0 évaluation0% ont trouvé ce document utile (0 vote)

3 vues116 pagesResearch paper

2015VERS007Vbis

© © All Rights Reserved

PDF, TXT ou lisez en ligne sur Scribd

Research paper

© All Rights Reserved

0 évaluation0% ont trouvé ce document utile (0 vote)

3 vues116 pages2015VERS007Vbis

Research paper

© All Rights Reserved

Vous êtes sur la page 1sur 116

networks

Amine Mohamed Adouane

Amine Mohamed Adouane. Dynamic management of spectral resources in LTE networks. Networking

and Internet Architecture [cs.NI]. Université de Versailles-Saint Quentin en Yvelines, 2015. English.

�NNT : 2015VERS007V�. �tel-01164507�

https://tel.archives-ouvertes.fr/tel-01164507

Submitted on 17 Jun 2015

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents

entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,

lished or not. The documents may come from émanant des établissements d’enseignement et de

teaching and research institutions in France or recherche français ou étrangers, des laboratoires

abroad, or from public or private research centers. publics ou privés.

THÈSE

Présentée à

Université de Versailles Saint-Quentin-en-Yvelines

pour obtenir le grade de

Mention ”Informatique”

par

Amine Mohamed Adouane

Directeur de Thèse :

Samir Tohme, Professeur des universités Université de Versailles Saint-Quentin-en-Yvelines

Co-encadrante :

Kinda Khawam, Maitre de conférences Université de Versailles Saint-Quentin-en-Yvelines

Rapporteurs :

Stefano Secci, Maitre de conférences HDR Université Pierre et Marie Curie

Bernard Cousin, Professeur des universités IRISA Université de Rennes 1

Examinateurs :

Steven Martin, Professeur des universités LRI, Université Paris-Sud

Djamal Zeglache, Professeur Telecom Sudparis

Invité :

Samer Lahoud, Maitre de conférences, IRISA, Université de Rennes 1.

To my parents

i

ii

Remerciements

Ce manuscrit conclut trois ans de travail, je tiens en ces quelques lignes à exprimer ma

reconnaissance envers tous ceux qui de prés ou de loin y ont contribué.

Je tiens à remercier en premier lieu Mr Samir Tohme, mon directeur de thèse pour m’avoir

donné la chance de faire cette thèse, ainsi que pour son encouragement, son expérience,

ses conseils et sa sympathie qui m’ont permis de mener à bien cette thèse.

Je tiens à exprimer mes plus vifs remerciements à Kinda Khawam qui fut pour moi une

encadrante de thèse attentive et disponible malgré ses nombreuses charges. Sa compétence,

sa rigueur scientifique et sa clairvoyance m’ont beaucoup appris. Ils ont été et resteront

des moteurs de mon travail de chercheur.

J’exprime tous mes remerciements à l’ensemble des examinateurs : Djamal ZEGHLACHE,

Paul MüHLETHALER et Steven MARTIN. Je remercie également Bernard COUSIN et

Stefano SECCI qui ont accepté d’être les rapporteurs de cette thèse, et pour l’effort qu’ils

ont fait pour lire mon manuscrit et l’intérêt qu’ils ont porté à mon travail. Je remercie

aussi Samer LAHOUD d’avoir accepté d’être présent pour ma soutenance.

Ce travail a été réalisé au sein du laboratoire PRISM de l’Université de Versailles dans le

cadre du projet SOAPS. Je tiens donc à remercier toutes les personnes que j’ai rencontrées

et qui m’ont aidé et encouragé tout au long de cette thèse.

Mes remerciements iront également à tous les membres de ma famille. Je pense surtout à

mes parents sans qui l’enfant que j’étais ne serait pas devenu l’homme que je suis. C’est

avec émotion qu’à mon tour je leur dévoile le fruit de mes efforts. J’espère être à la hauteur

de leur fierté. Une pensée profonde va directement vers mon père qui sans lui rien de cela

n’aurait existé et à maman pour toute son aide au fil de ses années. Je remercie également

mon oncle Seddiki Zakari pour ses conseils et orientations.

Sans oublier de remercier tous ceux qui ont participé de près ou de loin à l’accomplissement

de cette thèse.

iii

iv

Table des matières

1 Introduction 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 The LTE network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2.1 The LTE architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2.2 The access network E-UTRAN . . . . . . . . . . . . . . . . . . . . . 3

1.2.3 LTE frame structures . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.4 The X2 Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2.5 The LTE Uu interface . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.3 Inter-Cell Interference Coordination ICIC . . . . . . . . . . . . . . . . . . . 8

1.3.1 Frequency reuse based schemes . . . . . . . . . . . . . . . . . . . . . 10

1.3.2 Cell coordination based schemes . . . . . . . . . . . . . . . . . . . . 12

1.4 Downlink Power Control in LTE . . . . . . . . . . . . . . . . . . . . . . . . 16

1.5 Plan of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 System Framework 21

2.1 The Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.1.1 The Data Rate on the Downlink . . . . . . . . . . . . . . . . . . . . 22

2.1.2 The Bit Transfer Time . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.2 Non-Cooperative game for RBs selection . . . . . . . . . . . . . . . . . . . . 23

2.2.1 The Nash Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

tion 25

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.2 Non-Cooperative game for RBs selection . . . . . . . . . . . . . . . . . . . . 26

v

vi TABLE DES MATIÈRES

3.4 Distributed Learning of PNE . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.5.1 Speed of Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.5.2 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

tion 37

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4.2 The network model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.3 Non-cooperative congestion game for RBs selection . . . . . . . . . . . . . . 38

4.3.1 The load balancing game . . . . . . . . . . . . . . . . . . . . . . . . 39

4.4 Distributed learning of PNE . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.5 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.5.1 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.5.2 Performances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5.2 The network model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.3 Congestion game for RBs selection . . . . . . . . . . . . . . . . . . . . . . . 48

5.4 Distributed learning of PNE . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5.5 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5.5.1 Speed of convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5.5.2 Performances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5.5.3 Non homogeneous UEs distribution . . . . . . . . . . . . . . . . . . . 53

5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

6.2 The system model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

6.2.1 Data rate on the downlink . . . . . . . . . . . . . . . . . . . . . . . . 58

6.2.2 Cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

TABLE DES MATIÈRES vii

6.4 Non-cooperative game for power control . . . . . . . . . . . . . . . . . . . . 61

6.4.1 Sub-modular game . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

6.4.2 Attaining the Nash Equilibrium . . . . . . . . . . . . . . . . . . . . . 63

6.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

6.6 The Price of Anarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6.6.1 Optimal Centralized Approach . . . . . . . . . . . . . . . . . . . . . 69

6.6.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

6.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

7 General Conclusion 73

7.1 Summary of Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

7.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

viii TABLE DES MATIÈRES

Summary

The exponential growth in the number of communications devices has set out new am-

bitious targets to meet the ever-increasing demand for user capacity in emerging wireless

systems. However, the inherent impairments of communication channels in cellular systems

pose constant challenges to meet the envisioned targets. High spectral reuse efficiency was

adopted as a solution to higher data rates. Despite its benefits, high spectral reuse leads

to increased interference over the network, which degrades performances of mobile users

with bad channel quality. To face this added interference, OFDM (Orthogonal Frequency

Division Multiplexing) is used for the new 4th generation wireless network. Thanks to

its orthogonality OFDM eliminates the intra-cellular interference, but when the same re-

sources are used in two adjacent cells, the inter-cell interference becomes severe. To get

rid of the latter, several methods for Inter-Cell Interference Coordination (ICIC) have

been proposed. ICIC allows coordinated radio resources management between multiple

cells. The eNodeBs can share resource usage information and interference levels over the

X2 interface through LTE-normalized messages. Non-cooperative game theory was largely

applied where eNodeBs selfishly selects resource blocks (RBs) in order to minimize inter-

ference. In this thesis, we stress on ICIC for the downlink of a cellular OFDMA system in

the context of the SOAPS (Spectrum Opportunistic Access in Public Safety) project. This

project focuses on the improvement of frequency resource scheduling for Broadband Ser-

vices provision by PMR (Private Mobile Radio) systems using LTE technologies. Hence,

our first addresses the problem of downlink ICIC where the resource selection process is

apprehended as a potential game for which we propose a fully decentralized algorithm

based on replicator dynamics to attain the pure Nash equilibriums of the game. Extensive

simulations assessed the good performances of the algorithm for low to medium load. Re-

sults are in adequacy with the project need and with the system latency constraints. Our

second work is devoted to a more general solution for the ICIC problem not limited to

narrow band systems (with limited number of RBs). It portrays the problem of downlink

ICIC as a load balancing game. We adapt a stochastic version of a best-response dynamics

algorithm to attain the pure Nash equilibriums of the modeled game. Each eNodeB strives

to select a pool of favorable resources with low interference based on local knowledge only.

Proof of convergence is provided, and the efficiency of the tailored algorithm is proven

ix

x TABLE DES MATIÈRES

through extensive simulations. However, in this first solution the userś position in the cell

is not taken into account for ease of computation. The shortcoming of this first adaptation

is treated in a second version where a greedy algorithm is used to achieve faster conver-

gence times. The new simulation results show a significant improvement for both time

convergence and system performance even for larger system bandwidth. Finally, the ICIC

issue was treated through adequate power allocation on RBs. The power level selection

process of RBs is apprehended as a sub-modular game and a semi distributed algorithm

based on best response dynamics is proposed to attain the NEs of the modeled game,

striking a good balance between system performance and power economy.

Résumé

émergents fixe des objectifs toujours plus haut pour répondre à la demande de capacité

sans cesse croissante des utilisateurs. Cependant, les déficiences inhérentes des canaux de

communication dans les systèmes cellulaires posent des défis constants pour atteindre les

objectifs envisagés. Pour pallier à ce problème le spectre fréquentiel est utilisé avec une

forte efficacité (High efficiency spectral reuse) comme solution pour atteindre des débits

plus élevés. Malgré ses avantages, la réutilisation spectrale élevée conduit à des interfé-

rences accrues sur le réseau, ce qui dégrade dáutant plus les performances des utilisateurs

mobiles qui ont une mauvaise qualité de canal. Pour faire face à ces interférences, l’OFDM

(Orthogonal Frequency Division Multiplexing) est utilisé dans les réseaux de 4 ème généra-

tion. Grace à son orthogonalité, l’OFDM élimine l’interférence intra-cellulaire, mais quand

les mêmes ressources sont utilisées dans deux cellules adjacentes, l’interférence inter-cellule

devient importante. Pour se débarrasser de ces interférences, plusieurs méthodes connues

sous le nom d’Inter-Cell interferences coordination (ICIC) ont été proposées. L’ICIC per-

met la gestion des ressources radio coordonnée entre plusieurs cellules appelées eNodeB.

Ces eNodeB peuvent partager des informations sur l’utilisation des ressources et les ni-

veaux d’interférence grâce à l’interface X2 qui les relient, ces informations sont transmises

par des messages LTE normalisée. Lorsque les eNodeBs sélectionnent égoı̈stement les res-

sources blocs (RBS) afin de minimiser les interférences, la théorie des jeux non-coopératifs

est largement appliquée pour trouver un juste équilibre. Dans cette thése, nous mettons

l’accent sur l’ICIC pour la liaison descendante d’un système OFDMA cellulaire dans le

contexte du projet SOAPS (Spectrum Opportunistic Access in Public Safety). Ce projet

a pour but l’amélioration de la planification des ressources de fréquences pour fournir

des services à large bande dans les systèmes PMR (Private Mobile Radio) en utilisant les

technologies LTE. Nous adressons en premier le problème d’ICIC sur le lien descendent

où le processus de sélection des ressources est appréhendé comme un jeu potentiel, pour

cela nous proposons un algorithme entièrement décentralisé basé sur la dynamique de

réplication pour atteindre les équilibres de Nash purs du jeu. Des simulations approfon-

dies ont permis d’évalué les bonnes performances de l’algorithme à faible et à moyenne

charge. Les résultats ainsi obtenus sont en adéquation avec les besoins du projet et avec

xi

xii TABLE DES MATIÈRES

les contraintes de latence du systéme. Notre deuxième travail est consacré a une solution

plus générale au problème d’ICIC, ne se limitant pas aux systèmes à bande étroite (avec

un nombre limité de RBs). Le problème d’ICIC est décrit comme un jeu d’équilibrage de

charge (Load Balancing). Nous adaptons une version stochastique d’un algorithme dyna-

mique de best-response pour atteindre l’équilibre de Nash pur du jeu modélisé. Chaque

eNodeB s’efforce de sélectionner un pool de ressources favorables à faible interférence sur

la base des connaissances locales du système seulement. La preuve de convergence de cet

algorithme est apportée, et son efficacité est prouvée par de nombreuses simulations. Ce-

pendant, la position de l’utilisateur dans la cellule dans cette première solution n’est pas

prise en compte pour la facilité du calcul. La faiblesse de cette première adaptation est

traitée dans une seconde version où un algorithme glouton est utilisé pour obtenir des

temps de convergence plus rapide. Les nouveaux résultats de simulation montrent une

amélioration significative à la fois en temps de convergence, et en termes de performances

du système, même pour des systèmes à plus grande bande passante. Enfin, le problème

d’ICIC a été traité comme un problème d’allocation adéquate en puissance sur les RBs.

Le processus de sélection du niveau de puissance pour les RBS est appréhendé comme un

jeu sous-modulaire, ainsi, un algorithme semi distribué basé sur la dynamique de best-

response est proposé pour atteindre l’équilibre de Nash du jeu modélisé, un bon équilibre

entre les performances du système et l’économie d’énergie devant etre trouvé.

List of Publications

IEEE ISCC 2014, Replicator Dynamics for Distributed Inter-Cell Interference Coordina-

tion, Amine Adouane, Kinda Khawam, Johanne Cohen, Dana Marinca, Samir Tohmes,

University of Versailles PRISM Laboratory, University of Paris Sud LRI lab.

IEEE European Wireless 2014, Distributed Load Balancing Game for Inter-Cell Interfe-

rence Coordination, Amine Adouane, Lise Rodier, Kinda Khawam, Johanne Cohen, Samir

Tohme, University of Versailles PRISM Laboratory, University of Paris Sud LRI lab.

IEEE WCNC 2014, Game Theoretic Framework for InterCell Interference Coordination,

Amine Adouane, Lise Rodier, Kinda Khawam, Johanne Cohen, Samir Tohme, University

of Versailles PRISM Laboratory, University of Paris Sud LRI lab.

IEEE IFIP Networking 2014, Game Theoretic Framework for Power control in InterCell

Interference Coordination, Kinda Khawam, Amine Adouane, Samer Lahoud, Johanne Co-

hen, Samir Tohme, University of Versailles PRISM Laboratory, University of Paris Sud

LRI lab.

xiii

xiv TABLE DES MATIÈRES

Table des figures

1.2 LTE E-UTRAN architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 LTE Type-1 frame structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.4 LTE Type-2 frame structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.5 LTE RB composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.6 Inter-cell interference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.7 Interference Mitigation Classification . . . . . . . . . . . . . . . . . . . . . . 10

1.8 Frequency reuse 1 scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.9 Frequency reuse 3 scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.10 Partial Frequency Reuse Scheme . . . . . . . . . . . . . . . . . . . . . . . . 12

1.11 Soft Frequency Reuse scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.12 Downlink LTE FDD frame . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.2 Normal Mode vs. Accelerated Mode . . . . . . . . . . . . . . . . . . . . . . 31

3.3 Total Mean Cost : Random Algorithm vs. Replicator Dynamics Algorithm . 33

3.4 Relative Change : Random Algorithm vs. Replicator Dynamics Algorithm . 34

4.2 Convergence time as a function of load . . . . . . . . . . . . . . . . . . . . . 44

4.3 Random algorithm vs. Distributed Load Balancing algorithm as a function

of Load . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4.4 Random algorithm vs. Distributed Load Balancing algorithm as a function

of Frequency Band . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5.2 Speed of convergence as a function of load . . . . . . . . . . . . . . . . . . . 51

xv

xvi TABLE DES FIGURES

5.4 Trivial algorithm vs. Greedy algorithm as a function of frequency band . . . 53

5.5 Greedy algorithm with crowded cell-edge . . . . . . . . . . . . . . . . . . . . 54

5.6 DLB algorithm with crowded cell-center . . . . . . . . . . . . . . . . . . . . 55

6.1 Transfer time per zone as a function of power unitary cost for DBR vs. Max

Power Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6.2 Total Transfer Time as a function of power unitary cost for DBR vs. Max

Power Policy and Trivial Policy . . . . . . . . . . . . . . . . . . . . . . . . . 68

6.3 Power Economy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6.4 Convergence Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6.5 Global Transfer Time as a function of power unitary cost for DBR, Max

Power and Optimal policies . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

6.6 Power Economy as a function of power unitary cost for DBR and Optimal

policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

List of Abbreviations

ABSF Almost Blank Sub-Frame

BBU Base Band Unit

CN Core Network

CoMP Cooperative Multi-Point

CP Cyclic Prefix

CQI Channel State Indicator

CSB Cell Selection Bias

CSI Channel State Information

DCCH Dedicated Control Channel

EPC Evolved Packet network

FDD Frequency Division Multiplexing

FFR Fractional Frequency Reuse

FRF Frequency Reuse Factor

ICI Inter-Cell Interference

ICIC Inter-Cell Interference Coordination

i.i.d. independent and identically distributed

IMT International Mobile Telecommunications

IP Internet Protocol

ISI Inter Symbol Interference

ITU-R International Telecommunication Union-Radiocommunication

LTE Long Term Evolution

xvii

xviii TABLE DES FIGURES

MILP Mixed Integer linear Problem

MME Mobile Management Entity

ODE Ordinary Differential Equations

OFDMA Orthogonal Frequency Division Multiple Access

OI Overload Indicator

PAPR Peak to Average Power Ratio

PDSCH Physical Downlink Shared Channel

PER Paquet Error Rate

PFR Partial Frequency Reuse

PMI Precoding Matrix indicator

PMR Private Mobile Radio

PNE Pure Nash Equilibrium

PUCCH Physical Uplink Control Channel

PUSCH Physical Uplink Shared Channel

QAM Quadrature Amplitude Modulation

QPSK Quadrature Phase-Shift Keying

RAN Radio Access Network

RB Resource Block

RE Resource Elements

RI Rank Indicator

RNC Radio Network Coordinator

RNTP Relative Narrowband Transmit Power

RRH Radio Remote Head

RRM Radio Resource Management

RS Reference Signal

RSRP Reference Signal Received Power

RSRQ Reference Signal Received Quality

RSSI Received Strength Signal Indicator

SC-FDMA Single Carrier FDMA

SFFR Soft FFR

SFR Soft Frequency Reuse

S-GW Serving Gateway

SINR Signal to Interference-plus-Noise Ratio

TABLE DES FIGURES xix

TDD Time Division Multiplexing

TTI Transmit Time Interval

UE User Equipement

WIMAX Worldwide Interoperability for Microwave Access

ZFBF-SUS Zero Forcing BeamForming with Semi-orthogonal User Selection

xx TABLE DES FIGURES

Chapitre 1

Introduction

The 21st century is the century of communication that has set out new ambitious tar-

gets to meet the ever-increasing demand for UE (User Equipment) capacity in emerging

wireless systems. However, the physical channel of cellular networks has many constraints,

and meeting the requested target is very hard. In order to overcome the bounded capa-

city of such networks, there is a crucial need for higher spectral efficiency. Unfortunately,

this comes at the cost of increased interference level, which renders the expected solutions

even more challenging. Thanks to its various merits, Orthogonal Frequency Division Mul-

tiplexing (OFDM) has been adopted as the key physical layer technique in 4G wireless

system. Hence, to face the challenges ahead, the OFDM technology represents a key asset

for intra-cell interference mitigation owing to its orthogonality feature. Yet, for Inter-Cell

Interference (ICI), and in particular for dense frequency reuse, high interference on the

cell border can be encountered. Remedy to this ICI is paramount and can be treated with

efficient inter-cell interference coordination (ICIC) techniques which are the object of this

thesis. This chapter recaptures some key components of ICIC and highlights some of the

most relevant previous work. The plan of the thesis is given at the end of the chapter.

1.1 Introduction

The frequency reuse-1 is advocated by the Third Generation Partnership Project (3GPP)

[Pro06] for the LTE/LTE-Advanced networks. This allows to use the same resources all

over the network cells, taking benefits from the OFDM special features. OFDM has been

chosen as the 4G physical access due to its resistance to the Inter-Symbol Interference

(ISI), frequency selective fading, and to its high spectral efficiency. These benefits stem

from the orthogonality of sub-carriers composing the system frequency band, which allows

to virtually eliminate the intra-cell interference. Unfortunately, the UEs suffering from

poor channel quality will be exposed to ICI, and experience low throughput. Therefore, a

1

2 1. Introduction

lot of effort is being deployed to hinder ICI, the corresponding devised schemes are known

as Inter-Cell Interference Coordination (ICIC) techniques [DSZ12]. The main goal of ICIC

is to decrease the interference in order to provide better throughput for UEs suffering from

bad channel quality. We begin by reminding the reader of the main characteristics of LTE

then detail the various ICIC techniques used in the literature.

The long term evolution LTE was proposed as a solution to reduce data transfer time

in the cellular network, it was standardized in December 2008 and the first LTE service

was launched by TeliaSonera in Oslo and Stockholm on December 14, 2009 as a data

connection with a USB modem. The LTE main features are :

– Peak download rates up to 299.6 Mbit/s and upload rates up to 75.4 Mbit/s depending

on the UE category (with 4x4 antennas using 20 MHz of spectrum). Five different

terminal classes have been defined from a voice centric class up to a high end terminal

class that supports the highest peak data rates. All terminals will be able to process 20

MHz bandwidth.

– Low data transfer latency (sub-5 ms latency for small IP packets in optimal conditions),

lower latency for handover and connection setup time than with previous radio access

technologies.

– Improved support for mobility, exemplified by support for terminals moving at up to

350 km/h or 500 km/h depending on the frequency band.

– OFDMA (Orthogonal Frequency Division Multiple Access) for the downlink and SC-

FDMA (Single-Carrier FDMA) for the uplink to preserve power.

– Support for both Frequency and Time Division Duplexing FDD and TDD communica-

tion systems as well as half-duplex FDD with the same radio access technology.

– Support for all frequency bands currently used by IMT (International Mobile Tele-

communications) systems by ITU-R (International Telecommunication Union-Radio-

communication).

– Increased spectrum flexibility : 1.4 MHz, 3 MHz, 5 MHz, 10 MHz, 15 MHz and 20 MHz

wide cells are standardized.

– Support for cell sizes from tens of meters radius (femto and picocells) up to 100 km

radius (macro-cells).

– Supports of at least 200 active data clients in every 5 MHz cell.

– Simplified architecture : The network side of E-UTRAN is composed only of intelligent

eNodeBs

– Support for inter-operation and co-existence with legacy standards (e.g., GSM/EDGE,

UMTS and CDMA2000).

– Packet switched radio interface.

1.2. The LTE network 3

The goal of LTE is to increase the capacity and speed of wireless data networks, it was

then designed for full packet services ; it provides IP connection between the UE and the

network. Like the other cellular networks, the LTE network is formed by a wired Core

Network (CN) called Evolved Packet Core (EPC), and a wireless part called Evolved-

UTRAN (E-UTRAN), as we can see in Figure 1.1 [Cha13].

We are only concerned in this thesis with the access network detailed below in subsection

1.2.2.

The access network of LTE is called E-UTRAN, it consists of a batch of eNodeBs with

no centralized controller ; the E-UTRAN architecture is said to be flat as shown in Figure

1.2.

The eNodeBs are interconnected with each others over the X2 interfaces, each eNodeB

is connected with the MME (Mobility Management Entity) via the S1-MME interface

and with the S-GW (Serving Gateway) via the S1-U interface. In the E-UTRAN, the

eNodeB are the intelligent entity responsible of all the radio functions. Integrating the radio

controller functions into the eNodeB is one of the major upgrades of the LTE architecture

by virtue of which the system latency is reduced and efficiency increased.

The eNodeB functions can be summarized as follows :

– Radio Resource Management : it includes scheduling and dynamic allocation of resources

to UEs in both uplink and downlink, radio bearer control, radio mobility control and

radio admission control.

4 1. Introduction

better resource usage.

– Security : encryption of the data sent over the air.

– Positioning : help the E-SMLC (Evolved Serving Mobile Location Center) to position

the UE by sending relevant measurement.

– Connectivity to the EPC : providing the signaling towards the MME and the bearer

path towards the S-GW.

As it was done for the previous releases of cellular networks, the 3gpp has standardized

the frame structure of the wireless access network. The LTE frame is differently conceived

depending on the chosen division duplex : it is called Type 1 frame for the LTE FDD

(frequency division duplex mode) systems and Type 2 frame for the LTE TDD (time

division duplex) systems.

The Type 1 LTE frame has a length of 10 ms, it is divided into 20 individual slots as

shown in Figure 1.3. Each frame slot has a duration of 0.5 ms, in this case a slot represents

exactly a Resource Block (RB), which is the smallest radio resource we can allocate to a

UE [Com09].

1.2. The LTE network 5

The Type 2 frame is composed of 2 half frames of 5 ms as shown in Figure 1.4 ; each half

frame is divided into five sub-frames of 1ms each. According to the switch time, at least

one of the half frames contains three fields :

– GP - Guard Period

– UpPTS - Uplink Pilot Time Slot.

When the switch time is 10 ms, only the first half frame will carry this field ; whereas if

the switch time is 5ms, both half frames are used [Com09].

6 1. Introduction

A Physical Resource Block (PRB) is composed of six or seven symbols by twelve subcarriers

as displayed in Figure 1.5 [Com09], it is represented by a table of 6 or 7 rows and 12 lines,

each cell of the table is called a resource element and consist in a subcarrier of 15 kHz

for one symbol. This gives the RB a frequency band of 180 KHz. The number of symbols

is determined by the Cyclic Prefix (CP), if a normal CP is used the RB will have seven

symbols, if an extended CP is used it is composed of six symbols.

The X2 interface is used to interconnect the eNodeBs in order to exchange signaling infor-

mation when needed for load or interference management and for handover information.

We will focus on the load and interference messages since they are intimately related to

our work.

Since there is no central entity for Radio Resource Management (RRM) in the E-UTRAN,

the exchange of load and interference messages is essential among eNodeBs through the

so-called X2 interface. The physical characteristics of the X2 interface brings some latency

1.2. The LTE network 7

to the system, this latency imposes a message periodicity of approximately 200 ms. In the

following, we explain some of the important messages exchanged over the X2 link.

The eNodeB sends this message on the downlink in order to inform the neighboring eN-

odeBs about its power use on each RB. This message contains information relative to a

power threshold, when exceeded, the other eNodeBs should avoid using the corresponding

RB. The RNTP message is defined as follows [3gp12] :

1 if used power is greater than a given threshold,

RN T P = (1.2.1)

0 otherwise.

It indicates the PRB interference sensibility on the uplink, this allows the eNodeB to warn

other eNodeBs about the RB allocated with high power transmission ; in that case, the

neighboring eNodeBs will avoid allocating these resources to the mobile UEs with bad

channel quality [3gp11a].

The eNodeB sends information about the received interference to the adjacent eNodeBs

through the overload indicator message. Receiving an OI allows the eNodeB to manage

the resource scheduling to lower the interference received by mobile UEs with bad channel

quality [3gp11a].

This interface links the UE to the access network (the eNodeBs). In addition to carrying

the UEs data, the LTE-Uu interface allows the exchange of signaling messages between

the UE and the eNodeBs, some of these messages can be used to manage the interference

in the network. The UE sends to the eNodeB a CSI (Channel State Information) which

consists of channel quality indicator (CQI), precoding matrix indicator (PMI) and rank

indication (RI). Since in this thesis we consider only single antenna transmissions, the

PMI and RI messages are not considered. Hence, the most relevant message sent through

this interface for our work is the CQI (Chanel Quality Indicator) which has significant

impact on the system performance. It is sent by the UE to indicate to its serving antenna

the channel quality it perceives. In fact, downlink channel dependent scheduling in LTE

requires specific information to be sent by the terminals to the network. Such information

8 1. Introduction

is transmitted through channel state reports that contain CQI feedback. CQI represents

the highest modulation and coding scheme that guarantee a block error rate less than 10%

for physical downlink shared channel transmissions. There are two types of channel state

reports in LTE : periodic and aperiodic. The periodic CQI report is sent over the PUCCH

(Physical Uplink Control Channel) if no data is sent over the uplink, or in the PUSCH

(Physical Uplink Shared Channel) piggybacked with UE data in the same subframe, in this

case, the periodic PUCCH resource will be idle. The minimum periodicity could be 2 ms. If

the eNodeB needs more detailed CQI report, it can trigger an aperiodic CQI sent over the

PUSCH together with UL data or alone [3gp12]. Three levels of CQI report granularity can

be enumerated : wideband, UE selected subband, and higher layer configured subband. The

wideband report is responsible of the CQI information for the whole system bandwidth.

The UE selected CQI report provide wideband CQI information along with differential

CQI value, this CQI value indicates the report for multiple subband (the best M subbands

perceived by the UE). The highest layer configured subband is the most detailed report,

it provides wideband CQI value and multiple differential CQI for all the subbands. The

higher the CQI value (from 0 to 15) reported by UE, the higher the modulation scheme

(from QPSK to 64QAM ) and higher the coding rate will be used by eNodeB to achieve

higher efficiency. Having information on UE channel quality, the eNodeBs can give low

interfered RBs to UEs with bad channel quality for acceptable performances. On the

contrary, the eNodeB can reserve low interfered RBs to UEs with good channel quality to

enhance system performances.

In 4G networks, OFDM is used for the radio access as it allows increased spectral efficiency

while mitigating the intra-cell interference owing to its orthogonality feature. However, for

the inter-cell interference, when the frequency reuse one is deployed, close neighboring

UEs with same frequency will suffer from bad channel quality as highlighted by Figure

1.6. Inter-Cell Interference Coordination (ICIC) has proved to be one of the best solutions

to overcome this problem and has been adopted as a resource management mechanism to

enhance system performance. One of the first proposed solutions was the Fractional Fre-

quency Reuse (FFR) which manages statically the frequency resource distribution within

each cell in order to decrease the inter-cell interference [Ass08]. Instead of using all the

available RBs in adjacent cells, FFR gives them disjoint spectrum for their edge zones

(UEs at the cell periphery, usually suffering from bad channel quality). Hence, frequency

reuse 3 model is used for edge zones, and frequency reuse 1 for the cell center zone (UEs

close to the serving antenna, usually enjoying good channel quality). The ICI can also

be mitigated through power control on the allocated RBs. A good example is the well-

known Soft Frequency Reuse (SFR) [MMT08] which allocates lower power for cell-center

1.3. Inter-Cell Interference Coordination ICIC 9

RBs while cell-edge RBs get higher transmission power. Hence, interference mitigation is

achieved by either transmit power control [JL03] or by radio resource allocation schemes

[SAR09]. ICIC can also by a joint (complex) resource allocation and power adaptation

problem. Hence, FFR technique imposes restrictions on RB usage within each cell, while

SFR modifies both RB allocation and downlink power allocation. However, both schemes

are static schemes that fail to cope with a variable traffic.

In order to manage the inter-cell interference, different methods are used as shown in

Figure 1.7. They are divided into 2 main categories : Frequency Reuse-based schemes and

Cell Coordination-based schemes.

– The Frequency Reuse-based schemes are static methods characterized by the frequency

reuse factor (FRF) which is the rate at which the same frequency can be used in the

network. We can divide the frequency reused based schemes into :

1. The conventional frequency reuse which encompasses the frequency reuse 1 and 3,

2. The fractional frequency reuse (FFR) where we find the known Partial Frequency

Reuse (PFR), Soft Frequency Reuse (SFR), and Soft Fractional Frequency Reuse

(SFFR) among others.

– The Cell Coordination-based schemes are dynamic methods that include :

1. Centralized methods where a central entity takes decisions on behalf of a group of

eNodeBs using CoMP (Coordinated Multi-Point) [3GP11b], along with enhanced-

ICIC (e-ICIC) transmissions in LTE-A (LTE Advanced). This approach is optimal

but suffers from high complexity and long delays. Therefore, it is not suited for

quick network management.

10 1. Introduction

2. Semi-distributed networks where the different eNodeBs are the decision makers ;

however, their decisions are made by a central entity or aided by exchanged signa-

ling information among them. Hence, radio resource allocation, power allocation,

or both, are made using information sent over the X2 interface.

3. Distributed scheme are characterized by complete absence of cooperation among

eNodeBs. Hence, each eNodeB takes its decisions without relying on any assistance

or on any external signaling messages.

In the following section we will explain in more details all the methods just enumerated.

The main goal of the interference avoidance schemes is to control the allocation of radio re-

sources (frequency, time and power) in order to obtain higher SINRs (Signal to Interference

and Noise Ratio) and reduce interference between adjacent eNodeBs.

The LTE network is configured by default to be deployed with a Frequency Reuse Factor

(FRF) of 1. In this configuration, the whole frequency band is available in the entire cell

coverage area as shown in Figure 1.8. As already said, owing to the orthogonality feature

of OFDM, the intra-cell interference is mostly mitigated and can be ignored. However,

inter-cell interference is still problematic and UEs with bad channel quality can endure

high interference from adjacent cells.

1.3. Inter-Cell Interference Coordination ICIC 11

In order to face this problem, an easy solution is to adopt the FRF of 3. It consists in

dividing the entire frequency band into 3 parts, attributing each part to a cell in a way to

avoid using the same frequencies in any two adjacent cells as we see in Figure 1.9. This

method leads to a better SINR in the system and improves inter-cell interference, but since

the frequency band is divided by 3, the channel capacity decreases. This problem becomes

even more relevant when the system is loaded.

With the aim of improving the shortcomings of the FRF 3 scheme, the Fractional Fre-

quency Reuse FFR was developed. The main idea is to divide the cell into two geographical

areas, the area close to the eNodeB (with cell center UEs) and the area far away from the

eNodeB (with cell edge UEs). Each area will get a part of the frequency band : the cell-

edge UEs will get a fraction of the resources that are different from those of neighboring

cell edge UEs, while the rest of the frequency band will be allocated to the cell center UEs

(protected from adjacent interference by the edge area).

12 1. Introduction

Many derivative schemes from the FFR have been proposed, here are some of them :

1.3.1.2.1 Partial Frequency Reuse (PFR) FFR with full isolation (FFR-FI)

In the PFR scheme, the part of the spectrum allocated to the cell-edge UEs is divided into

three fractions as depicted in Figure 1.10. This results in an under-utilization of the radio

resource as some frequencies are not used in some sectors at all [KM11].

1.3.1.2.2 Soft Frequency Reuse (SFR) In order to overcome the deficiency of the

PFR scheme, the SFR scheme was proposed in [EE13] to provide some flexibility to the

latter. In the SFR scheme, the available bandwidth is divided into orthogonal segments,

and each neighboring cell is assigned a cell-edge band with a higher power allocation,

as shown in Figure 1.11 ; while the cell-center UEs can still have access to the cell-edge

bands selected by neighboring cells, but at a reduced power level. In this way, each cell can

utilize the entire bandwidth while reducing interference caused to the neighboring cells.

As a consequence, a lower ICI is achieved at cell-edges at the expense of reduced spectrum

utilization.

Unlike the frequency reuse schemes, the cell coordination schemes are dynamic and hence

able to adapt the resource allocation to the actual network state and traffic load at the

cost of increased complexity. The cell coordination schemes can be classified into three

main categories according to their operation mode.

The LTE architecture is characterized by the absence of a central entity that coordinates

base stations (such as the BSC Base Station Controller) and the RNC (Radio Network

1.3. Inter-Cell Interference Coordination ICIC 13

ced in [HDDM13] under the name Cloud-Radio Access Network (C-RAN), it allows the

coordination and management of the eNodeB over the cloud. This new solution allows the

connection of several aloof Radio Remote Heads (RRH) and Base Band Unit (BBU) to a

central point over the cloud. In the LTE-A [3GP08], a Coordinated Multi-Point (CoMP)

transmission and reception [3GP11b] is introduced, along with enhanced-ICIC (e-ICIC)

which brings a solution to the RB selection for the layered network (macro, small, pico and

femto cells). It also introduces the Almost Blank Subframe (ABSF) allowing time sharing

of resources between different cells [DMMS14].

Cooperative ICIC techniques benefit from the communication between network entities

to coordinate RBs distribution in a way to reduce ICI and improve system performance.

However, the cooperation between different base stations increases computational com-

plexity and generates an additional signaling load. More importantly, the resource mana-

gement issue may be related to a complex optimization problem where many constraints

are handled (interference level, SINR, ...). This problem can be seen as a multi-dimensional

allocation problem with resource restriction to mitigate the interference and it is proven

to be NP-HARD [RLLS98]. The issue of the NP-HARD complexity can be faced using

binary/integer linear programming [LCK08, RY07]. Nonetheless, the linearized problems

remain highly complex [LPJZ09].

In particular, in [NSV+ 14], the authors put forward a centralized ICIC management me-

thod applied to a LTE-A network, where a layered solution divides the network into groups

of few cells. This step is called ”Small-scale” coordination where the goal is to optimize the

total throughput, and then coordinate the optimized group to reach a better system per-

formance by lowering the interference. The authors showed that this problem is NP-HARD

complex and used a mixed-integer-linear problem (MILP) to bypass it. In [ENJA08], the

14 1. Introduction

matical model was proposed to mitigate the interference level. This is done by finding the

best routing and assigning slots to the end connection, the effect of the hidden terminal

and neighboring interference was considered. In [AADM07], the benefit of a centralized

schemes over distributed ones is highlighted, and an optimal resource allocation is propo-

sed with hard complexity. The proposed centralized scheme considers the constraints of

transmission quality and throughput while minimizing transmission power. In [BZGA10],

a solution for CoMP transmissions is proposed using the Medium Access Control (MAC)

scheduling in the LTE-A network. Each UE is served jointly by several cells that form

clusters, and each cluster is managed by its own central entity. These central entities are

aware of all the Channel State Information (CSI), Channel Quality Indicator (CQI), Pre-

coding Matrix Index (PMI) and Ranking Indicator (RI) of UEs served by their cluster.

The resource allocation for each cluster is done according to these messages. In [WWS+ ],

the problem of multi-layer network with macro and femto cells is treated. A dynamic cen-

tralized approach was proposed whose goal is to increase the forward link performance for

best effort and real time transmission in the femto cell network. The system is divided into

clusters, then a comprehensive research for optimal allocation between femto cells of the

same cluster is done. Finally, a conventional static scheme is applied (SFR, PFR, ...) at the

femto cell layer in order to reduce the radio Packet Error Rate PER. In [SSC11], the down-

link resource allocation problem is studied where the proposed solution comes into two

levels. The solution for the first level (the main level) is done using projected-subgradient

method. At the second level, minimum-cost network flow optimization was used to solve

each sub-problem. The authors showed through simulations that the obtained results are

comparable to those with frequency reuse 3 scheme for cell edge UEs.

that controls LTE base stations ; however, neighboring eNodeBs are connected via the

X2 interface through which they can exchange information relative to their load and

interference level. In such an approach, the intelligence of the network is dispatched on

many components as part of the decision is taken by the eNodeBs themselves. A central

entity can be responsible of giving the eNodeB the global framework of the allocation. This

approach allows working on different time scales, for example the central entity can give

a pool of resources to each eNodeB, and then the distributed algorithm on each eNodeB

will manage the allocation of these resources to their UEs.

In [SMB+ 14], a semi-distributed method has been used to overcome the ICI problem by

minimizing the total time needed by an eNodeB to send data over the network. An algo-

rithm was proposed based on the LTE-A new message (Almost Blank Sub-Frame) ABSF,

1.3. Inter-Cell Interference Coordination ICIC 15

which allows a sub-frame to be empty (Blank) for some eNodeBs, in order to avoid using

these sub-frames when the interference override a threshold. In [SEH14], the authors ad-

dressed the problem of interference with a new Location Aware multicellular Cooperation

(LAC) scheme where they combined two existing methods. The first method is a joint

zero-forcing beamforming with semi-orthogonal UE selection (ZFBF-SUS) transmission,

and the second method is a semi-distributed power allocation optimization. They used

CoMP transmission to serve the UEs with bad SINR, while the other UEs are served

using multiuser MIMO. A two steps algorithm is used : a distributed resource allocation is

done on each eNodeB for the non-CoMP UEs, and then a central entity serves the CoMP

UEs. In [MMT08], the uplink resource allocation is tackled through a semi-decentralized

dynamic soft frequency reuse scheme. The proposal tends to overcome the problem of

throughput harvesting for the cell edge UEs with the respect of RBs reuse minimization

via the use of a per frequency sub-band indicator sent over the X2 link to neighboring

cells. Hence, when a cell faces a need of supplementary RBs for cell edge UEs, it will send

a message over the X2 link to loan the requested RBs from adjacent cells. The coexistence

of macro and small cells is becoming a relevant problem with the growing demand for ca-

pacity, which induces an important traffic load on the system and co-channel interference.

For instance, in [LLK+ 11], the authors introduced the time domain ICIC using the ABS

for heterogeneous networks. A cooperative system was used to overcome the interference

caused by the small cells .

In the case of decentralized ICIC, the eNodeBs optimize their local parameters without

the help of central controllers or signaling exchange among peers. Decentralized ICIC can

react to relatively fast change of situation but it is difficult to avoid converging to local

optimums. The non-cooperative game theory is quite suitable to model the way eNodeBs

compete in a distributed manner for limited resources. The goal is to maximize a utility

function to match the best resource allocation for each cell. Planning a game theoretical

RBs selection scheme depends on the existence of Nash equilibriums for the modeled

game, where no player will take advantage from moving unilaterally from the attained

equilibrium.

In [MPS07], each UE tries to maximize the utility function which can be the transmit

power, carrier allocation strategy, transmission rate or modulation. A Nash equilibrium is

proved to be reached for this power control game. In [EHB08], a decentralized scheduling

algorithm is proposed aiming at lowering the interference with a local knowledge of the

system. Each eNodeB chooses a pool of low interfered resources, and then a fast allocation

algorithm will schedule them for active UEs. A Nash equilibrium is proven to be attained.

In [EASH09], the authors show that a decentralized game theoretical approach, where a

16 1. Introduction

Nash equilibrium is reached, gives better system performances, especially for non-uniform

traffic load compared to the frequency reuse 1 and 3 deployments.

In [MDV13], a distributed cognitive interference alignment method is proposed to address

the interference problem. Using only autonomous operations and local Channel State In-

formation (CSI), they optimized the spectral efficiency by means of a distributed one-shot

strategy. In [DMMS14], the author addressed the problem of interference with a decen-

tralized eICIC algorithm. They used Almost Blank Sub-frame (ABSF) and Cell Selection

Bias (CSB) to manage the interference between macro and pico cells. The algorithm was

implemented in a real network in New York city and showed promising results on the

deployed LTE network. In [PWSF13], the use of eICIC in a decentralized manner is also

discussed ; the problem of mixed deployment of macro and small cells is addressed using

eICIC and ABSF. Further, benefits of time domain resource sharing over network layers

are outlined.

To avoid converging to local optimum, many works rely on non-cooperative game theory,

it appears to be suitable to model the competition between eNodeBs for the scarce ra-

dio resources. In [AKG11], the interference problem in cognitive radio is portrayed as a

non-cooperative game to allocate resource in a distributed fashion while overcoming the

problem of co-channel interference. The goal is to provide autonomous behavior with dy-

namic resource allocation inside the femto cell component of the network. In [KZM12], the

uplink resource allocation in a cognitive network is tackled where a price based mecha-

nism is used to manage the interference among femto cells. A Stackelberg game is used to

address the sharing of resources between the macro and femto cells, with the macro cell

acting as the master and the femto cells as the followers. In [LYW+ 11], the ICI problem in

a LTE-A network is tackled using an evolutionary potential game, a jointly optimization

of resource allocation and power control is treated as a cross-layer ICIC framework. The

author used a Lagrangian multiplier method to optimize the constraints of the system,

then a potential game is introduced to allocate the RBs and tune their power to increase

the throughput. Simulation results highlighted the benefits of such a solution. Another

game to overcome co-channel interference in a heterogeneous macro/femto cells network

is presented in [ZCM+ 12]. A weighted technique is used, where the impact of the interfe-

rence of each allocation is given a price. Simulation results show that the proposed scheme

strikes a good balance between fairness and efficiency of resource allocation among both

macro and femto cells.

system interference, and hence on the achieved data throughput. The physical layer uses a

1.4. Downlink Power Control in LTE 17

special signal named Reference Signal (RS) that allows the UE to figure out the downlink

cell power. Two types of RS exist, the cell specific RS which is sent all over the bandwidth

and for all the sub-frames, and the UE specific RS which is sent over the dedicated UE RB.

The RS is sent on the downlink transmission every 6 subframe for the Frequency domain,

and every 4 OFDMA symbols for the time domain as shown in Figure 1.12 [AR11]. The

RS is sent as a clear reference for the power estimation, more precisely with the Reference

Signal Received Power (RSRP) and the Reference Signal Received Quality (RSRQ).

The RSRP signal is used to estimate the downlink power, it represents the average po-

wer of Resource Elements (RE) that carry cell specific Reference Signals over the entire

bandwidth. It is used for handover and cell selection/re-selection. The measurement of the

RSRP ranges from -44 to-140 dBm and the RSRP levels for usable signal typically range

from -75 dBm (for UEs close to the cell antenna) to -120 dBm (at the edge coverage).

Reference Signal Received Quality (RSRQ) is defined as the following ratio :

N ∗ RSRP

RSRQ = (1.4.1)

E − U T RAcarrierRSSI

Where N is the number of Physical Resource Blocks (PRBs) over which the Received

Strength Signal Indicator (RSSI) is measured, typically equal to system bandwidth. RSSI

is a parameter which provides information about total received wide-band power (measured

for all symbols) including total interference and thermal noise. Owing to the relevance of

the RS signal, it will be assigned the highest power during transmission, this power is

18 1. Introduction

cell-specific and must be constant over the whole bandwidth since all the other signals

(PDSCH, DCCH, synchronization, etc.) are calculated relatively to it. The RS power goes

from -60 to 50 dBm. Since the transmit power of an eNodeB is shared on all subcarriers,

the larger the bandwidth is, the lower the power on each subcarrier. Therefore, LTE uses

PA and PB parameters to adjust the power in order to have a constant power for all the

OFDMA symbols at the receiver :

– PA indicates the ratio of the data subcarrier power of OFDM symbols excluding pilot

symbols to the pilot subcarrier power.

– PB indicates the ratio of the data subcarrier power of OFDM symbols including pilot

symbols to the pilot subcarrier power.

Power allocation in OFDMA based networks has been widely studied and addressed in

different works. In [CLL13], the authors addressed the ICIC interference with a sequential

frequency reuse. Better SINR to cell edge UE is obtained by allocating more power for

them, while each cell is given sequentially a sub-channel of the frequency bandwidth to

avoid interfering with neighboring cells. In [YKS13], the problem of interference is studied

for both multicellular networks and layered networks with macro, pico and femto cells. A

power spectrum adaptation algorithm with a heuristic joint proportionally fair scheduling

and spatial multiplexing scheme was proposed. Authors showed that a relevant improve-

ment of system performances is realized in comparison with a fixed power allocation. In

[HMLN13], the authors presented an energy efficient power allocation for a point-to-point

multi-carrier link. A two level algorithm is adopted : first, a power optimization is perfor-

med without constraints ; second, the obtained power allocation is made in respect to the

system power constraints through a joint time and frequency optimization. In [SQ09], a

closed loop adaptive power control is proposed that aims at providing fair SINR from the

UEs point of view. Simulation results confirmed that power saving was possible while lo-

wering system interference and increasing fairness for UEs. In [CAM+ 09], the interference

problem between macro and femto cells is addressed where a distributed power allocation

algorithm is applied by femtocells in a way that the SINR is optimized without harming

the macro cell. In [BCRG14], the power control on the uplink is studied where authors

proposed a fully distributed algorithm where a limit on power is imposed to the UEs

to avoid interference with neighboring cells. In [PTLa14], a cognitive radio network was

studied where a decentralized beamforming algorithm where only local CSI information

to manage power allocation. In [LAV14], another beamforming decentralized approach is

proposed using the SRS message and the reciprocity of the radio channel. The results

show a significant improvement of the UEs throughput. In [BPVG06], the authors tried to

minimize the allocated power to UEs while guaranteeing a constant throughput. A non-

collaborative algorithm is proposed where each eNodeB performs its resource allocation

independently, taking into account the interference level among adjacent nodes.

Another important aspect of the OFDMA systems is the Peak-to-Average power Ratio

1.5. Plan of the Thesis 19

(PAPR). The authors of [AkI08] used a hybrid algorithm with an inverse relation between

power control and adaptive modulation (when the modulation order increases, the power

decreases). This leads to the lowest possible power while using high modulation order.

The authors in [SQ09] pointed out that the use of high power for center UEs leads to

performance degradation at the edge of the cell. Hence, they propose a solution with an

adaptive ower control scheme that aims at providing a fair SINR to the UEs.

Distributed power control schemes are sought for in 4G networks and are preferred to RRM

schemes with no power tuning due to lack of cooperation among serving nodes [KW11]. In

[SV09, WKSV10], dynamic Soft Frequency Reuse (SFR) schemes are put forward with pro-

ved efficiency. Furthermore, several solutions have been proposed to overcome interference

in OFDM systems using joint dynamic power and subcarrier allocation as in [KLL03].

Further, the joint power and RB allocation in [WCLM99, CKKL04] takes advantage of

the channel quality fluctuation [WCLM99, CKKL04]. This dynamicity permits to decrease

the power without degrading the BER (Bit Error Rate).

The main ICIC techniques are surveyed in Chapter 1. We discuss and classify a wide range

of methods that are either static or dynamic. A brief description of the LTE architecture

is provided for a full comprehension of the ICIC framework. The rest of the thesis is

organized as follows. The work in Chapter 3 has been done for the purpose of the SOAPS

project where downlink performances of OFDMA narrow band systems is enhanced. A

fully decentralized algorithm based on replicator dynamics was put forward to attain the

PNEs of the ICIC potential game. Chapters 4 and 5 address the problem of downlink

ICIC where the resource selection process is apprehended as a congestion game. A fully

decentralized algorithm was adapted to the paper context to attain the PNE of the modeled

game. Each eNodeB strives to select a pool of favorable resources with low interference

based on local knowledge only. Furthermore, performance improvement is realized in a

time coherent with the signaling period of the X2 interface. In chapter 6, the ICIC issue

is treated through adequate power allocation on selected resource blocks. The power level

selection process is apprehended as a sub-modular game and a semi distributed algorithm

based on best response dynamics is proposed to attain the NEs of the modeled game. Based

on local knowledge conveyed by the X2 interface, each eNodeB will first select a pool of

favorable resource blocks with low interference. Second, each eNodeB will strive to fix the

power level adequately on those selected RBs realizing performances comparable with the

Max Power policy that uses full power on selected RBs while achieving substantial power

economy. We summarize our contributions in the last chapter and give insights to future

research directions.

20 1. Introduction

Chapitre 2

System Framework

This thesis have been done in the context of the SOAPS (Spectrum Opportunistic

Access in Public Safety) project. The SOAPS project addresses low layer protocols issues

for Broadband Services provision by PMR (Private Mobile Radio) systems using LTE

technologies with particular focus on the improvement of frequency resource scheduling.

The work of this thesis goes beyond the project (dedicated for systems with only 6 RBs)

and propose more general solution for the ICI problem in wide band OFDMA systems (with

a large number of resource blocks). We introduce in this chapter the general framework we

have used along this thesis.

In order to validate our theoretic approach, our simulations are done in a cellular OFDM

based network model suitable for LTE, and hence for the SOAPS project. We present

hereafter the radio network framework :

– We model the system as a cellular network comprising N = {1, ..., n} hexagonal cells.

– The work focuses on the downlink scenario.

– The SOAPS project relays on LTE, hence OFDMA is used as the multiple access scheme.

– The time and frequency radio resources are grouped into time-frequency Resource Blocks

(RB).

– Each RB consists of Ns OFDM symbols in the time dimension and Nf sub-carriers in

the frequency dimension (in LTE, Ns = 7 and Nf = 12).

– The set of RBs are represented by M = {1, .., m} where m is the total number of RBs.

– We do not consider the MIMO configuration, therefore both eNodeBs and UE have

single antenna each.

– We consider without loss of generality, that each UE gets only one RB at each scheduling

iteration.

21

22 2. System Framework

One of the relevant parameters indicating a channel quality for the end UE is the data rate

perceived on the UE equipment. In order to calculate this data rate, we have to estimate

the SINR (Signal to Interference plus Noise Ratio) per RB.

P0 · Gi · ( di1 )β

u,i

SIN Ri,k,u = P 1 β (2.1.1)

P0 · j∈M, Gj · xj,k · ( di ) + PN

j6=i u,j

– P0 represents the maximal transmitted power per RB. No power control, other than

on/off with equal power levels, is assumed. As said, chapter 6 tackles power control for

ICI.

– PN represents the thermal noise power per RB.

– Gi is the antenna gain of eNodeB i.

– diu,j is the distance between eNodeB j and UE u served by eNodeB i.

– β is the path-loss factor varying between 2 and 6.

1 if RB k is used by eNodeB j,

xj,k = (2.1.2)

0 otherwise.

follows :

W

Di,k,u = · SIN Ri,k,u ,

Eb /N0

where W is the bandwidth per RB. Given a target error probability, it is necessary that

Eb /N0 ≥ γ, for some threshold γ which is UE specific.

Each cell will be logically divided into nz concentric discs of radii Rz and the area between

two adjacent circles of radii Rz−1 and Rz is called zone z ∈ Nz where Nz = {1, .., nz }

represents the set of zones. We consider that the UE belonging to the same zone z have

the same radio conditions leading to the same γ (denoted by γz ) and the same mean rate

2.2. Non-Cooperative game for RBs selection 23

W

R Rz

γz Rz−1 ρzmobile 2πrdr

rβ

· Gi · P0

Di,k,z = P 1

P0 · j∈M, Gj · xj,k · (di )β + PN

j6=i i,j

W 2−β 2−β

(2.1.3)

γz (Rz − Rz−1 ) · ρzmobile · Gi · xi,k

= P Gj

j∈M, xj,k · (δ z ·R β + η

i,j cell )

j6=i

where Rcell is the cell radius. As for interference, we consider mainly for simplification

the impact of eNodeB j on eNodeB i by replacing diu,j with diz,j = δi,j

z ·R

cell the distance

z depends on how far is eNodeB j from

between eNodeB i and eNodeB j (the value of δi,j

PN

zone z of eNodeB i). Finally, η = P0 . As η << 1, it will be neglected in what follows.

We denote by Ti,k,z the amount of time necessary to send a data unit through RB k in

eNodeB i for UEs in zone z. In fact, the delay needed to transmit a bit for a given UE is

the inverse of the data rate Di,k,z perceived by this UE :

1

Ti,k,z =

Di,k,z

P

j∈M, yj,k

z

· Hi,j (2.1.4)

j6=i

=

Hi,z

z = Gj

where Hi,j z ·R

(δi,j β captures distance-dependent attenuation between eNodeB j and

cell )

W 2−β 2−β

eNodeB i and for eNodeB i and Hi,z = γz (Rz − Rz−1 ) · ρzmobile · Gi captures distance-

dependent attenuation inside zone z.

Non-Cooperative game theory models the interactions between players competing for a

common resource. Hence, it is well adapted to distributed ICIC modeling. Here, eNodeBs

are the decision makers or players of the game. We define a multi-player game G between

the n eNodeBs. The eNodeBs are assumed to make their decisions without knowing the

decisions of each other. We present hereafter the general framework of the game used in

Chapters 3, 4 and 5. The final game is devoted to power control and will be presented in

details in the sequel of Chapter 6.

– The set of players is N = {1, .., n}.

– The set of available RBs is M = {1, .., m}.

24 2. System Framework

– Each eNodeB is a player that has to pick several RBs among the m available RBs.

– The strategy of eNodeB i is denoted by the m-dimensional binary vector xi ∈ Si whose

components are the xi,k variables defined in Equation (2.1.2) and where Si is the set of

eNodeB i’s strategies (Si = Bm ).

– X = (Xi )i∈N ∈ S = {1, 2, · · · , m}N is a pure strategy profile, where S is the set of

strategy profiles.

– In every chapter, we define ci (xi , X−i ) the cost function of eNodeB i that selects action

xi , where X−i denotes the vector of strategies played by all other eNodeBs except

eNodeB i.

The cost function varies from chapter to chapter and will be detailed in each chapter for

clarity.

a Nash Equilibrium (NE). A NE is a profile of strategies in which no player will profit

from deviating its strategy unilaterally. Hence, it is a strategy profile X ∗ ∈ S where each

player’s strategy is an optimal response to the other players’ strategies :

ci (X ∗ ) = ci (x∗i , X−i

∗

) ≤ ci (xi 0 , X−i

∗

), ∀i ∈ N, ∀xi 0 ∈ Si (2.2.1)

In general, finite games possess mixed NE but are not guaranteed to have pure NEs.

For mixed NE, each eNodeB has to continually change its RBs selection according to a

distribution probability over the strategy set. Implementing such solutions is cumbersome,

hence, for our portrayed games, we only want to allow deterministically chosen actions,

the so-called pure strategies.

Our games are distributed and need a local feedback from the system to compute the cost

of each allocation, we relies on the CQI sent back from the UE to calculate that cost.

2.3 Conclusion

The defined framework presented in this chapter will be used for all the upcoming chapters,

however some added aspects will be detailed for each contribution.

Chapitre 3

Distributed Inter-Cell Interference

Coordination

This chapter addresses the problem of ICIC in the downlink of cellular OFDMA sys-

tems where the resource selection process is apprehended as a load balancing game. Proving

the existence of Pure Nash equilibriums (PNE) shows that stable resource allocations can

be reached by selfish eNodeBs. We resort to a fully decentralized algorithm based on a re-

plicator dynamic to reach the PNE of the ICIC game. Each eNodeB will strive to select a

pool of favorable resources with low interference based on local knowledge only[AKC+ 14].

3.1 Introduction

In this chapter, we focus on the problem of inter-cell interference in the context of the

SOAPS. In this context, we propose an ICIC scheme modeled as an exact potential game.

Furthermore, we prove that a replicator dynamics algorithm converges to the pure NEs

of our game. The result of the devised coordination process in each cell will be a pool of

RBs that is not too highly interfered.

The rest of the chapter is organized as follows. The network model and cost characteri-

zation are given in the reference model in Section 2.1 presented in Chapter 2. The RBs

selection scheme is presented as a non-cooperative potential game in Section 3.3. The dis-

tributed learning algorithm based on replicator dynamics to reach pure NE is presented in

Section 3.4. Simulation results are portrayed in Section 3.5. Conclusion is given in Section

3.6.

25

26 3. Replicator Dynamics for Distributed Inter-Cell Interference Coordination

Non-Cooperative game theory is used to model the ICIC, where selfish eNodeBs share

common resources in a way to enhance their own local performance. We adopt the multi-

player game G between the n eNodeBs defined in Section 2.2. We will define the cost

function of a eNodeB i that selected strategy xi as follows :

X X

ci (xi , X−i ) = xi,k · Ti,k,z

k∈M z∈Nz

z (3.2.1)

P

j∈N, xj,k · Hi,j

j6=i

X X

= xi,k

Hi,z

k∈M z∈Nz

Where X−i denotes the vector of strategies played by all other eNodeBs except eNodeB

i. Note that the eNodeB will single out the strategy that minimizes the total bit transfer

time in its cell.

Potential games [Ros73] form a special class of normal form games where the unilateral

change of one UEs strategy xi to x0i results in a change of its utility function that is equal

to the change of a so-called potential function φ : S → R as follows :

A potential game [Ros73] admits at least one pure NE which is a desired property in this

context for practical reasons.

Démonstration. We will define our potential function which maps a profile X = (x1 , x2 , . . . , xn )

to a real : P z

j∈N, xj,k · Hi,j

j6=i

XX X

φ(X) = 1/2 xi,k (3.3.1)

Hi,z

i∈N k∈M z∈Nz

Now, we will prove that if X and X 0 are two pure profiles which only differ on the strategy

of one eNodeB `, then c` (x` , X−` ) − c` (x0` , X−` ) = φ(x` , X−` ) − φ(x0` , X−` ).

3.4. Distributed Learning of PNE 27

2φ(Y ) − 2φ(Y 0 ) =

X X X xi,k X

z z

( xj,k Hi,j + x`,k Hi,` )

Hi,z

i∈N, k∈M z∈Nz j∈N,

i6=` j6=i,

j6=`

X X X xi,k X

− ( z

xj,k Hi,j + x0`,k Hi,`

z

)

Hi,z

i∈N, k∈M z∈Nz j∈N,

i6=` j6=i,

j6=`

X X x`,k X

z

x0`,k X z

+

H`,z ( xj,k H`,j ) − ( x j,k H`,j )

H`,z

k∈M z∈Nz j∈N, j∈N,

j6=` j6=`

X X X xi,k

= H z · (x`,k − x0`,k )

Hi,z i,`

i∈N, k∈M z∈Nz

i6=`

X X X xi,k X X X xi,k

= x`,k z

Hi,` − x0`,k Hz

Hi,z Hi,z i,`

k∈M z∈Nz i∈N, k∈M z∈Nz i∈N,

i6=` i6=`

=2 · (c` (x` , X−` ) − c` (x0` , X−` ))

Implementing a practical distributed RBs selection policy to reach pure NE is not straight-

forward and must be carried out carefully. In this chapter, we resort to replicator dynamic

([BC13]) to learn Nash equilibriums.

pure strategies. In other words, pure strategy s is chosen with probability qi,s ∈ [0, 1], with

Pm

s=1 qi,s = 1. Let Ki be the simplex of mixed strategies for eNodeB i. Any pure strategy

s can be considered as a mixed strategy es , where vector es denotes the unit probability

vector with sth component being a unity component, hence a corner of Ki .

Let K = ni=1 Ki be the space of all mixed strategies. A strategy profile Q = (q1 , ..., qn ) ∈ K

Q

specifies the (mixed or pure) strategies of all players. Following classical convention, we

write Q = (qi , Q−i ), where Q−i denotes the vector of strategies played by all other eNodeBs

besides eNodeB i.

28 3. Replicator Dynamics for Distributed Inter-Cell Interference Coordination

Definition 3.1. The game mechanics work as follows : at t = 0, we begin with q(0) =

(q1 (0), ..., qn (0)) any random vector of probabilities. At each iteration t > 0 :

1. Each eNodeB i chooses an action xi (t) according to probability distribution qi (t).

2. Each eNodeB i learns the cost ci (t) resulting from the RBs allocation xi (t) it used

and the set of all actions of other players.

3. Each eNodeB i updates the probability vectors qi (t + 1) in the following way :

ci (t)

qi,s (t) + b(1 − cmax )(1 − qi,s (t))

if s = xi (t),

qi,s (t + 1) = ci (t)

(3.4.1)

qi,s (t) − b(1 − cmax )qi,s (t)

otherwise,

Theorem 3.1. Let G be an instance of game G. Let K∗ be a set of mixed profiles where

at most one player plays a pure strategy. The learning algorithm, for any initial condition

in K − K∗ , always weakly converges to a Nash Equilibrium.

Proof of convergence of Theorem 3.1 is given in the appendix A at the end of this thesis.

Algorithms of this form are fully distributed as decisions made by eNodeBs are completely

decentralized : at any iteration t, eNodeB i only needs to know its own cost ci (t) and mixed

strategy qi (t). To compute the cost ci (t), any eNodeB i makes use of signaling information

already present in the downlink of an LTE system. In fact, a UE assigned to a specific RB

measures its channel quality based on pilots, i.e. Cell Specific Reference Signals (CRS) that

are spread across the whole band independently of the individual UE allocation. Hence,

the cost function can be easily inferred through the CQI (Channel Quality Indicator) sent

every TTI (Transmit Time Interval) by serviced UEs. We consider that a time iteration

in our algorithm is equal to a TTI (1ms).

For the needs of the SOAPS project, we consider a bandwidth of 1.4 MHz with 6 RBs used

for PMR systems. The choice of the replicator dynamics is well adapted for the present

context to reach NE. In fact, the replicator dynamics algorithm is efficient but is known

to converge slowly to NE especially for!games with a large set of actions. With m RBs,

m m!

the number of possible actions is = for a eNodeB with k active UEs.

k k! · (m − k)!

With either 1 or 5 active UEs, we only have 6 possible actions. But with 50% of load (3

UEs), we reach 20 possible actions. With larger bandwidth systems, the set of strategies

3.5. Simulation Results 29

becomes very large. For instance, for a system with 15 RBs and 50% of load, we get

6435 possible actions rendering convergence times prohibitively long. For comparison, we

considered for all simulation results a system with 10 RBs and hence with a maximum of

252 actions at 50% of load.

Moreover, for LTE networks, a convergence time smaller than 200 TTI is sought for. In fact,

the RNTP (Relative Narrow-band Transmit Power) indicator, received from neighboring

eNodeBs every 200 TTI through the X2 interface, advertises on which RBs a neighboring

eNodeB will use high power. Hence, each eNodeB can aggregate information about trans-

mit power levels from adjacent cells and decides accordingly to preclude some RBs that

will be highly interfered (or alternatively allocate them to mobile UEs with good channel

quality) as already explained in the introductory chapter. Although this mechanism is

not taken into account as this thesis is dedicated to distributed ICIC, the convergence

time to NE should be nonetheless coherent with the RNTP signaling time to cope with

semi-distributed schemes.

We consider the following additional parameters listed in the 3GPP technical specifications

TS 36.942 :

– The mean antenna gain in urban zones is 12 dBi (900 MHz).

– Transmission power is 43 dBm (according to TS 36.814) which corresponds to 20 Watts

(on the downlink).

– eNodeBs have a frequency reuse of 1, with W = 180 KHz.

As for noise, we consider that the UE noise figure is 7.0 dB and the thermal noise is −104.5

dBm which gives a receiver noise floor of pN = −97.5 dBm.

We consider 10 hexagonal cells where the distance between two neighboring eNodeB is 2

Km. Two zones are taken into account : zone 1 which stands for cell-center UEs located

at a distance smaller than R0 = 0.5Km and zone 2 stands for cell-edge UEs located at

a distance ranging between R0 = 0.5Km and R1 = Rcell = 1Km. We assume that for

cell-center UEs 64-QAM modulation is used while for cell-edge UEs 16-QAM modulation

is used. The path-loss factor is set to 3. For each scenario, 400 simulations were run, where

in each cell a random number of UEs is chosen at the beginning of the simulation, they

are uniformly distributed among zones and correspond to a snapshot of the network.

In Figure 3.1, we depict the strategy dynamics qi of two randomly selected eNodeB (eN-

odeB 0 and eNodeB 3) as a function of the number of iterations.

All eNodeBs have only one center UE which gives them 6 possible actions (1, 0, 0, 0, 0, 0) ;

(0, 1, 0, 0, 0, 0) ; (0, 0, 1, 0, 0, 0) ; (0, 0, 0, 1, 0, 0) ; (0, 0, 0, 0, 1, 0) ; (0, 0, 0, 0, 0, 1). We can see

that the eNodeBs strategies converge to either 0 or 1, opting for one single action which can

be practical. We recorded this behavior through all the extensive simulations we performed

30 3. Replicator Dynamics for Distributed Inter-Cell Interference Coordination

(a) eNodeB 0

(b) eNodeB 3

We also notice the slow convergence of the algorithm (around 300 iterations are required

for the present case) which hinders the benefits of a distributed approach. As mentioned,

we wish to reach convergence faster than 200 iterations. However, we see through the

extensive simulations we ran that the convergence is relatively fast at the beginning of

the algorithm but slows down drastically half way through. At that point, the set of RBs

that will be ultimately selected by each eNodeB is clearly designated (we can see this

behavior in Figure 3.1 around 100 iterations) and it is useless to pursue the RB selection

process. Hence, we propose an accelerated mode of the algorithm such that whenever for a

3.5. Simulation Results 31

given eNodeB the probability of selecting any action surpasses 0.8, convergence is assumed

to be reached and the eNodeB chooses this particular action. In Figure 3.2, we run 100

simulations to compare the number of iterations of the accelerated mode (denoted REP

ACC) vs. the normal mode (denoted REP) for respectively 6 and 10 RBs as a function of

the mean number of active UEs per eNodeB.

(a) 6 RBs

(b) 10 RBs

We can see in Figure 3.2 the tremendous improvement in speed for the accelerated mode as

compared to the normal implementation of the replicator dynamics algorithm. However,

although for the 6 RBs system (which is the system of interest in this chapter), convergence

time is now coherent with the RNTP signaling time (limited to 220 iterations at 90% load),

32 3. Replicator Dynamics for Distributed Inter-Cell Interference Coordination

it largely surpasses this threshold for the 10 RBs system (around 5000 iterations at 90%

load). Note that the convergence time for the normal mode at 90% load is not reported in

the figure for clarity as it reaches a mean of 1.970.213.913 iterations.

In order to show the benefit of our solution, the performances are compared against a

Random algorithm where actions are chosen at random. For every snapshot, 400 runs of

P

the random algorithm were made. In Figure 3.3, we depicted the total cost ctotal = i∈N ci

(where ci is given in equation (4.3.1)) as a function of the mean number of UEs per eNodeB

for 6 RBs and 10 RBs. We reported the total cost for the random algorithm and the

replicator dynamics in normal mode and accelerated mode. Further, we captured results

for the replicator dynamics in normal mode at three different instants : at 100 iterations

(denoted REP 100IT), at 200 iterations (denoted REP 200IT) and at convergence (denoted

REP CONV). In Figure 3.4, the mean relative change of cost is displayed as a function of

the mean number of UEs per eNodeB where the relative change in eNodeB i is given by

cREP −cRandom

i i

cREP

∗ 100 with cRandom

i and cREP

i being the cost function according to Equation

i

(4.3.1) for the random algorithm and the replicator dynamics respectively.

3.5. Simulation Results 33

(a) 6 RB

(b) 10 RB

Figure 3.3: Total Mean Cost : Random Algorithm vs. Replicator Dynamics Algorithm

We can see that the main performance improvement is achieved during the first 100 itera-

tions in comparison with the random algorithm especially for the PMR system (6 RBs). All

proposed versions are roughly equivalent in terms of performance although the accelera-

ted version impedes slightly the performance improvement due to its aggressive approach.

Considering an iteration to be equivalent to a TTI, this result is quite attractive since we

obtain good results before eNodeBs exchange the RNTP messages among them through

the X2 interface. The latter signaling information should normally incite eNodeBs to ex-

clude certain actions from their strategy space (due to potentially high interference on

some of their RBs) before restarting the algorithm with new initial conditions set accor-

ding to the RNTP message content. Furthermore, as expected, the reduction in total cost

34 3. Replicator Dynamics for Distributed Inter-Cell Interference Coordination

decreases as load increases in comparison with the random algorithm (around 10% for the

maximum load for the 6 RBs system and falling below that for the 10 RBs system) as can

be seen in Figure 3.4. This behavior was expected, with the increase of system load, there

is fewer room for the algorithm to change the allocated strategy, most resources are being

already used rendering interference inevitable.

(a) 6 RB

(b) 10 RB

Figure 3.4: Relative Change : Random Algorithm vs. Replicator Dynamics Algorithm

3.6. Conclusion 35

3.6 Conclusion

In this chapter, the Inter Cell Interference problem was addressed for the special case of the

PMR network where RBs are astutely allocated in a distributed fashion. We put forward

a potential game to model the interference coordination among eNodeBs and propose a

fully distributed algorithm based on replicator dynamics to reach pure NEs. Numerical

simulations assessed the good performances of the proposed approach in comparison with a

random resource allocation. More importantly, improvement is realized in a time coherent

with the RNTP signaling time.

In future work, the RNTP messages will be taken into account to evaluate their impact

on the replicator dynamics algorithm performances. In fact, after receiving new RNTP

messages, the replicator dynamics algorithm will be restarted with an initial mixed strategy

that is no longer random but biased against actions comprising highly interfered RBs.

In this work, we put forward an efficient distributed algorithm suitable for narrow band

systems like PMR. However, it is of limited interest for more widespread OFDM-based

networks with larger number of spectral resources. Therefore, we put into light through the

next chapters more adapted game theoretic distributed algorithms with fast convergence

times.

36 3. Replicator Dynamics for Distributed Inter-Cell Interference Coordination

Chapitre 4

for Inter-Cell Interference

Coordination

This chapter reconsiders the problem tackled in Chapter 3 where the ICIC resource

selection process is apprehended as a potential game in the downlink of a cellular OFDMA

system. However, contrary to the work displayed in the previous chapter, there is no res-

triction on the system bandwidth size. We consider here a game theoretic approach where

the ICIC problem is modeled as a congestion game[ARK+ 14a]. We adapt the distributed

algorithm proposed by Berenbrinck et al. [AB12] to attain the PNE of the considered game.

Proof of convergence is obtained in a limited time.

4.1 Introduction

Turning to non-cooperative game theory is quite fit to model the way eNodeBs compete

in a distributed way for scarce and shared resources. We adapt a stochastic version of

the best-response dynamics proposed by Berenbrink et al. [AB12, BFHH12] where players

only have a local view of the system. Hence, all eNodeBs are able to simultaneously change

their strategy in a fully distributed way. The result of the devised coordination process in

each cell will be a pool of RBs that is not too highly interfered and which can then be

used for fast intra-cell scheduling by the eNodeB. More importantly, we provide proof of

convergence to PNE in a bounded time. Unfortunately, proof of convergence is obtained

at the cost of reduced model accuracy : in this chapter we overlooked the impact of UEs

position in a given cell. The rest of the chapter is organized as follows. The network model

and cost characterization are given in Chapter 2. The RB selection scheme is presented

as a non-cooperative congestion game in Section 4.3. Proof of convergence to PNE in a

37

38 4. Distributed Load Balancing Game for Inter-Cell Interference Coordination

bounded time is presented in Section 4.4. Simulation results are portrayed in Section 4.5.

Conclusion is given in Section 4.6.

We use the reference model of Section 2.1 presented in Chapter 2. To model the amount

of interference endured by eNodeBs, we use the model proposed in [EHB08]. A resource

allocation by any eNodeB consists in selecting a subset of RBs to be used in its cell. The

choice of eNodeB i is denoted by the m-dimensional vector Xi whose components xi,k are

defined in 2.1.2.

The different geographic positions of individual UEs are not taken into consideration : the

use of any RB k by eNodeB i is impacted by the concurrent use of that same RB by other

eNodeBs, the goal will be to lower the use of each RB over the adjacent cells, leading to

a load balancing situation. Furthermore, we consider that any eNodeB is interfered only

by its 6 neighboring eNodeBs. The impact of further away cells is neglected. Building on

that, we define Ii,k as the interference endured by eNodeB i on RB k as follows :

X

Ii,k = xj,k (4.2.1)

j∈N,

j6=i

Non-Cooperative game theory models adequately the interactions between selfish eNodeBs

competing for a common pool of RBs. The higher is the number of eNodeBs that select

concurrently the same RB k, the higher is the interference produced on the latter (conges-

tion impact). Hence, the highest will be the cost of eNodeBs that chose that RB k in

particular. Consequently, the ICIC problem at hand is adequately modeled by a conges-

tion game where eNodeBs (the decision makers or players of the game) will try on their

own to avoid RBs that are increasingly solicited by adjacent cells.

We adopt the multi-player game G between the n eNodeBs defined in Section 2.2. We will

define the cost function of a eNodeB i that selected strategy xi as follows :

X

ci (Xi , X−i ) = xi,k · Ii,k

k∈M

X X (4.3.1)

= xi,k · xj,k

k∈M j∈N,

j6=i

Where X−i denotes the vector of strategies played by all other eNodeBs except eNodeB

i. Every eNodeB i needs a certain amount of Ni RBs depending on the number of active

4.4. Distributed learning of PNE 39

Pm

UEs in its cell. The action set Xi is hence cell-specific as only actions with k=1 xi,k = Ni

are allowed in eNodeB i.

A load balancing game is a congestion game where every player chooses only one resource.

In our case, every eNodeB (player) chooses concurrently many RBs (resources). Hence, we

need to tailor our game so that it boils down to a load balancing game as follows : the

system starts with an initial state where Di RBs are randomly alloted in eNodeB i. In

each round, every eNodeB migrates randomly and uniformly the allocation of only one UE

from an occupied RB to an available RB. The goal is to adapt the algorithm proposed by

Berenbrinck et al in [AB12, BFHH12] where authors studied distributed load balancing

in networks with selfish players. They consider n resources and m > n selfish players that

unilaterally decide to move from one resource to another if this improves their experienced

load. Concurrent migration is possible based only on local information which is pretty

suitable for distributed ICIC. The adapted algorithm is sketched hereafter :

1: At each iteration

2: for i = 1 to n do

3: eNodeB i selects a RB l uniformly at random among the b used RBs

4: eNodeB i selects a free RB j uniformly at random in m − b (strategy x0i )

5: if (ci (xi , X−i ) > ci (x0i , X−i )) then

c (x0 ,X )

6: Change strategy (go to RB j) with probability 1 − cii (xii ,X−i

−i )

7: end if

8: end for

In each round in parallel, every eNodeB picks an available RB at random and decides

probabilistically whether or not to migrate a UE to that RB. To compute the cost of the

former strategy (xi ) and actual strategy (x0i ), any eNodeB i makes use of signaling infor-

mation already present in the downlink of an LTE system. In fact, as already explained,

a UE assigned to a specific RB l measures its channel quality based on pilots and sends a

report in the form of CQI to its serving eNodeB. Hence, the eNodeB can easily infer the

cost function based on the CQI sent every TTI by serviced UE to their eNodeBs.

This section is devoted to prove the convergence to PNE of the proposed algorithm in a

bounded time. Recall that n is the number of eNodeBs and m the number of RBs.

40 4. Distributed Load Balancing Game for Inter-Cell Interference Coordination

Theorem 4.1. A! Nash Equilibrium is reached in expected time O(m log(w̄)), with : w̄ =

1 X X

xi,k .

m

i∈N k∈M

We give in what follows a concise proof of the theorem. A detailed one can be found in

the Appendix B.

Démonstration. Let W (t) = (w1 , w2 , . . . , wm ) be the vector containing the number of UEs

P

on each resource RB at time t, with wk = i∈N xi,k corresponding to the number of UEs

on RB k. b is the number of active RB. Each UE to whom RB k is allocated will have a

weight equal to 1.

1 w`

b(m−(b−1)) 1 − wk if ` 6= k, wk > w` + 1

Pk,` = 0 if ` 6= k, wk ≤ w` + 1

(4.4.1)

X

1 − Pk,j if ` = k

j∈M,wk >wj

X

Φ(W ) = (wk − w̄)2

k∈M

m m

1 X X

= (wk − w` )2

m

k=1 `=k+1

1 P P

with w̄ = m i∈N k∈M xi,k the average load.

This potential function is a standard one according to [BFHH12], [Gol04]. We adapt the

analysis of convergence of such a potential function proposed by Berenbrink et al. for

load balancing of unweighted tasks [BFHH12]. The objective is to prove that for any

time t, E[Φ(W (t + 1))] < E[Φ(W (t))]. Indeed, the potential function Φ is the sum of

the difference between the load of each resource and the global mean w̄, all squared. For

a weight of W = {w̄, w̄, . . . , w̄} where all resources have a load that equals the global

mean w̄, the potential function attains its lowest value Φ(W ) = 0. Thus, if we prove that

E[Φ(W (t + 1))] < E[Φ(W (t))], it means that the potential function decreases on average

at each time step and since it is lower bounded by 0, convergence time is finite.

In order to prove this theorem, we first prove that E[Wk (t + 1)|W (t) = wk ] < E[W (t)].

4.4. Distributed learning of PNE 41

Xm

= E[ (Wk (t + 1) − w̄)2 |Wk (t) = wk ]

k=1

m

X (4.4.2)

= (E[(Wk (t + 1))|Wk (t) = wk ] − w̄)2

k=1

Xm

+ V ar[(Wk (t + 1))|Wk (t) = wk ]

k=1

According to this equation bounding the two sums on expectation and variance leads us

to a bound on E[Φ(W (t + 1))|W (t)].

The proof of Lemmas 1 and 2 are built on the model of Berenbrink et al. for load balancing,

details are displayed in the appendix. To overcome general understanding problems, we

define E[Wk,` ] to be the average total load of UEs which move from RB k to RB `. Without

loss of generality, we assume that w1 ≥ w2 ≥ · · · ≥ wm . Further, ∀k ∈ M , let Sk (t) be

the set of UEs using RB k at time t and let u be the index of a given UE, we have what

follows :

X

u wk − w`

E[Wk,` ] = Pk,` =

b(m − b + 1)

u∈Sk

wk − w`

≤

m

Lemma 1.

m

X

(E[(Wk (t + 1))|Wk (t) = wk ] − w̄)2 <

k=1

m X

m (4.4.3)

X

Φ(W ) − 2E[Wk,` ](wk − w` )

k=1 `=k+1

Lemma 2.

m

X

V ar[(Wk (t + 1))|Wk (t) = wk ]

k=1

m X

m (4.4.4)

X

≤2 E[Wk,` ]

k=1 `=k+1

42 4. Distributed Load Balancing Game for Inter-Cell Interference Coordination

Lemma 3.

m X

m

X (4.4.5)

− 2E[Wk,` ](wk − w` )

k=1 `=k+1

Recall that wk > w` + 1, otherwise E[Wk,` ] = 0. Hence, we get wk − w` > 1 and thus

wk − w` > (wk − w` − 1). From Equation (4.4.2) and both previous Lemmas, this lemma

holds.

Lemma 4.

8

E[Φ(W (t + 1))|W (t)] < Φ(W (t))(1 − ) (4.4.6)

(m + 2)

wk −w` 1 Pm Pm

By definition, E[Wk,` ] = b(m−b+1) . Since Φ(W ) = m k=1 `=k+1 (wk − w` )2 , we deduce

that

m X

m

X mΦ(W )

E[Wk,` ] (wk − w` ) =

b(m − b + 1)

k=1 `=k+1

Furthermore, from Lemma 3, we have that E[Φ(W (t + 1))|W (t)] < Φ(W (t)) − 2 mΦ(W (t))

b(m−b+1) .

Given that m ≥ b > 0 for any b since b ∈ M is the number of used RBs, by computation,

m 2

we have that b(m−b+1) ≥ (m+2) and hence this lemma holds.

8 τ

From Lemma 4, we have that E[Φ(W (t + τ ))|W (t)] ≤ Φ(W (t))(1 − (m+2) ) . From the

definition of Φ(W ), we have that 0 ≤ Φ(W ) ≤ w̄2 . Hence, after τ = m log(w̄) steps,

8 m log(w̄)

E[Φ(W (t + τ ))|W (t)] ≤ w̄(1 − (m+2) ) ≤ 2.

By Markov’s inequality, the probability that W (t + τ ) is a Nash equilibrium is greater

than 1/2. Now for each new run of τ steps, the probability to reach a Nash Equilibrium is

at least 1, hence the expected time to reach such a PNE is at most 2τ . Hence the theorem

holds.

eNodeBs with 6 surrounding eNodeBs and suppose that further away eNodeBs do not

interfere with them. UEs are uniformly disposed in every cell. eNodeBs have a frequency

reuse of 1, with PRBs bandwidth of 180 KHz. Simulations were conducted for several

LTE frequency bands : {1.4, 3, 5, 10, 15, 20} MHz. The distance between two neighboring

eNodeBs is 2 Km and transmission power is constant and set to 20 Watt (43 dBm). The

propagation loss factor is set to 3.

4.5. Simulation results 43

4.5.1 Convergence

PNE. We are interested in evaluating in what follows the speed of convergence of our

Distributed Load Balancing scheme (denoted DLB). In Figure 4.1, the mean number of

iterations to reach convergence is drawn as a function of the number of RBs for a constant

load of 50%. 400 simulations were run for every scenario. The results show that convergence

time increases linearly as a function of the number of RBs as proven in the previous section.

Since the number of possible strategies depend on the number of RB, when the number of

RBs becomes high, there are much more strategies for the algorithm to test, which leads

to a longer convergence time. Further, we deduce that our algorithm is much more suited

for relatively narrow band systems with a small amount of RBs.

In Figure 4.2, we evaluate the impact of load on convergence time. Simulations were run

with a frequency band of 5 MHz (25 RBs) and a load varying from 10% to 80%. The

results show that convergence time increases linearly as a function of load which again

verifies the proven bound on convergence time.

44 4. Distributed Load Balancing Game for Inter-Cell Interference Coordination

4.5.2 Performances

a trivial algorithm that selects randomly RBs in any cell once and for all. In Figure 4.3,

the mean cost, which is an image of average interference, is depicted for a bandwidth of 5

MHz and a load spanning from 10% to 80%. For our distributed algorithm, we reported

the mean cost after 100 iterations and after convergence is reached. We can clearly see that

the major part of improvement is achieved during the first 100 iterations in comparison

with the trivial algorithm. If we consider that an iteration is equivalent to a TTI (1ms),

this result is quite attractive since we obtain good results for any eNodeB before its

neighboring eNodeBs send new RNTP messages where certain RBs may be precluded due

to the advertised high power on them.

In the lower sub-figure, the mean relative change of cost is displayed where the rela-

cDLB −cRandom

tive change in eNodeB i is given by i i

cRandom

∗ 100 with cDLB

i and cRandom

i are

i

the cost function according to equation (4.3.1) for our distributed load balancing game

and the random algorithm respectively. We capture results at three different instants :

{100, 200, convergence} iterations. In particular, we notice an improvement greater than

20% for a load smaller than 40% in a time coherent with the RNTP signaling time, the

impact of load at high level is due to the fact that the majority of RBs are already selected,

and there is fewer chance for the algorithm to improve the resource allocation.

4.5. Simulation results 45

Figure 4.3: Random algorithm vs. Distributed Load Balancing algorithm as a function

of Load

Again, we conducted in Figure 4.4 further simulations to show the impact of the total

available number of RBs at a moderate load (50%). We conclude from those results that

the improvement achieved in comparison with the random approach decreases when the

available bandwidth increases. In fact, when the number of UEs is relatively small in

comparison with the available resources, there is limited latitude to achieve improvement.

46 4. Distributed Load Balancing Game for Inter-Cell Interference Coordination

Figure 4.4: Random algorithm vs. Distributed Load Balancing algorithm as a function

of Frequency Band

4.6 Conclusion

In this chapter, RBs are judiciously assigned in LTE or OFDM based networks in the

scope of Inter-Cell Interference coordination process. We proposed a fully distributed load

balancing algorithm to reach the PNE of the portrayed game where eNodeBs only rely

on local information. Proof of convergence in a finite time to PNE is provided and good

performances of the proposed approach are assessed through extensive simulations.

Chapitre 5

Load Balancing Game for ICIC

This chapter reconsiders the problem tackled in Chapter 4 where the ICIC resource

selection process is apprehended as a load balancing game in the downlink of a cellular

OFDMA system. However, contrary to the work displayed in the previous chapter, we

consider here the impact of UEs distance from their serving eNodeB. Unfortunately, the

increased accuracy in the network model is obtained at the cost of increased complexity.

Hence, we only provide proof of convergence of the proposed algorithm to PNE but without

any guarantee of convergence in a bounded time. However, we propose here a greedy algo-

rithm that gives interesting performance results, and reaches stability in a time coherent

with RNTP signaling time[ARK+ 14b]

5.1 Introduction

In this chapter, the model at hand is an exact potential game that is known to have at

least one PNE, since the potential function is strictly decreasing in any sequence of pure

best responses [Ros73]. A stochastic version of the best-response dynamics, where agents

only have a local view of the system, has been investigated by Berenbrink et al. [AB12,

BFHH12]. We adapt the proposed approach where eNodeBs are able to simultaneously

change their strategy in a distributed way with a local view of the system. The result of

the devised coordination process in each cell will be a pool of RBs that is not too highly

interfered and can then be used for fast intra-cell scheduling by the eNodeB. Extensive

simulations were conducted to assess the convergence of the adapted distributed algorithm

to PNEs. The rest of the chapter is organized as follows. The system model and cost

characterization are given in Section 5.2. The RB selection scheme is presented as a non-

cooperative congestion game in Section 5.3. The distributed learning algorithm is presented

47

48 5. A Greedy approach of Distributed Load Balancing Game for ICIC

in Section 5.4. Simulation results are portrayed in Section 5.5. Conclusion is given in

Section 5.6.

We use the reference model of Section 2.1 presented in Chapter 2. However, to model the

amount of interference endured by eNodeBs, we consider the SINR which is a much more

exact performance indicator. In particular, we consider the bit transfer time defined in

(2.1.4) where the position of UEs is accounted for.

Non-Cooperative game theory is a good fit to model the decentralized ICIC where selfish

eNodeBs share common resource blocks in a way to enhance their own local performances.

We adopt the multi-player game G between the n eNodeBs defined in Section 2.2 while

using the same cost function of Chapter 3. Recall that the cost function of eNodeB i that

selected strategy xi is as follows :

XX

ci (xi , X−i ) = xi,k · Ti,k,u

k∈M u∈U

(5.3.1)

P

j∈N, xj,k · Hi,j

j6=i

XX

= xi,k

Hi,u

k∈M u∈U

Any eNodeB will single out the strategy that minimizes the total bit transfer time in its

cell.

The potential game at hand is a congestion game where players may differ from one

another in their intrinsic preferences (the benefit they get from using a specific RB), their

contribution to congestion, or both. In fact, the cost sustained by a given eNodeB on any

selected RB depends indeed upon the congestion impact inflected by other BSs sharing

the same RB. In other words, the cost of each player depends on the selected resources

and the number of players choosing the same resources. In particular, Ti,k,u is increasing

in the number of UEs affected simultaneously to RB k.

with selfish players. The proposed algorithm proceed in a round-based fashion. Players

are initially assigned arbitrarily to the machines. In each round, every player picks a

5.4. Distributed learning of PNE 49

to that machine. Rapid convergence to PNE of such protocols is proved.

In our case, every eNodeB (player) may choose concurrently many RBs (machines). Hence,

we need to adapt our game so that we can implement one of the proposed protocols by

Berenbrinck. Accordingly, we put forward an adapted algorithm in chapter 4 deemed algo-

rithm 1 to reach the PNE : the system starts with an initial state where RBs are randomly

alloted in any eNodeB such as all UEs are served ; then, in each round, every eNodeB mi-

grates randomly and uniformly only one UE from an occupied RB to an available RB. In

that case, our game boils down to a load balancing game according to algorithm 2 sketched

hereafter.

Algorithm 2 Distributed Load Balancing Algorithm

1: At each iteration

2: for i = 1 to n do

3: eNodeB i selects a RB l uniformly at random among the b used RBs

4: eNodeB i selects a free RB j uniformly at random in m − b (strategy x0i )

5: if (ci (xi , X−i ) > ci (x0i , X−i )) then

c (x0 ,X )

6: Change strategy (go to RB j) with probability 1 − cii (xii ,X−i

−i )

7: end if

8: end for

To compute the cost of the former strategy (xi ) and actual strategy (x0i ), any eNodeB i

makes use of signaling information already present in the downlink of an LTE system.

Although eNodeBs update their strategies in parallel, we noticed through extensive si-

mulations portrayed in Section 5.5, that algorithm 2 is slow in convergence for systems

with more than 25 RBs. Therefore, we propose another greedy version deemed algorithm

3 (Greedy algorithm) where,in parallel and for each iteration, every eNodeB re-affects ran-

domly and uniformly all its UEs to the RBs according to the algorithm sketched below

(we denote by ni the number of UEs present in cell of eNodeB i).

1: At each iteration

2: for i = 1 to n do

3: for l= 1 to ni do

4: eNodeB i selects a RB uniformly at random among the m − (l − 1) available

RBs

5: end for

6: if (ci (xi , X−i ) > ci (x0i , X−i )) then

c (x0 ,X−i )

7: Change strategy with probability 1 − cii (xii ,X−i )

8: end if

9: end for

Algorithm 3 is still a load balancing game if we consider that UEs in each cell are the

50 5. A Greedy approach of Distributed Load Balancing Game for ICIC

actual players. In that case, one RB is alloted to a UE (player) at a time. The eNodeB

simply plays the role of a central entity that forbids the selection of the same RB by

different UEs. Algorithm 2 is a regular load balancing game and is deemed to converge to

PNE. However, for algorithm 3, we rely on extensive simulations to assess the convergence

to PNEs.

eNodeBs with 6 surrounding eNodeBs and suppose that further away eNodeBs do not

interfere with them. Two zones are taken into account : zone 1 which stands for cell-center

UEs located at a distance smaller than R0 = 0.5Km and zone 2 stands for cell-edge UEs

located at a distance ranging between R0 = 0.5Km and R1 = Rcell = 1Km. UEs are

uniformly distributed in every cell. We assume that for cell-center UEs 64-QAM modu-

lation is used while for cell-edge UEs 16-QAM modulation is used. The eNodeBs have a

frequency reuse of 1, with PRBs bandwidth of 180 KHz. Simulations were conducted for

several LTE frequency bands : {3, 5, 10, 15, 20} MHz. The distance between two neighbo-

ring eNodeBs is 2 Km and transmission power is constant and set to 20 Watt (43 dBm).

The propagation loss factor is set to 3.

to PNE. We are interested in evaluating in what follows the speed of convergence of both

algorithms. In Figure 5.1, the mean number of iterations to reach convergence is drawn

as a function of the number of RBs for a constant load of 50%. 400 simulations were run

for every scenario. We notice that although algorithm 2 converges faster for 15 RBs, the

greedy algorithm largely outruns it for wider system bandwidths. Changing several RBs at

the same time allows algorithm 3 to test relatively more strategies which leads to shorter

convergence time.

5.5. Simulation results 51

In Figure 5.2, we evaluate the impact of load on convergence time. Simulations were run

with algorithm 3 with a frequency band of 20 MHz (100RB) and a load varying from 10%

to 80%. The results show that convergence time increases linearly as a function of load

which is a desired property.

5.5.2 Performances

approach (deemed trivial algorithm) that selects randomly RBs in any cell once and for all.

In Figure 5.3, the mean cost (in ms) for algorithm 3 and trivial algorithm is depicted as a

function of load for a frequency band of 20 MHz. For algorithm 3, we show results after 100

ms and after convergence is reached. In the lower subfigure, the mean relative change of cost

52 5. A Greedy approach of Distributed Load Balancing Game for ICIC

is displayed where the relative change in eNodeB i is given by i

T

ci riv

i

∗ 100 with cAlgo3

i

and cTi riv are the cost function according to equation (4.3.1) for algorithm 3 and trivial

algorithm respectively. We captured results at three different instants : {100, 200, 1000}

iterations (ms).

We can clearly see that the major part of improvement is achieved during the first 100

iterations in comparison with the trivial algorithm. If we consider that an iteration is

equivalent to a TTI, this result is quite attractive since we obtain very good results before

that neighboring eNodeBs exchange the RNTP (Relative Narrow-band Transmit Power)

message where certain RBs may be precluded due advertised high interference on them.

This behavior represent the strength of our algorithm. The relative change in the lower

subfigure further highlights the good performances of our algorithm.

We conducted further simulations displayed in Figure 5.4 to show the impact of the total

5.5. Simulation results 53

available number of RBs at a load of 50%. We conclude from those results that the im-

provement achieved in comparison with the trivial approach decreases when the available

bandwidth increases, in fact when the number of available RBs in the cell grows, the num-

ber of possibility become so high that finding a strategy that satisfies every one is difficult,

hence, this leads to a lower improvement when the system bandwith grows.

Figure 5.4: Trivial algorithm vs. Greedy algorithm as a function of frequency band

We were interested in the behavior of the algorithm in a realistic non homogeneous UEs

distribution in any cell. Two cases were handled :

– In the first scenario, the edge zone of the eNodeB is crowded (the majority of UEs are

situated on the cell periphery),

– While in the second scenario, the center zone is crowded (the majority of UEs are close

54 5. A Greedy approach of Distributed Load Balancing Game for ICIC

to the eNodeB).

We show in this simulation the relative change between a random algorithm that selects

randomly RBs in any cell once and for all, and our greedy algorithm.

Figure 5.5 shows the first scenario where 75% of the UE are situated in the cell-edge.

In this case, the performance of the greedy algorithm decreased in comparison with the

previous results (Chapter 5.5.2), this can be explained as most of the UEs are interfering

with each other, which impacts negatively the channel quality of all the UEs. This case

corresponds to a loaded scenario, except there are more free RBs, and hence, the algorithm

manages to minimize the interference better than a fully loaded system.

Figure 5.6 shows the case where 75% of the UEs are in the cell-center. Since UEs are

far from neighboring cells, their channel quality is good, and hence, the greedy algorithm

manages to allocate the RBs in a way that strongly boosts the performance of the system

by decreasing the interference.

5.6. Conclusion 55

We can say that our algorithm is robust against any UE distribution with the advantage

of using frequency reuse 1 contrary to the SFR scheme that is static and performs badly

with non-uniform scenarios since there is no way to change the resource allocation along

with the users distribution [YDH+ 10].

5.6 Conclusion

In this chapter, we smartly allocated the resources in the system to avoid the inter-cell

interference. Using a load balancing game theoretic approach that is fully distributed shows

good improvement over basic allocation. Hence, the enhanced greedy algorithm allows to

obtain results in a time consistent with the signaling message latency, especially the RNTP

message. This can be used to obtain better performance in a semi-distributed system that

takes advantage from the RNTP message. Furthermore, using the RNTP message as an

initial state will be addressed in future work.

56 5. A Greedy approach of Distributed Load Balancing Game for ICIC

Chapitre 6

Inter-Cell Interference

Coordination

This chapter addresses the problem of power control for ICIC in the downlink of cel-

lular OFDMA systems. The power level selection process of resource blocks is apprehended

as a sub-modular game. Here again, the existence of Nash equilibriums for that type of

games shows that stable power allocations can be reached by selfish eNodeBs. We put for-

ward a semi distributed algorithm based on best response dynamics to attain the NEs of

the modeled game. Based on local knowledge conveyed by the X2 interface in LTE net-

works, each eNodeB will first select a pool of favorable RBs with low interference. Second,

each eNodeB will strive to fix the power level adequately on those selected RBs realizing

performances comparable with the Max Power policy that uses full power on selected RBs

while achieving substantial power economy. Finally, we compare the obtained results to an

optimal global solution, which is derived by solving an integer linear program. Hence, we

are able to quantify the efficiency loss of the distributed game approach. It turns out that

even though the distributed game results are sub-optimal, the low degree of system com-

plexity and the inherent adaptability make the decentralized approach promising especially

for dynamic scenarios[KAL+ 14].

6.1 Introduction

Contrary to previous chapters that stressed on RB selection for ICIC, we propose in this

chapter to reduce ICI through efficient Power Control. Power control does not only reduce

the impact of interfering signals by lowering their power level (signals usually belonging to

cell-center UEs), but it can increase the power level on resource blocks that suffer of bad

57

58 6. Power Control for Distributed Inter-Cell Interference Coordination

method for ICIC. Here again, we have recourse to non-cooperative game theory to model

the way selfish eNodeBs compete in a distributed manner for limited resources. Devising

an optimal power level selection scheme depends on the existence of Nash equilibrium

for the present game. In this chapter, we prove that the model at hand is a sub-modular

game (see [Top79, Yao95]). Such games have always a Nash Equilibrium and it can be

attained using a best response type algorithm (called algorithm I in both references). The

result of the devised coordination process in each cell will be the power tuning on the least

interfered RBs.

The rest of the chapter is organized as follows. The system model is given in Section 6.2.

In Section 6.3, the framework of the RB selection and power control schemes is described.

The power level selection scheme is presented as a non-cooperative sub-modular game in

Section 6.4 where a semi distributed learning algorithm based on best response dynamics

is presented in Subsection 6.4.2. Simulations results are portrayed in Section 6.5. The

optimal centralized approach is given in Section 6.6 as a benchmark to evaluate the price

of anarchy resulting from a decentralized approach. Conclusion is given in Section 6.7.

We consider the same system model described in Chapter 2, the UE distance is otherwise

considered with a dynamic power control achieved with the proposed algorithm.

P0 · Gi · pi,k · ( di1 )β

u,i

SIN Ri,k,u = P 1 β (6.2.1)

P0 · j∈M, Gj · pj,k · ( di ) + PN

j6=i u,j

where P0 represents the maximal transmitted power per RB, PN represents the thermal

noise power per RB, Gi the antenna gain of eNodeB i, diu,j is the distance between eNodeB

j and UE u served by eNodeB i and β is the path-loss factor varying between 2 and 6.

Finally, pi,k is the discrete variable that represents a fraction of the maximal transmitted

power P0 . Hence, pli,k = P0 · pi,k is one of the possible Nl power levels in eNodeB i on RB k

allotted to UE u. It varies between pli,k = 0.0 where RB k is not selected by eNodeB i and

pli,k = P0 where full power is transmitted on that RB. However, our power control starts

after the selection of a pool of RBs in any eNodeB i, denoted by Ni . Hence, pli,k 6= 0.0.

We denote by Di,k,u the data rate achieved by UE u on RB k in eNodeB i given by what

6.2. The system model 59

follows :

W

Di,k,u = · SIN Ri,k,u ,

Eb /N0

where W is the bandwidth per RB. Given a target error probability, it is necessary that

Eb /N0 ≥ γ, for some threshold γ which is UE specific.

Each cell will be logically divided into NZ concentric discs of radii Rz , z = 1, ..., NZ , and

the area between two adjacent circles of radii Rz−1 and Rz is called zone z, z = 1, ..., Nz .

UEs belonging to the same zone z have the same radio conditions leading to the same γz

and the same mean rate per zone Di,k,z according to what follows :

W

R Rz

γz Rz−1 ρzmobile 2πrdr

rβ

· Gi · P0 · pi,k

Di,k,z = P 1

P0 · j∈M, Gj · Pj,k · (di )β + PN

j6=i i,j

W 2−β 2−β

(6.2.2)

γz (Rz − Rz−1 ) · ρzmobile · Gi · P0 · Pi,k

= G

P0 · j∈M, Pj,k · (δz ·Rj )β + PN

P

i,j cell

j6=i

where Rcell is the cell radius. As for interference, we consider mainly for simplification

the impact of eNodeB j on eNodeB i by replacing diu,j by diz,j = δi,j

z ·R

cell the distance

z depends on how far is eNodeB j from

between eNodeB i and eNodeB j (the value of δi,j

zone z of eNodeB i).

We denote by Ti,k,z the amount of time necessary to send a data unit through RB k in

eNodeB i for UEs in zone z. In fact, the delay needed to transmit a bit for a given UE is

the inverse of the data rate perceived by this UE :

Ii,k,z

Ti,k,z = (6.2.3)

xi,k

z +P

P

j∈M, Pj,k · Hi,j N

j6=i

Ii,k,z = (6.2.4)

Hi,z

z = P0 G j

where Hi,j z ·R

(δi,j β captures distance-dependent attenuation of power between eN-

cell )

W 2−β 2−β

odeB j and eNodeB i and Hi,z = γz (Rz −Rz−1 )·ρzmobile ·Gi captures distance-dependent

attenuation of power inside zone z of eNodeB i.

We denote by Nzi the pool of RBs used by UEs in zone z. eNodeB i will pay an amount

αz per power unit for the use of a given RB k ∈ Nzi . This power unitary cost can decrease

with the zone index to further protect UEs that are far away from the antenna ; or it can

increase to favor cell-center UEs in order to enhance overall performances.

60 6. Power Control for Distributed Inter-Cell Interference Coordination

Accordingly, the goal of the power control scheme proposed in this chapter is to minimize

the following cost function in eNodeB i for RB k allotted to a UE in zone z :

κ · Ti,k,z + αz · P0 · yi,k , if RB k is used by eNodeB j,

ci,k,z = (6.2.5)

0 If RB k is not used in zone z.

In this chapter, we propose that each eNodeB aggregates information about transmit power

levels in adjacent cells and decides accordingly to choose the pool of RBs per zone. Recall

that the RNTP indicator, received from neighboring eNodeBs every 200 TTI through the

X2 interface, advertises on which RBs a neighboring eNodeB will use full power. In any

eNodeB i, we associate with every RB k a variable ai,k . This variable indicates the number

of neighboring eNodeBs (thus 0 ≤ ai,k ≤ 6) that advertised the use of that same RB k

with full power via the RNTP indicators. Thus, every eNodeB i updates its variables

ai,k approximately every 200 TTI and makes use of those variables to update the pool of

selected RBs per zone as described in algorithm 4 where niz is the number of UEs in zone

z of eNodeB i. Recall that Nzi is the pool of RBs reserved to UEs in zone z.

1: Every 200 TTI

2: for i=1 TO n do

3: for k=0 TO m do

4: eNodeB i updates the ai,k variables according to the RNTP indicators.

5: The updated ai,k variables are sorted in ascending order list Li .

6: for z=0 TO NZ − 1 do

7: The niz top values of Li are a reserved pool RBs for UEs in zone NZ − z

(denoted by Nzi ) and removed from the sorted list.

8: end for

9: end for

10: end for

The idea behind the algorithm is to reserve the pool of least interfered RBs to UEs who

are the furthest away from the eNodeB (the zones with the highest index). After selecting

Nzi , every eNodeB i proceeds to implementing the power control distributed algorithm

described in the following section and which is the focal point of the chapter. We assume

that the scheduler has already assigned UEs in a given zone z to RBs in Nzi . Afterwards,

our proposed power control scheme endeavors to find the optimal power levels on those

allocated RBs. After convergence of the latter, each eNodeB i will obtain the optimal

6.4. Non-cooperative game for power control 61

∗

power level Pi,k,z to be assigned to RB k in zone z.

We define a multi-player game G between the m eNodeBs that make their decisions without

knowing the decisions of each other.

The formulation of this non-cooperative game G = hN, S, Ci can be described as follows :

– A finite set of players M = (1, ..., m) and a finite set of RBs N = (1, ..., n).

– For each eNodeB i, the space of strategies Si is formed by the Cartesian product of

each set of strategies Si = Si,1 × ... × Si,n , where n is the total number of RBs. An

action of a eNodeB i is the amount of power xi,k sent on RB k. If RB k ∈ Nzi then

l −1 j

Si,k = {x0i,k , ..., xN

i,k } where xi,k is a fraction of P0 , else Si,k = ∅. The strategy chosen

by eNodeB i is then Xi = (xi,1 , ..., xi,n ). A strategy profile X = (X1 , ..., Xm ) specifies

the strategies of all players and S = S1 × ... × Sm is the set of all strategies.

– A set of cost functions C = (C1 (X), C2 (X), ..., Cm (X)) that quantify players costs for a

given strategy profile X where Ci = (ci,1,z , ci,2,z , ..., ci,n,z ) is the cost of eNodeB i where

the cost ci,k,z of using RB k in zone z is given in Equation (6.2.5).

As the frequencies allotted to different RBs are orthogonal, minimization of cost ci,k,z

on RB k is done independently of other RBs. Hence, we denote by x−i,k the strategies

played by all eNodeBs on the RB k except eNodeB i.

Sub-modularity was introduced into the game theory literature by Topkis [Top79] in 1979.

Sub-modular games are of particular interest since they have Nash equilibriums, and

there exists an upper and a lower bound on Nash strategies of each UE [OR94]. Fur-

thermore, these equilibriums can be attained by using a best response type algorithm

([Top79, Yao95]).

Definition 6.1. Consider a game G = hM, S, Ci with strategy spaces Si ⊂ Rm for all

i ∈ M and for all z ∈ NZ , k ∈ Nzi . G is sub-modular if for each i and k, Si,k is a sublattice 1

of Rm , and ci,k,z (xi,k , x−i,k ) is sub-modular in xi,k .

Definition 6.2. The cost function ci,k,z is sub-modular iff for all x, y ∈ Si,k ,

ci,k,z (min(x, y)) + ci,k,z (max(x, y)) ≤ ci,k,z (x) + ci,k,z (y) (6.4.1)

62 6. Power Control for Distributed Inter-Cell Interference Coordination

Proposition 2. The cost function ci,k,z is sub-modular for every eNodeB i and every selected

RB k.

Démonstration. We begin by defining the set A1 such that A1 = {xj,k , yj,k ∈ Sj,k |xj,k <

yj,k , j 6= i} and A2 such that A2 = {xj,k , yj,k ∈ Sj,k |xj,k > yj,k , j 6= i}. Accordingly, the

inequality in (6.4.1) gives the following for x = (xj,k , j = 1, ..., m) and y = (yj,k , j =

1, ..., m) where xj,k , yj,k ∈ Sj,k :

z +P

X min(xj,k , yj,k )Hi,j N

κ· + αz P0 min(xi,k , yi,k )

Hi,z min(xi,k , yi,k )

j6=i

z +P

X max(xj,k , yj,k )Hi,j N

+κ· + αz P0 max(xi,k , yi,k )

Hi,z max(xi,k , yi,k ) (6.4.2)

j6=i

X z

xj,k Hi,j + PN X z

yj,k Hi,j+ PN

≤κ· +κ·

Hi,z xi,k Hi,z yi,k

j6=i j6=i

+ P0 αz (xi,k + yi,k )

We notice in (6.4.2) that P0 αz (min(xi,k , yi,k ) + max(xi,k , yi,k )) = P0 αz (xi,k + yi,k ) and

1 1 1 1

PN · (m − 1)( min(xi,k ,yi,k ) + max(xi,k ,yi,k ) ) = PN · (m − 1)( xi,k + yi,k ) where m is the total

number of eNodeBs. Thus, inequality (6.4.2) simplifies to :

z

X min(xj,k , yj,k )Hi,j z

X max(xj,k , yj,k )Hi,j

+

min(xi,k , yi,k ) max(xi,k , yi,k )

j6=i j6=i

z z (6.4.3)

X xj,k Hi,j X yj,k Hi,j

≤ +

xi,k yi,k

j6=i j6=i

X z

xj,k Hi,j X z

yj,k Hi,j

+

min(xi,k , yi,k ) min(xi,k , yi,k )

j∈A1 j∈A2

X z

yj,k Hi,j X z

xj,k Hi,j

+

max(xi,k , yi,k ) max(xi,k , yi,k )

j∈A1 j∈A2

z

X xj,k Hi,j z

X xj,k Hi,j (6.4.4)

≤ +

xi,k xi,k

j∈A1 j∈A2

z

X yj,k Hi,j z

X yj,k Hi,j

+ +

yi,k yi,k

j∈A1 j∈A2

6.4. Non-cooperative game for power control 63

– Case 1 : xi,k < yi,k . In this case, inequality (6.4.4) gives the following :

z

X xj,k Hi,j z

X yj,k Hi,j

+

xi,k xi,k

j∈A1 j∈A2

z

X yj,k Hi,j z

X xj,k Hi,j

+

yi,k yi,k

j∈A1 j∈A2

z

X xj,k Hi,j z

X xj,k Hi,j (6.4.5)

≤ +

xi,k xi,k

j∈A1 j∈A2

z

X yj,k Hi,j z

X yj,k Hi,j

+ +

yi,k yi,k

j∈A1 j∈A2

X

z 1 1

(xj,k − yj,k )Hi,j ( − )<0 (6.4.6)

yi,k xi,k

j∈A2

The latter inequality is obviously true as xj,k > yj,k , ∀j 6= i and xi,k < yi,k .

– Case 2 : xi,k > yi,k . Similarly to case 1, inequality (6.4.4) simplifies to :

X

z 1 1

(yj,k − xj,k )Hi,j ( − )>0 (6.4.7)

yi,k xi,k

j∈A1

The latter inequality is obviously true as yj,k > xj,k , ∀j 6= i and xi,k > yi,k .

Since Si,k is a single dimensional finite set, Si,k is a compact sublattice of R. As we proved

that the cost function ci,k is sub-modular for every eNodeB i on every selected RB k, our

game is indeed sub-modular.

The best response strategy of player i is the one that minimizes its cost given other players’

strategies. A best response dynamics scheme consists of a sequence of rounds, each player i

chooses the best response to the other players’ strategies in the previous round. In the first

round, the choice of each player is the best response based on its arbitrary belief about

what the other players will choose. In some games, the sequence of strategies generated

by best response dynamics converges to a NE, regardless of the players’ initial strategies.

S-modular games are part of those games.

To reach the NE, [AA03] proposes the following best response algorithm built on an

64 6. Power Control for Distributed Inter-Cell Interference Coordination

algorithm called algorithm I in [Top79, Yao95] : there are T infinite increasing sequences

Tti for t ∈ T and i = 1, ..., m. Player i uses at time Tki the best response policy (a feasible

one) to the policies used by all other players just before Tki . This scheme includes in

particular parallel updates (when Tti does not depend on t). Once this UE updates its

strategy, the strategies of one or more other UEs need not be feasible anymore. In [AA03],

proof is given for the following two results :

– If each player i either initially uses its lowest or largest policy in Si , then the iterative

algorithm converges monotonically to an equilibrium (that may depend on the initial

state).

– If we start with a feasible policy, then the sequence of best responses monotonically

converges to an equilibrium : it monotonically decreases in all components in the case

of minimizing in a sub-modular game.

eNodeB i strives to find, for the pool of selected RBs in any zone z, the following optimal

power level :

∗

Pi,k,z = P0 · arg minxi,k ci,k,z (xi,k , x−i,k ),

l −1

for P0 · xi,k ∈ {x0i,k , ..., xN

i,k }.

∗

By definition, Pi,k,z is a best response of eNodeB i to the other eNodeBs strategies on RB

k in zone z.

In a real environment, a best response type algorithm as the one proposed in [Top79,

Yao95] cannot be practically applied as every eNodeB i needs to know the policy of

all other eNodeBs x−i,k on every used RB k which necessitates expensive signaling and

hinders the benefits of an efficient power control scheme. Fortunately, we can easily render

our algorithm distributed by making use of signaling information already present in the

downlink of an LTE system. In fact, x−i,k (or equivalently xj,k ∀j 6= i) only intervene in

the total interference Ii,k,z endured on RB k in zone z of eNodeB i in equation (6.2.4).

Interference can be easily inferred through the CQI sent every TTI by the UEs to which

RB k is attributed. However, the eNodeBs should update their transmission powers on

selected RBs sequentially in a predefined round robin fashion that need to be set once and

for all.

which is a power control scheme under best response dynamics in distributed mode for

our game.

6.5. Simulation Results 65

1: t = 0, conv = 0, stop = 0, N E = 0

2: while N E = 0 do

3: for i = 1 TO n do

4: while stopi = 0 do

5: for z = 0 TO NZ − 1 do

6: while convi,k = 0 do

7: for l = 0 TO Nl − 1 do

8: eNodeB i allocates in parallel all selected RB k ∈ Nzi with power

l

xi,k to a given UE,

9: Serviced UEs send back CQI related to their assigned RB k,

10: eNodeB i infers the corresponding value of Ii,k,z (t) and then com-

putes ci,k,z for xli,k at t,

11: eNodeB i computes Pi,k,z ∗ (t) and attributes it to RB k in zone z.

12: ∗ ∗

if |Pi,k,z (t + 1) − Pi,k,z (t)| < . where is a very small positive

quantity then

13: convi,k = 1

14: else

15: convi,k = 0

16: end if

17: end for

18: endPwhile

m

19: if k=1 convi,k = b . where b is the number of active RB then

20: stopi = 1

21: else

22: stopi = 0

23: end if

24: end for

25: endPwhile

26: if ni=1 stopi = n then

27: NE = 1

28: else

29: NE = 0

30: end if

31: end for

32: t=t+1

33: end while

34: Output : Each eNodeB i will obtain the optimal power level Pi,k,z ∗ to be assigned to

RB k ∈ Nz . i

We consider a bandwidth of 10 MHz with 50 RBs along with the following parameters

listed in the 3GPP technical specifications TS 36.942 : the mean antenna gain in urban

zones is 12 dBi (900 MHz). As for noise, we consider the following : UE noise figure 7.0

66 6. Power Control for Distributed Inter-Cell Interference Coordination

dB, thermal noise -104.5 dBm which gives a receiver noise floor of PN = −97.5 dBm. We

consider 10 hexagonal cells where each cell is surrounded with 6 other cells. The distance

between two neighboring eNodeB is 2 Km. Transmission power is 43 dBm (according to TS

36.814) which corresponds to 20 Watts (On the downlink). We set P0 = 10 Watts and xi,k

for any eNodeB i on RB k belongs to {0.1, 0.2, 0.35, 0.5, 0.6, 0.7, 0.85, 1.0}. Various power

unitary costs (α1 ;α2 ) were tested and for each scenario 400 simulations were run where in

each cell a random number of UEs is chosen in every zone corresponding to a snapshot of

the network state. Performances are compared against Max Power policy where full power

P0 is used on all RBs and against Trivial policy where power levels are set at random.

For every simulation, 100 runs of Trivial policy were made. Further, for each simulation

instance, the same pool of RBs per zone is given for the three policies : DBR, Max Power

and Random policy. Hence, results only assess the impact of power control. We consider

two zones : zone 1 which stands for cell-center UEs located at a distance smaller than

R0 = 0.5Km and zone 2 stands for cell-edge UEs located at a distance ranging between

R0 = 0.5Km and R1 = Rcell = 1Km. We assume that for cell-center UEs we use 64-QAM

modulation while for cell-edge UEs we use 16-QAM modulation.

6.5. Simulation Results 67

Figure 6.1: Transfer time per zone as a function of power unitary cost for DBR vs. Max

Power Policy

68 6. Power Control for Distributed Inter-Cell Interference Coordination

m n

In Figure 6.1, we depict the total bit transfer time per zone Tz = i=1 k=1 Ti,k,z for

cell-center and cell-edge UEs as a function of various power unitary costs (α1 ;α2 ) for DBR

and Max Power Policy. In most scenarios, we aimed at favoring cell-edge UEs by lowering

the power unitary cost in comparison to that of cell-center UEs. We notice as expected

that the improvement in one zone as compared to the Max Power policy is obtained at the

expense of degradation of the other zone. This fact is highlighted in the lowest sub-figure

TzDBR −TzM axP ower

where the relative deviation TzDBR

∗ 100 is displayed. Further, we see that the

improvement in one zone does not strictly depend on how low its power unitary cost is

but how low it is relatively to the other zone : despite the fact that no power unitary cost

is inflected on cell-edge UEs in scenario (1,0), the total transfer time is greater than that

for scenarios (2 ;0.2) or (4 ;0.2).

Figure 6.2: Total Transfer Time as a function of power unitary cost for DBR vs. Max

Power Policy and Trivial Policy

2 n m

In Figure 6.2, we depict the system bit transfer time T = z=1 i=1 k=1 Ti,k,z as a

function of power unitary cost for DBR, Max Power policy and Trivial policy. Except for

(0.2 ;3) and (4 ;0.2) where there is a large discrepancy between the power unitary cost of

one zone in comparison with the other, for all other scenarios performances of DBR and

Max Power policy are equivalent. However, DBR permits a considerable power economy in

comparison with Max Power policy as we can see in figure 6.3 where the relative deviation

between the total power in DBR and the Max Power policy is displayed as a function of

power unitary cost. We can clearly see that the highest the power unitary cost in a zone,

the highest the power economy and vice versa. The best performances are reached when

the same (high) power unitary cost is assigned for both zones in scenarios (2 ;2) and (3 ;3)

where power economy vary from 70% till 80% while the total transfer time is slightly lower

than that in the Max Power policy. As the Trivial policy, we can see that performances

are mediocre.

6.6. The Price of Anarchy 69

In figure 6.4, we report the mean convergence time as a function of power unitary cost.

We note that DBR attains NE faster than 110 TTI and hence before the exchange of new

RNTP messages (sent every 200 TTI).

In this section, we quantify the loss in efficiency suffered when a distributed scheme is

adopted rather than a centralized optimization.

Unlike the distributed approach where precedence is given to the interests of each indi-

vidual eNodeB, power control may be performed in a way that favors the overall system

70 6. Power Control for Distributed Inter-Cell Interference Coordination

assigns the power levels of each eNodeB in order to minimize the total network cost. To do

that, we simply relax the integrality constraints on xi,k , i.e., we assume that 0 < xi,k ≤ 1

for all i ∈ M and k ∈ N ). After relaxing the integrality constraints, the obtained optimi-

zation problem is a non-linear convex problem :

X Ii,k,z

Minimize ( + αz .P0 .xi,k )

xi,k

i,z (6.6.1)

Subject to : 0 ≤ xi,k ≤ 1

In Figure 6.5, we illustrate the mean time necessary to send a data unit for all UEs as a

function of the system load for the optimal policy, our algorithm based on Best Response

dynamics and the Max Power policy. We see that the performances of DBR and the

Optimal policy are equivalent while we notice an expected improvement in comparison

with the Max Power approach that systematically resorts to full power, especially for high

load.

Figure 6.5: Global Transfer Time as a function of power unitary cost for DBR, Max

Power and Optimal policies

However, the power economy made in the optimal approach as compared to DBR tempers

its benefits as we can see from figure 6.6, where the relative deviation between the total

power in DBR (respectively in the Optimal policy) and the Max Power policy is displayed

as a function of power unitary cost. It is obvious that the optimal policy saves up much

more power than the decentralized approach even in high load whereas the power economy

in DBR withers slowly as load increases. Nevertheless, the slight discrepancy between the

global transfer time in DBR and the Optimal policy which the primary goal sought for

6.7. Conclusion 71

and the low degree of system complexity of the decentralized approach makes it still an

attractive solution.

Figure 6.6: Power Economy as a function of power unitary cost for DBR and Optimal

policies

6.7 Conclusion

In this chapter, the power levels are astutely set as part of the LTE Inter Cell Interference

coordination process. We proposed a game based semi distributed algorithm based on best

response dynamics to reach NEs in a time coherent with the RNTP signaling time. Nume-

rical simulations assessed the good performances of the proposed approach in comparison

with a policy that services active UEs with full power. More importantly, considerable

power economy can be realized.

72 6. Power Control for Distributed Inter-Cell Interference Coordination

Chapitre 7

General Conclusion

This chapter presents the general conclusion of the different studies described in this

manuscript. We summarize the main contributions of our work and give the future research

directions that stem from it.

The exponential growth in the number of communications devices over the past decade

has set out new ambitious targets to meet the ever-increasing demand for UE capacity in

emerging wireless systems. However, the inherent impairments of communication channels

in cellular systems pose constant challenges to meet the envisioned targets. In order to

deal with the high cost and scarcity of suitable wireless spectrum, higher spectral reuse

efficiency is required across the cells, inevitably leading to higher levels of interference. To

combat interference, the ICIC concept is explored throughout this thesis in the downlink of

cellular OFDMA systems. ICIC allows coordinated radio resources management between

multiple cells. The eNodeBs can share resource usage information and interference levels

over the X2 interface through LTE-normalized messages. Non-cooperative game theory is

largely applied in this area were eNodeBs selfishly selects resource blocks (RBs) in order

to minimize interference. Hence, we have recourse to this mathematical tool to model the

interaction of eNodeBs over limited shared resources.

We focused our effort in this thesis on ICIC for the downlink of a cellular OFDMA system

in the context of the SOAPS (Spectrum Opportunistic Access in Public Safety) project.

This project introduces the LTE technologies for Broadband Services provision by PMR

systems, it addresses low layers protocols issues with a particular focus on the improve-

ment of frequency resource scheduling. Hence, our first work addresses the problem of

downlink ICIC in the context of the SOAPS network, where the resource selection process

is apprehended as a potential game for which we propose a fully decentralized algorithm

73

74 7. General Conclusion

based on replicator dynamics to attain the pure Nash equilibriums of the game. Exten-

sive simulations assessed the good performances of the algorithm for low to medium load,

especially for systems with a limited number of RBs, where our algorithm produces effi-

cient performances. These results comfort the adequacy of the proposed solution with the

project needs and with the system latency constraints.

For the rest of our contributions, we were interested in a more general solution to overcome

the ICIC problem for a full frequency band systems support (without a limited number of

RBs). The downlink ICIC has been seen as a load balancing game, where an adaptation

of a stochastic version of a best response dynamic algorithm is used to converge to the

PNE of the game. Each eNodeB strives to select a pool of favorable resources with low

interference based on local knowledge only. Proof of convergence is provided, and the

efficiency of the tailored algorithm is proven through extensive simulations. However, in

this first adaptation, the UE position in the cell is not taken into account for ease of

computation. Simulation results show better performances than the replicator dynamics

algorithm for systems using larger number of RBs.

The flaws of the previous proposed algorithm are treated in a second version using a more

realistic scenario that taken into account the UEs positions, where a greedy algorithm is

used to achieve faster convergence times. Simulation results of this new algorithm show a

significant improvement for both time convergence and system performance even for larger

system bandwidth.

As a final work, the ICIC issue was addressed through adequate power allocation on selec-

ted RBs. Finding a suitable power allocation allows to reduce both interference and power

consumption. The power level selection process of RBs is apprehended as a sub-modular

game and a semi distributed algorithm based on best response dynamics is proposed to

attain the PNE of the modeled game. Using the RNTP message exchanged over the X2

link, each eNodeB will first select a pool of favorable RBs with low interference. Then,

each eNodeB will strive to fix the power level adequately on those selected RBs realizing

performances comparable with the Max Power policy that uses full power on selected RBs

while achieving substantial power economy.

For our future work, we plan to enhance our RB selection algorithm by including position

of the UE, in such a solution, each eNodeB will divide the cell in zones according to the

adjacent cells, it will then compute the position of the UEs and locate them in one of

the previously determined cells zone, then the RBs will be allotted to the UEs according

to their positions. Each UE will get a RB in a way it minimizes the interference of that

RB between the serving eNodeBs and the adjacent eNodeBs close to the UE. This po-

7.2. Future Directions 75

sition aware algorithm allows to use even highly interfered RB, this is done by finding

the interferer eNodeB and avoiding using the RB in the shared region of both eNodeBs,

while locating a non-interferer eNodeB to allocate that same RB to the UE located in the

shared area. Another improvement of our game based algorithm is to determine an initial

strategy for the considered ICIC game, based on the received RNTP message. It will build

an initial strategy, taking into account the interference level of the adjacent eNodeB, this

helps run the algorithm from a favourable state, and such an approach will induce better

performances and quicker convergence time.

Another solution is to include a learning entity to our system, this algorithm will study

the global behaviour of the chosen strategies after each convergence of our load balancing

algorithm, it will then define some strategies as initial state or input for our algorithms in

order to get faster convergence and better performances.

Some networks can propose priority between the UEs, the SOAPS project due to its

critical aspect needs such prioritized UEs. hence, we are going to add this constraint to

our algorithm. In this approach, each UE will get an ID in order to create UEs groups,

and then the algorithm along with the position aware mechanism will take into account

this priority in the RBs allocation, UEs with high priority will get best RBs according to

their position to increase their throughput.

With the increasing demand of bandwidth, networks operator are using layered architec-

ture with the macrocell being surrounded with many pico and femto cells. Managing the

attribution of each UE to the cells become relevant, some macro cells could be loaded

where there is some free of charge pico cells. This problem is a load balancing problem

that we propose to tackle using a game theoretic approach. Using signaling between the

cells to locate the loaded ones, the algorithm will perform the UEs reallocation in a semi

distributed manner.

For the upcoming 5G networks, the main proposed solution is the heterogeneous wireless

network, where each UE can access different RAN ( Radio Access Network) technologies.

To optimize network performance while enhancing user experience (by providing high rates

with adequate QoS), efficient common radio resource management mechanisms need to be

defined. Typically, when a new or a handover session arrives, a decision must be made as

to what technology it should be associated with. This is known as the Radio Access Tech-

nology (RAT) selection. We propose a semi distributed approach where different networks

exchange signaling information in order to find the suitable allocation (CoMP).

We plan to extend this type of methods to a SON environment, taking into account

practical parameters expected for 5G and in a heterogeneous network with non-stationary

situations due to channel variations and change in the traffic pattern. More importantly,

resource allocation schemes with cooperation among eNodeBs will be investigated to avoid

inefficient equilibrium.

76 7. General Conclusion

Annexe A

the expected cost of eNodeB i with respect to a mixed profile Q.

According to Sastry and al. [SPT94], the Linear Reward-Inaction algorithm converges

weakly towards a replication dynamic :

dqi,x

(Q) = qi,x (E[ ci |Q ] − E[ ci |qi,x = 1, Q−i ]) (A.0.1)

dt

its limit points related to Nash equilibriums (through the so-called Folk’s theorem of

evolutionary game theory [HS03]). More precisely, we have the following theorem :

Theorem A.1. The following are true for the solutions of Equation (A.0.1) : (i) All

Nash equilibriums are stationary points. (ii) All strict Nash equilibriums are asymptotically

stable. (iii) All stable stationary points are Nash equilibriums.

From [BC13], the limit for b → 0 of the dynamics of stochastic algorithms is some Ordinary

Differential Equations (ODE) whose stable limit points, when t → ∞ (if they exist), can

only be Nash equilibriums. Hence, if there is convergence for the ordinary differential

equation, then one expects the replicator dynamic algorithm to reach an equilibrium.

Moreover, in [CTG09], Coucheney et al. proved that such Nash equilibriums are pure.

Let us see if the continuous dynamic converges with stability arguments. Given a pure

profile X = (x1 , x2 , . . . , xn ), we denote by π(X|Q) the probability that each eNodeB i

chooses the pure strategy xi according to the mixed profile Q :

n

Y

π(X|Q) = qi,xi . (A.0.2)

i=1

77

78 A. Proof for Theorem of Chapter 3

To prove the convergence of the learning scheme Eq. (3.4.1), we define the following func-

tion F : K → R :

X

F (Q) = π(X|Q)φ(X) (A.0.3)

X∈S

dF

Let us study the evolution of function F (Q) over time. We focus on dt (Q). By definition,

we have

n

X X ∂F dqi,s

dF

(Q) = (Q)

dt ∂qi,s dt

i=1 s∈Si

∂π(X|Q) ∂F

We will compute first ∂qj,x0 and then ∂qj,x0 (Q).

j j

n

Y

qi,xi if xj = x0j ,

∂π(X|Q)

= i=1,i6=j (A.0.4)

∂qj,x0j

0 otherwise.

∂π(X|Q)

= π(X|qj,x0j = 1, Q−j ).

∂qj,x0j

X ∂π(X|Q)

∂F

∂qj,x0 (Q) = φ(X)

j ∂qj,x0j

X∈S

X

= π(X|qj,x0j = 1, Q−j )φ(X)

X∈S

∂F

− ∂q∂F

∂qj,x0 j,xj

Xj

X∈S

X

= π(X|qj,x0j = 1, Q−j )(cj (x0j , X−j ) − cj (xj , X−j ))

X∈S

= E[ cj |qj,x0j = 1, Q−j ] − E[ cj |qj,xj = 1, Q−j ]

So,

∂F ∂F

− = E[ cj |qj,x0j = 1, Q−j ] − E[ cj |qj,xj = 1, Q−j ] (A.0.5)

∂qj,x0j ∂qj,xj

79

dF X X ∂F dqi,x

i

(Q) = (Q)

dt ∂qi,xi dt

i∈N xi ∈Si

X X ∂F

= qi,xi (E[ ci |Q ] − E[ ci |qi,xi = 1, Q−i ])

∂qi,xi

i∈N xi ∈Si

X X X ∂F

= qi,xi qi,x0i ·

∂qi,xi

i xi ∈Si x0i ∈Si

E[ ci |qi,x0i = 1, Q−i ] − E[ ci |qi,xi = 1, Q−i ]

!

X X ∂F ∂F

= qi,xi qi,x0i − ·

∂qi,xi ∂qi,x0i

i xi <x0i

E[ ci |qi,x0i = 1, Q−i ] − E[ ci |qi,xi = 1, Q−i ]

P P

= i xi <x0 −qi,xi qi,x0i ·

i

2

E[ ci |qi,x0i = 1, Q−i ] − E[ ci |qi,xi = 1, Q−i ]

dF

So, we can conclude that dt (Q) ≤ 0.

Thus F is non-decreasing along the trajectories of the replication dynamics. and, due

to the nature of the learning algorithm, all solutions of the ODE (A.0.1) remain in the

strategy space if initial conditions ∈ [0, 1]. From the previous computations, we know that

dF (Q∗ )

dt = 0 implies that ∀i ∈ N ∀xi , x0i ∈ Si :

∗

qi,x i

= 0, or (E[ ci |qi,xi = 1, Q−i ] = E[ ci |qi,xi = 1, Q−i ])

Since from Theorem A.1, all stationary points that are not Nash equilibriums are uns-

table, Theorem 3.1 holds. Thus all solutions have to converge to some stationary point

corresponding to Nash Equilibrium. We can deduce that the learning algorithm, for any

initial condition in K − K∗ , always converges to a Nash Equilibrium of instance G.

80 A. Proof for Theorem of Chapter 3

Annexe B

Let n be the total number of users and m the number of RBs, the resources and M =

{1, 2, . . . , m}. Let W (t) = (w1 , w2 , . . . , wm ) be the vector containing the weights of each

P

resource RB at time t, with wk = i∈N xi,k corresponding to the number of users on RB

k. We defined our potential function as follows :

1 X

Φ(W ) = (wk − w̄)2

2

k∈M

P xi,k

, with w̄ = i∈N m the average load.

This potential function is a standard one according to (ref Convergence Time to Nash

Equilibria, Bounds for the Convergence Rate of Randomized Local Search in a Multiplayer

Load-balancing Game, etc.). As stated in Chapter 4, the appendix is devoted to prove

n

that a Nash Equilibrium is reached in expected time O(m log( m )). We adapt the analysis

proposed by Berenbrink et al.

Algorithm 6

for each BS do

Select a RB i uniformly at random among the b used RBs

Select a free RB j uniformly at random in M − b (strategy x0i )

if (ci (xi , X−i ) > ci (x0i , X−i ) ) then

Go to RB j with probability 1 − ccii (x (xi ,X−i )

0 ,X )

−i

i

end if

end for

Let Pi,j be the random variable corresponding to the number of users moving from RB i

to RB j

81

82 B. Proof for Theorem of Chapter 4

1 1 wj

b × m−(b−1) 1− wi if j 6= i and wi ≥ wj + 1

Pi,j = 0 if j 6= i and wi ≤ wj + 1

X

1− Pi,k if j = i

k∈M,wi >wk

m m

1 X 2 1 X X

Φ(W ) = (wk − w̄) = (wi − wk )2

2 m

k∈M i=1 k=i+1

Xm

E[Φ(W (t + 1))|W (t)] = E[ (Wi (t + 1) − w̄)2 |Wi (t) = wi ]

i=1

m

X m

X

2

= (E[(Wi (t + 1))|Wi (t) = wi ] − w̄) + V ar[(Wi (t + 1))|Wi (t) = wi ]

i=1 i=1

(B.0.1)

Lemma 5.

m

X m X

X m

(E[(Wi (t + 1))|Wi (t) = wi ] − w̄)2 < Φ(W ) − 2E[Wi,k ] (wi − wk ) (B.0.2)

i=1 i=1 k=i+1

i−1

X n

X

E[Wi (t + 1)|W (t) = wi ] = wi + E[Wj,i ] − E[Wi,k ]

j=1 k=i+1

Let Si (t) be the set of BSs using resource i at time t. Let E[Wi,j ] be the average total

weight of players which move from resource i to j. Each player has a weight equals to 1.

X

E[Wi,j ] = Pi,j

`∈Si

wi wj wi − wj

= 1− ≤

b(m − b + 1) wi m

wi −wj

Let di,j = E[Wi,j ] be the average weight transferred from i to j. We have di,j ≤ m if

i < j, di,j = 0 otherwise.

m(m−1)

These transitions can be represented by a complete graph with m vertices S = 2

83

edges. Each of these edges has a weight corresponding to the average transferred weight

effected. Let e be an edge, e = (i, j) has a weight di,j for all i ∈ M , j ∈ M . Let E =

{e1 , e2 , . . . , eS } with e1 ≤ e2 ≤ · · · ≤ eS . The edges are activated sequentially from e1 to

eS with e1 the one of smaller weight.

Let W z = {w1z , w2z , . . . wSz } be the weight vector of each resource after activating z edges.

W 0 is the vector W without modification, ∀i ∈ M, wi0 = wi ⇒ Φ(W 0 ) = Φ(W ). Similarly

W S is the vector W after activating S edges which correspond to all the possible weight

transfers between the resources.

i−1

X n

X

wiS = wi + dj,i − di,k

j=1 k=i+1

Let ∆z (Φ) be the potential difference associated with the activation of the edge ez = (i, k).

PS

The objective is to measure the potential difference betweenΦ(W 0 ) et Φ(W S ) = z=1 Φ(W

z−1 ) − Φ(W z ) =

P

ez ∈E ∆z (Φ)

m

X m

X

= (wjz−1 − w̄)2 − (wjz − w̄)2

j=1 j=1

wi −wk

For all resources j activated before ez = (i, k) then di,j ≤ di,k ≤ m because j < k. The

resource i has (m − 2) resources different from k,so the average weight i could send to its

neighbors before activating ez = (i, k) is at most (m − 2)di,k ≤ m wi −w

m

k

− 2di,k

wiz−1 ≥ wi − (m − 2)di,k

Similarly, the resource k has (m − 2) neighbors different from resource i. The average

weight k may receive from them before the activation of ez = (i, k) is at most

wkz−1 ≤ wk + (m − 2)di,k

84 B. Proof for Theorem of Chapter 4

X

This implies Φ(W 0 )−Φ(W S ) ≥ 2di,k (wi − wk ). Since Φ(W S ) ≤ Φ(W 0 )−

P

ez ∈E 2di,k (wi − wk )

ez ∈E

and di,k = E[Wi,k ], the lemma holds.

Lemma 6.

m m m

!

X X X

V ar[(Wi (t + 1))|Wi (t) = wi ] ≤ 2 E[Wi,k ] (B.0.3)

i=1 i=1 k=i+1

`

Démonstration. Recall that Si (t) is the set of players on resource i at time t. Let Y(k,i)

be the unit vector indicating that player ` moves from resource k to i. Let ` et `0 be two

` ` 0

different players, then Y(k,i) and Y(k,i) are independents.

X

`

V ar[(Wi (t + 1))|Wi (t) = wi ] = V ar[ Y(k,i) ]

`∈N

X X

`

= V ar[Y(k,i) ]

`∈N `∈Sk (t)

X X X

= Pk,i (1 − Pk,i ) + Pi,i (1 − Pi,i )

k∈M `∈Sk (t) `∈Si (t)

k6=i

X X X

< Pk,i + (1 − Pi,i )

k∈M `∈Sk (t) `∈Si (t)

k6=i

X X

< (E[Wk,i ]) + E[Wi,k ]

k∈M k6=i

k6=i

m

X m

X X X

V ar[(Wi (t + 1))|Wi (t) = wi ] ≤ E[Wk,i ] + E[Wi,k ]

i=1 i=1 k6=i k6=i

m m

!

X X

≤2 E[Wi,k ] (B.0.4)

i=1 k=i+1

85

Lemma 7.

m X

X m

E[Φ(W (t + 1))|W (t)] < Φ(W (t)) − 4E[Wi,k ] (wi − wk ) (B.0.5)

i=1 k=i+1

. From Equation (B.0.1) and the both previous Lemmas, this lemma holds.

Lemma 8.

16

E[Φ(W (t + 1))|W (t)] < Φ(W (t))(1 − ) (B.0.6)

(m + 2)

Pm Pm

wi −wk i=1 k=i+1 (wi −wk )2

Démonstration. By definition, we have E[Wi,k ] = b(m−b+1) . Since Φ(W ) = m ,

we can deduce that

m X

m

X mΦ(W )

E[Wi,k ] (wi − wk ) =

b(m − b + 1)

i=1 k=i+1

So, from Lemma 7, we get E[Φ(W (t + 1))|W (t)] < Φ(W (t)) − 4 mΦ(W (t))

b(m−b+1) . By computation,

m 4

we can obtain b(m−b+1) ≥ (m+2) for any b and m such that m ≥ b > 0. So we can deduce

16

E[Φ(W (t + 1))|W (t)] < Φ(W (t))(1 − (m+2) ).

16

From Lemma 8, we deduce that E[Φ(W (t + τ ))|W (t)] ≤ Φ(W (t))(1 − (m+2) )τ . Recall

2 2

n 2

that Φ(W ) = 21 k∈M (wk − m n

. So, after τ = m log( nm ) steps,

P

) , thus 0 ≤ Φ(W ) ≤ 2m

2

n2 16 m log( nm )

E[Φ(W (t + τ ))|W (t)] ≤ 2m (1 − (m+2) ) ≤ 2 By Markov’s inequality, the probability

W (t + τ ) is a Nash equilibrium given W (t) = W is greater than 1/2. Now for each new run

of τ steps, the probability to reach a Nash Equilibrium is at least 1, hence the expected

time to reach such a equilibrium is at most 2τ .

86 B. Proof for Theorem of Chapter 4

Bibliographie

[3GP08] 3GPP, Requirements for further advancements for e-utra (lte-advanced) (re-

lease 8).

[3gp11a] 3gpp, evolved universal terrestrial radio access network (e-utran) x2 applica-

tion protocol (x2ap).

[3GP11b] 3GPP, Technical report : Coordinated multi-point operation for lte physical

layer aspects.

[3gp12] 3gpp, evolved universal terrestrial radio access (e-utra) ; physical layer proce-

dures, 2012.

[AA03] Eitan Altman and Zwi Altman, S-modular games and power control in wi-

reless networks, Automatic Control, IEEE Transactions on 48 (2003), no. 5,

839–842.

[AADM07] Andrea Abrardo, Alessandro Alessio, Paolo Detti, and Marco Moretti, Cen-

tralized radio resource allocation for ofdma cellular systems, 5738–5743.

[AB12] Clemens PJ Adolphs and Petra Berenbrink, Distributed selfish load balan-

cing with weights and speeds, Proceedings of the 2012 ACM symposium on

Principles of distributed computing, ACM, 2012, pp. 135–144.

[AKC+ 14] Amine Adouane, Kinda Khawam, Johanne Cohen, Dana Marinca, and Samir

Tohme, Replicator dynamics for distributed inter-cell interference coordina-

tion, Computers and Communication (ISCC), 2014 IEEE Symposium on,

IEEE, 2014, pp. 1–7.

[AKG11] Alireza Attar, Vikram Krishnamurthy, and Omid Namvar Gharehshiran, In-

terference management using cognitive base-stations for umts lte, Communi-

cations Magazine, IEEE 49 (2011), no. 8, 152–159.

[AkI08] Ibrahim Ismail Mohammed Al-kebsi and Mahamod Ismail, The impact of

modulation adaptation and power control on papr clipping technique in ofdm

of 4g systems, Telecommunication Technologies 2008 and 2008 2nd Malaysia

Conference on Photonics. NCTT-MCP 2008. 6th National Conference on,

IEEE, 2008, pp. 295–299.

87

88 BIBLIOGRAPHIE

in lte, http ://www.wirelessdesignmag.com/blogs/2011/02/understanding-

downlink-power-allocation-lte, 02 may 2011.

[ARK+ 14a] Amine Adouane, Lise Rodier, Kinda Khawam, Johanne Cohen, and Samir

Tohme, Distributed load balancing game for inter-cell interference coordina-

tion, European Wireless 2014 ; 20th European Wireless Conference ; Procee-

dings of, VDE, 2014, pp. 1–6.

[ARK+ 14b] , Game theoretic framework for inter-cell interference coordination,

Wireless Communications and Networking Conference (WCNC), 2014 IEEE,

IEEE, 2014, pp. 57–62.

[Ass08] Mohamad Assaad, Optimal fractional frequency reuse (ffr) in multicellular

ofdma system, Vehicular Technology Conference, 2008. VTC 2008-Fall. IEEE

68th, IEEE, 2008, pp. 1–5.

[BC13] Olivier Bournez and Johanne Cohen, Learning equilibria in games by stochas-

tic distributed algorithms, Computer and Information Sciences III, Springer,

2013, pp. 31–38.

[BCRG14] Erez Biton, Asaf Cohen, Guy Reina, and Omer Gurewitz, Distributed inter-

cell interference mitigation via joint scheduling and power control under noise

rise constraints, Wireless Communications, IEEE Transactions on 13 (2014),

no. 6, 3464–3477.

[BFHH12] Petra Berenbrink, Tom Friedetzky, Iman Hajirasouliha, and Zengjian Hu,

Convergence to equilibria in distributed, selfish reallocation processes with

weighted tasks, Algorithmica 62 (2012), no. 3-4, 767–786.

[BPVG06] Paola Bisaglia, Silvano Pupolin, Daniele Veronesi, and Michele Gobbi, Re-

source allocation and power control in a tdd ofdm-based system for 4g cellular

networks, Vehicular Technology Conference, 2006. VTC 2006-Spring. IEEE

63rd, vol. 4, IEEE, 2006, pp. 1595–1599.

[BZGA10] Stefan Brueck, Lu Zhao, Jochen Giese, and M Awais Amin, Centralized sche-

duling for joint transmission coordinated multi-point in lte-advanced, 177–184.

[CAM+ 09] Vikram Chandrasekhar, Jeffrey G Andrews, Tarik Muharemovict, Zukang

Shen, and Alan Gatherer, Power control in two-tier femtocell networks, Wi-

reless Communications, IEEE Transactions on 8 (2009), no. 8, 4316–4328.

[Cha13] Himanshu Chauhan, Lte network topology and architecture,

http ://worldtechie.blogspot.fr/2013/04/lte-network-topology-and-

architecture.html, 25 April 2013.

[CKKL04] Jung Min Choi, Jin Sam Kwak, Ho Seok Kim, and Jae Hong Lee, Adap-

tive subcarrier, bit, and power allocation algorithm for mimo-ofdma system,

BIBLIOGRAPHIE 89

vol. 3, IEEE, 2004, pp. 1801–1805.

[CLL13] Bumkwi Choi, Seyoun Lim, and Tae-Jin Lee, Sequential frequency reuse with

power control for ofdma systems, Wireless Communications and Mobile Com-

puting 13 (2013), no. 1, 37–46.

[Com09] Anritsu Company, Lte resource guide, 2009.

[CTG09] Pierre Coucheney, Corinne Touati, and Bruno Gaujal, Fair and efficient user-

network association algorithm for multi-technology wireless networks, INFO-

COM 2009, IEEE, IEEE, 2009, pp. 2811–2815.

[DMMS14] Supratim Deb, Pantelis Monogioudis, Jerzy Miernik, and James P Seymour,

Algorithms for enhanced inter-cell interference coordination (eicic) in lte het-

nets, IEEE/ACM Transactions on Networking (TON) 22 (2014), no. 1, 137–

150.

[DSZ12] A Daeinabi, K Sandrasegaran, and X Zhu, Survey of intercell interference

mitigation techniques in lte downlink networks, Telecommunication Networks

and Applications Conference (ATNAC), 2012 Australasian, IEEE, 2012,

pp. 1–6.

[EASH09] Jan Ellenbeck, Hussein Al-Shatri, and Christian Hartmann, Performance of

decentralized interference coordination in the lte uplink, Vehicular Technology

Conference Fall (VTC 2009-Fall), 2009 IEEE 70th, IEEE, 2009, pp. 1–5.

[EE13] Mohammed S ElBamby and Khaled MF Elsayed, Performance analysis of

soft frequency reuse schemes for a multi-cell lte-advanced system with carrier

aggregation, Personal Indoor and Mobile Radio Communications (PIMRC),

2013 IEEE 24th International Symposium on, IEEE, 2013, pp. 1528–1532.

[EHB08] Jan Ellenbeck, Christian Hartmann, and Lars Berlemann, Decentralized inter-

cell interference coordination by autonomous spectral reuse decisions, Wireless

Conference, 2008. EW 2008. 14th European, IEEE, 2008, pp. 1–7.

[ENJA08] Jad El-Najjar, Brigitte Jaumard, and Chadi Assi, Minimizing interference in

wimax/802.16 based mesh networks with centralized scheduling, 1–6.

[Gol04] Paul W Goldberg, Bounds for the convergence rate of randomized local search

in a multiplayer load-balancing game, Proceedings of the twenty-third an-

nual ACM symposium on Principles of distributed computing, ACM, 2004,

pp. 131–140.

[HDDM13] Mesud Hadzialic, Branko Dosenovic, Merim Dzaferagic, and Jasmin Musovic,

Cloud-ran : Innovative radio access network architecture, ELMAR, 2013 55th

International Symposium, IEEE, 2013, pp. 115–120.

90 BIBLIOGRAPHIE

[HMLN13] Amir Helmy, Leila Musavian, and Tho Le-Ngoc, Energy-efficient power

adaptation over a frequency-selective fading channel with delay and power

constraints.

[HS03] Josef Hofbauer and Karl Sigmund, Evolutionary game dynamics, Bulletin of

the American Mathematical Society 40 (2003), no. 4, 479–519.

[JL03] Jiho Jang and Kwang Bok Lee, Transmit power adaptation for multiuser ofdm

systems, Selected Areas in Communications, IEEE Journal on 21 (2003),

no. 2, 171–178.

[KAL+ 14] Kinda Khawam, Amine Adouane, Samer Lahoud, Johanne Cohen, and Samir

Tohme, Game theoretic framework for power control in intercell interference

coordination, Networking Conference, 2014 IFIP, IEEE, 2014, pp. 1–8.

[KLL03] Didem Kivanc, Guoqing Li, and Hui Liu, Computationally efficient bandwidth

allocation and power control for ofdma, Wireless Communications, IEEE

Transactions on 2 (2003), no. 6, 1150–1158.

[KM11] Bujar Krasniqi and Christoph F Mecklenbrauker, Efficiency of partial fre-

quency reuse in power used depending on user’s selection for cellular net-

works, Personal Indoor and Mobile Radio Communications (PIMRC), 2011

IEEE 22nd International Symposium on, IEEE, 2011, pp. 268–272.

[KW11] Martin Kasparick and Gerhard Wunder, Autonomous distributed power

control algorithms for interference mitigation in multi-antenna cellular net-

works, Wireless Conference 2011-Sustainable Wireless Technologies (Euro-

pean Wireless), 11th European, VDE, 2011, pp. 1–8.

[KZM12] Xin Kang, Rui Zhang, and Mehul Motani, Price-based resource allocation for

spectrum-sharing femtocell networks : A stackelberg game approach, Selected

Areas in Communications, IEEE Journal on 30 (2012), no. 3, 538–549.

[LAV14] Sandra Lagen, Adrian Agustin, and Josep Vidal, Decentralized beamforming

with coordinated sounding for inter-cell interference management.

[LCK08] Zhenyu Liang, Yong Huat Chew, and Chi Chung Ko, A linear programming

solution to subcarrier, bit and power allocation for multicell ofdma systems,

Wireless Communications and Networking Conference, 2008. WCNC 2008.

IEEE, IEEE, 2008, pp. 1273–1278.

[LLK+ 11] Lars Lindbom, Robert Love, Sandeep Krishnamurthy, Chunhai Yao, Nobu-

hiko Miki, and Vikram Chandrasekhar, Enhanced inter-cell interference coor-

dination for heterogeneous networks in lte-advanced : A survey, arXiv preprint

arXiv :1112.1344 (2011).

[LPJZ09] David López-Pérez, A Juttner, and Jie Zhang, Dynamic frequency planning

versus frequency reuse schemes in ofdma networks, Vehicular Technology

Conference, 2009. VTC Spring 2009. IEEE 69th, IEEE, 2009, pp. 1–5.

BIBLIOGRAPHIE 91

[LYW+ 11] Zhaoming Lu, Yan Yang, Xiangming Wen, Ying Ju, and Wei Zheng, A cross-

layer resource allocation scheme for icic in lte-advanced, Journal of Network

and Computer Applications 34 (2011), no. 6, 1861–1868.

[MDV13] Marco Maso, Merouane Debbah, and Lorenzo Vangelista, A distributed ap-

proach to interference alignment in ofdm-based two-tiered networks, Vehicular

Technology, IEEE Transactions on 62 (2013), no. 5, 1935–1949.

[MMT08] Xuehong Mao, Amine Maaref, and Koon Hoo Teo, Adaptive soft frequency

reuse for inter-cell interference coordination in sc-fdma based 3gpp lte uplinks,

Global Telecommunications Conference, 2008. IEEE GLOBECOM 2008.

IEEE, IEEE, 2008, pp. 1–6.

[MPS07] Farhad Meshkati, H Vincent Poor, and Stuart C Schwartz, Energy-efficient

resource allocation in wireless networks, Signal Processing Magazine, IEEE

24 (2007), no. 3, 58–68.

[NSV+ 14] Giovanni Nardini, Giovanni Stea, Antonio Virdis, Marco Caretti, and Dario

Sabella, Improving network performance via optimization-based centralized co-

ordination of lte-a cells, 18–22.

[OR94] Martin J Osborne and Ariel Rubinstein, A course in game theory, MIT press,

1994.

[Pro06] Third Generation Partnership Project, Physical layer aspects for evolved uni-

versal terrestrial radio access (utra) (release 7), 2006.

[PTLa14] Harri Pennanen, Antti Tolli, and Matti Latva-aho, Robust beamforming with

decentralized interference coordination in cognitive radio networks, Acoustics,

Speech and Signal Processing (ICASSP), 2014 IEEE International Conference

on, IEEE, 2014, pp. 7308–7312.

[PWSF13] Klaus I Pedersen, Yuanye Wang, Stanislaw Strzyz, and Frank Frederiksen,

Enhanced inter-cell interference coordination in co-channel multi-layer lte-

advanced networks, Wireless Communications, IEEE 20 (2013), no. 3.

[RLLS98] Ragunathan Rajkumar, Chen Lee, John P Lehoczky, and Daniel P Siewiorek,

Practical solutions for qos-based resource allocation problems, Real-Time Sys-

tems Symposium, 1998. Proceedings., The 19th IEEE, IEEE, 1998, pp. 296–

306.

[Ros73] Robert W Rosenthal, A class of games possessing pure-strategy nash equili-

bria, International Journal of Game Theory 2 (1973), no. 1, 65–67.

[RY07] Mahmudur Rahman and Halim Yanikomeroglu, Multicell downlink ofdm sub-

channel allocations using dynamic intercell coordination, Global Telecommu-

nications Conference, 2007. GLOBECOM’07. IEEE, IEEE, 2007, pp. 5220–

5225.

92 BIBLIOGRAPHIE

[SAR09] Sanam Sadr, Alagan Anpalagan, and Kaamran Raahemifar, Radio resource

allocation algorithms for the downlink of multiuser ofdm communication sys-

tems, Communications Surveys & Tutorials, IEEE 11 (2009), no. 3, 92–106.

[SEH14] Ahmed Hamdi Sakr, Hesham ElSawy, and Ekram Hossain, Location-aware

coordinated multipoint transmission in ofdma networks, Proc. of IEEE Inter-

national Conference on Communications, ICC, 2014.

[SMB+ 14] Vincenzo Sciancalepore, Vincenzo Mancuso, Albert Banchs, Shmuel Zaks,

and Antonio Capone, Interference coordination strategies for content update

dissemination in lte-a.

[SPT94] PS Sastry, VV Phansalkar, and M Thathachar, Decentralized learning of nash

equilibria in multi-person stochastic games with incomplete information, Sys-

tems, Man and Cybernetics, IEEE Transactions on 24 (1994), no. 5, 769–777.

[SQ09] Rainer Schoenen and Fei Qin, Adaptive power control for 4g ofdma systems

on frequency selective fading channels, Wireless Communications, Networking

and Mobile Computing, 2009. WiCom’09. 5th International Conference on,

IEEE, 2009, pp. 1–6.

[SSC11] Akram Bin Sediq, Rainer Schoenen, and Zhijun Chao, A novel distributed

inter-cell interference coordination scheme based on projected subgradient and

network flow optimization, Personal Indoor and Mobile Radio Communica-

tions (PIMRC), 2011 IEEE 22nd International Symposium on, IEEE, 2011,

pp. 1595–1600.

[SV09] Alexander L Stolyar and Harish Viswanathan, Self-organizing dynamic frac-

tional frequency reuse for best-effort traffic through distributed inter-cell co-

ordination, INFOCOM 2009, IEEE, IEEE, 2009, pp. 1287–1295.

[Top79] Donald M Topkis, Equilibrium points in nonzero-sum n-person submodular

games, SIAM Journal on Control and Optimization 17 (1979), no. 6, 773–

787.

[WCLM99] Cheong Yui Wong, Roger S Cheng, K Ben Lataief, and Ross D Murch, Mul-

tiuser ofdm with adaptive subcarrier, bit, and power allocation, vol. 17, IEEE,

1999, pp. 1747–1758.

[WKSV10] Gerhard Wunder, Martin Kasparick, Alexander Stolyar, and Harish Viswana-

than, Self-organizing distributed inter-cell beam coordination in cellular net-

works with best effort traffic, Modeling and Optimization in Mobile, Ad Hoc

and Wireless Networks (WiOpt), 2010 Proceedings of the 8th International

Symposium on, IEEE, 2010, pp. 295–302.

[WWS+ ] Jiao Wang, Jay Weitzen, Volkan Sevindik, Oguz Bayat, and Mingzhe Li, Dy-

namic centralized interference coordination in femto cell network with qos

provision.

BIBLIOGRAPHIE 93

[Yao95] David D Yao, S-modular games, with queueing applications, Queueing Sys-

tems 21 (1995), no. 3-4, 449–475.

[YDH+ 10] Yiwei Yu, Eryk Dutkiewicz, Xiaojing Huang, Markus Mueck, and Gengfa

Fang, Performance analysis of soft frequency reuse for inter-cell interference

coordination in lte networks, Communications and Information Technologies

(ISCIT), 2010 International Symposium on, IEEE, 2010, pp. 504–509.

[YKS13] Wei Yu, Taesoo Kwon, and Changyong Shin, Multicell coordination via joint

scheduling, beamforming, and power spectrum adaptation, Wireless Commu-

nications, IEEE Transactions on 12 (2013), no. 7, 1–14.

[ZCM+ 12] Haijun Zhang, Xiaoli Chu, Wenmin Ma, Wei Zheng, and Xiangming Wen,

Resource allocation with interference mitigation in ofdma femtocells for co-

channel deployment, EURASIP Journal on Wireless Communications and

Networking 2012 (2012), no. 1, 1–9.