poor wireless channel conditions to be starved of data. Thus a good
scheduler must ensure that those UEs with poor channel conditions receive adequate
data.
Another important factor in scheduling is the fact that transmissions from one
eNB can cause significant interference with the transmissions from neighbouring
eNBs. This is known as inter-cell interference. Thus the scheduling decisions made
by one eNB can have a dramatic effect on the performance of UEs being served by
other eNBs. In order to prevent or, more realistically, minimise this interference a
11
number of Inter-Cell Interference Coordination (ICIC) schemes have been developed.

These schemes generally use a combination of frequency allocation schemes and
schedulers to reduce the level of inter-cell interference and it is expected that within
this context self-organisation can offer significant improvements.
1.3 Self-Organisation
Within the context of LTE networks, self-organisation is concerned with automating

some aspect of the networks setup, optimisation or maintenance which was, in
previous networks, a manual process. Both 3GPP and Next Generation Mobile
Networks Alliance (NGMN) have published a list of use cases [7, 23] which highlight
a number of areas where self-organisation is expected to provide significant benefits.
Generally this takes the form of the automatic adaptation of network parameters so
that some aspect of the performance of the network is improved.
More generally, self-organisation can be thought of as the mechanism or the
process enabling a system to change its organisation without explicit external com-
mand during its execution time [42]. The essential concept for this comes from the
natural realm where numerous examples can be found, particularly of systems where
beneficial global behaviour emerges from simple interactions. For example groups of
birds are often seen displaying a level of self-organisation in their flocking behaviour.
This co-ordination of the movement of a large number of birds is achieved without
any central control, but merely by each bird following a number of simple rules.
Within computer science the ideas behind these natural systems have been used
as the basis of a number of problem solving techniques. For example capturing
the rules governing the flocking behaviour of birds [83] has been used to accurately
simulate the motion of flocking animals for animation. The foraging behaviour of
ants has also been adapted as an effective way of solving traditional graph based
problems [45] for example the travelling sales man problem [44]. Other examples of
self-organisation found in the biological world which have found applications in com-
puter science include brood sorting, web weaving and nest building [71] which have
been applied to areas as diverse as database organisation and clustering, network
12
routing and task cooperation in robots.

The focus of this research is how network performance in LTE can be improved
by utilising the interactions between neighbouring cells. The addition of the X2
interface to LTE networks, which allows for direct communication and interaction
between eNBs, may enable the eNBs to respond and adapt to their network envi-
ronment in a more organic and de-centralised way.
1.4 Contributions
This thesis offers a number of contributions in the area of self-organisation with

application to LTE networks in particular.
The first contribution of this thesis is the introduction of localised organisation

into the scheduling of transmissions in LTE networks, that is organisation that
is local to each eNB with no interactions with other eNBs. A parameterised
scheduling algorithm was developed that can be optimised to different network
environments using a Genetic Algorithm (GA). Based on simulations on a
number of network environments using standard parameters this approach was
seen to significantly improve on existing scheduling algorithms particularly on
networks which experienced a mixed traffic load.
The second contribution made by this thesis is the introduction of networked

organisation into the scheduling of transmissions in LTE networks, that is or-
ganisation that involves co-ordination between eNBs. A network aware sched-
uler was developed that enabled communication and collaborative scheduling
among eNBs using the X2 link. This collaborative algorithm was designed
such that it can be used with any existing scheduling algorithm to provide im-
proved performance. The algorithm was parameterised in a way that enables
a GA to optimise its performance for different networked environments. As
with the localised organisation a number of simulations were run to compare
this approach to existing techniques. These showed that under a number of
network conditions particularly when there were fewer UEs or when traffic was
13
light, this algorithm was able to provide significantly improved performance.
The third contribution made by this thesis is in the use of distributed organisa-
tion to further improve the scheduling of transmissions in LTE networks. This
builds on the previous contributions by enabling the performance optimisa-
tion to be done in a distributed, per-eNB basis rather than on a network-wide
level. This allows each eNB to optimise the scheduling for its own specific en-
vironment, while also collaborating with other eNBs. In order to compare this
approach with both the previous approaches and other existing approaches
simulations were again run which showed that this algorithm provided perfor-
mance gains across a number of network environments.
Another contribution made by this thesis is the creation and release of a phys-
ical layer LTE simulator. This simulator is capable of simulating the downlink
channel of an arbitrary number of eNBs and UEs with a number of different
traffic levels, user speeds and scheduling algorithms. The release of this simu-
lator should speed further research efforts in this area by avoiding the need to
re-implement such a simulator. The source code for the simulator is available
at http://downthewire.co.uk/projects/jLTE-Simulator/.
The final contribution of this thesis is the direct comparison of a number of

different scheduling algorithms for LTE. Five existing schedulers were com-
pared to the novel schedulers presented as part of this work. Each of the
schedulers was compared across twenty-seven different network environments
which included variations in the number of UEs, the traffic level of the UEs
and the speed of the UEs. The performance of each of the schedulers is com-
pared across two main axes allowing for the performance of each scheduler to
be compared across network conditions.
1.5 Outline of Thesis
The remainder of this thesis is divided into the following sections.
14
Chapter 2 provides a more detailed overview of LTE networks, in particular the

problem of inter-cell interference which is the main focus of this research. A number
of existing solutions are described and contrasted. Chapter 3 details a self-organising
method developed which operates locally to each eNB without coordination from
other eNBs. The working of this algorithm along with some comparative experi-
ments were used to show that this kind of optimisation can provide some level of
performance improvement over the existing approaches. A network simulator used
as the basis for the experiments is introduced and described. Chapter 4 reports
on a more coordinated approach to self-organisation in which eNBs communicate
with each other using the X2 link to achieve significantly improved performance.
Chapter 5 explores a more distributed approach to the ICIC problem by allowing
eNBs to continuously evolve their own configuration based on their local network
environment. Chapter 6 reviews the work that has been done and also offers some
suggestions for areas where further study would provide additional benefits.
15
Chapter 2
Literature Review
2.1 Introduction
Long Term Evolution (LTE) is an emerging wireless networking standard. As with

any new wireless technology, networks built using LTE face numerous optimisation
challenges. A particularly important problem in LTE networks is that of interfer-
ence. Since this research is focused on developing coordinated approaches to mitigate
the effects of interference, a number of optimisation problems which contribute to
the prevalence of interference in a network are first discussed.
The research described in this thesis presents a number of approaches to tackling
the interference problem in LTE networks using principles of self-organisation and
coordination. In particular Evolutionary Algorithms (EAs) in the specific form of
Genetic Algorithms (GAs) provide the means of self-organisation and are introduced
in Section 2.5.
2.2 Optimisation issues in LTE affecting Interfer-
ence
The design and optimisation of a wireless network is an on-going process throughout

the life-cycle of the network. It is concerned with finding the network setup which
maximises performance while usually minimising cost to the operator [75]. Some
16
Chapter 2: Literature Review
aspects of the networks setup must be decided upon before the network is deployed
and must remain fixed for the lifetime of the network. For example, the placement of
Evolved NodeBs (eNBs) is generally determined prior to network deployment as it
is expensive and impractical to move an existing eNB. Other aspects of the network
setup may be adjusted after the network had been deployed. For example cell size,
Quality of Service (QoS) constraints and some of the parameters of the antennae of
an eNB can all be adjusted to optimise the network after deployment. This section
presents some common optimisation problems faced by wireless networks such as
LTE.
2.2.1 Base Station Placement
As discussed, an LTE network consists of a number of fixed eNBs which serve the
users within their radius. In deploying an LTE network, operators generally want
to provide connectivity for the largest possible continuous area for the least possible
cost. This constraint requires that the number of eNBs used be kept to a minimum
since they represent a significant financial cost.
LTE networks are designed to provide mobile access which means that users or
User Equipments (UEs) must be able to move from cell to cell while still maintaining
any existing calls or data sessions. It is therefore important that the coverage area
provided by the eNBs is contiguous.
A simple LTE network is depicted in Figure 2.1. In moving directly from cell
1 to cell 3 UE1s network connection will be dropped due to a lack of overlapping
coverage between the cells. This will result in an interruption of the UEs service
and the dropping of any data connections or voice calls. In order to prevent this
there must be some area of overlap between the cells so that the UE can pass from
one cell to another without dropping the connection. For example UE2 can move
freely between cell 2 and cell 3 without dropping its connection as there is sufficient
coverage between the cells. However this overlapping area of coverage also introduces
interference between the cells which degrades their performance and so should be
kept to a minimum. This is discussed more fully in Section 2.3.2.
17
Figure 2.1: The importance of base station positioning for the movement of UEs
between cells.
Another factor which should be taken into account is the expected density of
UEs, and thus the expected traffic level, in the area of the cell. Each cell has a
finite amount of resources at its disposal and so it can only provide a certain level
of performance to the UEs that it is serving. If the number of UEs per eNB is too
high then the UEs will experience poor levels of connectivity. Due to this an urban
centre will have a much greater density of UEs, resulting in smaller cell sizes and
thus more eNBs, whereas cells can be much larger in rural areas and thus fewer
eNBs are required.
Several GAs have been developed in order to optimise the placement of eNBs.
The multi-objective GA proposed in [97] explicitly balances the network coverage
provided by the eNBs against the cost of the equipment, while [81] describes a GA
which finds near optimal placement for the eNBs of an unnamed 3rd generation
wireless network taking into account both traffic density and terrain. A flexible
encoding scheme is used which allows the genetic operators to alter both the number
and locations of the eNBs. However this is not practical as eNBs can often only be
placed at specific places such as the roof of a building. The GA was shown to provide
18
over 90% cover in both homogeneous and, more realistically, in heterogeneous traffic
environments. As an alternative to GAs the Adaptive search algorithm developed in
[24] provides good solutions to this problem for medium and large Universal Mobile
Telecommunications System (UMTS) networks.
2.2.2 Antenna Set-up
The physical parameters of the antennae used by the eNBs have a significant effect
not only on the performance of the cell but also on the level of interference expe-
rienced by the network. These parameters can be optimised early in the networks
life-cycle and are often combined with the eNB placement problem. For example
the multi-objective EA developed in [73] attempts to find both the optimal eNB lo-
cations along with the optimal antenna settings. However given a set of fixed eNBs,
various aspects of the antenna set-up can still be optimised.
There are a number of primary antenna parameters which can significantly im-
pact the performance of the network.
Antenna Downtilt The antenna downtilt is the angle of the main beam of an
antenna below the horizontal plane. This angle is determined by a combination of
the mechanical downtilt and electrical downtilt. These attributes should be consid-
ered separately, as opposed to modelling the total combined downtilt, due to the
fact that they affect the coverage of the antenna in different ways [53, 66, 80]. Me-
chanical downtilt involves altering the angle of the antenna on adjustable brackets
whereas electrical downtilt allows the antenna tilt to be changed without the need
to physically adjust the angle. This is done by altering the strength of the currents
in the elements of the antenna which causes the radiation pattern of the antenna to
be altered [91].
Azimuth The azimuth of the antenna is the horizontal angle between north and
the direction of the antenna. It can have a significant impact on both the coverage
provided by the antenna and on the sector overlap. Typically altering the azimuth
must be done manually at the eNB site [91].
19
Antenna Height Antenna height contributes to the determination of the coverage

area and performance of the antenna due to its effect on the path loss and delay
spread [52] (these terms will be defined more fully in Section 2.3.1.2). Generally the
antenna height is only considered at the initial network planning stage as it is costly
to change this once a network has been deployed.
Transmission Power The transmission power of the antenna determines its sig-
nal strength and coverage area. Altering the transmission power of the antenna is
primarily used to compensate for the effects of pathloss and shadowing [77]. Increas-
ing the transmission power of the antenna can be used to overcome some of these
effects by increasing the range of the signal and therefore the size of the sectors
coverage area, however it can also have a significant and negative effect on the in-
terference experienced by neighbouring cells. The transmission power is commonly
optimised with a number of constraints. For example a GA has been developed [78]
to minimise the transmission power of base stations in a Code Division Multiple
Access (CDMA) network while maintaining a sufficient bit rate to meet the QoS
requirements of its UE.
2.3 Interference in LTE Networks
The main problem addressed by this research is interference in LTE networks. This
section describes what interference is and how it arises in the context of an LTE
network. Several existing approaches to mitigating interference are described.
2.3.1 Downlink Transmission in LTE Networks
In order to understand how interference arises in LTE networks it is important to

understand how the eNBs and UEs communicate. This section therefore describes
the downlink transmission scheme used by LTE networks in order to provide the
context for the discussion of interference in Section 2.3.2 and of existing approaches
to reducing interference in Section 2.4.
20
Figure 2.2: The time domain structure of LTE downlink physical resource.
2.3.1.1 Downlink Resources
The radio resources which must be allocated by a eNBs to transmit data to UEs
are frequency/time/space blocks. The spatial dimension is dependent on utilis-
ing multiple-antenna transmission and reception which is beyond the scope of this
project.
The structure of the time-frequency resources is defined in the 3rd Generation
Partnership Project (3GPP) standard documents [4, 15]. In the time domain the
largest structure is a radio frame which has a duration of 10ms. This is further
divided into 10 subframes each 1ms in duration, which can themselves be further
divided into 2 slots each of 0.5ms in duration. Each slot can further be broken down
into 7 Orthogonal Frequency Division Multiple Access (OFMDA) symbols. This
can be more clearly seen in Figure 2.2.
The resources in the frequency domain are known as sub-carriers and are spaced
at 15kHz each, as defined in [15]. These sub-carriers are aggregated into groups
of 12, which for the duration of one slot are known as a Resource Block (RB) and
occupy a total of 180kHz. See Figure 2.3 for a visual explanation of these divisions.
The number of RBs available to each eNB is dependent on the range of frequency
available to the network operator. See Table 2.1 for details.
Each RB can be subdivided into a number of Resource Elements (REs), typically
84 (12 sub-carriers 7 OFMDA symbols) [15]. Some of these REs are reserved for
specific purposes, for example reference signals, control signals and synchronisation
signals which are not used to transmit data to the UE directly but rather are required
21
Channel Bandwidth (MHz) Number of RBs

1.4 6
3 15
5 25
12 50
15 75
20 100
Table 2.1: The relationship between the operators channel bandwidth and the
number of available RBs [3]
Figure 2.3: The frequency structure of the LTE downlink physical resource.
22
for the smooth running of the network [15]. The remainder of the REs are used to
transmit data directly to the UEs.
The amount of information transmitted in each RE varies depending on the
Modulation and Coding Scheme (MCS) used by the eNB to encode the data for
transmission [17]. The modulation scheme used determines the amount of redun-
dant data that is included in the transmission to aid the receiving UE in decoding
it. This is dependent on the signal quality currently being experienced. This is dis-
cussed more fully in the following section. If the signal received by the UE is strong
then more bits can be transmitted per symbol but if the UE is experiencing heavy
interference or is otherwise experiencing poor signal quality then more redundant
information is required.
2.3.1.2 Signal Quality
The strength of the signal can be considered as its power. Each signal is transmitted
by the eNB at a certain power level. The signal strength as experienced by a UE
is determined by how much of that transmitted power it receives. In the following
sections the various phenomena which lead to the dissipation of this power as it is
transmitted are discussed.
Pathloss, or attenuation, is the reduction in power which is experienced by every
electromagnetic wave as it travels through a medium such as air. Essentially the
further the original signal travels from the transmitter the weaker it gets as the
energy in it dissipates into the environment [82, 31]. Thus, aside from any other
phenomena, as a UE moves further from its eNB, its signal will inevitably get weaker.
Figure 2.4 shows the relationship between the pathloss experienced by a UE and its
distance from the eNB.
Shadow fading, or simply shadowing, is the dissipation of power in a signal due to
the effect of the environment through which it travels [61]. For example the strength
of a wireless signal can be degraded due to being reflected by buildings or absorbed
by trees. A UE which has a direct line of sight to its eNB will have significantly
stronger signal than a UE which is behind a building.
23
Figure 2.4: The effect of distance on the propagation loss experienced by a UE using
the propagation model outlined in [6].
Figure 2.5: Shadowing causes signal strength to diminish due to environmental

factors
24
Figure 2.6: Multi-path fading causes multiple copies of the transmitted signal to be
received by the UE due to interactions with the environment.
Another phenomenon which affects the strength of the received signal is known
as multi-path fading [61]. This is caused by the reception at a UE of multiple copies
of the same signal. These multiple versions of the original signal are caused when
the signal is reflected or refracted by objects in the environment such as buildings
or vegetation, see Figure 2.6. These reflections may arrive almost simultaneously, if
the effect is due to local scattering, or after longer intervals due to multiple paths
between the eNB and UE. As a UE moves the magnitude and phase of each reflected
signal changes. This can result in either constructive or destructive interference,
which can strengthen or weaken the signal.
2.3.1.3 Scheduling of Resources
This section examines how RBs are allocated to UEs given that data is transmitted
to UEs in RBs and a variety of factors affect the quality of the signal received over
time. It is the job of the eNBs scheduler to determine which RB to allocate to
which UE at any given time. An eNB maintains a data queue for each UE which it
serves. Given other input data, such as the reported quality of a UEs channel or
the current traffic load and priority, it must determine how to allocate the available
25
RBs in the most efficient manner.

As the eNB receives data to be transmitted to a UE, the data is stored in a buffer.
This data is then divided into a number of RBs and transmitted to the UE at a
given power level. The amount of data which can be transmitted to a UE on an RB
is determined by the MCS employed by the eNB. This is determined dynamically
by the eNB and is based on the signal quality experienced by the UE.
Thus, every 0.5ms an eNB will transmit zero or more RBs to each UE which
it is serving. Typically the scheduler will attempt to satisfy the QoS constraints
of as many UEs as possible by allocating the RBs and transmission powers of each
sub-frame.
The eNB can transmit each RB at a different power level. This affects the quality
of the signal received by the UE along with the interference experienced by other
UEs
2.3.1.4 CQI and Link Adaptation
In order to utilise the available resources most effectively eNBs attempt to match a
suitable MCS to the signal quality received by the UE. This process is known as link
adaptation [38]. In order to aid the eNB in this process the UEs can be configured to
report the channel quality conditions (known as Channel Quality Indicators (CQIs))
to its serving eNB. This involves the UE measuring the strength of the signal it
receives on a number of RBs and reporting it back to the eNB.
2.3.1.5 Existing Schedulers
Maximum Rate Scheduling Maximum rate scheduling attempts to maximise

the amount of data sent to UEs at the expense of fairness. This involves the eNB
scheduling each RB to that UE which is currently experiencing the best channel
conditions. This allows it to transmit the most amount of data possible per RB,
thus maximising resources. However there is a clear cost in that UEs which are
experiencing sub-optimal channel conditions may never be scheduled.
26
Proportionate Fair Scheduling Proportional Fair (PF) scheduling aims to pro-

vide a fairer approach to scheduling. It schedules a UE when its instantaneous
channel quality is high relative to its own average channel condition over time.
2.3.2 Interference
To understand Inter-Cell Interference Coordination (ICIC) it is important to first

understand how interference arises in LTE networks and how this affects perfor-
mance. Each eNB serves a number of UEs within its coverage area. This area is
determined by the transmission power of the eNB. This can be thought of as anal-
ogous to volume of the human voice, the louder a person talks the greater the area
in which others can hear their voice.
When designing and deploying a network, one of the important tasks of a network
operator is to determine where each eNB will be placed. In particular it is important
that between eNBs there is an area where their coverage overlaps. This is to allow
UEs to move between cells whilst still maintaining a single call or data session. For
example in Figure 2.7 if UE X moves from cell A to cell B they will move through
an area where there is no coverage and so their call will be dropped. In contrast
Figure 2.8 shows that UE X can move freely between cell A and cell B when there
is an area of overlap between them.
While this overlapping coverage is essential for allowing UEs to move seamlessly
between cells it also gives rise to interference. Interference is experienced by a UE
if, when receiving a transmission from its serving eNB on a given RB, it additionally
receives transmissions on the same RB from one or more other eNBs. To illustrate
Figure 2.9 represents two network cells, each of which is serving three UEs. Each
cell has a total of 100 RBs available to serve these UEs and has split them into
three groups each represented by a color. It is important to note that there is a
small area of overlap between the cells. In this simple network setup UEs 1.B and
2.B will both experience interference. This is because they are both located in the
area of coverage overlap and because they are being scheduled on the same RBs.
In comparison, although they are being scheduled on the same group of RBs, UEs
27
Figure 2.7: Network cells without overlapping coverage. The UE moving from cell
A to cell B will lose its connection in the process.
Figure 2.8: Network cells with overlapping coverage allowing the UE to move from
one cell to the other without losing connection.
28
Figure 2.9: Mobile 1.B and 2.B will experience interference since they both occupy
an area of overlapping coverage and are both being scheduled on the same group of
RBs.
1.A and 2.A will not experience interference because they are not located in the
area of coverage overlap and therefore will not receive the transmissions from the
neighbouring eNB.
Interference is only experienced when the transmissions occur over the same RB.
This is because RBs are orthogonal in the frequency domain and so adjacent RBs
do not cause interference.
The primary effect of interference is to reduce the Signal to Interference plus
Noise Ratio (SINR) of the UEs on the network, which has a direct effect on their
throughput. The throughput of a UE can be modelled using the following equation
[6]:

0 if SIN R < SIN Rmin

datarate(bps/Hz) = .S(SIN R) if SIN Rmin < SIN R < SIN Rmax

dataratemax if SIN R > SIN Rmax

Where: S(SIN R) is the Shannon bound of the signal-to-interference-plus-noise

ratio: S(SIN R) = log2 (1 + SIN R) bps/Hz. The parameter, which represents im-
plementation losses, takes the value 0.6, while the SIN Rmin is -10dB and SIN Rmax
29
4.5
4.0
3.5
3.0
Data Rate (bps/Hz)
2.5
2.0
1.5
1.0
0.5
0.015 10 5 0 5 10 15 20 25 30
SINR (dB)
Figure 2.10: Relationship between SINR and UE datarate.
is 25db. The maximum datarate which can be achieved under this model is 4.4bp-
s/Hz. Figure 2.10 shows the relationship between the SINR experienced by a UE
and its maximum achievable data-rate.
The key component of this model is the SINR. This is the ratio of the strength of
the signal received by the UE to the combined strength of the noise and interference
received by the UE. It can be seen that an increase in the interference experienced
by a UE causes the throughput of that UE to decrease.
2.4 Inter-Cell Interference Mitigation Techniques
The three main approaches to mitigating inter-cell interference which have been
proposed [19] are described below.
30
2.4.1 Inter-cell Interference Randomisation
Inter-cell interference randomisation techniques incorporate some level of random-

ness to the transmission process in an effect to reduce the level of interference ex-
perienced by the network.
One approach that has been explored in the literature is that of cell-specific
scrambling [47, 27]. In this approach a unique scrambling code is assigned to each
eNB which is then used as part of the encoding of its transmissions. When these
transmissions are received by a UE being served by that eNB the cell specific scram-
bling code is used to decode the transmission. As part of this process, provided the
scrambling codes used are orthogonal, the interference from neighbouring cells will
have been whitened or de-correlated. This refers to the fact that the power spectrum
of the interfering signal will become more similar to that of white noise which has
well understood statistical properties. This allows the UE to better differentiate the
actual signal transmitted by its serving eNB from the interference transmitted by
neighbouring eNBs.
An alternative approach to whitening the interfering signals which aims to im-
prove the decoding performance of the system is known as Interleave Division Multi-
ple Access (IDMA) [85]. A cell specific interleaving code is used by each eNB which
achieves the same result as the cell specific scrambling code discussed above. The
key improvement available in an IDMA system is that of iteratively coded multi-
user detection. This allows the UE to improve their suppression of the interfering
transmissions over time and thus can provide better performance than that of cell
specific scrambling.
Another approach known as frequency hopping has also been proposed as a way
to reduce the interference in a network [49]. It uses a randomised pattern to alternate
the sub-carriers that are used by the UEs. Rather than continuing to schedule a
UE on the same set of sub-carriers for the duration of its session frequency hopping
dictates that the set of sub-carriers used by each UE change over time. Since a
unique hopping pattern is assigned to each cell this means that the interference
experienced will be averaged across all UEs rather than reducing the interference
31
experienced by a specific set of UEs.

One basic scheme simply randomises the sub-carriers that are assigned to each
UE [35]. Thus each UE transmission is randomly spread across the available time-
frequency resources. Since this randomisation takes place at each eNB the interfer-
ence experienced by UEs is spread randomly rather than being concentrated in a
single section of the available time-frequency resources. This simple approach was
shown to provide some performance improvement under a heavily loaded network.
Despite the research highlighted above it has been suggested that inter-cell in-
terference randomisation is unlikely to provide the required improvement to the
performance of cell edge UEs since it cannot provide any SINR gain to the UEs [20].
2.4.2 Inter-cell Interference Cancellation
Inter-cell interference cancellation is generally concerned with using antenna tech-

niques to enhance the signal received by the UEs from the serving eNB and reduce
the level of interference received from other eNBs.
One approach to inter-cell interference cancellation uses a combination of beam-
forming and multiple antenna technology. Beam-forming allows the eNB to shape
the overall antenna beam of its transmission in the direction of the receiving UE and
can improve the signal strength received by the UE in proportion to the number of
antenna used by the eNB. For small networks of 2 to 3 cells this has been shown to
provide significant throughput gains [102].
Another approach to inter-cell interference cancellation which has been shown
to reduce the impact of interference is the combined use of sub-carrier based virtual
multiple-input multiple-output with a virtual signature randomiser [70]. Multiple-
input multiple-output techniques can increase the throughput of an LTE network
by using multiple antenna to transmit and receive the wireless signals. The multiple
antenna can be used to increase the performance of the network in various ways such
as pre-coding, spatial multiplexing or diversity coding, depending on the network
environment. Virtual multiple-input multiple-output attempts to utilise the same
concepts as multiple-input multiple-output but operates across the sub-carriers used
32
to transmit the data rather than operating across multiple physical antenna. The
virtual signature randomiser attempts to improve the UEs ability to distinguish the
signals transmitted from its serving eNB from those transmitted from neighbouring
eNBs. It does this by introducing an eNB specific random weight vector into the
scrambling block included in the transmissions. This allows the UE to better differ-
entiate the signals coming from the serving eNB from those originating from other
eNBs.
2.4.3 Inter-cell Interference Coordination
ICIC seeks to apply restrictions to the downlink scheduling resources in a way which
is coordinated with neighbouring cells. Typically this involves restricting which RBs
are available or limiting the transmission power of certain RBs. These restrictions
can actually reduce the level of interference in the system, rather than simply sup-
press the existing interference, but are more complex to implement.
Proposed ICIC schemes are classified as either static or semi-static schemes, de-
pending on the level of communication between nodes required to coordinate the
resources. Static schemes require no or at most very limited communication between
eNBs, typically only at the initial setup phase. The coordination in these schemes
comes primarily from the way the frequency is divided up when the network is
initially set up. Semi-static ICIC schemes usually have more frequent communica-
tion between nodes in order to reconfigure the scheme depending on the network
environment.
2.4.3.1 Static Inter-Cell Interference Coordination Schemes
A static ICIC scheme proposed in [46, 48], known as Soft Frequency Reuse (SFR),
divides the total available spectrum into a number of distinct groups known as
sub-bands. These sub-bands are then allocated to the networks cells such that
neighbouring cells are allocated bands which are orthogonal, thus minimising inter-
ference. The users at the centre of the cells can be served using the whole of the
available spectrum, but at a lower transmission power which reduces the amount of
33
Figure 2.11: Frequency allocation in a static SFR scheme.
interference they cause to those UEs in neighbouring cells. The users at the cell edge
are served at full power but can only be allocated resources within the sub-band of
the cell.
Figure 2.11 illustrates this concept. It can be seen that the edge of each cell is
only using a portion of the available spectrum while the centre of each cell is able to
use the whole available spectrum. In the central cell the edge users can only make
use of the red spectrum band, while the edge users of the neighbouring cells use the
blue and green bands. This means that any transmissions to the edge users of the
central cell will not cause interference to the edge users of the neighbouring cells.
This is a simple ICIC scheme which requires little overhead to implement and
is therefore often used as the basis for more complex semi-static schemes. It does
however limit the frequency utilisation of the edge users to 1/3 thus reducing their
maximum throughput. The performance of this scheme will also depend greatly
on the load of the network. A proposal by Alcatel [21, 22] improves the frequency
utilisation of this method to 3/7 by sub-dividing each cell into 3 sectors.
2.4.3.2 Semi-Static Inter-Cell Interference Coordination Schemes
In order to improve the flexibility of the static ICIC schemes, Siemens developed a
semi-static scheme which builds on the previous proposals while allowing for greater
flexibility depending on the network load [90, 88]. In this scheme the available
spectrum is subdivided into 16 sub-bands which can be allocated in a flexible way
34
between edge and centre users. The number of edge and centre users is then calcu-
lated dynamically and the sub-bands divided between them in a proportionate way.
This allows the cells to alter their frequency allocation depending on the traffic load
and user distribution.
Another scheme has been proposed which adapts to the ratio of edge to centre
users along with the traffic demands of each group [99]. In this scheme the edge
users are allocated a number of sub-bands depending on their load while the centre
users can be allocated the entire available spectrum. Transmissions to the centre
users are sent at a lower power level while those to the edge users are transmitted at
full power. If there is heavy load in the edge area of one cell, it can borrow from the
edge sub-bands of neighbouring cells provided they are under a much lighter load.
This scheme provides greater spectral efficiency but does require signalling between
nodes to facilitate the sharing of resources.
Softer Frequency Reuse (SerFR) is a scheme which builds on the traditional SFR
scheme and utilises a modified PF scheduling algorithm [103]. In this scheme the
entire available frequency spectrum is used by the whole cell. In each cell a sub-band
of frequencies are transmitted at full power. These sub-bands are allocated to cells
in an orthogonal manner. In order to determine which users are scheduled on which
RBs, the scheduler calculates a priority value for each of the users. This priority
value is based mainly on the current achievable data rate and the average data rate
for that user, but is also modified by a value. This value depends on whether
the RB being scheduled is part of the full power frequency sub-band and whether
the user is an edge or centre user. These values are configurable and so allow
the algorithm to be flexible. For example certain values would prevent any centre
users being scheduled on full power RBs while others would not distinguish at all
between edge and centre users.
2.4.3.3 De-centralised Inter-Cell Interference Coordination Schemes
The ICIC schemes discussed above can be described as centralised, that is the pa-
rameters are set as part of the network planning process and are not adjusted dy-
35
namically. De-centralised schemes by contrast are designed to be more adaptive to

the network environment and as such can alter their behaviour without any central
control. For example one approach to de-centralised ICIC is described in [93]. In
this scheme the available RBs are split into 3 non-overlapping sub-groups each of
which is assigned to a separate cell. When an eNB requires extra resources it can
request the use of one of the RBs from its neighbouring eNBs by using the X2 link.
The average throughput achieved by this scheme is reported as being slightly higher
than that of SFR however the throughput of the cell edge UEs remains similar to
SFR. It is worth noting however that the use of the X2 link was independent of the
scheduler used, in this case Round Robin, which should allow the scheme to be used
with any scheduling algorithm.
2.5 Evolutionary Algorithms
In this thesis extensive use of EAs, in the form of GAs, will be made to find optimal
solutions to various problems. Optimisation problems are concerned with finding
the optimal solution to a given problem. Many interesting problems, such as the
travelling salesman problem, can be formulated as optimisation problems. Optimi-
sation problems themselves can also be seen as a search problem where a search
space containing all possible solutions must be searched for the optimal solution.
This firstly assumes that there exists some function, called the objective function,
which can ascertain the quality, or fitness, of a solution.
The combination of all possible solutions and the objective function can be vi-
sualised as a graph, like that in Figure 2.12, known as a fitness landscape. Here the
y-axis represents the fitness, as determined by the objective function, and the x-axis
represents all possible solutions. This graph contains a number of local maxima and
a single global maximum. The purpose of an optimisation algorithm is to search
this space for the solution that represents the global maximum. However the more
complex the landscape the more difficult this task proves. The two optimisation
algorithms used in this project are introduced in the following sections.
36
Figure 2.12: An example one-dimensional fitness landscape.
2.5.1 Single-Objective Genetic Algorithms
Traditional GAs [76, 57, 86, 58] offer a robust approach to search and optimisation
problems inspired by genetics and natural selection. This is achieved by maintaining
a population of possible solutions to the given problem. The population is initialised
by creating a number of randomly generated solutions. During each iteration a
new population is created by selecting solutions from the existing population as
parents and recombining them to produce offspring using crossover operators. These
offspring are then mutated with a given probability and added to the new population.
This process is depicted in Figure 2.13.
GAs tend to be more robust and less likely to become trapped in a local maximum
than more basic search algorithms such as hill-climbing, in part due to the diversity
gained by maintaining a number of different solutions.
In order to optimise a problem a GA must be given a means to represent solutions
for the problem as a chromosome. This is known as the encoding of the problem.
Often this takes the form of a list of values where each value represents some variable
in the solution. This list is known as a chromosome and each value is known as
a gene. Each chromosome is then encoded as a list of values such that the list
37
Figure 2.13: Overview of the main steps taken by a GA during each iteration.
represents completely the solution to the problem. For example if a solution to

some problem could be represented by five binary digits then a randomly generated
chromosome could be represented as follows (Figure 2.14):
Figure 2.14: Structure of a GA chromosome.
There are a number of operators available for GAs to utilise. Operators are the
means by which the GAs create random solutions from their existing populations.
The main types of operator are mutation operators, crossover operators and selection
mechanisms.
Mutation Operators The mutation operator produces a single new child chro-
mosome from an existing parent chromosome. There are a number of possible oper-
ators which can be used for this purpose [28], however a commonly used approach is
simply to select a random gene and choose a new random value for it. This approach
is known as single point mutation and is based on biological point mutation [63].
38
Given our example chromosome illustrated above such a mutation would operate as
seen in Figure 2.15.
Figure 2.15: Mutation on a GA chromosome.
Crossover Operators The crossover operator produces two or more child chro-
mosomes from two or more parents. As with mutation operators, there are a number
of possible operators which can be used [29], several of which are briefly outlined
below.
Single point crossover makes use of a single, randomly chosen, mark point be-
tween the first and the last gene of a chromosome. To construct a child chromosome,
genes from index 0 to the mark point are taken from the first parent and the re-
maining genes are taken from the second parent. This can easily be used to generate
two children as seen in Figure 2.16.
Figure 2.16: One point crossover operating on two GA chromosomes.
Two point crossover operates in a similar manner except that two points are
chosen randomly. To construct a child chromosome genes from index 0 to the first
point are taken from the first parent, genes from the first point to the second point
39
are taken from the second parent with the remaining genes being taken again from
the first parent. This is illustrated in Figure 2.17.
Figure 2.17: Two point crossover operating on two GA chromosomes.
Uniform crossover operates slightly differently. A child chromosome is created

by iterating over each gene of the parents. With a given probability (usually 50%)
the gene is chosen from the first parent or else it is taken from the second. This is
illustrated in Figure 2.18.
Figure 2.18: The operation of uniform crossover on a GA chromosome.
There is no clear consensus regarding which of the crossover operators is most

effective [32]. The contribution that each operator makes to the performance of a
GA can be understood in terms of the schemata theorem and the building block
hypothesis.
The schemata theorem [64] is an attempt to explain the problem solving power
of GAs in terms of schemata. Schemata are particular gene patterns which may be
contained in a candidate solution. The binary coded candidate solution 1101, for
example, contains the schemata 11##, 1#01, #1## and 11#1 where the
40
# symbol matches any value. The order of a schemata is the number of non-#
digits it contains and the defining length is the distance from the outermost non-
# digits in the schemata. It assumed that the fitness of a candidate solution is
directly tied to its schemata such that fitter candidate solutions must contain better
schemata. Since these fitter candidate solutions contribute more of their genes to the
next generation more good schemata are passed on to the next generation improving
its fitness. The theorem states that shorter, fitter schemata increase with subsequent
generations. The purpose of crossover operators is to combine parents in such a way
as to combine the good schemata in each where possible.
Each crossover operator has a different effect on the propagation or destruction
of the schemata present in a candidate solution [32]. Propagated schemata are those
that are passed on to the child during crossover whereas schemata that experience
disruption are not passed on. Under uniform crossover, schemata of all orders are
equally likely to be disrupted. Under single point crossover and two point crossover
however the defining length of a schemata is directly proportional to its likelihood
of destruction. So the choice of crossover operator determines the likelihood of
schemata with certain characteristics being passed on to subsequent generations.
The building block hypothesis [59] states that the power of GAs is in their
ability to find good building blocks. These are low order schemata with a short
defining length which complement each other when present in a candidate solution.
The formation of these building blocks can be encouraged by the use of a coding
scheme that meets two criteria. Firstly related genes are located close together
in the chromosome and secondly there is little interaction between genes. This
interaction between individual genes is known as epistasis and can cause a reduction
in the effectiveness of crossover operators. For example is has been found that the
performance of two point crossover drops if these coding scheme recommendations
are not met [33] while uniform crossover still performs well.
Selection Mechanisms The selection mechanism determines which of the par-

ents in a population will be selected as parents in the production of the next genera-
tion. One common approach to selection is known as fitness proportional or roulette
41
wheel selection [30]. In this approach the probability of a chromosome being selected
is directly proportional to their fitness. The fitter a chromosome is the more likely
it is to be chosen for mutation or crossover. This ensures that in general the fitter
solutions remain in the population while at the same time there is some level of
diversity, in that some less fit solutions are also likely to be included.
Another common approach is known as tournament selection [74]. This involves
running several tournaments between randomly chosen chromosomes. The tourna-
ment size determines the number of chromosomes that are randomly selected to
be entered into the tournament pool. The fittest of the chromosomes within the
tournament pool are then selected.
A key factor in the choice of selection mechanisms is the trade-off that exists
between selection pressure and genetic diversity [34, 60]. Selection pressure is the
tendency to select the best members of the current generation to propagate to the
next generation while genetic diversity is required to ensure that the search space
is adequately explored. There is a tension between these two factors since a higher
selection pressure will likely lead to a decrease in the genetic diversity of the popu-
lation.
Tournament size allows for adjustment of selection pressure by adjusting the
tournament size. The larger the tournament size the more likely that the fitter can-
didate solutions will be selected and thus the greater the selection pressure. Larger
tournament sizes however are also associated with a loss of genetic diversity in the
population. In contrast the selection pressure and genetic diversity resulting from
use of the proportional selection mechanism are dependent on the fitness function
itself allowing for less transparent understanding of its performance.
2.5.2 Multi-Objective Genetic Algorithms
A traditional GA can only search for a single optimal solution, which in many cases is
all that is required. That is, a traditional GA can only use a single fitness measure or
objective function. However in more complex cases, such as those under investigation
in this thesis, there are a number of possible fitness measures or objective functions
42
which must all be optimised. Often these are even competing measures such that an
improvement in one measure will only be obtained at the cost of another measure. In
these cases a more advanced version of the GA, known as a Multi-Objective Genetic
Algorithm (MO-GA), can be used.
Figure 2.19: Main steps taken by a Multi-Objective Genetic Algorithm
An MO-GA operates in much the same way as a traditional GA (see Figure

2.19) but with the addition of a Pareto front. The Pareto front is a set where the
non-dominated candidate solutions discovered by the algorithm are retained. A non-
dominated candidate solution is one where no other candidate solution achieves a
better fitness across all objectives. A candidate solution is said to dominate another
solution if it is considered fitter according to all objective functions.
When plotted the candidate solutions contained in the Pareto front often result
in a graph like that in Figure 2.20 where there are a number of solutions which
perform well. In this case some of the candidate solutions perform better in fitness
43
measure than y and there are a number in between. The end user can then choose
which candidate solution is optimal depending on the priority they assign to each
fitness measure.
Figure 2.20: An example of a Pareto front.
To more fully describe the operation of an MO-GA the standard NSGA-II [41]
algorithm will be discussed before highlighting the differences between it and the
algorithm used in this research. The acronym NSGA-II stands for Non-dominated
Sorting Genetic Algorithm. A diagrammatic overview of its operation can be seen
in Figure 2.21.
Initially the NSGA-II algorithm creates a random population of candidate solu-
tions P0 of size N which are then sorted by non-domination level. Solutions with
a non-domination level of zero are known as Pareto Optimal solutions since they
are not dominated by any other candidate solutions. An initial offspring population
Q0 of size N is then produced using the standard crossover and mutation opera-
tors. Once this setup phase of the algorithm is complete the general procedure for
a tth generation continues as follows. First a combined population Rt = Pt Qt is
constructed. This combined population is now of size 2N . The candidate solutions
44
Figure 2.21: Overview of NSGA-II algorithm.
contained in Rt are then sorted using a fast, non-dominating sorting function which
sorts the candidate solutions into a number of fronts according to their domination
count. Those solutions which are not dominated by any other solution are sorted
into front F0 , while those that are dominated by only a single other solution appear
in front F1 .
Beginning at front F0 each front is added to the next generations population
Pt+1 until it has size N . If a given front is too big to fit into the new population its
members are sorted by the crowded comparison operator and the highest ranked so-
lutions are chosen for inclusion. The crowded comparison function prefers candidate
solutions which are in a less crowded region of the front which helps to maintain a
diverse set of candidate solutions in the new population.
Once the new population has been populated with N candidate solutions it is
used to produce a new offspring population Qt+1 by using the standard crossover and
mutation operators. Binary tournament selection is used for parent selection with
the selection criterion being the crowded comparison operator. The new population
Pt+1 and offspring population Qt+1 are then used as the input to the next generation.
The algorithm used as part of this research shares similarities from the NSGA-II
algorithm described above but also takes inspiration from other approaches such as
the Pareto Archived Evolutionary Strategy described in [69]. A graphical overview
of its operation can be seen in Figure 2.22. As with NSGA-II an initial population
45
Figure 2.22: Overview of the MO-GA used in this research.
of size N is created. During each generation two candidate solutions are selected
for crossover using fitness proportional selection along with another for mutation.
This results in three new candidate solutions being added to the existing population.
The domination counts for each solution are then calculated and any non-dominated
solutions are added to a separate Pareto front archive. Since the size of the current
population is N + 3 selection is again used to select those candidate solutions which
will survive to the next generation. This algorithm lacks the crowding distance
sorting that is used in the NSGA-II algorithm and is therefore likely to lead to less
diverse solutions over time.
2.5.3 Evolutionary Algorithms in Telecommunication
EAs have been widely used within the telecommunications research community.
For example antennae play an increasingly key role in wireless networks. The
design of antennae is a challenging problem since the antennae themselves must
generally be cheap and as simple as possible to implement at the hardware level but
must satisfy certain electrical requirements. For example NASA have used EAs to
develop antennae specifically for communication with satellites [55]. The resulting
designs not only improved on the performance of those antennae designed in a
conventional fashion but the speed of producing the designs was both considerably
faster and cheaper. Parallel EA were show-cased by Villegas et al. [95] as a robust
46
and fast way to produce designs for a patch antenna to be used in wireless networks.
EAs have also been used to optimise the design of multiple-antenna arrays [67] that
are becoming increasingly important as LTE adoption continues.
Within wireless networking in particular there are several areas where EAs have
proved be a valuable optimisation approach. The placement of eNBs (or equivalent)
within a wireless network is an important design decision. It affects not only the
cost of infrastructure for the network operator but also the area covered by the
network and the quality of the service provided to the users. A number of EAs
have been developed to tackle this problem both in its general form [36] and in
its application to specific network technologies such as 3G [72]. Related to this is
the problem of configuring the antennae on each eNB in an optimal fashion. As
previously discussed there are a number of parameters which can have a significant
effect on the performance of the network such as the transmission power and the
antenna downtilt and beam-width [101].
Within the specific area of scheduling and ICIC in LTE networks EAs have
been developed that optimise the uplink scheduling of UEs. The centralised scheme
presented in [92] uses an EA to uplink scheduling of UEs in LTE networks which
make use of multiple antenna technology (known as Multiple-Input Multiple-Output
(MIMO)). It was found that EAs could provide almost optimal results.
Another EA has been developed to optimise the scheduling of frequency resources
in LTE networks with a particular focus on the frequency of channel quality reports
which are sent from the UEs to the eNB [98]. In LTE the UEs can report back the
channel quality they experience on each RB to their serving eNB. This allows the
eNB to adapt its transmission encoding scheme to the channel conditions that each
UE experiences. The higher signalling rate that is required is costly and so it is
important to optimise the reporting frequency to maximise the opportunity of the
eNB to encode transmissions (and as such improve the data rate of the UE) while at
the same time minimising the signalling cost by minimising the number of reports
sent.
Another EA has been developed to optimise LTE networks using semi-smart
47
antennae [100]. Semi-smart antennae allow the amplitude and phase of their trans-
missions to be altered which allows the coverage pattern produced to be varied thus
allowing the size and shape of each cell to be optimised. The semi-smart anten-
nae were optimised using an EA with two key objectives: to provide coverage for all
users, and to maximise the total throughput of those users. Since the size and shape
of the cells coverage area was being optimised it was possible for a network to be
generated which had coverage gaps. This is undesirable since any user in a coverage
gap would not be able to send or receive any traffic and thus their throughput would
effectively be zero. Since the coverage patterns were being optimised it was also im-
portant to optimise the transmission power simultaneously in order to minimise the
interference experienced by the users and thus maximise their throughput.
These highly constrained problems illustrate how GAs can be used to optimise
several parameters simultaneously within a set of constraints.
2.6 Conclusion
This chapter has described aspects of the downlink transmission of LTE networks,
in particular those aspects that give rise to the problem of interference. A number
of approaches to mitigating interference were discussed with a specific focus on
ICIC. Some existing ICIC schemes presented in the literature were described and
contrasted. These schemes will be referenced in the following chapters as a number
of alternative approaches to ICIC are developed and introduced which are influenced
by them. EAs and their existing applications in the wireless networks have been
introduced and discussed as these techniques will be incorporated into the new ICIC
schemes presented in the subsequent chapters.
48
Chapter 3
Localised Organisation at the

Level of a Single eNodeB
In Chapter 2 the problem of interference in Long Term Evolution (LTE) networks

was introduced along with the concept of an Inter-Cell Interference Coordination
(ICIC) scheme to mitigate the effects of interference. Two existing ICIC schemes,
Soft Frequency Reuse (SFR) and Softer Frequency Reuse (SerFR), were also de-
scribed. The main theme of both these existing approaches was to divide the
available frequency into a number of bands which could then be assigned to User
Equipments (UEs) using different power levels. In each of these approaches a new
scheduler was developed that was able to take advantage of the different power levels
available to schedule the bands in a way that reduced the interference and improved
the performance of the network.
This chapter describes a new ICIC scheme which is based on the SerFR described
in the previous chapter. A parameterised scheduler is presented which can be op-
timised by an Genetic Algorithm (GA) to allow it to be adapted to any network
environment. We refer to this as a local optimisation in that the scheduler has no
knowledge of the neighbouring Evolved NodeBs (eNBs).
A number of experiments are described detailing the performance of this new
algorithm and comparing it with the existing ones. A network simulator is also
described that was developed in order to fairly compare the new algorithm with
those in the literature.
49
Chapter 3: Localised Organisation at the Level of a Single eNodeB
3.1 Adaptive Softer Frequency Reuse
The Adaptive Softer Frequency Reuse (aSerFR) is based on the SerFR scheme dis-
cussed in Section 2.4.3.2. As with the SerFR scheme the available frequency is split
into a number of bands some of which are transmitted at full power and the re-
mainder at a reduced power level. This allows the central UEs to be scheduled on
the reduced power Resource Blocks (RBs) thus reducing the interference that they
cause in neighbouring cells. The cell edge UEs can then be scheduled on the full
power RBs in order to increase their throughput.
SerFR uses a proportional fairness scheduling algorithm to schedule UEs to RBs.
To determine which UE to schedule on a given RB a priority value, Pk,j (n), is
calculated for each UE on that RB. The UE with the maximum priority value is
scheduled on that RB. The priority value is defined as
DRCk,j (n)
Pk,j (n) = k,j , (3.1)
Rk (n)
where DRCk,j (n) is the achievable data rate the the k th UE on the j th RB at
time n and Rk (n) is the low-pass filtered averaged data rate of the k th UE at time
n. Rk (n) is defined as
Rk (n + 1) = (1 )Rk (n) + DRCk0 (n) (3.2)
where is the forgetting factor and DRCk0 (n) is the average data rate of the
k th UE at time n. The k,j in equation 3.1 is used to modify the priority value
depending on on whether the k th UE is at the centre or edge of the cell and whether
the j th RB is a high or low power RB. The value of k,j can be adjusted to ensure
that despite having lower average data rates UEs on the cell edge are more likely to
get scheduled on high power RBs.
The aSerFR scheme utilises the same proportional fair scheduling algorithm dis-
cussed above but modifies its behaviour based on a number of parameters. This
allows the aSerFR scheduler to be optimised by an GA for use in particular network
environments. The optimisation is performed by changing a number of parameters
50
each of which are described below.

As with SerFR UEs are classified as either edge or centre users. This classification
of UEs is based on a signal quality threshold which is one of the parameters which can
be optimised by the GA. Since edge users typically experience greater interference
from neighbouring cells it is a safe assumption that those users with the poorest
signal quality are edge users, see [50, 40, 54]. This signal quality threshold ranged
from -15dB to +15dB and as such was encoded using real value encoding. To mutate
this value uniform mutation was used to generate a uniform random value between
the upper and lower bounds of the desired range.
The eNB will transmit at full power to edge users and at a reduced power to
central users. The level of reduced power which is used for the central users is
another parameter of the scheduler which can be optimised by the GA. The edge
users are transmitted to using 100% of the available power while central users will
be transmitted with only some proportion of that depending on the parameter value
set by the GA. This reduced power factor was represented in the chromosome using
a real value between 1.0, representing full power, and 0.0 representing no power and
as such was encoded using real value encoding. Uniform mutation was again used
to generate new uniform random values for this gene during mutation.
In a similar fashion to SerFR a certain number of the available RBs are transmit-
ted using full power and the remainder are transmitted at a reduced power level. The
number of RBs which are transmitted at full power is also a configurable parameter
in this scheduler. It is expected that the number of RBs required by the edge users
will vary depending on the traffic conditions. The proportion of full power RBs is
represented as a real value ranging between 1.0, representing the state that all RBs
are transmitted a full power, and 0.0, representing the state that none of the RBs
are transmitted at full power. As with the previous real value encoded parameters
uniform mutation was used to generate a uniform random value for mutation.
The scheduling mechanism adopted by aSerFR is the same Proportional Fair
(PF) scheduler used in the SerFR scheme, with one modification. On occasion
the scheduler will schedule its users randomly rather than using the PF scheme.
51
Short-name Description Values

RandomT rigger The chance of the scheduler oper- 0.0 - 1.0
ating randomly
P roportionF ullP owerRBs The proportion of RBs to trans- 0.0 - 1.0
mit at full power
ReducedP owerF actor The proportion of full power to 0.0 - 1.0
use when transmitting to central
users
EdgeT hreshold The threshold at which UEs are -15dB -
considered edge users 15dB
RBAllocationScheme The scheme used for determining Orthogonal
which RBs will be transmitted at or
full power Random
Table 3.1: The GA parameters for the aSerFR scheduler.
The probability of this random scheduling occurring is governed by a parameter

which will be optimised by the GA. The probability of this random behaviour being
triggered was modelled as a real value ranging from 1.0, representing continuous
random scheduling, and 0.0, representing no random scheduling at all. As with the
previous real values uniform mutation was used to generate new random values for
this parameter.
The other difference between the SerFR scheme and the aSerFR scheme is the
orthogonality of the full power RBs. The SerFR scheme assigns these in a strictly
orthogonal manner such that neighbouring cells transmit at full power on different
RBs. The aSerFR supports the orthogonality of the full power RBs but also al-
lows the option of randomly distributing these such that there is a possibility that
neighbouring cells will transmit at full power on some of the same RBs. The orthog-
onality of the RB allocation scheme is represented in the chromosome as a single
bit where 1 represents the full power RBs being allocated in a strictly orthogonal
manner and 0 represents them being allocated by random distribution. When this
gene was mutated the bit was simply flipped.
These five parameters govern the behaviour of the aSerFR scheduler. They
are listed in Table 3.1 for easy reference, along with their possible values, and are
referenced using the shortened form shown for the remainder of this document.
52
3.2 LTE Network Simulator
In order to model the behaviour of each of the ICIC schemes and allow for the de-
velopment of new ones an LTE network simulator was built. The source code for the
simulator is available at http://downthewire.co.uk/projects/jLTE-Simulator/.
The design and operation of the simulator are described in the follow sections.
3.2.1 Overview of Simulator Execution
The simulator is designed to simulate specific aspects of the Radio Access Network
(RAN) of an LTE network. In particular it is mainly concerned with the calcu-
lation of interference, data rates and the scheduling of RBs to UEs. In addition
the simulator is only concerned with the downlink direction and therefore consid-
eration of the uplink direction is outside the scope of this research. This is due to
two considerations. The first is the additional computational cost of calculating the
interference in the uplink direction. The computation cost of calculating the inter-
ference in either direction is directly related to the number of entities which may
interfere with each other. In the downlink direction there are only 19 eNBs whose
signals may cause interference, while in the uplink direction there can be as many as
1150 UEs causing interference. This represents a significant increase in the compu-
tational cost of calculating the uplink interference. The second factor in focusing on
the downlink is its greater relative importance to network users. Downlink speeds
are typically more important in the applications which network users value. For
example streaming video and online gaming can require significant downlink speeds
and there is growing demand for these services, whereas there is very little demand
for uplink intensive services.
The network to be simulated consists of 19 eNBs each with 3 sectors as recom-
mended by 3rd Generation Partnership Project (3GPP) [5]. A number of UEs are
then randomly distributed across the network area. The specific number of UEs
generated can be varied in order to simulate networks under different levels of user
density.
At a high level the simulator simply runs the following algorithm:
53
initialise the network
setup the network elements: eNBs, sectors, UEs
determine which sector is to serve which UE
for each iteration:
calculate the downlink power between each sector and each UE
schedule the RBs to UEs using the given scheduling algorithm
calculate the signal power and interference experienced by each UE
calculate the datarate of each UE
Initialisation A number of initialisation steps must take place before the simu-
lation can begin. The most important aspect of this is the creation of a channel
between each UE and each sector. This channel models the quality of the radio
channel between a UE and a sector by calculating a number of things such as the
pathloss experienced by the signal between the sector and the UE, the shadowing
experienced, and the location of the UE with respect to the sector. In order to
determine which sector should serve a given UE the channel between the UE and
each sector is calculated and the channel with the best downlink gain is chosen.
The pathloss, which also takes shadowing effects into account, is modelled using
the propagation model detailed in Section 4.5.2 of [6].
3.2.2 Datarate Model
The datarate model used by the simulator to calculate the datarate of individual
UEs is defined in [6]. Using this model the datarate of a given UE can be calculated
from its Signal to Interference plus Noise Ratio (SINR) using the following equation:

0 if SIN R < SIN Rmin

datarate(bps/Hz) = .S(SIN R) if SIN Rmin < SIN R < SIN Rmax

dataratemax if SIN R > SIN Rmax

54
Where: S(SIN R) is the Shannon bound: S(SIN R) = log2 (1 + SIN R) bps/Hz.

The parameter, which represents implementation losses, takes the value 0.6, while
the SIN Rmin is -10dB and SIN Rmax is 25db. The maximum datarate which can
be achieved under this model is 4.4bps/Hz.
Using this model we can determine the maximum datarate that can be achieved
in our simulator. The total available bandwidth in our simulation is 18,000,000 Hz
(100 RBs 180,000 Hz per RB). Since the maximum datarate is 4.4bps/Hz under
this model, the maximum total datarate is 75.5 Mbps (4.4 bps/Hz 18,000,000
Hz).
3.2.3 Channel Estimation Model
In order to model the wireless channel between the UEs and eNBs to a reasonable
degree of accuracy it is necessary to implement a channel model. There are two
main types of channel model which can be used:
empirical channel models
statistical channel models
Empirical channel models are based on empirical measurements taken in a va-

riety of wireless propagation environments. They are considerably more detailed
than statistical models as they calculate precisely a large number of paths which a
transmission may take along with details such as the angle of departure, angle of
arrival and phase change of each individual path from the eNB to the UE.
While empirical models do provide a more realistic approach to simulating wire-
less channels the level of detail of the calculations required to utilise them in our
simulator proved to be slow for practical purposes.
Statistical channel models are generally simpler and are therefore more useful for
simulations. The simulator makes use of a statistical model know as the Rayleigh
Fading model which has been found to approximate closely the channel capacity in
urban areas [37, 9, 89, 87, 43, 56].
Due to the computation load in calculating the gain values resulting from the
channel gain model these are produced in batch before any simulations are run. This
55
reduces the time required to run the simulations since the fading values do not need
to be calculated each time a simulation is run.
The speed of the UEs mainly effects the fading values generated by the channel
model. The faster a UE is moving the faster their signal quality will vary. For
example see Figure 3.1.
0
Pathloss (dB)
10
3 km/h
30 km/h
120 km/h
150 20 40 60 80 100
Time (ms)
Figure 3.1: The effect of UE speed on the variation of signal strength.
3.2.4 Traffic Model
The UEs modelled by the simulator are categorised into four traffic groups: Light,
Medium, Heavy and Full. The category a UE is placed in determines both the
probability that there is traffic to be transmitted to it each iteration but also the
number of RBs which are to be scheduled for that UE. The specific values for these
parameters can be found in Table 3.2.
Traffic Category Probability of Data Max RBs to transmit

Light 20% 5
Medium 50% 15
Heavy 80% 100
Full 100% 100
Table 3.2: The probabilities associated with the traffic classes.
The number of UEs grouped into each class is then determined by the overall
traffic profile used by the simulator. There are four profiles: Light, Mixed, Heavy
56
and Full. The proportion of UEs in each traffic class is shown in Table 3.3.
This traffic model is not based on any empirical data about traffic loads but was
instead designed as an approximation of the various levels of load that could be
placed on the resources of the network. From the perspective of the ICIC problem
the particular type of traffic, whether voice data or packet data, and its latency
requirement are not of interest. Rather the number of RBs required to serve the
traffic is the most important consideration. This traffic model captures the fact
that, from the networks perspective, less demanding UEs will require less RBs to
be served while more demanding UEs will require more RBs to be served. This
is reflected in the Full traffic mode which is analogous to the the full buffer traffic
model found in the literature [26, 22, 6, 103].
Traffic Category Proportion Proportion Proportion Proportion

of Light of Medium of Heavy of Full UEs
UEs UEs UEs
Light 70% 20% 10% 0%
Mixed 20% 60% 20% 0%
Heavy 0% 0% 100% 0%
Full 0% 0% 0% 100%
Table 3.3: The proportion of UEs created in each traffic class.
The aim of this traffic model is to allow for a fuller understanding of the perfor-
mance of the scheduling algorithms by seeing how their performance changes with
network load. As such this traffic model is not designed to be a realistic model of
user behaviour. Rather it is intended to be a simple heuristic approach to modelling
the level of load placed on the the network and therefore the level of contention for
the radio resources. For example the heaviest possible load is when all the UEs on
the network constantly need to be scheduled which is represented by the Full traffic
level. In this case there is maximum contention for the networks radio resources
and the scheduling algorithms will be operating under the most constrained envi-
ronment. The other levels reflect network environments in which the contention for
resources is decreasing.
57
3.2.5 Performance Metrics
During its operation the simulator gathers detailed statistics about the performance
of the network. These metrics are used to evaluate and compare the relative perfor-
mance of each of the ICIC algorithms that run on the network.
The most important metrics for this research are the average UE throughput
and the 5th percentile UE throughput. These are the measurements typically used
in the literature to evaluate ICIC schemes [39, 103, 48, 40]. It is important to note
that both of these metrics are calculated using the average throughput of the UEs
during the course of the simulation. The 5th percentile UE throughput is obtained
by sorting the UEs by their average throughput. The average is then taken of the
average throughput of the worst performing 5% of UEs. In the literature the 5th
percentile UE throughput is used as a proxy for measuring the performance of the
cell edge UEs. It is assumed that those UEs which experienced the worst throughput
are those that are most affected by interference which are most likely to be those
UEs at the cell edge [50, 40, 54].
It is important to note that these two metrics are competing constraints. This is
due to the fact that there are a finite number of RBs which can be scheduled during
each iteration and since the amount of data transmitted across each RB varies with
the network conditions of the UEs to which it is scheduled.
As the 5th percentile UE throughput increases it is at the expense of the average
UE throughput. In order to increase the 5th percentile throughput more RBs must
be assigned to the cell edge UEs. These RBs will transmit less information to
those UEs than they would to cell centre UEs due to the Modulation and Coding
Scheme (MCS). Thus to increase the throughput of the edge UEs necessarily involves
decreasing the throughput of the central UEs.
3.2.6 Validation
Unfortunately there are no standardised benchmarks against which to compare our

simulator. Therefore in order to verify that our simulator behaves as expected a
number of simple test cases were created and the output of the simulator compared
58
to that of the theoretical output.

In order to verify the implementation of our simulator a number of tests were run.
For example it was noted earlier that the maximum datarate that could be achieved
with the simulator was 75.5 Mbps. To verify that the simulator was indeed capable
of achieving this datarate a simple network setup was created which consisted of
a single eNB and a single UE. This provided near optimal network conditions and
allowed the UE to achieve the maximum datarate. This confirmed that the model
implemented by the simulator was behaving as expected.
The other main verification step taken was the comparison of the performance of
the baseline schedulers. While a direct comparison using the absolute performance
was impossible (since there are no standardised simulator implementation, or even
standardised simulation parameters) a relative comparison was done. Regardless
of the implementation of the simulator the relationship between the performance
of several of the baseline schedulers is well understood and should be observable.
For instance the maximum rate scheduler should always achieve the best average
UE throughput but will always perform very poorly in terms of 5th percentile UE
throughput. The SFR and SerFR schedulers should generally outperform the other
schedulers in terms of 5th percentile UE throughput since they are designed to
optimise this metric [103, 46, 48].
3.3 Baseline LTE Schedulers
These existing schedulers are described in the literature and are described briefly
here. They are used to compare against the results from the aSerFR scheduler.
3.3.1 Random Scheduler
The Random scheduler is a very simple scheduler which is designed to show the effect
of scheduling when channel conditions are not taken into account by the scheduler.
The schedulers used as part of the ICIC schemes discussed previously all make use of
UE feedback regarding the channel conditions to inform their scheduling decisions.
The random scheduler does not take into account such information when scheduling
59
RBs to UEs. It is used in the experiments in this and the following chapters as
a benchmark against which to compare those schedulers which are sensitive to the
channel conditions experienced by the UEs.
3.3.2 Round Robin Scheduler
Round robin scheduling [39] assigns the available RBs in a fair though simplistic way.
It is fair in that the same number of RBs are made available to all the UEs. However
due to the fact that it does not take into account the variations in channel quality
it can also be considered an unfair way of scheduling since it will not provide the
same service quality to all the UEs. This is because of the fact that those UEs which
experience better signal quality will be able to use their RB more efficiently thus
enjoying a better service quality than that those UEs which experience degraded
signal quality.
Since round robin scheduling does not take advantage of the channel quality
conditions it will lead to lower system performance. It does however guarantee that
no UEs are starved of resources since all UEs will get an equal share.
3.3.3 Maximum Rate Scheduler
Maximum rate scheduling (also known as Max-C/I scheduling) [39] is a type of

scheduling which takes into account the channel conditions being experienced by
each of the UEs. Since the channel quality experienced by UEs within a cell typically
varies independently due to effects like multi-path fading (fading caused by the
reception of multiple copies of the same signal at the UE) there is generally at least
one UE whose channel quality is near its peak. By scheduling the UE with the best
channel quality the eNB maximises its throughput because the UEs with the best
channel quality can make the most efficient use of the radio resources.
Maximum rate scheduling however cannot be considered fair in many circum-
stances. Whether maximum rate scheduling can be considered fair depends primar-
ily on the causes of the channel quality variation. If all UEs have a similar channel
quality which varies only due to effects like fast fading or multi-path fading then
60
they will experience a very similar average throughput. However if the variations
in channel quality are due to factors such as distance, shadow fading or fading due
to speed then it is very unlikely that maximum rate scheduling will behave fairly.
If there are UEs whose average channel quality is always less than others then they
will never be scheduled and will essentially be starved of radio resources resulting
in zero throughput.
3.3.4 Proportionate Fair Scheduler
PF [39, 65] scheduling attempts to find a balance between the approaches taken by
maximum rate scheduling and round robin scheduling. Maximum rate scheduling
makes the most efficient use of the radio resources available at the expense of the
performance of UEs with poor channel conditions while round robin is scrupulously
fair in its allocation of radio resources but does not deliver equal throughput to all
UEs. PF scheduling makes use of the channel quality information available while
also being fair in its distribution of radio resources. It does this by scheduling the
UEs with the relatively best channel quality conditions. It does this by monitoring
the channel conditions of each UE over a particular time window. When making a
scheduling decision the relative channel quality for each UE is calculated by com-
paring the average channel quality of that UE to its current channel quality. This
approach means that even those UEs who experience significantly worse channel
conditions in general will receive radio resources when their channel quality peaks
relative to its own average.
3.3.5 Soft Frequency Reuse Scheduler
As discussed in Section 2.4.3.1 SFR is an ICIC scheme that attempts to reduce in-
terference in neighbouring cells by dividing up the available spectrum into a number
of sub-bands which are served with different power levels. The full power sub-bands
are reserved for use by the cell-edge UEs while the remaining reduced power sub-
bands are used to serve the cell-centre UEs. These sub-bands are allocated to cells
on an orthogonal basis such that no neighbouring cell will have the same sub-bands
61
assigned to its cell-edge UEs.

While no specific scheduler is discussed as part of this ICIC scheme a modified
PF scheduling algorithm has been proposed [103] which takes into account the status
of the UE being scheduling, whether it is a cell-edge or cell-centre UE, and which
sub-band it should be scheduled on.
3.3.6 Softer Frequency Reuse Scheduler
As discussed in Section 2.4.3.2 SerFR is an ICIC scheme which is built on the SFR
scheme. Like SFR, it divides the available frequency in sub-bands some transmitted
with full power and some with a reduced power factor; unlike SFR it is more flexible
in the allocation of these sub-bands to UEs.
The scheduler used by the SerFR scheme is also a modified PF algorithm which
takes into account the status of the UE as either cell-edge or cell-centre and also the
power level of the sub-band.
3.4 Experiments
These experiments were designed to compare the performance of the optimised

aSerFR scheduler, developed as part of this research, with that of the baseline
schedulers described in Section 3.3.
In order to compare each of the algorithms thoroughly the experiment was run
on a number of different network environments. In practice this meant a different
number of UEs (see Table 3.4), different movement speeds of the UEs (see Table
3.5) and also different traffic patterns of the UEs (see Table 3.6). An experiment
was run for each different combination of these parameters, resulting in 27 different
network environments.
The number of UEs in each network environment was chosen to provide a reason-
able number of UEs per cell. The simulated network was composed of 19 cells (see
Table 3.7) and so 285 UEs represent lightly loaded cells with an average of only 15
UEs per cell, whereas the network environment with 1150 UEs represents a network
where the cells are very heavily loaded. The traffic levels represent both the proba-
62
bility and amount of data arriving at an eNB for a given UE as described in Section
3.2.4. The UE speeds were chosen based on the standard simulation parameters set
out by 3GPP [5].
Number of UEs Average number of UEs per cell

285 15
680 35
1150 60
Table 3.4: The possible number of UEs in each network environment and the result-
ing average number of UEs per cell
Traffic Levels
Light
Mixed
Heavy
Full
Table 3.5: The possible traffic levels of the UEs in each network environment
Speed (km/h)
3
30
120
Table 3.6: The possible speeds of the UEs in each network environment
In order to evaluate the performance of the baseline algorithms each of them

was run ten times for each network environment with different random seeds. The
results for each of these runs were then averaged and plotted in the graphs seen in
Section 3.5.
To evaluate the performance of the aSerFR scheduler a Multi-Objective Genetic
Algorithm (MO-GA) was run in order to optimise its parameters for each network
environment. The MO-GA was designed to optimise two metrics: 5th percentile
throughput, and average throughput. The MO-GA produces a number of results
which together form a Pareto front based on these two metrics. The solutions
produced can all be considered equal in terms of fitness although they will have
different characteristics in terms of their 5th percentile and average throughput
values. This makes explicit the trade-offs that must be made between 5th percentile
throughput and average throughput since they are competing constraints.
63
The simulator described above was used to evaluate the chromosomes produced
by the MO-GA. Due to the time required to run the MO-GA (approximately 5 days)
it was only run once for each network environment. However, in order to ensure
that the simulator produced a representative result for any given chromosome, each
chromosome was evaluated 10 times using a different random seed each time. These
10 values were then averaged to give the fitness value for that chromosome which
was used by the MO-GA.
3.4.1 Simulation Parameters
The simulation parameters shown in Table 3.7 are mainly taken from [39, 5].
Parameter Value
Number of eNBs 19
Number of UEs 252, 680, 1150
UE speed 3kmh, 30kmh, 120kmh
Data generation Light, Mixed, Heavy Traffic profiles
Distance attenuation L = 128.1 + 37.6 log(distance) [6]
Shadow fading Log normal, 8dB standard deviation
Multi-path fading Rayleigh fading
Cell layout Hexagonal grid, 3-sector sites, 57 sectors in total
Cell Radius 500m inter-site distance
Simulation Time 500 ms
Table 3.7: The parameters used for the simulator in experiment 1.
The layout of the 19 cell network can be seen in Figure 3.2.

Note that the simulation run time shown in Table 3.7 was arrived at after a pilot
study had been conducted to determine the required number of iterations which
were needed to achieve a stable and accurate result. Each of the baseline algorithms
was run under a particular network environment (1150 UEs, full traffic, 120km/h)
for 2500 iterations (2500 ms of network time).
As can be seen from Figures 3.3 and 3.4 while the average throughput did not
stabilise after 500 iterations in all cases, the 5th percentile throughput did reach a
stable state after 500 iterations. Since the 5th percentile throughput is the more
important measure for determining the effectiveness of an ICIC solution and since
it would have been prohibitive to run the simulation for longer, 500 iterations was
chosen as a reasonable simulation time.
64
Figure 3.2: The 19 cell network used in experiment 1.
0.20
0.15
Tput (Mbps)
0.10
SerFR
0.05 Random
SFR
MaxCI
ProportionalFair
0.000 500 1000 1500 2000 2500
Iterations
Figure 3.3: The convergence of 5th percentile UE throughput for each algorithm.
3.4.2 MO-GA Parameters
The parameters for the MO-GA can be seen in Table 3.8.

The population size and the number of generations were limited as a result of
the number of fitness evaluations required to complete an experiment along with
65
2.0
1.5
Tput (Mbps)
1.0
SerFR
0.5 Random
SFR
MaxCI
ProportionalFair
0.00 500 1000 1500 2000
Iterations
Figure 3.4: The convergence of average UE throughput for each algorithm.
Parameter Value
Population Size 10
Generations 100
Mutation Strength 1
Mutation Operator Point Mutation
Crossover Operator Uniform Crossover
Table 3.8: The parameters used for the MO-GA in experiment 1.
the time taken for each fitness evaluation. With a population of 10 chromosomes
and 100 generations a total of 3100 fitness evaluations were required to complete
each experiment: 100 evaluations to initialise the population, along with 30 every
iteration, 10 for mutation and 20 for the child chromosomes produced by uniform
crossover i.e. 100 + (100 (10 + 20)) = 3100
Since the time to execute a single fitness evaluation averaged 3 minutes this
results in each MO-GA taking approximately 7 days. Since there were 27 network
environments this meant that 27 MO-GAs had to be run. It was decided that this
was the upper limit on the amount of time that could be reasonably spent running
experiments and so the population and generations were fixed at these values.
The mutation strength indicates how many genes in each chromosome are to be
66
mutated during each iteration of the MO-GA. For further details about the operation
of point mutation and uniform crossover refer to Section 2.5.
3.5 Results
Due to the fact that a number of algorithms are being compared a mixed ANOVA
was run on the results in order to determine the statistical significance of the differ-
ences between the algorithms. The between-subjects factor was algorithm and the
within-subjects factors were traffic level, speed and number of UEs. This resulted
in a 6 (algorithm) 4 (traffic level) 3 (speed) 3 (number of UEs) ANOVA.
The average UE throughput and the 5th percentile UE throughput were analysed
separately for clarity.
3.5.1 Overall Analysis
The overall ANOVA for average UE throughput demonstrated a significant main

effect for algorithm, (F(5,5191) = 1447.98, p 0.001). The F value captures the
ratio of variance explained by the factor algorithm to unexplained variance, meaning
that a lot of the variance can be attributed to the algorithm. This means that across
all conditions the algorithm has a significant effect on the average throughput.
This analysis only shows that different algorithms have different effects on the
average throughput. To see which algorithms differed from each other and how sig-
nificant the differences between each algorithm were a Bonferroni post-hoc analysis
was carried out. This analysis performs multiple comparisons between each of the
algorithms and adjusts the significance value of any of the differences to account for
the fact that there were multiple comparisons. For each comparison the Bonferroni
analysis provides a mean difference and the significance of that difference.
As suggested by Figure 3.5 the maximum rate algorithm achieved a significantly
higher average UE throughput when compared to all the other algorithms with an
average mean difference between the maximum rate algorithm and each of the re-
maining algorithms of 498kbps (all p 0.001). The next best performing algorithms
were SerFR, random scheduler, the PF algorithm and the SFR algorithm which were
67
3.5
Median
3.0 Mean
Average Tput (Mbps) 2.5
2.0
1.5
1.0
0.5
0.0
ir
CI
dom
SFR
FR
R
alFa
eSF
Max
Ser
Ran
ptiv
tion
Ada
por
Pro
Figure 3.5: Boxplot showing the average UE throughput of each algorithm across
all the network environments and traffic levels.
statistically indistinguishable. The aSerFR algorithm performed worst in terms of

average UE throughput with an average mean difference between it and the other
algorithms, excluding the maximum rate algorithm, of 77kbps (all p 0.001).
This overall ANOVA for average UE throughput also demonstrated a number of
interaction effects. The interaction effects mean that different algorithms perform
differently under different network environments. The strongest interaction was be-
tween algorithm and traffic level (F(15,5191) = 275.21, p 0.001). This called for
follow-up simple effects analyses within each level of traffic. These subsequent anal-
yses (ANOVAs investigating the effect of algorithm on average throughput within
each traffic level) are detailed in separate sections below.
The same ANOVA was carried out to investigate the effects of algorithm and
environment on 5th percentile UE throughput. This demonstrated a significant
main effect for algorithm (F(5,5191) = 1801.82, p 0.001). The Bonferroni post-
hoc analysis was then carried out to determine the relative performance of each
of the algorithms. The ranking of the algorithms in terms of 5th percentile UE
throughput can be seen in Figure 3.6. The performance of each of the algorithms was
significantly different from each other, meaning that each algorithm was significantly
68
0.7
Median
0.6 Mean
5th Percentile Tput (Mbps) 0.5
0.4
0.3
0.2
0.1
0.0
ir
CI
dom
SFR
FR
R
alFa
eSF
Max
Ser
Ran
ptiv
tion
Ada
por
Pro
Figure 3.6: Boxplot showing the 5th percentile UE throughput of each algorithm
across all the network environments and traffic levels.
better than the next best one. As can be seen the aSerFR outperformed each of the
other algorithms, with an average mean difference of 59kbps (all p 0.001).
Once again the strongest interaction effect was between algorithm and traffic level
calling for separate analyses within each traffic level as detailed below (F(15,5191)
= 243.19, p 0.001).
3.5.2 Light Traffic
The results of each algorithm when run under the light traffic network environ-
ments are plotted in Figure 3.7. It can be seen from this figure that the average
UE throughput of most of the algorithms are very similar, in fact the average UE
throughput of the SFR, SerFR, PF and random algorithms are not statistically
distinguishable.
A one-way ANOVA looking at the effect of algorithm on average UE throughput
within light traffic network environments revealed a significant effect of algorithm
(F(5,1534) = 6.02, p 0.001). As suggested by Figure 3.7 the Bonferroni post-hoc
comparisons showed that the only significant difference was between the maximum
rate algorithm and the aSerFR algorithm with the maximum rate algorithm per-
69
1.6
1.4 Median
Mean
1.2
Average Tput (Mbps)
1.0
0.8
0.6
0.4
0.2
0.0
ir
CI
dom
SFR
FR
R
alFa
eSF
Max
Ser
Ran
ptiv
tion
Ada
por
Pro
Figure 3.7: Boxplot showing the average throughput of each algorithm across all the
network environments under light traffic.
forming significantly better than aSerFR (mean difference = 138kbps, p = 0.01).

No other differences in average UE throughput were significant.
Figure 3.8 shows the average performance of each algorithm in terms of the 5th
percentile UE throughput. It can be seen that the 5th percentile throughput of
the aSerFR algorithm outperformed the other algorithms. Indeed the average 5th
percentile throughput was over 100% better than the other algorithms. It can also
be seen that the results obtained by the SFR algorithm, while not achieving the
performance of the aSerFR algorithm, do outperform those achieved by the other
algorithms.
A one-way ANOVA looking at the effect of algorithm on 5th percentile UE
throughput within light traffic network environments revealed a significant effect
of algorithm (F(5,1534) = 225.07, p 0.001). As suggested by Figure 3.8 the
Bonferroni post-hoc comparisons showed that the aSerFR outperformed each of
the other algorithms, with average mean difference of 67kbps (all p 0.001). The
maximum rate algorithm performed significantly worse than all the other algorithms,
with an average mean difference of 36kbps (all p 0.001). The remaining algorithms
were statistically indistinguishable.
70
0.20
Median
Mean
5th Percentile Tput (Mbps) 0.15
0.10
0.05
0.00
ir
CI
dom
SFR
FR
R
alFa
eSF
Max
Ser
Ran
ptiv
tion
Ada
por
Pro
Figure 3.8: Boxplot showing the 5th percentile throughput of each algorithm across
all the network environments under light traffic.
The results achieved by each algorithm on the light traffic network environments
can be seen in Figure 3.9. This figure shows the nine different network environments
run under light traffic. These are combinations of three UE density levels (285 UEs,
680 UEs, 1150 UEs) and three UE speeds (3km/h, 30km/h, 120km/h). These re-
sults show that although the aSerFR algorithm performed best on average in terms
of 5th percentile UE throughput there are variations in performance depending on
the specific network environment. To further investigate the significance of differ-
ences between algorithms under specific network environments unpaired t-tests were
carried out where appropriate.
It can be seen that in every network environment the aSerFR algorithm outper-
forms the other algorithms by at least 40kbps (Plot 8) in terms of 5th percentile
throughput and in some cases up to 125kbps (Plot 0). In all cases the 5th percentile
throughput is improved by at least 100% and in some cases over 300% (Plot 0).
However this increase in 5th percentile throughput comes at a cost to the average
UE throughput. The aSerFR chromosomes that result in the best performance in 5th
percentile throughput also show a very significant reduction in average throughput,
up to 70kbps in some cases.
71
0.18 Plot: 0, Speed: 3km/h, UEs: 285 0.16 Plot: 1, Speed: 3km/h, UEs: 680 0.10 Plot: 2, Speed: 3km/h, UEs: 1150
0.16 0.14
0.08
5th Percentile Tput (Mbps)

0.14 0.12
0.12 0.10 0.06
0.10
0.08
0.08
0.06 0.04
0.06
0.04 0.04
0.02
0.02 0.02
0.000.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.00
0.25 0.30 0.35 0.40 0.45 0.50
Average Tput (Mbps) Average Tput (Mbps) Average Tput (Mbps)
0.16 0.14 0.10

0.14 0.12
0.12 0.10 0.08
0.10
0.08 0.06
0.08
0.06 0.06 0.04
AdaptiveSFR
0.04
MaxCI
0.04 Pareto front
0.02
0.02 0.02 ProportionalFair
0.000.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00 0.00 Random
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.25 0.30 0.35 0.40 0.45 0.50 SFR
SerFR
0.16 0.14 0.07

0.14 0.12 0.06
0.12 0.10 0.05
0.10
0.08 0.04
0.08
0.06 0.06 0.03
0.04 0.04 0.02
0.02 0.02 0.01
0.000.4 0.6 0.8 1.0 1.2 1.4 0.00
0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.00
0.30 0.32 0.34 0.36 0.38 0.40 0.42 0.44 0.46
Figure 3.9: The 5th Percentile UE throughput vs average UE throughput in light

traffic across nine network environments. Across the nine sub-plots the number of
UEs increases from left to right while the UE speed increases from top to bottom.
In some cases the aSerFR algorithm performs better in both 5th percentile
throughput and average throughput. This is most clearly seen when there are 285
UEs on the network. For example when there are 285 UEs on the network with
speed of 3 km/h (Plot 0) the aSerFR improves on the SFR algorithm (the next best
performing algorithm) by 30kbps in 5th percentile throughput with no significant
change in average throughput. An independent t-test showed the difference between
the aSerFR and SFR algorithms in 5th percentile throughput to be statistically very
significant when there are 285 UEs on the network with a speed of 3km/h (t(18) =
60.3669, p = 0.0001).
It can also be seen from these results that in almost every case the results pro-
72
duced by the aSerFR scheme form a Pareto front which dominates the results pro-
duced by all the other schemes. This means that for each result produced by the
baseline algorithms there was a result produced by the aSerFR which improved on
both the 5th percentile throughput and the average throughput. The exception
to this is the maximum rate scheduler which achieves the best average throughput
when the network is under the heaviest UE load.
3.5.3 Mixed Traffic
As with the light traffic results averages of the results from each algorithm under
mixed traffic were taken. The average of the average UE throughput can be seen
in Figure 3.10. The average performance of each algorithm under mixed traffic
networks displays a similar pattern to the average performance under light traffic
loads. In particular the maximum rate algorithm again achieves an improved average
UE throughput, with the aSerFR algorithm performing below the other algorithms.
The remaining algorithms performed at a very similar level.
1.8
1.6 Median
1.4
Mean
Average Tput (Mbps)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
r
CI
dom
SFR
FR
R
lFai
eSF
Max
Ser
a
Ran
ptiv
tion
Ada
por
Pro
Figure 3.10: Boxplot showing the average throughput of each algorithm across all
the network environments under mixed traffic.
As with the light traffic results algorithm was found to have a significant effect on
average /glsue throughput (F(5,1406) = 3.07, p = 0.001) by the one-way ANOVA. As
73
suggested in Figure 3.10 the only significant difference found by the Bonferroni post-
hoc analysis was between the maximum rate algorithm and the aSerFR algorithm.
The maximum rate algorithm performing significantly better than aSerFR (mean
difference = 148kbps, p = 0.002). No other differences in average UE throughput
were significant.
The results in terms of 5th percentile UE throughput can be seen in Figure
3.11. As with the average UE throughput results these are also very similar to
those achieved under light traffic loads. The aSerFR algorithm achieves the best
performance by a large margin (over 100% or 40kbps), with the SFR algorithm
achieving the next best result. The remaining algorithms, with the exception of the
maximum rate achieve a very similar level of performance, while the maximum rate
algorithm performs much worse than the others.
0.25
Median
0.20 Mean
0.15
0.10
0.05
0.00
air
CI
dom
SFR
FR
R
eSF
Max
lF
Ser
ona
Ran
ptiv
ti
Ada
por
Pro
all the network environments under mixed traffic.
Algorithm was also found to have a significant effect on the 5th percentile UE
throughput (F(5,1406) = 297.31, p 0.001). with the post-hoc analysis showing
that the aSerFR outperformed each of the other algorithms, with average mean
difference of 109kbps (all p 0.001). The maximum rate algorithm performed sig-
nificantly worse than all the other algorithms, with an average mean difference of
74
47kbps (all p 0.001). The remaining algorithms were statistically indistinguish-

able.
The individual results for each network environment can be seen in Figure 3.12
which shows the results for each algorithm when run under mixed traffic. The
results show that, as with light traffic, in each network environment the aSerFR
algorithm performed better than the baseline algorithms in terms of 5th percentile
throughput with an increase of between 5kbps (Plot 8) up to 169kbps (Plot 0). The
improvements are particularly pronounced when either 285 UEs or 680 UEs are on
the network. When 1150 UEs are on the network the improvements in 5th percentile
throughput are much smaller (approximately 6kbps).
0.14 0.040
0.20

0.12 0.035
0.10 0.030
0.15
0.025
0.08
0.020
0.10 0.06 0.015
0.04 0.010
0.05
0.02 0.005
0.000.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.000
0.35 0.40 0.45 0.50 0.55 0.60
0.12 0.040
0.20
0.035
0.10
0.030
0.15 0.08 0.025
0.06 0.020
0.10 AdaptiveSFR
0.015 MaxCI
0.04
0.05 0.010 Pareto front
0.000.6 0.00 0.000 Random
0.8 1.0 1.2 1.4 1.6 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.35 0.40 0.45 0.50 0.55 0.60 SFR
0.25 Plot: 6, Speed: 120km/h, UEs: 285 0.14 Plot: 7, Speed: 120km/h, UEs: 680 0.045 Plot: 8, Speed: 120km/h, UEs: 1150 SerFR
0.12 0.040
0.20
0.035
0.10
0.030
0.15 0.08 0.025
0.06 0.020
0.10
0.015
0.04
0.05 0.010
0.02 0.005
0.000.6 0.8 1.0 1.2 1.4 1.6 0.00
0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.000
0.35 0.40 0.45 0.50 0.55 0.60
Figure 3.12: The 5th Percentile UE throughput vs average UE throughput in mixed

traffic environments. Across the nine sub-plots the number of UEs increases from
left to right while the UE speed increases from top to bottom.
75
The improvements in 5th percentile throughput are again associated with a re-
duction in average throughput. The larger the gains in 5th percentile throughput
the greater the reduction in average throughput. When the gain in 5th percentile is
highest (169kbps in Plot 0) the loss in average throughput is over 800kbps.
It can also be seen that in most cases the results obtained from the aSerFR
scheme produced a Pareto front which dominated the solutions produced by the
other schemes. The exceptions to this are the maximum rate scheduler which
achieves the best average throughput and the SFR scheme contributes to the Pareto
front in particular as the network comes under heavier UE load.
3.5.4 Heavy Traffic
The graph of the average throughput for each algorithm under Heavy traffic can be
seen in Figure 3.13. It can be seen that the maximum rate algorithm achieves the
highest performance in terms of the average UE throughput by over 500kbps.
2.5
Median
2.0 Mean
Average Tput (Mbps)
1.5
1.0
0.5
0.0
r
CI
dom
SFR
FR
R
lFai
eSF
Max
Ser
a
Ran
ptiv
tion
Ada
por
Pro
the network environments under heavy traffic.
Algorithm was again found to be a significant main effect (F(5,1190) = 34.86, p

0.001). As suggested by Figure 3.13 the Bonferroni post-hoc comparisons showed
that the maximum rate algorithm performed significantly better than each of the
76
other algorithms (average mean difference = 702kbps, all p 0.001). All the re-
maining algorithms are statistically indistinguishable in terms of their average UE
throughput.
Figure 3.14 shows the results of each algorithm in terms of the 5th percentile
UE throughput. In this case both the SFR and SerFR algorithms performed best.
They improved on the results achieved by the other algorithms by at least 25kbps.
The aSerFR and random algorithms both achieved a similar level of performance in
terms of their 5th percentile UE throughput. The performance of the PF algorithm
was about 30kbps less than that of the aSerFR and random algorithms. Finally the
maximum rate algorithm achieved an average of 0kbps in terms of 5th percentile
throughput.
0.5
Median
0.4 Mean
0.3
0.2
0.1
0.0
r
CI
dom
SFR
FR
R
lFai
eSF
Max
Ser
a
Ran
ptiv
tion
Ada
por
Pro
all the network environments under heavy traffic.
The one-way ANOVA again confirmed that algorithm had a significant effect on
the results (F(5,1196) = 71.96, p 0.001). As suggested by Figure 3.14 the Bonfer-
roni post-hoc comparisons showed that the SFR and SerFR algorithms performed
significantly better than all the other algorithms, but themselves were statistically
indistinguishable. The aSerFR algorithm performed at a similar level to the PF and
random algorithms, significantly better than the maximum rate algorithm (mean dif-
77
ference of 219kbps, p 0.001) but significantly worse than SFR and SerFR (mean
difference of 50kbps, both p 0.002).
As under the previous traffic conditions the results for each algorithm are shown
for each network environment in Figure 3.15. As can be seen the results are less
clear cut than previously.
0.5 Plot: 0, Speed: 3km/h, UEs: 285 0.30 Plot: 1, Speed: 3km/h, UEs: 680
0.16 Plot: 2, Speed: 3km/h, UEs: 1150
0.25 0.14
0.4

0.12
0.20 0.10
0.3
0.15 0.08
0.2 0.06
0.10
0.04
0.1 0.05 0.02
0.01.4 1.6 1.8 2.0 2.2 2.4 0.00
0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 0.00
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1
0.25 0.14
0.4
0.12 5th Percentile Tput (Mbps)

0.20 0.10
0.3
0.15 0.08
0.2 0.06 AdaptiveSFR
0.10 MaxCI
0.04
0.02 ProportionalFair
0.01.2 1.4 0.00 0.00 Random
1.6 1.8 2.0 2.2 2.4 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 SFR
SerFR
0.25 0.14
0.4
0.12
0.20 0.10
0.3
0.15 0.08
0.2 0.06
0.10
0.04
0.1 0.05 0.02
0.01.4 1.6 1.8 2.0 2.2 0.00
0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1
Figure 3.15: The 5th Percentile UE throughput vs average UE throughput in heavy

When there are 285 UEs on the network the aSerFR algorithm improves the 5th
percentile UE throughput by 25kbps but at a reduction in average UE throughput
of 20kbps.
When there are 680 UEs on the network the improvements in 5th percentile
throughput achieved by the aSerFR were much smaller, about 5kbps, but still sig-
78
nificant. For example when the speed is 120km/h (Plot 7) the aSerFR algorithm
performs significantly better than the SFR algorithm (t(18) = 3.1817, p = 0.0052).
The decrease in average UE throughput of the aSerFR by 30kbps compared to the
SFR in this case is also statistically significant (t(18) = 2.9616, p = 0.0084).
With 1150 UEs on the network the aSerFR no longer shows any improvement
in terms of either 5th percentile or average throughput. For example in Plot 8 the
SFR algorithm performs better than the aSerFR algorithm by a significant 27kbps
(t(18) = 7.4401, p = 0.0001).
In each case the Pareto front receives contributions from most of the algorithms
suggesting that no one approach is able to provide overall optimal performance. In
every case the PF algorithm is dominated by another solution on the Pareto front
showing that in every case there is at least one solution which achieves both better
5th percentile throughput and better average throughput than the PF algorithm.
3.5.5 Full Traffic
The average UE throughput results obtained under Heavy traffic can be seen in
Figure 3.16. It can be seen that the maximum rate algorithm outperforms the other
algorithms by a very large margin, over 100% or 1Mbps. The other algorithms
achieve a very similar level of performance.
Algorithm was again found to be the main effect (F(5,1253) = 69.35, p 0.001).
As suggested by Figure 3.16 the Bonferroni post-hoc comparisons showed that the
maximum rate algorithm performed significantly better than each of the other al-
gorithms (average mean difference = 1.083Mbps, all p 0.001). All the remaining
algorithms are statistically indistinguishable in terms of their average UE through-
put.
The 5th percentile throughput results, seen in Figure 3.17 show that the best
performing algorithms, aSerFR, SFR and SerFR achieve a very similar level of per-
formance with the random and PF algorithms achieving worse results by 20kbps
and 50kbps respectively.
As before algorithm was found to be the main effect(F(5,1253) = 72.82, p
79
3.5
Median
3.0 Mean
2.0
1.5
1.0
0.5
0.0
ir
CI
dom
SFR
FR
R
alFa
eSF
Max
Ser
Ran
ptiv
tion
Ada
por
Pro
the network environments under full traffic.
0.7
Median
0.6 Mean
0.5
0.4
0.3
0.2
0.1
0.0
r
CI
dom
SFR
FR
R
lFai
eSF
Max
Ser
a
Ran
ptiv
tion
Ada
por
Pro
all the network environments under full traffic.
0.001). As suggested by Figure 3.17 the aSerFR provided the best performance.
The Bonferroni post-hoc comparisons showed that the aSerFR was only significantly
better than the maximum rate algorithm (mean difference of 272kbps, p 0.001)
and the PF (mean difference of 63kbps, p 0.001). As before the maximum rate
80
performs significantly worse than all the other algorithms with an average mean
difference of 250kbps (all p 0.001).
As before Figure 3.18 shows the results achieved by each algorithm under each
network environment.
0.40 0.30 0.14

0.35 0.12
0.25
0.30 0.10
0.25 0.20
0.08
0.20 0.15
0.15 0.06
0.10 0.04
0.10
0.05 0.05 0.02
0.001.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 0.000.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 0.000.2 0.4 0.6 0.8 1.0 1.2 1.4
0.5 0.30 0.14


0.25 0.12
0.4 0.10
0.20
0.3 0.08
0.15
0.10 0.04
MaxCI
0.1 Pareto front
0.01.0 Random
1.5 2.0 2.5 3.0 0.000.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.2 0.4 0.6 0.8 1.0 1.2 1.4 SFR
SerFR
0.40 0.14
0.25
0.35 0.12
0.30 0.20 0.10
0.25
0.15 0.08
0.20
0.15 0.10 0.06
0.10 0.04
0.05 0.02
0.05
0.001.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 0.000.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.2 0.4 0.6 0.8 1.0 1.2 1.4
Figure 3.18: The 5th Percentile UE throughput vs average UE throughput in full

When 285 UEs are present on the network the aSerFR performs significantly
better in terms of 5th percentile throughput than the baseline algorithms. When
the speed is either 3km/h (Plot 0) the performance of the aSerFR is significantly
better than that of the SerFR algorithm in terms of 5th percentile throughput
(t(18) = 2.4178, p 0.0264) with no statistically significant reduction in average
UE throughput (t(18) = 1.0725, p = 0.2977). However when the speed is 30km/h
81
(Plot 3) the performance of the aSerFR is significantly greater, by 173kbps, than the
performance achieved by SerFR (t(18) = 13.0365, p = 0.0001), however this comes
with a significant reduction of 400kbps in average UE throughput between SerFR
and aSerFR (t(18) = 10.5249, p = 0.0001).
When 680 UEs are present on the network the aSerFR performs significantly bet-
ter in terms of 5th percentile throughput by providing an improvement of 50kbps
over the baseline algorithms. For example in (Plot 7) the aSerFR algorithm per-
formed significantly better than the SFR algorithm (t(18) = 6.0910, p = 0.0001).
This comes with a statistically significant reduction in average throughput from
665kbps with SerFR to 620kbps with aSerFR (t(18) = 3.3006, p = 0.004).
With 1150 UEs on the network both the SFR and SerFR algorithms perform best.
For example in Plot 8 the SFR algorithm improves on the performance achieved by
the aSerFR algorithm by a significant 23kbps (t(18) = 5.8800, p = 0.0001).
The Pareto front produced in this case shows that when the traffic is full the
SerFR algorithm dominates the results produced by a number of the other algo-
rithms including, in some cases, the aSerFR algorithm. As with the mixed traffic
scheme the results produced by the PF algorithm are always dominated by those
produced by another algorithm under the full traffic load.
3.5.6 Summary
The results described in the preceding sections demonstrate that the aSerFR algo-
rithm was in general able to improve significantly on the performance of the 5th
percentile UE throughput in most cases. Further analysis showed that the improve-
ments made under light and mixed traffic levels were both significant and substan-
tial. Under heavy traffic the aSerFR algorithm matched the performance of the
best existing baseline algorithms in terms of the 5th percentile UE throughput, but
under heavy traffic the aSerFR was not able to achieve the same performance as the
baseline algorithms.
In terms of the average UE throughput the aSerFR performed significantly worse
overall. In general the lighter the traffic level the worse the relative performance of
82
the aSerFR was. Under full traffic environments the aSerFR was indistinguishable
from the other baseline algorithms, except for the maximum rate algorithm which
consistently performed best in terms of average UE throughput.
3.6 Discussion
This chapter introduced and described a new approach to ICIC in LTE networks.
This approach, known as aSerFR, was based on the existing scheme SerFR and
allowed a GA to optimise its parameters in order to tailor its behaviour to various
network environments. This was discussed as an example of localised organisation
where the eNBs do not have any knowledge of their neighbours behaviour and
there is no active coordination. An experiment was described which compared this
approach to existing ICIC schemes and some baseline schedulers in order to see if it
offered any improvement in the performance of the network.
The results described in the previous section demonstrate that particular im-
provements were noted when the network was under one of the lower traffic loads,
either light or mixed. In these cases it was observed that aSerFR achieved sig-
nificantly better 5th percentile throughput. In many cases aSerFR dominated all
the other algorithms, achieving both better 5th percentile throughput along with
better average throughput. Under these lighter traffic loads there is less contention
on the networks resources and under these conditions aSerFR utilises the available
resources more effectively than the baseline algorithms.
The improvements achieved by aSerFR under lighter traffic loads can also be seen
to a lesser extent when the network is under heavier traffic loads. When there are less
UEs on the network the aSerFR does achieve some improvement in 5th percentile
throughput, however the improvements are more modest than under lighter traffic
loads. However when there are more UEs active on the network the SFR and SerFR
schemes dominate by a significant margin in every case.
The results imply that localised organisation demonstrated in the aSerFR scheme
can indeed offer improvements in the performance of the network. The improvements
are greatest when the radio resources of the network are relatively uncontested
83
showing that this type of organisation can achieve good results when the constraints
are looser. When the constraints on the network resources are more severe however
it can be seen that this type of organisation does not improve the 5th percentile
throughput but rather the existing SFR and SerFR schemes are best suited to this
environment.
In summary it can be seen that this type of local organisation can lead to signif-
icant improvements in 5th percentile UE throughput, particularly under light and
mixed traffic loads. Under heavier traffic loads this approach was shown to match
the performance of the existing ICIC approaches.
84
Chapter 4
Networked Organisation for

eNodeB Co-ordination
In Chapter 3 a new Inter-Cell Interference Coordination (ICIC) scheme was intro-

duced which incorporated the idea of localised organisation to reduce the interference
in Long Term Evolution (LTE) networks. This was achieved by using ac Genetic
Algorithm (GA) to optimise it for use on a variety of network environments. It
was observed that this type of local organisation offered some improvements in net-
work performance when compared to existing ICIC schemes such as Soft Frequency
Reuse (SFR) and Softer Frequency Reuse (SerFR), in particular under light and
mixed traffic loads. In the local organisation demonstrated in the previous chapter
the network nodes, Evolved NodeBs (eNBs), did not have any knowledge of their
neighbours behaviour, but rather optimised their own behaviour.
One of the key features of LTE networks in the context of this research is the
X2 link which can be used by eNBs to communicate directly with their neighbours.
This chapter explores the possibilities of utilising this X2 link to enable a more net-
worked type of organisation. It is expected that allowing the eNBs to communicate
with their immediate neighbours to share scheduling information will lead to better
network performance as the eNBs will be able to take into account the performance
of their neighbours when making scheduling decisions.
This chapter describes the X2 protocol which operates in LTE networks and a
new ICIC scheme which is built on the Adaptive Softer Frequency Reuse (aSerFR)
85
Chapter 4: Networked Organisation for eNodeB Co-ordination
scheme presented in the previous chapter and utilises this X2 link to network the
eNBs so that they can schedule their transmissions in a way that attempts to reduce
the interference experienced by the network.
4.1 Overview of the X2 Protocol
The X2 protocol is a point-to-point protocol which enables eNBs to directly com-

municate with each other. This direct communication allows the eNBs to exhibit
a greater level of autonomy than was available under Universal Mobile Telecom-
munications System (UMTS) by sharing information which could allow for self-
organisation and self-optimisation. The main use cases envisaged by the 3rd Gener-
ation Partnership Project (3GPP) are for mobility support, load management, ICIC
and error reporting [7, 23].
Mobility management is concerned with the movement of User Equipments (UEs)
from one cell to another. By communicating with each other it is expected that the
eNBs can manage handovers in a more seamless manner.
The load management features allow the eNBs to share information about the
interference levels experienced on their Resource Blocks (RBs).
The details of the X2 protocol are defined in a number of 3GPP standards
documents [11, 12, 13, 18, 14]. It is important to note that the standards that define
these protocols are still under active development and that while 3GPP is responsible
for establishing the protocol standards, in this case what messages will be passed
along the X2 protocol, they do not enforce the behaviour an eNB should take in
each instance. These implementation details are left to the vendor to determine.
There is also a possibility of developing additions to the X2 functionality which may
or may not be incorporated into future standards. For example an eNB vendor may
develop a proprietary message which will run over the X2 link and which will be
usable only by their eNBs.
86
4.2 Adaptive Softer Frequency Reuse with X2
The ICIC scheme presented in this section, known as Adaptive Softer Frequency
Reuse + X2 (aSerFR+X2) utilises the X2 link to share information about the
downlink interference. aSerFR+X2 defines a new type of message which can be
used to share information regarding downlink interference between eNBs. As with
the aSerFR scheme a number of parameters have been specified so that this scheme
can be optimised in a similar way. It functions in the following way.
When an eNB becomes heavily loaded it can send a message to its neighbours
containing a list of those RBs which are experiencing the greatest level of interfer-
ence. The number of RBs included in such a message is dictated by the numRBs
parameter. Upon receiving such a message an eNB will avoid scheduling any trans-
missions on those RBs. Such a message can be sent by an eNB when its load level
rises above a certain level which is determined by the loadLevel parameter. The
message will either be sent to a randomly chosen neighbour or all neighbours de-
pending on the value of the neighbours parameter. Each X2 message sent by an
eNB will have a certain lifetime during which it will be in effect. This is governed by
the msgLif etime parameter. To prevent a heavily loaded eNB from continuously
sending out messages to its neighbours the msgW aitT ime parameter defines the
length of time an eNB must wait before sending another X2 message.
4.3 Experiments
In order to evaluate the performance of the aSerFR+X2 scheme presented in this

chapter and to compare it against existing approaches a number of experiments were
run. As in Chapter 3 the ICIC schemes were compared across 27 different network
environments in order to build up a more complete picture of their performance
characteristics. The network environments varied by the number of UEs present on
the network, the speed of the UEs on the network and the level of traffic requested
by the UEs.
As in experiment 1 described in Section 3.4 an Multi-Objective Genetic Algo-
87
rithm (MO-GA) was run for each algorithm in order to optimise the parameters of
the aSerFR+X2 scheme and achieve the best performance. The MO-GA produced a
number of results for each network environment which, taken with the results of the
other algorithms, produce a Pareto front where all the solutions can be considered
equally fit, but with varying values of 5th percentile and average throughput.
The MO-GA was run for 100 generations and the simulator described in Section
3.2 was used to evaluate the fitness of each of the chromosomes in the population.
In order to ensure that the fitness value calculated for any given chromosome was
representative of that chromosomes fitness the simulator was run ten times with
different random seeds and the results averaged to obtain the fitness. The remaining
parameters of the MO-GA are shown in Table 4.1. The parameters used in the
simulator are identical to those used in Chapter 3. See Table 4.2 for reference.
Parameter Value
Population Size 10
Generations 100
Mutation Strength 1
Crossover Operator Uniform Crossover
Table 4.1: The parameters used for the MO-GA in experiment 2.
Parameter Value
Number of eNBs 19
Number of UEs 252, 680, 1150
UE speed 3kmh, 30kmh, 120kmh
Data generation Light, Mixed, Heavy Traffic profiles
Distance attenuation L = 128.1 + 37.6 log(distance) [6]
Shadow fading Log normal, 8dB standard deviation
Multi-path fading Rayleigh fading
Cell layout Hexagonal grid, 3-sector sites, 57 sectors in total
Cell Radius 500m inter-site distance
Simulation Time 500 ms
Table 4.2: The parameters used for the simulator in experiment 2.
4.4 Results
As with the results of the previous experiment this section will report on the per-
formance of the aSerFR+X2 algorithm in comparison to those algorithms already
88
presented.
As in the previous chapter due to the fact that a number of algorithms are
being compared an mixed ANOVA was run on the results in order to determine
the statistical significance of the differences between the algorithms. The between-
subjects factor was algorithm and the within-subjects factors were traffic level, speed
and number of UEs. This resulted in a 6 (algorithm) 4 (traffic level) 3 (speed)
3 (number of UEs) ANOVA. The average UE throughput and the 5th percentile
UE throughput were analysed separately for clarity.
The overall ANOVA for average UE throughput demonstrated a significant main

effect for algorithm, (F(6,6620) = 1292.11, p 0.001). This shows that much of
the variance in the results can be attributed to the algorithm which means that the
algorithm has a significant effect on the average UE throughput.
As before a Bonferroni post-hoc analysis was carried out to determine which of
the algorithms differed from each other and how significant the differences between
each algorithm were. This analysis performs multiple comparisons between each of
the algorithms and adjusts the significance value of any of the differences to account
for the fact that multiple comparisons have taken place. For each comparison the
Bonferroni analysis provides a mean difference and the significance of that difference.
higher average UE throughput when compared to all the other algorithms with an
average mean difference between the maximum rate algorithm and each of the re-
maining algorithms of 518kbps (all p 0.001). The next best performing algorithms
were SerFR, random scheduler, the Proportional Fair (PF) algorithm and the SFR
algorithm which were statistically indistinguishable. The aSerFR algorithm per-
formed worse again that the above algorithms in terms of average UE throughput
with an average mean difference between it an the other algorithms, excluding the
maximum rate algorithm, of 124kbps (all p 0.001). Finally the aSerFR+X2 algo-
rithm performed significantly worse than the aSerFR algorithm. The mean difference
89
3.5
Median
3.0 Mean
2.0
1.5
1.0
0.5
0.0
ir
CI
dom
2
SFR
FR
R
alFa
R_X
eSF
Max
Ser
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
all all traffic levels.
was 62kbps (p 0.001).

This overall ANOVA for average UE throughput also demonstrated a number of
interaction effects. The interaction effects mean that different algorithms perform
differently under different network environments. The strongest interaction was be-
tween algorithm and traffic level (F(18,6620) = 207.98, p 0.001). As before this
called for follow-up simple effects analyses within each level of traffic. These subse-
quent analyses (ANOVAs investigating the effect of algorithm on average throughput
within each traffic level) are detailed in separate sections below.
The results of the algorithms in terms of 5th percentile UE throughput can be
seen in Figure 4.2. The same ANOVA was carried out to investigate the effects of
algorithm and environment on 5th percentile UE throughput. This demonstrated a
significant main effect for algorithm (F(6,6620) = 5030.61, p 0.001). The Bon-
ferroni analysis showed that most of the algorithms produced significantly different
results. In particular the aSerFR+X2 algorithm performed significantly better than
all the other algorithms, with an average mean difference of 66kbps (all p 0.001).
Once again separate analyses were conducted within each traffic level in view of
the significant interaction between traffic level and algorithm (F(18,6620) = 131.83,
90
0.8
0.7 Median
Mean
0.6
0.5
0.4
0.3
0.2
0.1
0.0
ir
CI
dom
2
SFR
FR
R
alFa
R_X
eSF
Max
Ser
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
across all all traffic levels.
p 0.001).
4.4.2 Light Traffic
The results of each algorithm when run under the light traffic network environments
can be seen in Figure 4.3. It can be seen from this figure that, as expected, the
relative performance of each of the algorithms is very similar to that described in
Chapter 3. The aSerFR+X2 algorithm introduced in this chapter performs worse
than all the other algorithms tested.
Algorithm was again found to be a significant effect on the results (F(6,1958) =
20.20, p 0.001). As suggested by Figure 3.7 the Bonferroni post-hoc comparisons
showed that the aSerFR+X2 algorithm performed significantly worse than all the
other algorithms with an average mean difference of 237kbps (all p 0.001). The
aSerFR algorithm performed significantly better than the aSerFR+X2 with a mean
difference of 146kbps (p 0.001) and significantly worse than the maximum rate
algorithm with a mean difference of 138kbps (p = 0.006). No other differences were
significant.
It should be noted here that the level of variance in the results has changed
91
1.6
1.4 Median
Mean
1.2
Average Tput (Mbps)
1.0
0.8
0.6
0.4
0.2
0.0
ir
CI
dom
2
SFR
FR
R
alFa
R_X
eSF
Max
Ser
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
all light traffic network environments.
from the previous chapter due to the addition of another algorithm (aSerFR+X2)
into the multiple comparisons. In the previous chapter the aSerFR was found to
be significantly worse than the baseline algorithms in this case, but with the ad-
dition of the aSerFR+X2 algorithm to the comparisons, which performs worst of
all, the difference between the aSerFR and the baseline algorithms was found to be
insignificant.
percentile UE throughput. It can be seen that the 5th percentile throughput of the
aSerFR+X2 algorithm outperformed the other algorithms. It can also be seen that
the results obtained by the SFR algorithm, while not achieving the performance of
the aSerFR+X2 or aSerFR algorithm, do outperform those achieved by the other
algorithms.
As with the average throughput results algorithm was again found to be a sig-
nificant effect (F(6,1958) = 239.06, p 0.001). As suggested by Figure 4.4 the
Bonferroni post-hoc comparisons showed that the aSerFR+X2 outperformed each
of the other algorithms, with average mean difference of 69kbps (all p 0.001).
As in the previous chapter the results achieved by each algorithm on the light
92
0.25
Median
0.20 Mean
0.15
0.10
0.05
0.00
ir
CI
dom
2
SFR
FR
R
alFa
R_X
eSF
Max
Ser
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
Figure 4.4: Boxplot showing the average 5th percentile UE throughput of each
algorithm across all light traffic network environments.
traffic network environments can be seen in Figure 4.5. This figure shows the nine
different network environments run under light traffic. These are combinations of
three UE density levels (285 UEs, 680 UEs, 1150 UEs) and three UE speeds (3km/h,
30km/h, 120km/h). It can be seen that when the network is under a light traffic
load and there are relatively few UEs the aSerFR+X2 scheme generally improves on
the performance achieved by the aSerFR and other algorithms. For example in Plot
6 the aSerFR+X2 algorithm outperforms the aSerFR algorithm by a statistically
significant 23kbps (t(18) = 6.3196, p = 0.0001).
With 680 users the 5th percentile throughput values achieved by the aSerFR+X2
do not surpass those achieved by the aSerFR. However in several cases it can achieve
a similar 5th percentile throughput with a greater average UE throughput. For
instance when there are 680 UEs on the network with speed 120km/h (Plot 7)
the aSerFR can achieve a 5th percentile throughput of 104kbps with an average
throughput of 430kbps. In comparison the aSerFR+X2 can achieve the same 5th
percentile throughput while improving the average throughput by 70kbps.
When the user load is increased to 1150 UEs the aSerFR+X2 achieves the best
absolute 5th percentile throughput only when the UE speed is 120km/h (Plot 8).
93
0.14
0.08

0.15 0.12
0.10 0.06
0.10 0.08
0.06 0.04
0.05 0.04
0.02
0.02
0.000.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.00
0.25 0.30 0.35 0.40 0.45 0.50
0.14 0.10

0.15 0.12
0.10 0.08
0.10 0.08 0.06 AdaptiveSFR
0.06 0.04
AdaptiveSFR_X2
0.05 0.04
MaxCI
0.02 Pareto front
0.000.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00 0.00 Random
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.20 0.25 0.30 0.35 0.40 0.45 0.50 SFR
SerFR
0.14
0.08

0.15 0.12
0.10 0.06
0.10 0.08
0.06 0.04
0.05 0.04
0.02
0.02
0.000.4 0.6 0.8 1.0 1.2 1.4 0.00
0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.00
0.25 0.30 0.35 0.40 0.45 0.50

In this case the 5th percentile throughput is 20kbps better than aSerFR (t(18) =
7.6648, p = 0.0001). When the speed is less than 120km/h, the aSerFR+X2 is
not statistically distinguishable from the aSerFR algorithm. For example in Plot
2 the aSerFR algorithm achieves a 5th percentile throughput of 100kbps while the
aSerFR+X2 algorithm achieves 98kbps. The difference between these values was
found to be not significant (t(18) = 1.1630, p = 0.26).
4.4.3 Mixed Traffic
The results of the average UE throughput can be seen in Figure 4.6. The average
throughput of each algorithm under mixed traffic networks displays a similar pattern
94
to the average throughput under light traffic loads. In particular the maximum rate
algorithm again achieves an improved average UE throughput, with the aSerFR
algorithm performing below the other algorithms and the aSerFR+X2 algorithm
achieving the worst performance of all. The remaining algorithms performed at a
very similar level.
1.8
1.6 Median
1.4
Mean
Average Tput (Mbps)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
ir
CI
dom
2
SFR
FR
R
alFa
R_X
eSF
Max
Ser
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
all mixed traffic network environments.

within mixed traffic network environments revealed a significant effect of algorithm
As with previous results algorithm was found to be a significant effect(F(6,1842)
= 31.18, p = 0.001). As suggested by Figure 4.6 the Bonferroni post-hoc comparisons
showed that that the aSerFR+X2 algorithm performed significantly worse than all
the other algorithms with an average mean difference of 276kbps (all p 0.001).
The aSerFR algorithm performed significantly better than the aSerFR+X2 with a
mean difference of 236kbps (p 0.001) and significantly worse than the maximum
rate algorithm with a mean difference of 148kbps (p = 0.002). No other differences
were significant.
The results in terms of 5th percentile UE throughput can be seen in Figure
4.7. The best performing algorithm in this case was the aSerFR followed by the
95
aSerFR+X2 algorithm. The remaining algorithms performed significantly worse in

each case. The performance achieved by the aSerFR and aSerFR+X2 algorithms
are at least twice as good as any of the baseline algorithms.
0.25
Median
0.20 Mean
0.15
0.10
0.05
0.00
ir
CI
dom
2
SFR
FR
R
alFa
R_X
eSF
Max
Ser
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
algorithm across all mixed traffic network environments.
As with the average throughput results algorithm was again found to be a sig-
nificant effect (F(6,1835) = 239.82, p 0.001). As suggested by Figure 4.7 the
Bonferroni post-hoc comparisons showed that the aSerFR outperformed each of the
other algorithms, with average mean difference of 96kbps (all p 0.001). The dif-
ference in performance between the aSerFR and the aSerFR+X2 was a significant
decrease of 27kbps (p 0.001). While the aSerFR+X2 algorithm was outperformed
by the aSerFR algorithm it in turn outperformed the remaining algorithms by an
average mean difference of 82kbps (p 0.001).
The results for each algorithm under each network environment with mixed traffic
can be seen in Figure 4.8 which shows that the aSerFR+X2 algorithm is often domi-
nated by the aSerFR algorithm such that it aSerFR generally performs best in terms
of both 5th percentile throughput and average throughput and that aSerFR+X2 of-
fers no performance improvement. The exception to this case is when there are 1150
UEs on the network. In these cases the aSerFR+X2 algorithm achieves significantly
96
0.14 0.07
0.20

0.12 0.06
0.15 0.10 0.05
0.08 0.04
0.10 0.06 0.03
0.04 0.02
0.05
0.02 0.01
0.000.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.00
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
0.12 0.08
0.20

0.07
0.10
0.06
0.15 0.08 0.05
0.10 AdaptiveSFR_X2
0.03 MaxCI
0.04
0.000.6 Random
0.8 1.0 1.2 1.4 1.6 0.00
0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.00
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 SFR
SerFR
0.12 0.08
0.20

0.07
0.10
0.06
0.15 0.08 0.05
0.06 0.04
0.10
0.03
0.04
0.05 0.02
0.02 0.01
0.000.6 0.8 1.0 1.2 1.4 1.6 0.00
0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.00
0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60

improved 5th percentile throughput.

When the speed of the 680 users is 3km/h the aSerFR+X2 algorithm gener-
ally achieves a very similar level of performance to the aSerFR algorithm. When
the speed is greater than 3km/h the aSerFR+X2 does improve the 5th percentile
throughput by 10kbps with an associated reduction in average UE throughput of
40kbps.
When 1150 users are present on the network the aSerFR+X2 algorithm achieves
better 5th percentile throughput, improving on aSerFR by at least 40kbps. For
example in Plot 8 the aSerFR+X2 algorithm achieves a 5th percentile throughput
of 85kbps which is significantly better than the 41kbps achieved by the aSerFR
97
(t(18) = 28.4917, p = 0.0001).

It can also be seen that the results produced by the aSerFR+X2 algorithm
generally feature heavily in the Pareto front in particular as the number of UEs
increases. The only baseline algorithm to contribute to defining the Pareto front
was the SFR algorithm.
4.4.4 Heavy Traffic
The result for the average UE throughput under Heavy traffic can be seen in Figure
4.9. It can be seen that the maximum rate algorithm achieves the highest perfor-
mance in terms of the average UE throughput by approximately 600kbps. It can
also be seen that the aSerFR+X2 algorithm achieves better performance than the
aSerFR by over 100kbps.
2.5
Median
2.0 Mean
Average Tput (Mbps)
1.5
1.0
0.5
0.0
r
CI
dom
2
SFR
FR
R
lFai
R_X
eSF
Max
Ser
a
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
all heavy traffic network environments.
As with the other traffic levels algorithm proved to be a significant effect (F(6,1461)
= 29.81, p 0.001). As suggested by Figure 4.9 the Bonferroni post-hoc compar-
isons showed that the maximum rate algorithm performed significantly better than
each of the other algorithms (average mean difference = 692kbps, all p 0.001).
The aSerFR+X2 algorithm while performing significantly worse than the maximum
98
rate algorithm, with a mean difference of 636kbps (p 0.001), did provide a signif-
icant improvement when compared to the aSerFR algorithm with a mean difference
of 126kbps (p = 0.01). No other differences were significant.
Figure 4.10 shows the 5th percentile UE throughput results. In the previous
chapter it was found that the aSerFR algorithm did not perform as well as the base-
line algorithms under heavy traffic. However with the addition of the X2 protocol
the aSerFR+X2 algorithm performs best in terms of 5th percentile throughput.
0.8
0.7 Median
Mean
0.6
0.5
0.4
0.3
0.2
0.1
0.0
ir
CI
dom
2
SFR
FR
R
alFa
R_X
eSF
Max
Ser
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
algorithm across all heavy traffic network environments.
A one-way ANOVA looking at the effect of algorithm on 5th percentile UE

throughput within heavy traffic network environments revealed a significant effect
of
Algorithm was again found to be a significant effect (F(6,1461) = 58.68, p
that the aSerFR algorithm performed worse than the SFR and SerFR algorithms
with an average mean difference of 50kbps (p = 0.02). The aSerFR+X2 algorithm
in comparison is statistically indistinguishable from the SFR and SerFR algorithms
and performs significantly better than the remaining algorithms with an average
mean difference of 125kbps (p 0.05).
99
Under heavy traffic conditions, as can be seen in the individual results in Figure
4.11, the results are somewhat varied. When there are only 285 UEs on the net-
work the aSerFR+X2 algorithm clearly performs the best in terms of 5th percentile
throughput, providing an improvement of 200kbps over the aSerFR algorithm with
an associated reduction in average UE throughput of 100kbps.
0.6 0.25 0.14

0.5 0.12
0.20 0.10
0.4
0.15 0.08
0.3
0.10 0.06
0.2 0.04
0.1 0.05 0.02
0.01.2 1.4 1.6 1.8 2.0 2.2 2.4 0.00
0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 0.00
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1
0.6 0.25 0.14
0.5 0.12 5th Percentile Tput (Mbps)

0.20 0.10
0.4
0.3
0.10 0.06 AdaptiveSFR_X2
0.2 0.04
MaxCI
0.05 Pareto front
0.01.2 0.00 0.00 Random
1.4 1.6 1.8 2.0 2.2 2.4 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 SFR
SerFR
0.6 0.25 0.14
0.5 0.12
0.20 0.10
0.4
0.15 0.08
0.3
0.10 0.06
0.2 0.04
0.1 0.05 0.02
0.01.2 1.4 1.6 1.8 2.0 2.2 2.4 0.00
0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1

In comparison when there are 680 or 1150 UEs on the network the aSerFR+X2
algorithm is largely dominated by the aSerFR algorithm and offer no performance
improvements. This shows that the improvements offered by the aSerFR+X2 algo-
rithm, at least under heavy traffic loads, are due in most part to its performance
under a smaller number of UEs.
100
In network environments under heavy traffic load the aSerFR+X2 algorithm

contributed significantly to the Pareto front along with the aSerFR, SFR and SerFR
algorithms.
4.4.5 Full Traffic
The results under Full traffic can be seen in Figure 4.12. As can be seen from the
average UE throughput results the maximum rate algorithm outperforms the other
algorithms by a very large margin, over 100% or 1Mbps. As under heavy traffic
conditions the aSerFR+X2 algorithm provides some improvement in performance
over the aSerFR algorithm.
3.5
Median
3.0 Mean
Average Tput (Mbps)
2.5
2.0
1.5
1.0
0.5
0.0
r
CI
dom
2
SFR
FR
R
lFai
R_X
eSF
Max
Ser
a
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
all full traffic network environments.
A significant effect of algorithm was discovered (F(6,1590) = 61.96, p 0.001).

As suggested by Figure 4.12 the Bonferroni post-hoc comparisons showed that the
maximum rate algorithm performed significantly better than each of the other al-
gorithms (average mean difference = 1.06Mbps, all p 0.001). As suggested by
the figure the aSerFR+X2 improved upon the performance of the SFR algorithm
with a mean difference of 122kbps (p = 0.005). All the remaining algorithms are
statistically indistinguishable in terms of their average UE throughput.
101
The 5th percentile throughput results, seen in Figure 4.13 show that the aSerFR+X2
algorithm provides large performance gains over the other algorithms. The results
in the previous chapter showed that the aSerFR offered no improvement on the
existing algorithms. This shows that the addition of the X2 protocol enhances its
performance.
0.8
0.7 Median
Mean
0.6
0.5
0.4
0.3
0.2
0.1
0.0
ir
CI
dom
2
SFR
FR
R
alFa
R_X
eSF
Max
Ser
Ran
eSF
ptiv
tion
ptiv
Ada
por
Ada
Pro
algorithm across all full traffic network environments.
Again a significant effect of algorithm was found (F(6,1590) = 65.92, p 0.001).

As suggested by Figure 4.13 the aSerFR+X2 provided the best performance. The
Bonferroni post-hoc comparisons showed that the aSerFR+X2 was significantly bet-
ter than all the other algorithms with an average mean difference of 147kbps (all p
0.001).
The individual results from the experiment when run with the full traffic profile
can be seen in Figure 4.14. It can be seen that then there are only 285 UEs on
the network the aSerFR+X2 algorithm generally performs significantly better than
the aSerFR, except in the case where the speed is 30km/h. When the speed is
3km/h (Plot 0) the aSerFR+X2 scheme improves the 5th percentile throughput of
the aSerFR algorithm by a 250kbps (t(18) = 14.6331, p = 0.0001), with a reduction
in average UE throughput of 300kbps. When the speed is 120km/h (Plot 6) the
102
aSerFR+X2 scheme again improves the 5th percentile throughput by 200kbps (t(18)
= 13.1272, p = 0.0001) with a reduction in average throughput of 60kbps.
0.7 0.30 0.14

0.6 0.25 0.12
0.5 0.10
0.20
0.4 0.08
0.15
0.3 0.06
0.2 0.10 0.04
0.1 0.05 0.02
0.01.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 0.000.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.2 0.4 0.6 0.8 1.0 1.2 1.4
0.5 0.30 0.14


0.25 0.12
0.4 0.10
0.20
0.15
0.10 0.04
MaxCI
0.1 Pareto front
0.01.0 Random
1.5 2.0 2.5 3.0 0.000.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.2 0.4 0.6 0.8 1.0 1.2 1.4 SFR
SerFR
0.6 0.25 0.14
0.5 0.12
0.20 0.10
0.4
0.15 0.08
0.3
0.10 0.06
0.2 0.04
0.1 0.05 0.02
0.01.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 0.000.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.2 0.4 0.6 0.8 1.0 1.2 1.4

When there are 680 or 1150 UEs on the network the aSerFR+X2 scheme does
not improve the performance of the aSerFR algorithm.
It can also be seen from these results that the aSerFR+X2 algorithm contributes
to the Pareto front in each case. Often, however, the results obtained from the
aSerFR+X2 algorithm are dominated by others, in particular those produced by
the aSerFR algorithm. As discussed under heavy traffic the aSerFR+X2 offers the
greatest performance improvements when there are smaller numbers of UEs on the
network.
103
4.4.6 Summary
As reflected by the overall analysis of the results, the aSerFR+X2 algorithm was
able to improve significantly on the performance of both the aSerFR and baseline
algorithms in terms of 5th percentile UE throughput. In every traffic level, with
the exception of heavy traffic, the aSerFR+X2 algorithm outperformed the baseline
algorithms by a significant margin. When the traffic was heavy the aSerFR+X2
algorithm was able to match the best performing of the baseline algorithms. The
improvements were particularly pronounced under the light and full traffic levels
with the aSerFR+X2 algorithm performing on average 147kbps better than aSerFR
under full traffic conditions.
4.5 Discussion
This chapter introduced a new ICIC scheme that built on the aSerFR scheme pre-
sented in the previous chapter and incorporated the X2 link. This scheme was
presented as an example of networked organisation, that is using the X2 link to
allow direct communication and information sharing between eNBs. A number of
experiments were described which compared the performance of the aSerFR+X2
scheme with the aSerFR and other schemes in order to see if this form of networked
organisation offered any performance improvements.
The results described in the previous section demonstrate that in general the net-
worked approach can provide significant improvements in 5th percentile throughput.
Improved results were achieved under light, mixed and full traffic loads indicating
that this approach is perhaps less dependent on the resource constraints of the
network that the aSerFR was found to be. In the one case where the aSerFR+X2
scheme was not able to exceed the results produced by the existing ICIC schemes it
was able to match their 5th percentile UE throughput.
It was also observed that when the aSerFR+X2 scheme did not improve the
5th percentile throughput it was generally dominated by the results of the aSerFR
and offered no improvement in terms of either 5th percentile throughput or average
104
throughput. However it was noted that in those cases when the aSerFR+X2 scheme
did increase the 5th percentile throughput achieved by the network the gains were
generally very significant, in some cases doubling the best performance achieved by
the other algorithms.
The results imply that this form of networked organisation can be used to signif-
icantly improve the 5th percentile UE throughput of the network. It is also implied
that networked organisation of this type is not as limited by the contention on the
radio resources of the network as the local organisation approach was found to be.
105
Chapter 5
De-centralised Organisation
Previous chapters have described a number of Inter-Cell Interference Coordination

(ICIC) schemes, both existing and new, which are designed to reduce the level of
interference in Long Term Evolution (LTE) networks. The new ICIC schemes de-
veloped as part of this research have been developed around the idea of localised
organisation (Adaptive Softer Frequency Reuse (aSerFR)) and networked organisa-
tion (Adaptive Softer Frequency Reuse + X2 (aSerFR+X2)). Local organisation
functions at the level of the individual network nodes, Evolved NodeBs (eNBs),
without any communication between nodes or knowledge of neighbouring nodes.
Networked organisation made use of the X2 link available in LTE to enable com-
munication and knowledge sharing between eNBs. It was demonstrated through
a series of experiments that these approaches could provide significantly improved
network performance over existing ICIC schemes in the literature.
This chapter introduces the concept of de-centralised ICIC schemes. A de-
centralised ICIC scheme is one where the process of computing the optimisation
takes place on the Radio Access Network (RAN) itself, that is the eNBs themselves
are used to perform the optimisation calculation.
In the previous ICIC schemes, all of which could be described as centralised
schemes, the optimisation process takes place on some central network node, not
itself involved in the operation of the RAN. This centralised node is responsible for
calculating the optimal network configuration to be used by the eNBs. These con-
figuration parameters were then applied uniformly to all the eNBs on the network.
106
Chapter 5: De-centralised Organisation
A de-centralised approach, facilitated by the X2 link, allows the optimisation itself

to be performed by the eNBs themselves in real-time.
5.0.1 Central and Distributed Organisation
The optimisation schemes discussed in previous chapters can all be described as

centralised Self-Organising/Optimizing Network (SON) systems. A centralised SON
is an approach to optimisation where the optimisation process itself takes place in
an offline fashion [2]. That is the computation required to determine the optimal
network configuration takes place on a node that is not directly involved with the
running of the network. This network node then pushes the calculated configuration
to those nodes that are involved in the running of the network, i.e. the eNBs, so
that they may make use of the calculated optimisation.
There are several potential problems with this approach. Firstly the parameters
used by each eNB are the same across the network; this means that even though
eNBs will be experiencing different levels of load and User Equipment (UE) be-
haviour they will all have the same scheduling parameters as determined by the
optimisation. Such a scheme is unlikely to be optimal since it is much more likely
that individually tailoring the parameters of each eNB to its own network environ-
ment will lead to improved performance as the parameters will better reflect the
reality of the network.
Secondly such a centralised scheme is not very responsive to changes in the
network environment. For example networks typically see different usage levels and
patterns at different times during the day and one set of parameters is unlikely to
be optimal for each condition. In order to adapt to the changing environment the
optimisation would need to be re-run when the network environment changes. This
means there will be an inevitable disparity between the parameters used by the
eNBs and the network environment they are operating in.
An alternative to the centralised approach is to use a de-centralised or distributed
scheme. Rather than running the optimisation on a central network node, it is run
over the network elements themselves. This means that the eNBs can optimise
107
themselves and adapt to the network changes that they experience in a much more
responsive way. It also means that the parameters obtained by the optimisation are
tailored specifically to the eNB which is running the optimisation. It does however
mean that the algorithm must be computationally cheap to run since the eNBs have
little in the way of spare computational capacity.
5.0.2 A De-centralised ICIC Algorithm
The de-centralised ICIC scheme presented in this section, known as De-centralised

Softer Frequency Reuse (dSerFR), is based on those described in previous chapters,
specifically aSerFR. aSerFR provided a parameterised scheduler which could be
optimised for different network environments with the use of a Genetic Algorithm
(GA). However, rather than running this GA as a centralised process dSerFR runs
the GA on the eNBs themselves.
In the previously described ICIC schemes the GA is run on a central process.
This process stores a population of chromosomes which represent different configu-
rations of the scheduler. This is mutated over time and evaluated using the network
simulator described in Section 3.2. In dSerFR there is no central process to run
the GA but instead each eNB maintains a single chromosome which represents its
own current scheduler configuration. At start-up this configuration is initialised with
random values. The eNB itself then mutates this configuration over time while mea-
suring the effect it has on its own, and possibly its neighbours, performance. This
allows each eNB to independently evolve their own set of scheduling parameters
which are tailored to the specific network environment in which they are located.
In order for this approach to be effective it is important that all of the eNBs
on the network do not attempt to mutate their chromosomes and measure their
performance change at the same time. If this were to be allowed to happen it
would be difficult to determine whether the performance change of an eNB was
due to its own parameters changing or those of its neighbours. Ideally while an
eNB is mutating its chromosome and measuring the change in its performance all
other variables would be kept constant. In reality this is not possible since the
108
network environment will always be changing around the eNBs as UEs leave and
join cells, initiate calls and initiate data sessions. However it is still important that
the algorithm not conflate this further by mutating multiple neighbouring eNBs at
the same time.
In order to avoid this the eNBs are conceptually divided into clusters called
neighbourhoods. Within each neighbourhood only one eNB is able to evolve its
configuration at any one time. Each eNB periodically communicates with those
eNB in its neighbourhood using the X2 link to determine whether any of them is
currently mutating. If any of the other eNBs in the neighbourhood are mutating
then none of the other eNBs will begin a mutation, but will wait for a period of
time before querying their neighbours status again. However if none of the eNBs
are mutating then the eNB which began the query will begin to evolve.
There are two different mutation strategies used by dSerFR: greedy and consen-
sus based. The mutation operator used is in both cases is the basic single point
mutation as described in Section 2.5.1.
Greedy Selection Prior to mutation the eNB stores the value of the current 5th
percentile throughput of the UEs that it is serving as a fitness benchmark. One
of the genes in the chromosome, a single configuration value, is chosen randomly
and its value is mutated. The eNB then continues serving its UEs for some time
window using this mutated configuration. After this window has passed the eNB
again takes a measure of the 5th percentile throughput of its UEs and compares
that value with the value measured at the beginning of the mutation. If this fitness
value has improved over this period then mutation is kept, if not it is discarded and
the configuration is reverted.
Consensus Selection The consensus selection strategy differs from the greedy
strategy mainly in that not only does the mutating eNB record its changing fitness
during the mutation but all the eNBs in the neighbourhood record the changes to
their fitness. When an eNB begins to mutate it alerts those eNBs in its neigh-
bourhood using the X2 link so that all the eNBs take a measure of their current 5th
109
percentile throughput. The mutating eNB then mutates its configuration and begins
to use this mutated configuration to serve its UEs. Once the mutation window has
passed all the eNBs in the neighbourhood measure their 5th percentile throughput
a second time and communicate to the mutating eNB whether their fitness has in-
creased or decreased. The mutating eNB will then take these responses into account
to determine whether it should accept the new configuration value. A parameter
of the dSerFR algorithm specifies what proportion of the neighbourhood must have
seen an increase in their fitness value before the mutating eNB can accept the new
configuration. If the proportion of eNBs in the neighbourhood that observed an
increase in their fitness value is less than the dSerFR parameter, meaning that a
number of the eNBs saw a decrease in their performance, then the mutating eNB
will discard the mutation as it has not been beneficial to the neighbourhood.
This scheme is computationally simple. Each eNB needs only keep track of
its own fitness over a time window along with some basic information about its
chromosome. Currently however this algorithm will only optimise a single objective
function, in this case 5th percentile throughput. It would be possible to optimise
in a multi-objective way but care would have to be taken to develop a function to
determine the fitness across competing objectives.
5.1 Experiments
In order to determine the performance of this de-centralised approach a number of

experiments were undertaken and the results compared to the approaches described
in the previous chapters.
In contrast to previous simulation runs the simulator was run for 200,000 it-
erations which represents approximately 5 seconds of real network time. This is
important since the performance of the optimisation changes over time. However
it is important to ensure that this does not provide this approach with a benefit
not given to the previous algorithms. A pilot study was carried out to determine
whether this extended run-time would confer any advantage to the dSerFR algo-
rithms. The graphs shown in Figures 5.1 and 5.2 show that each of the existing
110
algorithms generally stabilises after approximately 100 iterations while the dSerFR
algorithm continues to change, in terms of both 5th percentile UE throughput and
average UE throughput.
2.0
1.5
Tput (Mbps)
1.0
AdaptiveSFR_X2
DistributedSFR
AdaptiveSFR
Random
SerFR
SFR
MaxCI
0.00 500 1000 1500 2000
Iterations
Figure 5.1: The convergence of the average UE throughput values in the simulator
over time.
The parameters used for the dSerFR scheme can be seen in Table 5.1.
Parameter Value
Mutation Window 100
Consensus Proportion 0.75
Mutation Strength 1
Table 5.1: The parameters used for the dSerFR.
5.2 Results
The results of this experiment can be seen in Figures 5.7 to 5.16. Note that for the
sake of clarity only the best performing algorithms from the previous chapters have
been shown.
111
0.20
0.15
Tput (Mbps)
0.10
AdaptiveSFR_X2
DistributedSFR
AdaptiveSFR
0.05 Random
ProportionalFair
SerFR
SFR
MaxCI
0.000 500 1000 1500 2000 2500
Iterations
Figure 5.2: The convergence of the 5th percentile UE throughput values in the
simulator over time.
As in previous chapters an ANOVA for average UE throughput was run which

demonstrated a significant main effect for algorithm, (F(8,7258) = 1036.65, p
0.001). Bonferroni analysis were then undertaken to compare the results of each
algorithm taking into account the fact that multiple comparisons have taken place.
higher average UE throughput when compared to all the other algorithms with
an average mean difference between the maximum rate algorithm and each of the
remaining algorithms of 492kbps (all p 0.001). The next best performing algo-
rithms were dSerFR algorithms, both greedy and consensus based, followed closely
by the baseline algorithms. The aSerFR and aSerFR+X2 again performed signifi-
cantly worse than all the other algorithms in terms of average UE throughput. Of
particular interest in the context of this chapter the dSerFR algorithms, both the
greedy and consensus based, performed significantly better than both the aSerFR
and aSerFR+X2 (all p 0.001).
112
3.5
Median
3.0 Mean
2.0
1.5
1.0
0.5
0.0
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
all traffic levels.
As in previous analyses a number of interaction effects were discovered. In par-

ticular the strongest interaction was between algorithm and traffic level (F(24,7258)
= 181.96, p 0.001) which called for follow-up simple effects analyses within each
level of traffic. These subsequent analyses (ANOVAs investigating the effect of al-
gorithm on average throughput within each traffic level) are detailed in separate
sections below.
The same process was carried out to investigate the effects of algorithm and
environment on 5th percentile UE throughput. This demonstrated a significant
main effect for algorithm (F(8,7259) = 934.31, p 0.001). The results for each
algorithm in terms of 5th percentile UE throughput can be seen in Figure 5.4.
As in the previous chapter the aSerFR and aSerFR+X2 algorithms achieved the
best 5th percentile throughput. The dSerFR algorithms, while not matching the
performance of the aSerFR and aSerFR+X2, were able to match the performance
of the Soft Frequency Reuse (SFR) and Softer Frequency Reuse (SerFR) algorithms
with an average mean difference of 2kbps (all p = 1).
Once again separate analyses were conducted within each traffic level in view of
the significant interaction between traffic level and algorithm (F(24,7258) = 113.16,
113
0.8
0.7 Median
Mean
0.6
0.5
0.4
0.3
0.2
0.1
0.0
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
across all traffic levels.
p 0.001).
5.2.2 Light Traffic
The results in terms of average UE throughput can be seen in Figure 5.5. It can
be seen from this figure that, as expected, the relative performance of each of the
algorithms is very similar to that described in Chapter 4. The average throughput
achieved by the dSerFR algorithms was able to match that achieved by the baseline
algorithms and thus improve on that achieved by the aSerFR and aSerFR+X2.
within light traffic network environments was again carried out. This revealed a
significant effect of algorithm (F(8,2132) = 17.27, p 0.001). As suggested by
Figure 5.5 the Bonferroni post-hoc comparisons showed that the dSerFR algorithms,
both the greedy and consensus based, were significantly better than the aSerFR+X2,
with a mean difference of 264kbps (p 0.001) and 258kbps (p 0.001) respectively,
and was able to match the performance of all of the baseline algorithms, including
the maximum rate algorithm, with an average mean difference of 7kbps (all p = 1.0)
and 0.5kbps (all p = 1.0) respectively.
114
1.8
1.6 Median
1.4
Mean
Average Tput (Mbps)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
all light traffic network environments.
Figure 5.6 shows the results of each algorithm in terms of the 5th percentile UE
throughput. It can be seen that the 5th percentile throughput of the two dSerFR
algorithms, while not achieving the same performance as either the aSerFR or the
aSerFR+X2, do improve on the 5th percentile throughput achieved by the baseline
algorithms.
The ANOVA found a significant effect of algorithm (F(8,2132) = 205.35, p
0.001). As suggested by Figure 5.6 the post-hoc comparisons showed that the
dSerFR algorithms, both greedy and consensus based, were significantly worse than
both the aSerFR and the aSerFR+X2 with an average mean difference of 48kbps
(both p 0.001) and 48kbps (both p 0.001) respectively. While there was no
significant difference with the SFR algorithm and either of the dSerFR algorithms,
the dSerFR algorithms did perform significantly better than all the other baseline
algorithms with an average mean difference of 28kbps (all p 0.001).
As in the previous chapters the results achieved by each algorithm on the light
traffic network environments can be seen in Figure 5.7. This figure shows the nine
different network environments run under light traffic. These are combinations of
three UE density levels (285 UEs, 680 UEs, 1150 UEs) and three UE speeds (3km/h,
115
0.25
Median
0.20 Mean
0.15
0.10
0.05
0.00
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
algorithm across all light traffic network environments.
30km/h, 120km/h).
It can be seen from Figure 5.7 that the dSerFR algorithm has mixed results under
light traffic. When there are 285 or 680 UEs on the network the dSerFR algorithm
achieves better 5th percentile throughput results than the baseline results. In these
cases the average throughput is at least as good as the baseline algorithms.
In those cases where there are 1150 UEs on the network the dSerFR algorithm
matches the performance of the SFR algorithm in terms of 5th percentile throughput,
but appears to achieve an improved average throughput, however this was shown
to just fall short of statistical significance. For example in Plot 8, the dSerFR
algorithms achieved comparable 5th percentile UE throughput to the SFR algorithm,
33kbps in both cases. However the average UE throughput achieved by the dSerFR
algorithms was 407kbps compared to the 388kbps achieved by the SFR algorithm
which was found to be a statistically significant improvement (t(18) = 2.1026, p =
0.0498).
It can also be seen that in each network environment the dSerFR did not achieve
better results than the aSerFR+X2, either in terms of 5th percentile throughput or
average throughput. However it can be seen that in most cases the results from
116
0.14
0.08

0.15 0.12
0.10 0.06
0.10 0.08
0.06 0.04
0.05 0.04
0.02
0.02
0.000.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.00
0.25 0.30 0.35 0.40 0.45 0.50
0.14 0.10

0.15 0.12
AdaptiveSFR_X2
0.10 0.08 0.06 DistributedSFR-Consensus
0.06 0.04
DistributedSFR-Greedy
0.05 0.04
MaxCI
0.02 Pareto front
0.000.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00 0.00 Random
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.20 0.25 0.30 0.35 0.40 0.45 0.50 SFR
SerFR
0.14
0.08

0.15 0.12
0.10 0.06
0.10 0.08
0.06 0.04
0.05 0.04
0.02
0.02
0.000.4 0.6 0.8 1.0 1.2 1.4 0.00
0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.00
0.25 0.30 0.35 0.40 0.45 0.50

the dSerFR do appear as part of the Pareto front and thus are not dominated by
any of the other results which shows that the results do offer an improved trade-off
between the average UE throughput and the 5th percentile UE throughput.
5.2.3 Mixed Traffic
The results for average UE throughput can be seen in Figure 5.8. The average per-
formance of each algorithm under mixed traffic networks displays a similar pattern
to the average performance under light traffic loads. In particular the dSerFR again
provide improved performance over the aSerFR and aSerFR+X2 algorithms and
achieve a similar level performance to the baseline algorithms.
117
1.8
1.6 Median
1.4
Mean
Average Tput (Mbps)
1.2
1.0
0.8
0.6
0.4
0.2
0.0
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
all mixed traffic network environments.
The ANOVA again found a significant effect of algorithm (F(8,2009) = 23.50,

p = 0.001). As suggested by Figure 5.8 the Bonferroni post-hoc showed that the
dSerFR algorithms, both the greedy and consensus based, were significantly better
than the aSerFR+X2, with a mean difference of 300kbps (p 0.001) and 303kbps (p
0.001) respectively, and was able to match the performance of all of the baseline
algorithms, including the maximum rate algorithm, with an average mean difference
of 15kbps (all p = 1.0) and 45kbps (all p = 1.0) respectively.
percentile UE throughput. It can be seen that in similar fashion the results found
under light traffic the 5th percentile throughput of the two dSerFR algorithms, while
not achieving the same performance as either the aSerFR or the aSerFR+X2, do
achieve a similar level of 5th percentile throughput as the best baseline algorithm.
The ANOVA again revealed a significant effect of algorithm (F(8,2009) = 226.29,
p 0.001). As suggested by Figure 5.9 the Bonferroni post-hoc comparisons showed
that the dSerFR algorithms, both greedy and consensus based, were significantly
worse than both the aSerFR and the aSerFR+X2 with an average mean difference
of 81kbps (both p 0.001) and 80kbps (both p 0.001) respectively. There was
118
0.25
Median
0.20 Mean
0.15
0.10
0.05
0.00
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
algorithm across all mixed traffic network environments.
no significant difference between the dSerFR algorithms and any of the baseline
algorithms, apart from the maximum rate algorithm which as expect performed
significantly worse than all of the other algorithms in terms of the 5th percentile UE
throughput, with an average mean difference of 56kbps (all p 0.003).
As with the light traffic results the individual results obtained by each algorithm
under mixed traffic network environments can be seen in Figure 5.10.
It can be seen that when there are 285 UEs on the network the dSerFR dominates
most of the baseline algorithms, that is it performs at least as well in both the 5th
percentile throughput and average throughput metrics. The only exception to this is
the maximum rate algorithm which optimises the average throughput at the expense
of 5th percentile throughput. It can also be seen that none of the aSerFR results,
while providing better absolute results, dominate the dSerFR results.
When there are 680 UEs on the network the dSerFR algorithm is outperformed
by the SFR algorithm in terms of 5th percentile throughput. However the dSerFR
does achieve better average throughput. It can also be seen that the dSerFR main-
tains a position on the Pareto front, often dominating a number of other solutions,
including the baseline algorithms.
119
0.14 0.07
0.20

0.12 0.06
0.15 0.10 0.05
0.08 0.04
0.10 0.06 0.03
0.04 0.02
0.05
0.02 0.01
0.000.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.00
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60
0.12 0.08
0.20

0.07
0.10
0.06 AdaptiveSFR
0.05
0.06 0.04 DistributedSFR-Consensus
0.10 DistributedSFR-Greedy
0.03 MaxCI
0.04
0.000.6 Random
0.8 1.0 1.2 1.4 1.6 0.00
0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.00
0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 SFR
SerFR
0.12 0.08
0.20

0.07
0.10
0.06
0.15 0.08 0.05
0.06 0.04
0.10
0.03
0.04
0.05 0.02
0.02 0.01
0.000.6 0.8 1.0 1.2 1.4 1.6 0.00
0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.00
0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60

In the case where there are 1150 UEs on the network the SFR algorithm again
achieves better 5th percentile throughput than the dSerFR algorithms. In these
cases however it can be seen that the dSerFR results are dominated by those achieved
by the aSerFR+X2 algorithm. In these cases the aSerFR+X2 algorithm achieves
better 5th percentile throughput and average throughput than the dSerFR algo-
rithm.
5.2.4 Heavy Traffic
The average UE throughput results can be seen in Figure 5.11. The average perfor-
mance of each algorithm under heavy traffic networks appears very similar with the
120
exception of the maximum rate algorithm which achieves a much better average UE
throughput.
2.5
Median
2.0 Mean
Average Tput (Mbps)
1.5
1.0
0.5
0.0
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
r i
r
Dist
Ada
Pro
all heavy traffic network environments.
The one-way ANOVA found a significant effect of algorithm (F(8,1637) = 22.46, p

= 0.001). As suggested by Figure 5.11 the Bonferroni post-hoc shows that the only
significant difference in performance is due the improved average UE throughput
achieved by the maximum rate algorithm. The average mean difference between the
maximum rate algorithm and all the other algorithms was 684kbps (all p 0.001).
Figure 5.12 shows the results for each algorithm in terms of the 5th percentile
UE throughput. It can be seen that there appears to be less variation in the results
achieved by each algorithm than when under mixed traffic and light traffic loads.
In particular it appears that the performance of the dSerFR algorithms achieved a
similar level of performance to the aSerFR and aSerFR+X2 along with the baseline
algorithms.
The one-way ANOVA again revealed a significant effect of algorithm (F(8,1637)
= 42.82, p 0.001). As suggested by Figure 5.12 the Bonferroni post-hoc compar-
isons showed that there was no significant difference between the results achieved
by the dSerFR algorithms and any of the other algorithms, excepting the maximum
121
0.8
0.7 Median
Mean
0.6
0.5
0.4
0.3
0.2
0.1
0.0
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
algorithm across all heavy traffic network environments.
rate algorithm which again performed worse than any of the others. In contrast
to the mixed traffic environments the performance of the dSerFR algorithms was
statistically indistinguishable from the aSerFR and aSerFR+X2 algorithms.
As can be seen from the individual results under heavy traffic in Figure 5.13
when 285 UEs are present on a heavily loaded network the dSerFR algorithm pro-
vides statistically significant improvements over the baseline algorithms in terms of
5th percentile throughput. For example in Plot 0, the dSerFR algorithms achieve a
5th percentile UE throughput of 469kbps, compared to the best performing of the
baseline algorithms which was the random algorithm with 424kbps. This difference
was found to be statistically significant (t(18) = 2.2053, p = 0.0407). While the
baseline algorithms, in particular the SerFR algorithm, do appear to achieve better
average UE throughput, this was not found to be statistically significant. For exam-
ple in Plot 0 the SerFR algorithm achieved an average UE throughput of 1.672Mbps
compared to the 1.596Mbps achieved by the dSerFR algorithms. However this was
discovered to not be statistically significant (t(18) = 1.2873, p = 0.2143).
When 680 UEs are on the network both SFR and SerFR algorithms achieve
the best absolute 5th percentile UE throughput, the dSerFR algorithms are able to
122
0.6 0.25 0.14

0.5 0.12
0.20 0.10
0.4
0.15 0.08
0.3
0.10 0.06
0.2 0.04
0.1 0.05 0.02
0.01.2 1.4 1.6 1.8 2.0 2.2 2.4 0.00
0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 0.00
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1
0.6 0.25 0.14

0.5 0.12
0.4 AdaptiveSFR_X2
0.3
0.10 0.06 DistributedSFR-Greedy
0.2 0.04
MaxCI
0.05 Pareto front
0.01.2 0.00 0.00 Random
1.4 1.6 1.8 2.0 2.2 2.4 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 SFR
SerFR
0.6 0.25 0.14

0.5 0.12
0.20 0.10
0.4
0.15 0.08
0.3
0.10 0.06
0.2 0.04
0.1 0.05 0.02
0.01.2 1.4 1.6 1.8 2.0 2.2 2.4 0.00
0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 0.00
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1

dominate some of the solutions discovered by the other algorithms, thus expanding
the Pareto front.
In each network environment the dSerFR also appears on the Pareto front often
dominating a number of solutions produced by the aSerFR, aSerFR+X2, SFR and
SerFR algorithms.
5.2.5 Full Traffic
The average UE throughput results under Full traffic can be seen in Figure 5.14.
The average performance of each algorithm under full traffic networks appears very
similar with the exception of the maximum rate algorithm which achieves a much
123
better average UE throughput.
3.5
Median
Average Tput (Mbps) 3.0 Mean
2.5
2.0
1.5
1.0
0.5
0.0
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
all full traffic network environments.
The one-way ANOVA looking at the effect of algorithm on average UE through-

put within full traffic network environments revealed a significant effect of algorithm
(F(8,1768) = 47.32, p = 0.001). As suggested by Figure 5.11 the Bonferroni post-hoc
shows that the only significant difference in performance is due the improve aver-
age UE throughput achieved by the maximum rate algorithm. The average mean
difference between the maximum rate algorithm and all the other algorithms was
1.06Mbps (all p 0.001).
Figure 5.15 shows each algorithms performance in terms of the 5th percentile UE
throughput. It can be seen that the aSerFR+X2 algorithm appears to perform better
than the others algorithms amount which there seems to be very little variation in
results achieved.
The ANOVA confirmed a significant effect of algorithm (F(8,1768) = 50.41, p
that the dSerFR algorithms, both greedy and consensus based, performed signifi-
cantly worse than the aSerFR+X2 with mean differences of 104kbps (p 0.001)
and 106kbps (p 0.001) respectively. They also performed significantly better than
124
0.8
0.7 Median
Mean
0.6
0.5
0.4
0.3
0.2
0.1
0.0
ir
CI
dom
s R
edy SFR
2
SFR
FR
R
alFa
R_X
sen dSF
eSF
Max
Ser
Dist us
Griebuted
Ran
eSF
Con bute
ptiv
tion
ptiv
Ada
por
ri
r
Dist
Ada
Pro
algorithm across all full traffic network environments.
the maximum rate algorithm which itself performed significantly worse than all the
other algorithms (average mean difference was 263kbps, all p 0.001).
The results achieved by each algorithm when the network is under full load can be
seen in Figure 5.16. It can be seen that when there are only 285 UEs on the network
the dSerFR achieves better 5th percentile throughput that the baseline algorithms.
While the baseline algorithms do appear to achieve better average throughput the
improvement is not statistically significant.
When there are either 680 UEs or 1150 UEs on the network both SFR and SerFR
algorithms achieve statistically improved 5th percentile throughput when compared
to the dSerFR algorithm. However, the dSerFR algorithm does achieve statistically
improved average throughput when compared with the baseline algorithms.
It can also be seen that the results obtained from the dSerFR algorithm con-
tribute in each case to the Pareto front defined by those results which are not
dominated by any other.
125
0.7 0.30 0.14

0.6 0.25 0.12
0.5 0.10
0.20
0.4 0.08
0.15
0.3 0.06
0.2 0.10 0.04
0.1 0.05 0.02
0.01.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 0.000.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.2 0.4 0.6 0.8 1.0 1.2 1.4
0.5 0.30 0.14


0.25 0.12
0.20 AdaptiveSFR_X2
0.15
0.2 0.06 DistributedSFR-Greedy
0.10 0.04
MaxCI
0.1 Pareto front
0.01.0 Random
1.5 2.0 2.5 3.0 0.000.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.2 0.4 0.6 0.8 1.0 1.2 1.4 SFR
SerFR
0.6 0.25 0.14

0.5 0.12
0.20 0.10
0.4
0.15 0.08
0.3
0.10 0.06
0.2 0.04
0.1 0.05 0.02
0.01.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 0.000.6 0.8 1.0 1.2 1.4 1.6 1.8 0.00
0.2 0.4 0.6 0.8 1.0 1.2 1.4

5.2.6 Summary
The results reported above demonstrate that the dSerFR algorithms were, in every
case, able to match the performance of the existing baseline algorithms in terms of
5th percentile UE throughput. In most cases the dSerFR algorithms were not able
to match the performance of the aSerFR+X2 algorithm in terms of 5th percentile
throughput but in general did achieve a similar level of average UE throughput.
126
5.3 Discussion
This chapter introduced a new ICIC scheme built on the aSerFR scheme presented
in Chapter 3 and incorporated the X2 link to facilitate communication between
the network nodes. This scheme was presented as an example of de-centralised
organisation, that is allowing the eNBs themselves to run the optimisation in a
distributed manner using the X2 link rather than requiring a centralised process. An
experiment was described which compared the performance of this dSerFR scheme
with the aSerFR and other schemes in order to see if this form of de-centralised
organisation offered any performance improvements.
It can be seen in the results described in the preceding section that the dSerFR
algorithm was not able to match the performance of the aSerFR or aSerFR+X2
algorithms in terms of 5th percentile UE throughput, though it was able to match
the best performing of the baseline algorithms, the SFR and SerFR.
The fact that the performance of the dSerFR algorithms was not able to match
that of theaSerFR or aSerFR+X2 algorithms may be because the only means of
mutation in the dSerFR algorithm was the single point mutation operator which
mutates a single gene in the chromosome. This means that the rate of mutation of
the genes of the chromosomes, and consequently the exploration of the search space,
will be slow. In contrast the Multi-Objective Genetic Algorithm (MO-GA) used
to optimise the aSerFR and aSerFR+X2 algorithms also made use of a crossover
operator which was able to introduce a greater level of genetic diversity into the
chromosomes by combining two chromosomes to produce two distinct children. This
process allows for a greater rate of genetic change and thus a quicker traversal of
the search space. This lower rate of genetic change, due to the lack of crossover
operation, could mean that the dSerFR algorithm took longer to explore the search
space and therefore did not arrive at as good a solution as the aSerFR and aSerFR
algorithms. This is also suggested by the fact that the performance of the dSerFR
algorithm continued to improve over the duration of the simulation. This implies
that the algorithm is still exploring the search space and has not arrived at a final
optimal solution for each eNB.
127
It can also be noted from the above results how similar the results were between
the greedy and consensus based dSerFR. This similarity suggests that both of these
approaches to determining the fitness of a set of chromosomes do equally as well. It
is possible that when taking the greedy approach, which attempts to optimise the
5th percentile throughput of the eNB without regard to its neighbours performance,
the results are generally good for the neighbouring eNBs also. This implies that an
eNB optimising its own 5th percentile throughput may have a positive effect on its
neighbours even without considering the mutations effect on them.
Despite not improving the absolute 5th percentile throughput the dSerFR was
able in many cases to extend the Pareto front by dominating some of the solutions
created by the aSerFR and aSerFR+X2 schemes. This shows that the dSerFR
scheme presented here was able to improve, in some areas of the search space, on
those results generated by the other algorithms.
128
Chapter 6
Conclusion
The research described in this thesis was designed to investigate a number of ap-
proaches to using Self-Organising/Optimizing Network (SON) to enable network
operators to better manage their networks. The particular focus of this work has
been on the Inter-Cell Interference Coordination (ICIC) problem and using SON to
optimise the scheduling of transmissions in such a way as to reduce the level of inter-
ference experienced by the network and thus automatically improve the performance
of the network.
6.1 Main Findings
6.1.1 Local Organisation
The initial stage of the investigation attempted to observe the performance of some
existing approaches to ICIC optimisation while also investigating whether some basic
local optimisation could be used to improve the performance of the network.
A simulator was developed in order to investigate the performance of the existing
ICIC schemes. An new ICIC scheme, Adaptive Softer Frequency Reuse (aSerFR),
was developed based on the existing Softer Frequency Reuse (SerFR) scheme de-
scribed in the literature. A number of additional parameters were added to the
SerFR scheme along with a more dynamic approach to the allocation of the fre-
quency bands. These parameters were then optimised using a Multi-Objective Ge-
129
Chapter 6: Conclusion
netic Algorithm (MO-GA) which focused on optimising the parameters for improved
5th percentile throughput and average throughput.
The first experiment, which tested the aSerFR algorithm across 36 network envi-
ronments showed that in general the developed aSerFR scheme was able to improve
the 5th percentile throughput achieved by the network by significant amounts. In-
deed typically the aSerFR scheme produced the results with the best 5th percentile
throughput, in particular when run on network environments which were not heavily
loaded. Under the more heavily loaded network environments the aSerFR scheme
generally matched the performance of the Soft Frequency Reuse (SFR) and SerFR
schemes.
The Pareto front produced by the results from all the algorithms was greatly
expanded by the addition of the aSerFR results. Due to the fact that the aSerFR
scheme was optimised using a MO-GA the trade-offs between 5th percentile through-
put and average throughput were made explicit. This would allow a network opera-
tor who wanted to use the aSerFR scheme to run the optimisation on their network
and choose the set of parameters which best suited their requirements.
This shows that under some situations significant improvements in network per-
formance can be obtained by using a local organisation strategy, such as an MO-GA,
to tailor the configuration of the ICIC to the specific network environment.
6.1.2 Networked Organisation
The next stage of the investigation, presented in Chapter 4, was to determine

whether a more network aware approach to organisation using the X2 interface
to enable communication between the Evolved NodeBs (eNBs) could provide addi-
tional improvements to the performance of Long Term Evolution (LTE) networks.
A new ICIC scheme, Adaptive Softer Frequency Reuse + X2 (aSerFR+X2),
was introduced which made use of the X2 link to facilitate direct communication
between the eNBs on the network. In particular the eNBs were able to share per-
formance information regarding the level of interference currently being experienced
on particular Resource Blocks (RBs). When an eNB was under heavy load it could
130
communicate to its neighbours a list of those RBs which were experiencing the most
interference. This could then be used by the neighbouring eNBs to inform their
scheduling decisions with the aim of reducing the interference experienced.
The results showed that in general the aSerFR+X2 scheme was able to further
improve upon the gains made by the aSerFR scheme. In a number of cases the
results produced achieved the best 5th percentile throughput when compared to all
the other algorithms and in general the average 5th percentile throughput achieved in
each traffic level either matched or exceeded that achieved by the other algorithms.
The improvements were generally most pronounced when the network was under
lighter loads rather than the heavier loads. It was also observed that in every case
the results produced by the aSerFR+X2 scheme contributed to the Pareto front
which contained the most fit results. Even in those cases where the 5th percentile
throughput of the aSerFR+X2 scheme was only able to match the performance of
the existing schemes, showing no showing no statistically significant improvement,
there were often a number of solutions which did advance the Pareto front and
dominate a number of the solutions produced by the existing algorithms.
The use of the X2 link in the aSerFR+X2 scheme and the results obtained
demonstrate that significant improvements can be gained in terms of network per-
formance by using a more network aware approach to self-organisation. In general
the results obtained using this approach achieved better performance than those
produced with only the local organisation embodied in aSerFR scheme.
6.1.3 De-centralised Organisation
The final stage of the investigation, described in this Chapter 5, considered the case
of a more de-centralised approach to self-organisation. All of the previous approaches
discussed could be considered centralised systems, where an optimisation process
takes place on a central network node with the results pushed down to the actual
nodes involved in running the Radio Access Network (RAN).
A simple distributed Genetic Algorithm (GA), De-centralised Softer Frequency
Reuse (dSerFR), was developed which was designed to run on the eNBs themselves
131
as opposed to some central network node. Each eNB was responsible for evolving its
own set of configuration parameters in coordination with those eNBs in its immediate
neighbourhood. This allowed each eNB to maintain a set of parameters which more
accurately reflected the network environment in that specific location independently
of its neighbours. The use of the X2 link was crucial in facilitating the organisation
of this approach so that each eNB was aware of when it should begin evolving its
configuration so as to avoid conflicts with its neighbours.
The results obtained from the final experiment described in Chapter 5 demon-
strate that the dSerFR scheme was able to match the performance of the base-
line algorithms but was not able to offer further improvements on the aSerFR and
aSerFR+X2 algorithms in terms of 5th percentile. It was observed however that the
majority of results produced by the dSerFR scheme did contribute to the Pareto
front formed by the fittest of all the results taken together. This shows that the
dSerFR scheme was able to successfully explore areas of the search space where the
other algorithms had not. It was also observed that the performance of the dSerFR
scheme continued to improve as the network ran. Since the optimisation progresses
over time, in contrast to the more centralised schemes presented, this should mean
that if run over longer periods of time the dSerFR algorithm could produce even
better results.
This demonstrates that a more de-centralised approach to ICIC can provide addi-
tional options to network operators looking to balance the 5th percentile throughput
of their network with the average throughput. The other main benefit of this type
of de-centralised approach in general is that it is more autonomous that the more
centralised approaches. In the centralised approach the network operator must run
the optimisation process periodically on a network node and push the resulting pa-
rameters out to the eNBs on the network. In comparison, a de-centralised approach
can run continuously on the eNBs themselves with essentially no external input
required.
132
6.2 Future Work
There are a number of areas in this research which would benefit from further
investigation. These additions can be classified in three main ways: improvements to
the network simulator, further investigation and development into the ICIC schemes
themselves and further investigation into alternative optimisation techniques.
6.2.1 Improvements to the Network Simulator
There are a number of features which could be added to the network simulator to
build a more realistic picture of the ICIC schemes evaluated in this research.
6.2.1.1 Improved Traffic Model
The simulators traffic model is used to determine when and how much data arrives
for a User Equipment (UE) during the course of the simulation. A UE with a light
traffic level will only request a small amount of data from the eNB occasionally,
whereas a UE operating with a full traffic level will continuously be requesting
the maximum amount of data possible from the eNB. The current traffic model is
probabilistic: each traffic level has a certain probability of requesting data and the
amount of data requested is determined probabilistically. These specific probabilities
are discussed in section 3.2.4. The traffic model used by the simulator in this research
could model four traffic levels. This allowed the simulator to evaluate the ICIC
schemes on a number of different network environments in order to build a more
complete picture of their relative performance than that found in the literature.
However further improvements could be made to the traffic model used. For
example the traffic model does not distinguish between different types of traffic. In
reality there are a number of different types of traffic use that can be observed on
a live network. 3rd Generation Partnership Project (3GPP) describes the following
distinct traffic classes: conversational class, streaming class, interactive class, and
background class [8]. Each of these traffic classes has different behaviour both in
terms of the amount of data they require and also the sensitivity of that data to
delay. For example the conversational traffic class, used to model the transmission of
133
voice calls, has very stringent latency requirements while not necessarily requiring a
large amount of data to be transmitted. In comparison the background traffic class
which is used to model the delivery of email or the downloading of files, is much
more insensitive to the latency but may request a much larger volume of data over
time.
A number of concrete models have been proposed in the literature which model
the data and latency requirements of these types of traffic classes. Implementing
these in the network simulator would provide a more realistic view of the performance
of the various ICIC schemes. For example [68] a number of models are presented
that could be used to model Voice Over Internet Protocol (VoIP), that is traditional
voice calls, video streaming, on-line gaming and HTTP based internet browsing.
These were not implemented in the simulator as part of this research since most
of the existing studies are not concerned with the performance of algorithms on
different traffic levels. The equivalent of the full traffic level used in this research,
often called full buffer traffic mode, is generally used to assess the performance of
ICIC algorithms in the literature [26, 22, 6, 103]. However, it has been found that the
choice of traffic model, whether full buffered or not, can have a significant impact
on the type of improvements that can be expected from ICIC schemes [51] and
so further investigations with more realistic traffic models should enable a better
understanding of the performance improvements offered by ICIC schemes under
different network environments.
6.2.1.2 Changes in Network Environment
Another possible improvement to the network simulator is the addition of more

dynamic network environments. The network environments currently modelled by
the network simulator can be thought of as static environments. The traffic level
and the number of UEs on the network are fixed at the start of the simulation and
do not change. This is only an approximation of a realistic network where UEs enter
and leave the network at various times and where a single UE can change traffic
level by making a new voice call or starting a new internet browsing session. In
134
particular there are certain areas of operator networks, such as those serving heavy
commuting routes, which typically experience significant changes in the number of
UEs and traffic demand during the course of a day. Implementing such a dynamic
network environment would allow the simulator to more accurately represent an
LTE network. This would also allow the ICIC schemes to be evaluated on how well
they performed under changing traffic loads as this is an area of research where
further study would be helpful.
6.2.2 Further Investigation into ICIC Schemes
The investigations into improved ICIC described in this thesis has yielded signif-
icantly improved results in many cases. However there remain further areas of
investigation which may enable further performance increases.
It was seen from the results described in the preceding chapters that the area
where least improvements were observed was under those network environments
which were heavily loaded either in terms of the number of UEs on the network or
the traffic level of those UEs. It has been suggested in the literature [35] that under
these heavily loaded network environments Inter-Cell Interference Randomisation
(ICIR) may provide improved performance over ICIC. While ICIC was the focus
of this research and as such ICIR has not been discussed at length or investigated,
if it promised to improve upon the results delivered by ICIC then some further
investigation would indeed be warranted.
There is also scope for further investigation of the performance of the ICIC
presented in this thesis. In particular the dSerFR scheme described in Chapter 5
could potentially be improved in a number of ways. Due to time constraints the effect
of a number of parameters of the algorithm have not been explored. For example
the choice of selection algorithm used to determine which eNB should mutate could
be explored to determine whether this has any significant effect on the algorithms
performance. In addition parameters like the mutation window could be further
explored: what affect do different lengths of mutation window have on the results
achieved by the dSerFR scheme?
135
6.2.3 Alternative Optimisation Techniques
As described in the previous chapters the MO-GA developed as part of this thesis
provided the optimisation approach used to develop improved ICIC schemes. Two
further promising related approaches could be investigated in this area to build on
this research.
6.2.3.1 Cellular GAs
Cellular GAs are are GAs in which the population takes the form of a grid with
each vertex of the grid being occupied with a single candidate solution. During each
iteration of the GA a new population is produced by applying the recombination
and mutation operators to the existing population. The recombination and muta-
tion operators do not apply to the whole population as in a standard GA. Rather
when creating a new child solution for a given vertex, only those candidates in the
immediate neighbourhood can be selected as parents. This leads to a sense of local-
ity within the population know as isolation by distance as similar individuals tend
to cluster into sub-populations in particular parts of the grid.
Multi-objective cellular GAs have also been developed [79] where the non-dominated
candidate solutions are retained in a Pareto front. Experimental results suggest that
when compared to standard MO-GAs the Pareto front obtained from a cellular MO-
GA is significantly more diverse providing a more diverse range of Pareto optimal
solutions.
Considering the spatial nature of the cellular MO-GA in light of the spatial
nature of cellular networks such as LTE, using a cellular MO-GA may present an
advantage. Allowing each eNB to represent a candidate solution at one point on
the grid and allowing it to recombine with those eNBs in its neighbourhood could
allow for a more diverse Pareto front to develop and potentially allow for areas of
the network with different traffic patterns to develop solutions tailored to that area.
136
6.2.3.2 Embodied Evolution
Embodied evolution [96] is a distributed, asynchronous GAs which was initially de-
veloped to solve three main problems in the field of evolutionary robotics by allowing
the optimisation algorithm to run on a population of physical robots. Firstly, pre-
vious approaches did not scale well to many robots due many interactions between
them. Secondly a large number of candidate solutions needed to be evaluated but
previous methods of evaluation were too slow to evaluate these serially on a real
robot. Yet thirdly, the evaluations needed to be done on a real robot since the low
fidelity of simulations led to transference problems where simulation results were
not repeatable in reality.
The embodied evolution scheme functions in the following way. Each robot at-
tempts to complete a specific task and measure its own performance on the task.
Prior to carrying out the task the control software for each robot is initialised with
a random candidate solution which governs its behaviour and thus determines its
fitness. As the robots go about their tasks they probabilistically broadcast a ran-
domly chosen and mutated gene from their own chromosome. The rate of a robots
broadcast is proportional to its fitness with fitter robots broadcasting more often
than their less fit counterparts. Robots receiving a broadcast accept it with a prob-
ability inversely proportional to their own fitness so that less fit robots are more
likely to accept broadcasts from surrounding fitter robots.
These mechanisms allowed the embodied evolution approach to overcome the
problems associated with previous approaches as follows. The population based
approach allowed it to scale to the many robots required while also allowing for a
large number of evaluations since these could be computed across a number of robots
in parallel. The use of multiple real robots also avoided transference problems by
running on real robots rather than on a simulation.
This work could be applied directly to the ICIC schemes developed in this thesis.
An embodied scheme could be used to evolve the dSerFR ICIC scheme presented in
chapter 5. In this case each of the eNBs would evaluate their own fitness and broad-
cast their randomly selected genes at a rate proportional to their own fitness. This
137
approach could provide a low-bandwidth, low-complexity approach to developing a

more distributed ICIC scheme.
6.3 Concluding Remarks
This thesis has described research into using self-organisation techniques as part
of an ICIC approach to reduce interference and thus improve the performance of
LTE networks. The original ICIC schemes presented in this work have been shown
to significantly improve the performance of LTE networks under a large number of
different network conditions. A number of further avenues for investigation have
been described which could be used to further improve the performance of LTE
networks going forward. It is hoped that the research presented in this thesis can
not only inform industry regarding the current state of ICIC, but can serve as a
foundation upon which further developments can be built.
138
Bibliography
[1] 3GPP TR 25.913. Requirements for evolved UTRA (E-UTRA) and evolved
UTRAN (E-UTRAN), December 2008.
[2] 3GPP TS 32.500. Telecommunication management; Self-Organising networks

(SON): concepts and requirements, July 2008.
[3] 3GPP TS 36.104. TS 36.104: Base Station (BS) radio transmission and re-
ception, May 2008.
[4] 3GPP TS 36.201. TS 36.201: LTE Physical Layer - General Description,

November 2007.
[5] 3GPP. TR 36.814: Further advancements for E-UTRA physical layer aspects,
2010.
[6] 3GPP. TS 36.942: Radio Frequency (RF) system scenarios, 2010.
[7] 3GPP. TR 36.902: Self-configuring and self-optimizing network (SON) use

cases and solutions, 2011.
[8] 3GPP. Quality of Service (QoS) concept and architecture, 2012.
[9] 3GPP. TS 25.996: Spatial channel model for Multiple Input Multiple Output
(MIMO) simulations, September 2012.
[10] 3GPP. TS 36.214: Physical Layer - Measurements, December 2012.
[11] 3GPP. TS 36.420: X2 General Aspects and Principles, September 2012.
[12] 3GPP. TS 36.421: X2 Layer 1, December 2012.
139
BIBLIOGRAPHY
[13] 3GPP. TS 36.422: X2 Signalling Transport, September 2012.
[14] 3GPP. TS 36.424: X2 Data Transport, September 2012.
[15] 3GPP. TS 36.211: Physical Channels and Modulation, March 2013.
[16] 3GPP. TS 36.212: Multiplexing and Channel Coding, March 2013.
[17] 3GPP. TS 36.213: Physical Layer Procedures, March 2013.
[18] 3GPP. TS 36.423: X2 Application Protocol (X2AP), March 2013.
[19] 3GPP TR 25.814. Physical layer aspects for evolved Universal Terrestrial
Radio Access (UTRA), 2006.
[20] 4G++. 4G++: Advanced Performance Boosting Techniques in 4th Generation

Wireless Systems: Inter-Cell Interference Coordination, 2012.
[21] Alcatel. Multi-cell simulation results for interference co-ordination in new

OFDM DL. Technical Report R1-050694, 29th August 2nd September 2005,
London, UK, September 2005.
[22] Alcatel. R1-060209: System simulation results for downlink interference coor-
dination. Technical report, 3GPP, Helsinki, Finland, January 2006.
[23] NGMN Alliance. Informative list of SON use cases, April 2007.
[24] E. Amaldi, A. Capone, and F. Malucelli. Discrete models and algorithms for
the capacitated location problems arising in umts network planning. In Pro-
ceedings of the 5th international workshop on Discrete algorithms and methods
for mobile computing and communications, DIALM 01, pages 18, New York,
NY, USA, 2001. ACM.
[25] Jeffrey G. Andrews, Arunabha Ghosh, and Rias Muhamed. Fundamentals

of WiMAX, Understanding Broadband Wireless Networking. Prentice Hall,
February 2007.
140
BIBLIOGRAPHY
[26] D. Astely, E. Dahlman, A. Furuskar, Y. Jading, M. Lindstrom, and S. Parkvall.

LTE: the evolution of mobile broadband. Communications Magazine, IEEE,
47(4):4451, April 2009.
[27] G. Auer. On modeling cellular interference for multi-carrier based communica-

tion systems including a synchronization offset. In WPMC03, pages 290294,
2003.
[28] T Back, D.B Fogel, and Z Michalewicz, editors. Mutation Operators. Institute
of Physics Publishing, Bristol, 2000.
[29] T Back, D.B Fogel, and Z Michalewicz, editors. Mutation Operators. Institute
of Physics Publishing, Bristol, 2000.
[30] Thomas Back. Evolutionary algorithms in theory and practice: evolution

strategies, evolutionary programming, genetic algorithms. Oxford University
Press, Oxford, UK, 1996.
[31] Constantine A. Balanis. Antenna Theory: Analysis and Design, 3rd Edition.
Wiley-Interscience, 3 edition, April 2005.
[32] David Beasley, David R. Bull, and Ralph R. Martin. An Overview of Genetic
Algorithms: Part 2, Research Topics. University Computing, 15(4):170181,
1993.
[33] David Beasley, David R. Bull, and Ralph R. Martin. An overview of genetic
algorithms: Part 1, fundamentals, 1993.
[34] Tobias Blickle and Lothar Thiele. A comparison of selection schemes used
in genetic algorithms. Technical report, Gloriastrasse 35, CH-8092 Zurich:
Swiss Federal Institute of Technology (ETH) Zurich, Computer Engineering
and Communications Networks Lab (TIK, 1995.
[35] R. Bosisio and U. Spagnolini. Interference coordination vs. interference ran-

domization in multicell 3gpp lte system. In Wireless Communications and
Networking Conference, 2008. WCNC 2008. IEEE, pages 824829, 2008.
141
BIBLIOGRAPHY
[36] Patrice Calegari, Frederic Guidec, Pierre Kuonen, and Frank Nielsen. Com-
binatorial optimization algorithms for radio network planning. Theoretical
Computer Science, 263(12):235 245, 2001.
[37] D. Chizhik, J. Ling, P.W. Wolniansky, R.A. Valenzuela, N. Costa, and K. Hu-
ber. Multiple-input-multiple-output measurements and modeling in manhat-
tan. Selected Areas in Communications, IEEE Journal on, 21(3):321 331,
apr 2003.
[38] Brian Classon, Ajit Nimbalker, Stefania Sesia, and Issam Toufik. Channel
Coding and Link Adaptation, pages 207241. John Wiley and Sons, Ltd, 2009.
[39] E. Dahlman, H. Ekstrom, A. Furuskar, Y. Jading, J. Karlsson, M. Lunde-

vall, and S. Parkvall. The 3G Long-Term evolution - radio interface concepts
and performance evaluation. In Vehicular Technology Conference, 2006. VTC
2006-Spring. IEEE 63rd, volume 1, pages 137141, 2006.
[40] Erik Dahlman, Stefan Parkvall, Johan Skold, and Per Beming. 3G Evolution,
Second Edition: HSPA and LTE for Mobile Broadband. Academic Press, 2
edition, 2008.
[41] Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T. Meyarivan. A fast
elitist multi-objective genetic algorithm: Nsga-ii. IEEE Transactions on Evo-
lutionary Computation, 6:182197, 2000.
[42] Giovanna Di Marzo Serugendo, Marie-Pierre Gleizes, and Anthony Karageor-

gos. Self-organization in multi-agent systems. Knowl. Eng. Rev., 20:165189,
June 2005.
[43] S.N. Donthi and N.B. Mehta. Joint performance analysis of channel quality
indicator feedback schemes and frequency-domain scheduling for lte. Vehicular
Technology, IEEE Transactions on, 60(7):30963109, 2011.
[44] M. Dorigo, V. Maniezzo, and A. Colorni. Ant system: optimization by a colony

of cooperating agents. Systems, Man, and Cybernetics, Part B: Cybernetics,
IEEE Transactions on, 26(1):29 41, feb 1996.
142
BIBLIOGRAPHY
[45] Marco Dorigo. Ant colony optimization: A new meta-heuristic. In Proceedings

of the Congress on Evolutionary Computation, pages 14701477. IEEE Press,
1999.
[46] Ericsson. Inter-cell interference handling for E-UTRA. Technical Report R1-
050764, 3GPP, London, UK, September 2005.
[47] Ericsson. R1-050764: Inter-cell Interference Handling for E-UTRA. Technical

report, Ericsson, London, UK, 2005.
[48] Ericsson. Downlink inter-cell interference co-ordination/avoidance evaluation

of frequency reuse. Technical report, 3GPP, May 2006.
[49] Ericsson. Frequency hopping for E-UTRA uplink. Technical report, Ericsson,
Seoul, Korea, 2006.
[50] Ericsson. On inter-cell interference coordination schemes without/with traffic

load indication. Technical report, 3GPP, October 2006.
[51] Ericsson. R1-081093: Further considerations on the use of Inter-Cell interfer-

ence indication, 2008.
[52] M.J. Feuerstein, K.L. Blackard, T.S. Rappaport, S.Y. Seidel, and H.H. Xia.
Path loss, delay spread, and outage models as functions of antenna height for
microcellular system design. Vehicular Technology, IEEE Transactions on,
43:487498, 1994.
[53] I. Forkel, A. Kemper, R. Pabst, and R. Hermans. The effect of electrical and
mechanical antenna down-tilting in UMTS networks. In 3G Mobile Commu-
nication Technologies, 2002. Third International Conference on (Conf. Publ.
No. 489), pages 8690, 2002.
[54] R. Ghaffar and R. Knopp. Interference suppression strategy for cell-edge users
in the downlink. Wireless Communications, IEEE Transactions on, 11(1):154
165, 2012.
143
BIBLIOGRAPHY
[55] Al Globus, Greg Hornby, Derek Linden, and Jason Lohn. Automated antenna
design with evolutionary algorithms.
[56] I.A. Glover and P.M. Grant. Digital communications. Pearson Prentice Hall,
2010.
[57] David E. Goldberg. Genetic Algorithms in Search, Optimization and Machine

Learning. MAddison Wesley, 1989.
[58] David E. Goldberg. Genetic Algorithms in Search, Optimization, and Machine

Learning. Addison-Wesley Professional, January 1989.
[59] David E. Goldberg. Genetic Algorithms in Search, Optimization and Machine

Learning. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA,
1st edition, 1989.
[60] David E. Goldberg and Kalyanmoy Deb. A comparative analysis of selection

schemes used in genetic algorithms. In Foundations of Genetic Algorithms,
pages 6993. Morgan Kaufmann, 1991.
[61] Andrea Goldsmith. Wireless Communications. Cambridge University Press,

New York, NY, USA, 2005.
[62] B. Hagerman, D. Imbeni, J. Barta, A. Pollard, R. Wohlmuth, and P. Cosi-

mini. Wcdma 6-sector deployment - case study of a real installed umts-fdd
network. In Vehicular Technology Conference, 2006. VTC 2006-Spring. IEEE
63rd, volume 2, pages 703 707, may 2006.
[63] John H. Holland. Adaptation in natural and artificial systems. MIT Press,
Cambridge, MA, USA, 1992.
[64] John H. Holland. Adaptation in Natural and Artificial Systems: An Introduc-

tory Analysis with Applications to Biology, Control and Artificial Intelligence.
MIT Press, Cambridge, MA, USA, 1992.
144
BIBLIOGRAPHY
[65] J.M. Holtzman. Asymptotic analysis of proportional fair algorithm. In Per-

sonal, Indoor and Mobile Radio Communications, 2001 12th IEEE Interna-
tional Symposium on, volume 2, pages F33F37 vol.2, 2001.
[66] T. Isotalo, J. Niemela, and J. Lempiainen. Electrical antenna downtilt in

UMTS network. Proc. 5th European Wireless Conference, Barcelona, pages
265271, 2004.
[67] P.D. Karamalis, N.D. Skentos, and A.G. Kanatas. Selecting array configu-
rations for mimo systems: an evolutionary computation approach. Wireless
Communications, IEEE Transactions on, 3(6):1994 1998, nov. 2004.
[68] Farooq Khan. LTE for 4G Mobile Broadband. Cambridge University Press,
2009.
[69] J. Knowles and D. Corne. The pareto archived evolution strategy: a new
baseline algorithm for pareto multiobjective optimisation. In Evolutionary
Computation, 1999. CEC 99. Proceedings of the 1999 Congress on, volume 1,
pages 105 Vol. 1, 1999.
[70] Kyu In Lee, Kyung-Soo Woo, Yo Han Ko, Jae Young Ahn, and Yong-Soo Cho.
An inter-cell interference cancellation method for ofdm cellular systems using
a subcarrier-based virtual mimo. In Vehicular Technology Conference, 2006.
VTC-2006 Fall. 2006 IEEE 64th, pages 15, 2006.
[71] Marco Mamei, Ronaldo Menezes, Robert Tolksdorf, and Franco Zambonelli.
Case studies for self-organization in computer science. J. Syst. Archit., 52:443
460, August 2006.
[72] C. Maple, Liang Guo, and Jie Zhang. Parallel genetic algorithms for third
generation mobile network planning. In Parallel Computing in Electrical En-
gineering, 2004. PARELEC 2004. International Conference on, pages 229
236, sept. 2004.
145
BIBLIOGRAPHY
[73] H. Meunier, E.-G. Talbi, and P. Reininger. A multiobjective genetic algo-

rithm for radio network optimization. In Evolutionary Computation, 2000.
Proceedings of the 2000 Congress on, volume 1, pages 317324 vol.1, 2000.
[74] Brad L. Miller and David E. Goldberg. Genetic algorithms, tournament se-
lection, and the effects of noise. Complex Systems, 9:193212, 1995.
[75] Ajay R. Mishra. Fundamentals of Cellular Network Planning and Optimisa-

tion: 2G, 2.5G, 3G... Evolution to 4G. John Wiley and Sons, 2004.
[76] T. Mitchell. Machine Learning. McGraw-Hill Higher Education, 1997.
[77] M. Moustafa, I. Habib, and M.N. Naghshineh. Efficient radio resource con-
trol in wireless networks. Wireless Communications, IEEE Transactions on,
3:23852395, 2004.
[78] M. Moustafa, I. Habib, and M.N. Naghshineh. Efficient radio resource con-
trol in wireless networks. Wireless Communications, IEEE Transactions on,
3:23852395, 2004.
[79] Antonio J. Nebro, Juan J. Durillo, Francisco Luna, Bernab Dorronsoro, and
Enrique Alba. Mocell: A cellular genetic algorithm for multiobjective opti-
mization. International Journal of Intelligent Systems, pages 2536, 2007.
[80] J. Niemela and J. Lempiainen. Impact of mechanical antenna downtilt on per-

formance of WCDMA cellular network. In Vehicular Technology Conference,
2004. VTC 2004-Spring. 2004 IEEE 59th, volume 4, pages 20912095 Vol.4,
2004.
[81] Byoung-Seong Park, Jong-Gwan Yook, and Han-Kyu Park. The determination
of base station placement and transmit power in an inhomogeneous traffic
distribution for radio network planning. In Vehicular Technology Conference,
2002. Proceedings. VTC 2002-Fall. 2002 IEEE 56th, volume 4, pages 2051
2055 vol.4, 2002.
146
BIBLIOGRAPHY
[82] Theodore Rappaport. Wireless Communications: Principles and Practice.

Prentice Hall PTR, Upper Saddle River, NJ, USA, 2nd edition, 2001.
[83] Craig W. Reynolds. Flocks, herds and schools: A distributed behavioral model.
In Proceedings of the 14th annual conference on Computer graphics and inter-
active techniques, SIGGRAPH 87, pages 2534, New York, NY, USA, 1987.
ACM.
[84] D.H. Ring. Mobile telephony - wide area coverage. Technical Memo TA4 47-1
60-37, Bell Laboratories, December 1947.
[85] RITT. Inter-cell Interference Mitigation based on IDMA. Technical report,

RITT, Sophia Antipolis, France, June 2005.
[86] Stuart J. Russell and Peter Norvig. Artificial Intelligence: A Modern Approach
(2nd Edition). Prentice Hall, December 2002.
[87] Farag Sallabi and K. Shuaib. Modeling of downlink wireless fading channel
for 3gpp lte cellular system. In GCC Conference and Exhibition (GCC), 2011
IEEE, pages 421424, 2011.
[88] Siemens. Aspects of interference mitigation by coordination. Technical report,

3GPP, November 2005.
[89] Siemens. Interference mitigation considerations and results on frequency

reuse. Technical report, 3GPP, September 2005.
[90] Siemens. R1-050738: Interference mitigation - Considerations and Results on

Frequency Reuse. Technical report, 3GPP, London, UK, September 2005.
[91] Iana Siomina, Peter Varbrand, and Di Yuan. Automated optimization of

service coverage and base station antenna configuration in UMTS networks.
Wireless Communications, IEEE [see also IEEE Personal Communications],
13:1625, 2006.
[92] Fanglei Sun, Mingli You, Jin Liu, Zhenning Shi, Pingping Wen, and Jianguo
Liu. Genetic algorithm based multiuser scheduling for single- and multi-cell
147
BIBLIOGRAPHY
systems with successive interference cancellation. In Personal Indoor and Mo-

bile Radio Communications (PIMRC), 2010 IEEE 21st International Sympo-
sium on, pages 1230 1235, sept. 2010.
[93] A. Triki and L. Nuaymi. Intercell interference coordination algorithms in

ofdma wireless systems. In Vehicular Technology Conference (VTC Spring),
2011 IEEE 73rd, pages 1 6, may 2011.
[94] A.A. Triolo. Directions in cellular telephony and implications for power sys-
tems. In Telecommunications Energy Conference, 2000. INTELEC. Twenty-
second International, pages 811, 2000.
[95] F.J. Villegas, T. Cwik, Y. Rahmat-Samii, and M. Manteghi. A parallel electro-

magnetic genetic-algorithm optimization (ego) application for patch antenna
design. Antennas and Propagation, IEEE Transactions on, 52(9):2424 2435,
sept. 2004.
[96] Richard A. Watson, Sevan G. Ficici, and Jordan B. Pollack. Embodied evo-
lution: Distributing an evolutionary algorithm in a population of robots.
Robotics and Autonomous Systems, 39:118, 2002.
[97] R.M. Whitaker, L. Raisanen, and S. Hurley. A model for conflict resolution
between coverage and cost in cellular wireless networks. In System Sciences,
2004. Proceedings of the 37th Annual Hawaii International Conference on,
page 8 pp., 2004.
[98] Mehmet E. Aydin ; Raymond Kwan ; Wei Ding ; Joyce Wu. A genetic algo-
rithm approach for multiuser scheduling on the lte downlink. In Lecture Notes
in Engineering and Computer Science, volume 2198, 2012.
[99] Fan Xiangning, Chen Si, and Zhang Xiaodong. An Inter-Cell interference
coordination technique based on users ratio and Multi-Level frequency al-
locations. In Wireless Communications, Networking and Mobile Computing,
2007. WiCom 2007. International Conference on, pages 799802, 2007.
148
BIBLIOGRAPHY
[100] Xu Yang, Yapeng Wang, Dapeng Zhang, and Laurie Cuthbert. Resource
allocation in lte ofdma systems using genetic algorithm and semi-smart an-
tennas. In Wireless Communications and Networking Conference (WCNC),
2010 IEEE, pages 1 6, april 2010.
[101] O.N.C. Yilmaz, S. Hamalainen, and J. Hamalainen. Analysis of antenna pa-

rameter optimization space for 3gpp lte. In Vehicular Technology Conference
Fall (VTC 2009-Fall), 2009 IEEE 70th, pages 1 5, sept. 2009.
[102] Jun Zhang and J.G. Andrews. Adaptive spatial intercell interference cancella-
tion in multicell wireless networks. Selected Areas in Communications, IEEE
Journal on, 28(9):14551468, 2010.
[103] Xunyong Zhang, Chen He, Lingge Jiang, and Jing Xu. Inter-cell interference
coordination based on softer frequency reuse in OFDMA cellular systems. In
Neural Networks and Signal Processing, 2008 International Conference on,
pages 270275, 2008.
149

Self-Organisation in LTE Networks An Inv PDF

Uploaded by

Copyright:

Available Formats

Self-Organisation in LTE Networks An Inv PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Self-Organisation in LTE Networks An Inv PDF

Uploaded by

Copyright:

Available Formats

SELF-ORGANISATION IN LTE NETWORKS:

Submitted for the degree of

Department of Computer Science

School of Mathematical and Computer Sciences

of the quotation or information.

Mobile telecommunications networks based on Long Term Evolution (LTE) technol-

This research, funded by an EPSRC CASE studentship in collaboration with Mo-

3 Localised Organisation at the Level of a Single eNodeB 49

4 Networked Organisation for eNodeB Co-ordination 85

5 De-centralised Organisation 106

1.1 The relationship between channel bandwidth, number of sub-carriers

4.1 The parameters used for the MO-GA in experiment 2. . . . . . . . . 88

1.1 A conceptual representation of two wireless telecommunications net-

3.1 The effect of UE speed on the variation of signal strength. . . . . . . 56

5.1 The convergence of the average UE throughput values in the simulator

3GPP 3rd Generation Partnership Project.

ARQ Automatic Repeat reQuest.

aSerFR Adaptive Softer Frequency Reuse.

aSerFR+X2 Adaptive Softer Frequency Reuse + X2.

CDMA Code Division Multiple Access.

CQI Channel Quality Indicator.

dSerFR De-centralised Softer Frequency Reuse.

eNB Evolved NodeB.

ICIC Inter-Cell Interference Coordination.

IDMA Interleave Division Multiple Access.

LTE Long Term Evolution.

MCS Modulation and Coding Scheme.

MO-GA Multi-Objective Genetic Algorithm.

MOEA Multi-Objective Evolutionary Algorithm.

NGMN Next Generation Mobile Networks Alliance.

OFMD Orthogonal Frequency Division Multiple.

OFMDA Orthogonal Frequency Division Multiple Access.

QoS Quality of Service.

RAN Radio Access Network.

RNC Radio Network Controller.

SerFR Softer Frequency Reuse.

SFR Soft Frequency Reuse.

SINR Signal to Interfernce plus Noise Ratio.

SON Self-Organising/Optimizing Network.

UMTS Universal Mobile Telecommunications System.

This thesis describes an investigation into self-organisation in Long Term Evolution

in order to provide the necessary information to understand the wider context of

1.2 Overview of LTE Networks

Figure 1.1: A conceptual representation of two wireless telecommunications net-

Conceptually wireless telecommunication networks can be viewed as consisting

the operators network or to external networks. As such it maintains connections

1.2.1 Radio Access Network Architecture

Figure 1.3: A simple cellular network demonstrating frequency reuse

1.2.2 UMTS Radio Access Network Architecture

Figure 1.5: UMTS radio access network architecture

varies across deployments, but several hundred is not uncommon.

1.2.3 LTE Radio Access Network Architecture

Figure 1.6: LTE radio access network architecture

1.2.4 Transmission Scheduling in LTE

Figure 1.7: The time structure of the LTE physical resource.

sub-frame in the time domain. In the frequency domain a RB spans 12 sub-carriers.

Figure 1.8: The frequency structure of the LTE physical resource.

The amount of data that can be transmitted in an RB is dependent on the

Channel Bandwidth (MHz) 1.4 3 5 10 15 20