5G AND BEYOND WIRELESS NETWORKS OPTIMIZATION Thesis

5G AND BEYOND WIRELESS
NETWORKS OPTIMIZATION
THROUGH UPLINK AND
DOWNLINK DECOUPLED ACCESS
A THESIS SUBMITTED TO THE U NIVERSITY OF M ANCHESTER

FOR THE DEGREE OF D OCTOR OF P HILOSOPHY
IN THE FACULTY OF S CIENCE AND E NGINEERING
2021
By
Yao Shi
School of Electrical and Electronic Engineering
Contents
List of Tables 7
List of Figures 8
List of Abbreviations 11
Abstract 14
Declaration 16
Copyright 17
Acknowledgements 18
1 Introduction 19
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.2 Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.4 Publication List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2 Background and State of the Art 28

2.1 Heterogeneous Networks . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2 Small Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2.1 The Advantages of SCells . . . . . . . . . . . . . . . . . . . 32
2.2.2 The Challenges of SCell . . . . . . . . . . . . . . . . . . . . 32
2.3 Cell Association Criteria . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Signal to Interference Plus Noise Power Ratio (SINR) based
Cell Association . . . . . . . . . . . . . . . . . . . . . . . . 34
2
2.3.2 Reference Signal Received Power (RSRP) based Cell Associ-
ation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.3 Reference Signal Received Quality (RSRQ) based Cell Asso-
ciation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.4 Cell Range Extension . . . . . . . . . . . . . . . . . . . . . . 35
2.4 Downlink and Uplink Decoupling . . . . . . . . . . . . . . . . . . . 38
2.5 Millimeter Wave (mmWave) Communications . . . . . . . . . . . . . 39
2.6 Multi-Connectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.7 Device to Device Communication . . . . . . . . . . . . . . . . . . . 42
2.7.1 Short-Range Communication Technologies . . . . . . . . . . 42
2.7.2 Advantages of D2D communication . . . . . . . . . . . . . . 43
2.7.3 D2D Applications . . . . . . . . . . . . . . . . . . . . . . . 43
2.7.4 Key Technologies and Challenges of D2D . . . . . . . . . . . 45
2.8 Cellular-Enabled Unmanned Aerial Vehicular Communication . . . . 49
2.9 Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . 50
2.10 Mobile Edge Computing . . . . . . . . . . . . . . . . . . . . . . . . 53
2.11 Optimization Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.12 Simulation Process . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3 Capacity Enhancement Decoupled Access 58

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.1.1 Main Contributions . . . . . . . . . . . . . . . . . . . . . . . 60
3.1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2.1 Heterogeneous network model . . . . . . . . . . . . . . . . . 60
3.2.2 Cell Association model . . . . . . . . . . . . . . . . . . . . . 61
3.2.3 Uplink and Downlink Interference Model . . . . . . . . . . . 62
3.2.4 Resource Allocation . . . . . . . . . . . . . . . . . . . . . . 63
3.2.5 Propagation Model . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.6 Energy Efficiency Model . . . . . . . . . . . . . . . . . . . . 64
3.3 Capacity-Based Cell Association . . . . . . . . . . . . . . . . . . . . 64
3.3.1 Coupling/Decoupling Association . . . . . . . . . . . . . . . 65
3.3.2 Single/Multi-BS Association . . . . . . . . . . . . . . . . . . 67
3.3.3 Joint Cell Association and Network Capacity Maximization . 70
3.4 Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.4.1 Decoupling Access in Three-Tier HetNet . . . . . . . . . . . 71
3
3.4.2 Decoupling Access in Two-Tier HetNet . . . . . . . . . . . . 80
3.4.3 Reduced Density Scenario . . . . . . . . . . . . . . . . . . . 83
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4 Interference Management for D2D based on DUDe 86

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.2.1 Transmission Data Rate . . . . . . . . . . . . . . . . . . . . 92
4.2.2 Path-loss and Energy-Efficiency Models . . . . . . . . . . . . 94
4.3 Resource Allocation in D2D-Underlay HetNets with DUDe . . . . . . 95
4.3.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . 95
4.3.2 Cell-Association . . . . . . . . . . . . . . . . . . . . . . . . 96
4.3.3 Resource Block Allocation . . . . . . . . . . . . . . . . . . . 97
4.3.4 Power Control . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5 DUDe Cellular-Enabled UAV Communications 114

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.2.1 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.2.2 Propagation Model . . . . . . . . . . . . . . . . . . . . . . . 117
5.2.3 Antenna Elements . . . . . . . . . . . . . . . . . . . . . . . 117
5.2.4 Blockage Model . . . . . . . . . . . . . . . . . . . . . . . . 117
5.2.5 Resource Block Allocation . . . . . . . . . . . . . . . . . . . 118
5.2.6 Transmission Data Rate . . . . . . . . . . . . . . . . . . . . 118
5.3 DUDe Access in Cellular Networks Serving UAVs and GUEs . . . . . 119
5.4 Energy-Efficiency Maximization . . . . . . . . . . . . . . . . . . . . 121
5.4.1 GUEs Energy-Efficiency Maximization . . . . . . . . . . . . 122
5.4.2 UAVs Energy-Efficiency Maximization . . . . . . . . . . . . 126
5.4.3 Computational Complexity . . . . . . . . . . . . . . . . . . . 128
4
5.5.1 Proposed DUDe Access Scheme Simulations . . . . . . . . . 129
5.5.2 Energy-Efficiency Maximization . . . . . . . . . . . . . . . . 132
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6 Q-Learning and DQL based Power Control for UAVs 136

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.1.1 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.1.3 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.2 RL-Based Optimization of Energy-Efficiency . . . . . . . . . . . . . 139
6.2.1 Q-Learning (QL) . . . . . . . . . . . . . . . . . . . . . . . . 139
6.2.2 Deep Q-Learning (DQL) . . . . . . . . . . . . . . . . . . . . 141
6.3 Implementation of QL and DQL . . . . . . . . . . . . . . . . . . . . 144
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
7 DUDe Access in Mobile Edge Computing 154

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.2.1 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.2.2 Transmission Model . . . . . . . . . . . . . . . . . . . . . . 158
7.2.3 MEC Model . . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.2.4 Latency Model . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.2.5 Communication Model . . . . . . . . . . . . . . . . . . . . . 159
7.2.6 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . 161
7.3 JCASA Based on SPA . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.3.1 Student-Project Allocation Matching Model . . . . . . . . . . 163
7.3.2 Swap Matching . . . . . . . . . . . . . . . . . . . . . . . . . 165
7.4 Power Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
8 Conclusions and Future Research Directions 183

8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
5
8.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Bibliography 188
A Proof in Chapter 5 210

A.1 Proof of Lemma 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
A.2 Proof of Lemma 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
Word Count: 40480
6
List of Tables
2.1 The Parameters of Various Types of Nodes in the LTE-A Network . . 31

2.2 Qualitative Comparison among Existing Cell Association Schemes for
Multi-tier Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3 Comparison of Short Range Technologies . . . . . . . . . . . . . . . 42
3.1 Cell Association Schemes . . . . . . . . . . . . . . . . . . . . . . . . 71

3.2 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.1 Resource Allocation Schemes . . . . . . . . . . . . . . . . . . . . . 105

5.1 Antenna parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

5.2 Transmission Parameters of the Different Schemes . . . . . . . . . . 119
5.3 Complexity of Different Schemes . . . . . . . . . . . . . . . . . . . 128
6.1 QL Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

6.2 DQL Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.1 Summary of MEC Scenarios and Objectives . . . . . . . . . . . . . . 155

7.2 Resource Allocation Schemes . . . . . . . . . . . . . . . . . . . . . 174
7
List of Figures
1.1 Evolution of the mobile communication system. . . . . . . . . . . . . 20
2.1 A multi-tier network model. . . . . . . . . . . . . . . . . . . . . . . 29

2.2 LTE-A heterogeneous network topology. . . . . . . . . . . . . . . . . 30
2.3 The phenomenon of the Uplink and downlink imbalance. . . . . . . . 36
2.4 The change of BS coverage after introducing a bias. . . . . . . . . . . 37
2.5 A scenario of downlink and uplink decoupling. . . . . . . . . . . . . 38
2.6 Combining DL on C-band with UL on sub-3 GHz band. . . . . . . . . 39
2.7 Dual connectivity used for DUDe. . . . . . . . . . . . . . . . . . . . 41
2.8 D2D application scenarios. . . . . . . . . . . . . . . . . . . . . . . . 44
2.9 Schematic representation of overlay inband, underlay inband, and out-
band D2D. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.10 Interference situation in the cell when share the UL and DL resources. 48
2.11 Flappy Bird. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.12 The locations of MEC in end-to-end network. . . . . . . . . . . . . . 54
2.13 Typical use cases and requirements of MEC. . . . . . . . . . . . . . . 54
3.1 Illustration of SMBA association scheme. . . . . . . . . . . . . . . . 67

3.2 Average number of UEs per cell in the UL for different association
schemes in single and dual connectivity scenarios (β = 2, γ = 0.5). . . 72
3.3 Ratio of UEs associated with different BSs at different ratios of SCell
density to MCell density and association schemes (γ = 0.5). . . . . . 73
3.4 Ratio of UEs decoupled in the UL and DL vs. ratio of SCell density to
MCell density (γ = 0.5). . . . . . . . . . . . . . . . . . . . . . . . . 74
3.5 10th, 20th, 50th, 80th, 90th percentile UE UL data rates based on dif-
ferent cell association schemes in single and dual connectivity scenar-
ios (β = 5, γ = 0.5). . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
8
3.6 Network sum-rate in the UL vs. ratio of SCell density to MCell density
in single and dual connectivity scenarios (γ = 0.5). . . . . . . . . . . 77
3.7 Network sum-rate in the UL vs. ratio of UHF SCell density to SCell
density in single and dual connectivity scenarios (β = 5). . . . . . . . 78
3.8 Normalized network energy efficiency when the UHF SCells make up
different proportions of all the SCells (β = 5). . . . . . . . . . . . . . 79
3.9 Number of handovers in the UL vs. ratio of SCell density to MCell
density in single and dual connectivity scenarios (γ = 0.5). . . . . . . 80
3.10 Ratio of UEs associated with mmWave SCells in the UL vs. ratio of
SCell density to MCell density with different ratios of mmWave SCells. 81
3.11 Network data rate in the UL vs. ratio of SCell to MCell density with
different ratios of mmWave SCells. . . . . . . . . . . . . . . . . . . . 82
3.12 Ratio of UEs associated with different BSs at different ratios of SCell
density to MCell density and association schemes (less density sce-
nario, γ = 0.5). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.13 Network sum-rate in the UL vs. ratio of SCell density to MCell density
in single and dual connectivity scenarios (less density scenario, γ = 0.5). 84
4.1 Interference between MUEs, SUEs and DUEs . . . . . . . . . . . . . 93

4.2 Weighted bipartite graph matching. . . . . . . . . . . . . . . . . . . . 99
4.3 Network sum-rate vs. ratio of SCell to MCell density. . . . . . . . . . 106
4.4 Network sum-rate vs. ratio of SCell to MCell density. . . . . . . . . . 108
4.5 CUE sum-rate vs. ratio of SCell to MCell density. . . . . . . . . . . . 109
4.6 10th , 20th , 50th , 80th , and 90th percentile UE data rates based on dif-
ferent schemes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.7 DUE sum-rate vs. maximum transmission power of DUEs. . . . . . . 110
4.8 Normalized energy-efficiency vs. number of UEs. . . . . . . . . . . . 111
4.9 Average number of iterations of Algorithms 1 and 2. . . . . . . . . . 112
5.1 Cell Association and Interference map of the proposed scheme. . . . . 119
5.2 10th , 20th , 50th , 80th , and 90th percentile data rate per mmWave link
and UHF link in the UL and DL. . . . . . . . . . . . . . . . . . . . . 130
5.3 UL UAV sum-rate and normalized EE vs. UAV altitude. . . . . . . . . 131
5.4 Network sum-rate in the UL and DL vs. number of UAVs/GUEs. . . . 132
5.5 UL GUE sum-rate and normalized EE vs. ratio of SCell density to
MCell density. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
9
5.6 UL UAV sum-rate and normalized EE vs. ratio of SCell density to
MCell density. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.7 10th , 20th , 50th , 80th , and 90th percentile data rate per UAV and GUE
in the UL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.1 Proposed power control scheme using DQL . . . . . . . . . . . . . . 144

6.2 DQL Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.3 10th , 30th , 50th , 70th , and 90th percentile data rate per user in the UL:
(a) mmWave band and (b) UHF band - SBS to MBS ratio = 4 . . . . 148
6.4 UL GUE sum-rate vs. ratio of SBS to MBS. . . . . . . . . . . . . . . 149
6.5 UL UAV sum-rate vs. ratio of SBS to MBS. . . . . . . . . . . . . . . 150
6.6 UL mmWave UEs normalized EE vs. ratio of SBS to MBS. . . . . . . 151
6.7 UL UHF UEs normalized EE vs. ratio of SBS to MBS. . . . . . . . . 151
6.8 Network UL sum-rate vs. ratio of SBS to MBS. . . . . . . . . . . . . 152
7.1 A DUDe MEC model. . . . . . . . . . . . . . . . . . . . . . . . . . 156

7.2 Flow-chart of the proposed scheme. . . . . . . . . . . . . . . . . . . 174
7.3 UL sum transmission latency. . . . . . . . . . . . . . . . . . . . . . . 176
7.4 UL sum transmission latency. . . . . . . . . . . . . . . . . . . . . . . 176
7.5 Sum computation latency of Min-PL-G-FPC, CUDA and SPA-FPC
schemes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7.6 Sum-latency. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
7.7 Sum-latency of Min-PL-G-FPC, SPA-SM-FPC and SPA-SM-OPA. . . 179
7.8 Jain’s Fairness Index of UL Transmission Latency vs. number of SBSs. 179
7.9 Jain’s Fairness Index of Computation Latency vs. number of SBSs. . . 180
7.10 Normalized UL transmission EE vs. number of SBSs. . . . . . . . . . 181
7.11 10th , 20th , 50th , 80th , and 90th percentile MD data rates based on dif-
ferent schemes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
10
List of Abbreviations
AMC Adaptive Modulation and Coding

AI Artificial Intelligence
AWGN Additive White Gaussian Noise
BER Bit Error Rate
BS Base Station
CA Carrier Aggregation
C&C Command and Control
CNPC Control and Non-Payload Communication
CPU Central Processing Unit
CDA Coupling/Decoupling Association
CQI Channel Quality Indicator
CRE Cell Range Extension
CSI Channel State Information
CUDA Coupled Uplink/Downlink Access
CUE Cellular User Equipment
DC Dual Connectivity
D.C. Local Difference of Two Convex Functions
DL Downlink
DNN Deep Neural Network
DQL Deep Q-Learning
DQN Deep Q-Network
DRL Deep Reinforcement Learning
DUDe Downlink and Uplink Decoupling
DUE Device-to-Device User Equipment
D2D Device-to-Device
EE Energy Efficiency
FDD Frequency Division Duplexing
11
FPC Fractional Power Control
GCS Ground Control Station
GUE Ground User Equipment
HARQ Hybrid Automatic Repeat Request
HCP Hardcore Processes
HD High-Definition
HetNet Heterogeneous Network
HPPP Homogeneous Poisson Point Process
ICIC Inter-Cell Interference Coordination
LOS Line-of-Sight
IoT Internet of Things
LA Link Adaptation
LAN Local Area Network
LTE-A Long Term Evolution Advanced
MAC Media Access Control
Max-RSS Maximum Received Signal Strength
MBS Macro Base Station
MCell Macro Cell
MCS Modulation and Coding Schemes
MD Mobile Device
MDP Markov Decision Process
MEC Mobile Edge Computing/Multi-Access Edge Computing
MIMO Multiple-Input multiple-Output
MINLP Mixed-Integer Non-Linear Programming
Min-PL Minimum Path Loss
mmWave Millimeter Wave
MUE Macro Cell User Equipment
NLOS Non-Line-of-Sight
NOMA Non-Orthogonal Multiple Access
OFDM Orthogonal Frequency Division Multiplex
OFDMA Orthogonal Frequency Division Multiple Access
PAN Personal Area Network
PS Packet Scheduling
QL Q-Learning
QoS Quality of Service
12
RB Resource Block
RE Resource Element
RL Reinforcement Learning
ReLU Rectified Linear Unit
RRM Radio Resource Management
RSSI Received Signal Strength Indicator
RSRP Reference Signal Received Power
RSRQ Reference Signal Received Quality
RUE Reuse User Equipment
SBS Small Base Station
SCell Small Cell
SE Spectrum Efficiency
SINR Signal to Interference Plus Noise Power Ratio
SM Swap Matching
SMBA Single/Multi-BS Association
SNR Signal to Noise Ratio
SP Strauss Process
SPA Student-Project Allocation
SUE Small Cell User Equipment
TDD Time Division Duplexing
UAV Unmanned Aerial Vehicle
UE User Equipment
UHF Ultra High Frequency
UL Uplink
UPF User Plane Function
U2G UAV-to-Ground
3GPP Third Generation Partnership Project
13
Abstract
In 3G and 4G eras, many mobile communication services were initially created for
users to consume content rather than generate it. Thus, the network traffic in the down-
link (DL) tended to be much larger than that in the uplink (UL). As such, traditional
networks were designed to mainly maximize DL capacity. However, with the rise of
a new trend of user-centric wireless services and applications, the demand for the UL
capacity is expected to intensify [1]. A prominent factor that restricts the UL capacity
in heterogeneous networks (HetNets) is the UL and DL imbalance problem. As there
is a clear disparity between the transmission powers of the macro cells (MCells) and
small cells (SCells), the best serving cell per user may be different in the UL and DL
directions; hence, if the UL and DL associations are coupled, the UL capacity may
be severely limited, and this problem will become even worse when the operating fre-
quency increases. A potential solution to alleviate such a predicament is the DL/UL
decoupling (DUDe) technique [2], which allows users to connect to different BSs at
different frequency bands in the UL and DL to maximize the link capacity in each
direction. DUDe not only shortens the distance between the user equipments (UEs)
and serving base stations (BSs), but also makes better use of the spectral resources of
SCells.
Most studies of DUDe are based on the minimum path-loss (Min-PL) cell associa-
tion scheme, which chooses the BS with the minimum path-loss in the UL. However,
it does not take the cell loads into consideration and cannot make full use of the net-
work resource. Thus, this thesis first considers decoupling the UL and DL BS from the
perspective of maximizing network capacity. Moreover, the increasing desire to incor-
porate millimeter-wave (mmWave) communications and multi-connectivity in future
networks further enriches the possibilities to achieve higher capacities, and they have
potential to combine with DUDe to get better performance. MmWave communication
is restricted by the high penetration and path loss, but DUDe can shorten the distance
between UEs and their serving BSs in the UL, thus enriching the coverage of mmWave
cells. As for multi-connectivity, it is highly controversial to adopt dual connectivity in
the UL since the transmission power of a UE is much lower than that of a BS. How-
ever, DUDe makes it more power-efficient in the UL and brings a solution to this prob-
lem. Considering all those things, this thesis investigates the merits of adopting DUDe
capacity-based multi-association in the ultra-high frequency (UHF) and mmWave hy-
brid networks, where mobile users may simultaneously connect to multiple different
14
UHF SCells, mmWave SCells and/or UHF MCells.
Apart from DUDe, mmWave communications, and multi-connectivity, another
way to improve the network capacity is device-to-device (D2D) communication. It is
more flexible than traditional cellular communication, and is a potential enabler for the
Internet of Things (IoT) networks. However, D2D communication also increases the
complexity and heterogeneity of the network structure, and bring challenges for inter-
ference management. As such, DUDe facilitates a more benign environment for D2D
receivers by lowering the cellular user (CUE) UL transmission power, which results
in less interference, and enables more D2D transmissions. To this end, we investigate
the application of DUDe in D2D-underlay heterogeneous networks, and propose an
efficient joint cell-association, subchannel allocation, and power control scheme for
network sum-rate maximization.
This thesis also investigates the potential of DUDe in cellular-enabled unmanned
aerial vehicle (UAV) communication networks. Integrating UAVs with cellular net-
works is considered pivotal to tapping into new business opportunities for cellular op-
erators, especially as the smartphone market is almost saturated. In this thesis, we
propose a DUDe scheme for efficiently integrating UAVs within 5G cellular systems,
where the UAVs’ control and non-payload communication (CNPC) links, as well as
the ground user (GUE) uplinks and downlinks, are decoupled from the perspectives
of serving BSs and operating frequency bands. Moreover, as battery life is a major
constraint for UL transmissions, an optimal power allocation algorithm based on frac-
tional programming and successive convex approximation, and two algorithms based
on Q-learning (QL) and deep Q-learning (DQL) are proposed for optimizing the EE of
this DUDe scheme.
Lastly, this thesis analyzes the application of DUDe to mobile edge computing
(MEC). MEC is a key player in low latency 5G networks, particularly for resolving the
conflict between computationally-intensive mobile applications and resource-limited
mobile devices (MDs). As such, there has been intense interest in this topic. Gener-
ally, computational task offloading is limited by the type of MD-BS association with
almost all previous works considering offloading an MD’s computational task to the
MEC servers attached to its serving BS. In multi-BS association, or DUDe scenarios,
however, one MD can have multiple serving BSs, and hence more offloading choices
can be exploited. Motivated by this, the thesis considers the communication and com-
putational disparity of small BS (SBS) and macro BS (MBS) cloudlets with the objec-
tive to optimize the system performance subject to certain quality-of-service require-
ments. Specifically, a joint BS association and subchannel allocation algorithm, based
on a student-project allocation (SPA) matching model and an optimal power allocation
scheme, are proposed to minimize the network sum-latency.
15
Declaration
No portion of the work referred to in this thesis has been
submitted in support of an application for another degree or
qualification of this or any other university or other institute
of learning.
16
Copyright
i. The author of this thesis (including any appendices and/or schedules to this the-
sis) owns certain copyright or related rights in it (the “Copyright”) and s/he has
given The University of Manchester certain rights to use such Copyright, includ-
ing for administrative purposes.
ii. Copies of this thesis, either in full or in extracts and whether in hard or electronic
copy, may be made only in accordance with the Copyright, Designs and Patents
Act 1988 (as amended) and regulations issued under it or, where appropriate,
in accordance with licensing agreements which the University has from time to
time. This page must form part of any such copies made.
iii. The ownership of certain Copyright, patents, designs, trade marks and other in-
tellectual property (the “Intellectual Property”) and any reproductions of copy-
right works in the thesis, for example graphs and tables (“Reproductions”), which
may be described in this thesis, may not be owned by the author and may be
owned by third parties. Such Intellectual Property and Reproductions cannot
and must not be made available for use without the prior written permission of
the owner(s) of the relevant Intellectual Property and/or Reproductions.
iv. Further information on the conditions under which disclosure, publication and
commercialisation of this thesis, the Copyright and any Intellectual Property
and/or Reproductions described in it may take place is available in the Univer-
sity IP Policy (see http://documents.manchester.ac.uk/DocuInfo.aspx?
DocID=487), in any relevant Thesis restriction declarations deposited in the Uni-
versity Library, The University Library’s regulations (see http://www.manchester.
ac.uk/library/aboutus/regulations) and in The University’s policy on pre-
sentation of Theses
17
Acknowledgements
First and foremost, I would like to express my sincere gratitude to my supervisor, Prof.
Emad Alsusa. He has been supportive and patient throughout my Ph.D., and I have
learned a lot of crucial characters of becoming a qualified researcher from him, such
as curiosity, passion, and hard-working. He not only offered valuable advice on my
research, but also introduced to me knowledgeable peers and professors, in particular
Prof. Mohammed W. Baidas to whom I must express my sincere appreciation and
gratitude for his expert advice on various optimisation techniques. I also wish to thank
Dr. Khairi Hamdi and Dr. Yuanwei Liu for accepting the invitation to be my Ph.D.
examiners.
I would also like to thank my collaborators, Mutasem Q Hamdan. He is enthusiastic

about my research on UAV communications and was essential in helping me complete
the Q-Learning and deep Q-Learning parts in this thesis. I have learned many advanced
techniques from him.
I want to extend my thanks to my friends and colleagues who provided help and en-
couragement. I would like especially to thank Dr. Runze Hu, Dr. Murat Temiz, Ruowei
Yin, Han Wu, and other colleagues in the Microwave and Communication Group at the
University of Manchester.
My acknowledgment would not be complete without thanking my parents for their

relentless support and endless love.
18
Chapter 1
Introduction
1.1 Background
In just under four decades, mobile telecommunications have evolved from first genera-
tion (1G) to the fifth generation (5G). At the same time, the traditional communication
network, which mainly focuses on voice services, has been gradually revolutionised
into a multi-functional system that provides mobile data, mobile computing and mobile
multimedia services. Compared with previous generations of mobile communication
technologies, the fourth generation of broadband cellular network technology (4G) has
greatly improved the transmission rate, latency and user experience. However, in the
face of the rapid development of the Internet of Things (IoT) and the rise of new data-
hungry applications such as mobile video, telemedicine and virtual reality, it will be
difficult for 4G networks to support the massive data access and ever-increasing data
traffic demands of future mobile communications for which 5G comes into being.
5G provides not only better capacity in the downlink (DL), but also higher capacity
and lower latency in the uplink (UL). In 3G and 4G eras, most services involved in
mobile communications were created for users to consume content rather than generate
it. Therefore, compared with the network traffic in the UL, the traffic in the DL is
much larger. In such a context, traditional networks are designed to maximize the
DL capacity based on the DL performance parameters. However, new services like
augmented reality, wireless body area network and vehicular network require higher
and faster UL transmission as well. Under these circumstances, a large number of
studies have been carried out to improve the UL performance (including those that
improve the performances in both UL and DL directions), which can be categorized as
network-side and non-network-side.
19
20 CHAPTER 1. INTRODUCTION
Figure 1.1: Evolution of the mobile communication system.
From the network side, the solution is to deploy low-power nodes such as femto-
cells and pico cells at the edge of the macro cells (MCells) to increase network capacity
and improve network performance. These low power nodes are also known as small
cells (SCells). The large deployment of SCells in the network has been considered to
be one of the most efficient and cost-effective solutions for meeting the capacity and
coverage needs of future mobile communication networks. However, the change in
network architecture brings corresponding problems. In conventional wireless cellular
networks, a specific user equipment (UE) is connected to the base station (BS) with
the maximum DL received signal strength in both the UL and DL [1]. It is called cou-
pled uplink/downlink access (CUDA). In homogeneous networks, this CUDA mode
is a nearly optimal access approach, since the best serving BS is the same in the UL
and DL. However, with the deployment of more and more small low-cost cells, the
traditional homogeneous networks become heterogeneous networks (HetNets). Due to
the transmission power disparity of SCells and traditional MCells, the DL coverage of
MCells is usually much larger than that of SCells. Consequently, more UEs are associ-
ated with MCells in the DL. However, the situation in the UL is completely different.
As UEs are battery powered with approximate transmission power, their coverage is
1.1. BACKGROUND 21
almost the same. If the associated cell in the UL is consistent with that in the DL, the
link quality of the MCell edge UEs will be poor. This problem is called the UL and DL
imbalance, which increases simultaneously with the growth of SCells [3]. In this situa-
tion, CUDA may no longer be the optimal user access strategy. For the seek of reduced
path-loss, it would be a better choice for some UEs to connect to a geometrically closer
SCell in the UL. This architecture is called DL and UL decoupling (DUDe) or the UL
and DL split [2], which allows UEs to connect to different BSs in the two directions.
Similarly, the UL and DL operating frequency bands can also be decoupled. As the
frequency of C-band is higher than existing systems, using the existing 4G spectrum
of sub-3 GHz to carry 5G NR UL services becomes a natural option. At Mobile World
Congress 2019, Huawei presented a ”5G Uplink and Downlink Decoupling” scheme
which enables C-Band and 1.8GHz co-site deployment with the same coverage [4].
DUDe has numerous advantages. BS decoupling shortens the distance between

UEs and their serving BSs, thus reduces the path-loss and improves the signal to in-
terference plus noise power ratio (SINR). The reduced path-loss allows lower trans-
mission power, so the interference to neighboring BSs is reduced. Furthermore, BS
decoupling pushes more UEs to underutilized SCells in the UL, which in turn allows
more efficient resource utilization of SCells and achieves higher data rates. Frequency
decoupling increases the reach of 5G in C-band to enhance the customer experience,
and could reduce investment in additional sites. Due to the advantages above, DUDe
promotes green communications. From the user perspective, since 5G mobile phones
are equipped with more antennas and can support wider bandwidths than 4G equip-
ment, they may cost more energy. Under these circumstances, BS decoupling allows
lower transmission power, which helps reduce energy consumption and extend battery
life. From the mobile operator perspective, over 90% of network costs are spent on en-
ergy, consisting mostly of fuel and electricity consumption. Most of this spend powers
the radio access network (RAN), with data centres and fibre transport accounting for
a smaller share [5]. New techniques such as sleep node and energy harvesting have
attracted more and more attention [6]. In this context, frequency decoupling widens
the coverage of BSs in the UL, thus mobile operators can deploy fewer new sites in
the 5G era, which reduces energy consumption. Another technology that can increase
resource utilization of SCells is cell range extension (CRE) or biased cell associa-
tion [7]. It adds a cell selection offset to the reference signals of the SCells to increase
their coverage and offload more traffic from the MCells. However, since CRE does not
consider the UL and DL separately, it will bring some problems. For example, a large
offset may cause high interference in the DL for SCell edge UEs, and the resource of
MCells may not be fully utilized. By contrast, DUDe brings in the benefits of having
very high offsets in the UL without the interference effects in the DL.
For the non-network side, millimeter wave (mmWave) communication, carrier ag-
gregation (CA), multi-connectivity, device-to-device (D2D) communication, mobile
edge computing (MEC), multiple input multiple output (MIMO) [8], etc., are utilized
to improve network performance. Among them, mmWave communication, multi-
connectivity, D2D communication and MEC can be combined with DUDe to get better
UL performances.
MmWave communication and multi-connectivity: Wireless communications
are reaching the bottleneck of in-sufficient traditional spectrum resources, mmWave
frequency spectrum is envisaged to offer an essential lifeline for future broadband cel-
lular systems in the quest to satisfy the anticipated explosive resources demand. How-
ever, the characteristics such as high near-field path-loss, severe oxygen and gaseous
water molecules absorption have posed significant challenges to mmWave communi-
cation. A recent study on electromagnetic field exposure [9] showed that to satisfy
applicable exposure limits at frequencies above 6 GHz, the maximum transmission
power in the UL might have to be several dB below the power levels used for current
cellular technologies. As the transmission power of mobile devices (MDs) is limited,
in order to keep the UL SNR unchanged, the distance between UEs and serving BSs
should be shortened, and DUDe can achieve this goal. Moreover, as millimeter waves
are sensitive to blockage, they are best suited for indoor hotspots and outdoor small
cell scenarios. It would be more realistic to jointly deploy mmWave BSs and ultra-
high frequency (UHF) BSs. UHF BSs can provide seamless coverage and support
high mobility, while mmWave BSs can provide high data rates. In this case, multi-
connectivity can allow a user being simultaneously served by multiple BSs [10, 11].
Take dual connectivity (DC) as an example, each UE is allowed to simultaneously uti-
lize the spectrum from two BSs connected together via non-ideal backhaul links. But
as the transmission power of UEs is limited, adopting in the UL is highly controversial.
Nevertheless, we might as well combine dual connectivity with DUDe since DUDe re-
duces the distance between users making it more power-efficient in the UL. The work
has been published in the IEEE Transactions on Vehicular Technology [12].
Device-to-device communication: D2D communication refers to direct commu-
nication between two mobile users without traversing the BS or core network. In a
traditional cellular network, all communications must go through the BS, which is
1.1. BACKGROUND 23
suitable for conventional low data rate mobile services such as voice call and text mes-
saging in which UEs are far away from each other. However, with the development
of some data rate services, such as gaming, video sharing, proximity-aware social net-
working, UEs could be close enough for direct communications (i.e., D2D). In such
scenarios, D2D communications can not only increase the spectral efficiency of the
network, but energy efficiency, delay, and fairness as well. Most of the current research
focuses on in-band underlay D2D, in which D2D re-uses the licensed spectrum; hence,
the spectrum resources are simultaneously used by the cellular UEs (CUEs) and D2D
UEs. A few studies consider D2D in HetNets, where SCells and MCells co-exists. In
this case, not only the co-channel interference between the CUE and D2D UE (DUE),
SCell UE (SUE) and CUE, SUE and DUE should be considered, but the interference
between the DUEs/SUEs that share the same subchannels should be considered as
well. As the complex interference among UEs, it is necessary to control the inter-
ference, apart from various resource allocation and power control methods, DUDe is
a new promising technique to solve this problem. DUDe facilitates a more benign
environment for D2D receivers by lowering the CUE UL transmission power, which
results in less interference, and enables more D2D transmissions. The work has been
submitted to the IEEE Transactions on Vehicular Technology.
Mobile edge computing: Low latency is also an important requirement for 5G

and beyond networks. With the development of 5G ultra-dense networks, MEC has
become a promising technique to minimize the increasing latency and satisfy capac-
ity requirements of mobile communications, where edge servers are deployed at the
cellular base-stations, such as macro base-stations (MBSs) and/or small base-stations
(SBSs) to finish the computationally-intensive tasks and workloads offloaded from
MDs. Generally speaking, computational task offloading is potentially limited by the
type of MD-BS association, and almost all the previous works consider offloading an
MD’s computation task to the MEC servers available at its serving BS. According to
CUDA, a specific UE is connected to the BS with the maximum DL received signal
strength in both the UL and DL the BS, but the BS with rich communication resources
may not have enough computation resources. Fortunately, in DUDe scenario, one MD
can have multiple serving BSs, hence more offloading choices can be exploited, and
the serving BS can be chosen from the perspective of both communication capacity
and computing power. The work has been submitted to the IEEE Transactions on
Communications.
From the above analysis, it is clear that investing DUDe in 5G and beyond networks
is necessary. Apart from these, DUDe can be also be applied to other 5G scenarios, for
example, unmanned aerial vehicle (UAV) communications. The global market for
UAVs has grown substantially over the past decade and has become a high point for
economic growth in many countries. Hence integrating UAVs with cellular networks
is considered pivotal to tapping into new business opportunities for cellular operators.
It is worth noting that UAV communication has a significant character which is the
UL and DL data rate imbalance. UAV uplinks are dominated by data transmission,
requiring high data rates up to hundreds of Mbps. In contrast, UAV downlinks are
dominated by control and non-payload communications (CNPC). They are crucial to
the UAV operation requiring low latency, ultra-reliability, and high security, but could
not be guaranteed due to the interference from the ground users (GUEs) [13]. To over-
come this problem, DUDe can be applied to split the UAV data communications from
the CNPC communications, as well as the GUE communications from UAV commu-
nications. Part of the work has been published in the IEEE Systems Journal [14] and
part of the work has been submitted to IEEE Internet of Things Journal.
1.2 Main Contributions

The primary aim of this report is to study how to decouple the access work in HetNets
and explore the feasibility of combining DUDe with other cutting-edge communication
techniques, such as mmWave communication, multi-connectivity, D2D, UAV commu-
nication, mobile edge computing, etc.. The main contributions of this report are listed
as follows, more detailed contributions are provided in the introduction part of each
chapter.
• In Chapter 3, we investigate the feasibility of decoupling the UL and DL serv-

ing BSs from the perspective of maximizing the total network capacity in both
single and multi-connectivity scenarios. Specifically, we propose an adaptive
decoupling and multi-BS association scheme for mmWave/UHF hybrid HetNets
where mmWave SCells, UHF SCells and MCells coexist, and compare its perfor-
mance for a wide range of metrics with benchmark single and dual connectivity
alternatives.
• In Chapter 4, we investigate the performance of DUDe in D2D-underlay Het-

Nets. In turn, a joint cell-association, subchannel allocation, and power control
(J-CA-SCA-PC) problem for UL network sum-rate maximization is formulated,
1.3. THESIS ORGANIZATION 25
which happens to be non-convex and NP-hard, and thus is computationally-

expensive. Thus, in this thesis, we decouple the J-CA-SCA-PC problem and
solve it via a low-complexity and near-optimal scheme.
• In Chapter 5, we propose a novel DUDe based access scheme for cellular-

enabled UAV communications, in which both dimensions of serving BSs and
operating frequencies are utilized. The proposed DUDe scheme eliminates inter-
ference between the UAVs and GUE, as well as part of the inter-cell interference
among GUEs, leading to: (1) interference-limited GUE communications, and
(2) noise-limited UAV and line-of-sight (LOS) GUE communications, which is
fundamentally different from the conventional scenario where UAVs and GUEs
share the same resources. Then, we devise an optimal low-complexity power
allocation algorithm, based on fractional programming and successive convex
approximation, to maximize the UL communication energy efficiency (EE) in
tandem with the proposed DUDe scheme. Furthermore, in Chapter 6, two power
allocation schemes based on Q-learning (QL) and deep Q-learning (DQL) are
also proposed to optimize the EE of the DUDe network, and are compared with
the fractional power control scheme which is applied in 4G and 5G networks, as
well as the optimal power allocation scheme we proposed.
• In Chapter 7, we analyze the application of DUDe on MEC, the communication

and computational disparity of SBS and MBS cloudlets are taken into consid-
eration. In contrast to most existing works that consider BS association and
subchannel allocation separately, a joint BS association and subchannel alloca-
tion algorithm based on a student-project allocation (SPA) matching is proposed
in this thesis. Then, an optimal power allocation algorithm is proposed to mini-
mize the sum-latency of the network MDs. The formulated problem is a sum-of-
ratios problem, which is non-convex and NP-hard. To efficiently and optimally
tackle it, the proposed algorithm tightly approximates the problem as a convex
optimization problem, and successively solves it until convergence to the global
optimal power allocation solution.
1.3 Thesis Organization

The rest of this thesis is organized as: Chapter 2 introduces the theoretical background
and some representative published works. Chapter 3 investigates the compatibility
of mmWave communication and multi-connectivity in DUDe. Chapter 4 proposes

a DUDe joint BS association, subchannel allocation and power control scheme for
D2D-underlay HetNets. Chapter 5 investigate the performances of DUDe in cellular-
enabled UAV networks. Chapter 6 analyzes the application of QL and DQL to UAV
power control in the DUDe scenario. Chapter 7 analyzes the application of DUDe on
MEC. Chapter 8 provides the conclusion and future work.
1.4 Publication List

• Journal Papers
1. Shi Y, Alsusa E, Ebrahim A, and Baidas M W, ”Uplink Performance En-
hancement Through Adaptive Multi-Association and Decoupling in UHF-mmWave
Hybrid Networks”, IEEE Transactions on Vehicular Technology, 2019, 68(10):
9735-9746.
2. Shi Y, Alsusa E, and Baidas M W, ”Joint DL/UL Decoupled Cell-Association
and Resource Allocation in D2D-Underlay HetNets”, submitted to IEEE Trans-
actions on Vehicular Technology on December 31, 2020, under review with ma-
jor revisions.
3. Shi Y, Alsusa E, and Baidas M W, ”An Energy-Efficient Decoupled Ac-
cess Scheme for Cellular-Enabled UAV Communication Systems”, IEEE Sys-
tems Journal, Accepted.
4. Shi Y, Mutasem Q Hamdan, Emad Alsusa, Khairi A Hamdai and Baidas M
W, ”Decoupled Access and Reinforcement Machine Learning Power Control for
Cellular-Enabled UAV network”, submitted to IEEE Internet of Things Journal
on January 6, 2021, under review.
5. Shi Y, Alsusa E, and Baidas M W, ”On the Application of Uplink/Downlink
Decoupled Access in Heterogeneous Mobile Edge Computing”, submitted to
IEEE Transactions on Communications on October 20, 2020, under review with
major revisions.
• Conference Papers
1. Shi Y, Alsusa E, and Ebrahim A, ”Single and Dual Connectivity for Decou-
pled Uplink and Downlink Access in UHF-mmWave Hybrid Networks”, 2018
IEEE Global Communications Conference (GLOBECOM), IEEE, 2018: 1-6.
1.4. PUBLICATION LIST 27
2. Shi Y, Alsusa E, and Baidas M W, ”Downlink-Uplink Decoupled Access

in Heterogeneous Cellular Networks with UAVs,” 2020 IEEE 31st Annual In-
ternational Symposium on Personal, Indoor and Mobile Radio Communications
(PIMRC), IEEE, 1-6.
Chapter 2
Background and State of the Art
2.1 Heterogeneous Networks

Mobile communication has experienced tremendous changes in recent decades. From
2G to 4G, the data traffic over networks has increased significantly and will continue to
grow with the maturity of new concepts, such as augmented reality, wireless body area
network, vehicular network, smart city, and so on. To meet the growing data rate re-
quirement, network operators will have to improve the capacity of networks constantly.
In traditional homogeneous networks, where each node, i.e., MCell, has a similar wire-
less transmission mode, BS type, coverage, topology, etc., the main method to increase
network capacity is to increase the number of BSs and expand transmission bandwidth.
However, sometimes it is unavailable to add more MCells due to the lack of available
sites, for example, in crowded public places and terrain complex areas. Thus, more
small cells are introduced to the network. These cells are flexible-deployment, low-
cost and low-coverage. They can provide extra connections in hot spots and change
the topology of cellular networks. Networks that overlay MCells and SCells are called
HetNets.
In fact, there are two definitions of HetNet. One refers to the network with dif-
ferent wireless access schemes, where traditional cellular networks (2G, 3G, 4G),
satellite communication, Ad Hoc network, or broadband wireless connection (WLAN,
WiMAX) coexist. The other refers to the network with different types of network
element nodes, such as femtocells and picocells. For one thing, these nodes can be
deployed by users, which makes a random topology of the entire network different
from the traditional cellular structure. For another, the BS type, parameter settings,
transmission power, and antenna configuration of these nodes can be different from
28
2.2. SMALL CELL 29
Figure 2.1: A multi-tier network model.
each other, which provides users with diverse communication methods and ubiquitous
access services. The definition of HetNet in this thesis is a special case of the lat-
ter one, where BSs can transmit over different frequencies with different transmission
powers. A HetNet with N different types of BSs is called an N-tier network. A typ-
ical structure of HetNet is a Long Term Evolution Advanced (LTE-A) network [15],
where MCells provide essential coverage and seamless connections; dense SCells pro-
vide supplementary coverage and increase the network capacity. Fig. 2.1 illustrates
such a multi-tier network with an MCell overlaid by relays, picocells and femtocells.
Arrows indicate wireless links, and the dashed lines denote the backhaul connections.
The LTE-A HetNet adopts advanced wireless and wired backhaul technologies, that is,
some SCells can configure their backhaul interfaces, or several cells can form a cluster
to aggregate and forward network data. SCells can also use relay nodes, i.e.Donor eNB
(DeNB), to set up their backhaul links. Fig. 2.2 shows the network topology [16].
2.2 Small Cell
SCell is a general term for a series of low-powered cellular radio access nodes. They
are small compared to a mobile MCell, partly because they have a shorter range and
30 CHAPTER 2. BACKGROUND AND STATE OF THE ART
Figure 2.2: LTE-A heterogeneous network topology.
partly because they typically handle fewer concurrent calls or sessions. They can in-
crease the capacity of indoor and outdoor hotspot scenarios and cope with the explo-
sive growth of mobile traffic. Generally, SBS can be divided into two types: fully-
functioning BSs (picocells and femtocells) and macro-extension access points (relays
and Remote Radio Heads (RRHs)) [17]. The fully-functioning SBS is capable of
performing all the functions of a MCell with lower power and a smaller coverage
area. Specifically, all functions of the entire protocol stack can be run on the fully-
functioning base station [18]. A macro-extension access node is an extension for the
MCell to effectively extend the signal coverage, and it performs all or some of the phys-
ical layer functions only. In addition, according to the coverage, transmission power,
and deployment scenarios, SCells can be classified into the following types [19]:
• Picocells are SBSs deployed and installed by operators, and their coverage is
about 300 meters. They are usually deployed in indoor or outdoor hotspots such
as shopping centers, bus stops, etc., and can serve dozens of UEs. The trans-
mission power of picocells is around 23-30dBm, and they are mainly used to
increase network capacity. The backhaul of picocells is similar to MBS and can
provide ideal high-speed, low-latency links over fiber or microwave.
• Femtocells are indoor BSs installed by UEs with a coverage of about tens of
2.2. SMALL CELL 31
meters. They are usually deployed to cover indoor spots, such as homes, offices,
conference rooms, etc., and to serve a small group of UEs. Femtocells can be
divided into three categories based on their access control mode: open access,
closed access, and hybrid access [20]. In the open-access mode, any UE can
get access to the femtocell; in the closed access mode, only the UEs belong to
the closed subscriber group (CSG) can get access to the femtocell; in the hybrid
access mode, the UEs that not belong to the CSG may camp and acquire some
level of service on these cells. However, UEs subscribed to the femtocell may
get preferential charging and treatment in comparison with UEs not subscribed
to the cell that receive service from it. The transmission power of the femtocells
is usually less than 23dBm, and they are mainly used to guarantee the signal
strength of home UEs. The femtocells can be connected to the network via any
of the users’ broadband connections such as digital subscriber line (DSL), cable,
or fiber.
• Relays are the access point deployed by operators, which can transmit the UEs’
data back and forth from and to the MCell, featuring what is considered as wire-
less backhauls. They are usually used to extend the coverage of MCells, cover
the dead zones, and improve the performance of MCell edge UEs. The trans-
mission power and coverage of the relay node and picocell are the same. The
main difference lies in three aspects: first, the picocell is a fully-functioning BS,
and the relay is an extension for the MCell; second, the picocell is mainly used
to increase the network capacity, and the relay is used to extend the coverage;
third, picocell has an ideal backhaul link, and the backhaul is a wireless in-band
or out-of-band backhaul.
Table 2.1 summarizes the basic characteristics of the deployment scenarios, coverage,
transmission power, and backhaul links of the above-mentioned SCells, and compares
them with the MCell [19].
Table 2.1: The Parameters of Various Types of Nodes in the LTE-A Network
Parameters Max transmission power Coverage radius Backhaul interface
Macrocell 46dB few km S1 interface
Relay 30dB 200m wireless
Femtocell <23dB <50m Internet IP
Picocell 23-30dBm <300m X2 interface
2.2.1 The Advantages of SCells

To meet the explosion of data and shortage of spectrum resources, there is a real beauty
to going small. The advantages of SCells are as follows:
• Flexible to deploy: SCells are smaller in size and hence they are easy to install.
Compared to MCells, SCells can be located at hotpots without erecting towers,
such as shopping malls and office buildings. They can also be deployed in weak
coverage areas such as MCell edges, indoors, tunnels, etc., so as to provide
reasonably contiguous coverage.
• Low transmission power: The distance between the user and the SCell is rela-
tively close, so the path-loss is lower and energy efficiency is higher. In addi-
tion, since the coverage and transmission power of SCells are low, the spectral
efficiency can be improved through the dense deployment of SCells, thereby
improving the network capacity.
• Low cost: The deployment of SCells does not require high costs of tower con-
struction, equipment maintenance, etc., so the cost of networking is relatively
low. In addition, SCells consume less power. The energy consumption of an
MCell is equivalent to 30 Picocells. The deployment of SCells can save 76% of
the energy compared with the traditional cellular network [21].
• Better support high-frequency resources: In order to achieve large-scale cover-

age of the network, the traditional macro cellular network generally uses UHF
spectrum resource with low penetration loss and path-loss. For signals above
3 GHz, it will experience severe penetration loss and path-loss during propaga-
tion. As the coverage of SCells is small, most of the links are line-of-sight, so
the high-frequency resources can be better utilized.
The advantages of the SCell show its strong vitality, which will inevitably become an
indispensable part of the future 5G network, but it will also introduce new challenges
as follows.
2.2.2 The Challenges of SCell

Although introducing SCells into the homogeneous network can increase network cov-
erage and capacity, it also brings some new challenges as follows:
2.2. SMALL CELL 33
• Interference management: After deploying SCells to the MCell network, the

HetNet can be divided into two layers: MCell layer and SCell layer. There are
generally two ways to allocate spectrum resources in this network: the first is
that the SCell layer uses the same or partially the same spectrum resource as the
MCell layer, called the co-channel deployment; the second is the SCell layer and
the MCell layer use orthogonal spectrum resources, which is called orthogonal
channel deployment. In the co-channel deployment scenario, there is inter-layer
interference between the MCell layer and the SCell layer, and intra-layer inter-
ference among cells in each layer. In the orthogonal channel deployment sce-
nario, there is only intra-layer interference, but the spatial bandwidth efficiency
is lower than that in the co-channel deployment scenario. Introducing SCells
into the traditional homogeneous network causes more interference. Therefore,
it is necessary to research the interference management technology in the MCell-
SCell hybrid network.
• Mobility management: Mobility management in MCell-Scell hybrid networks

involves issues such as cell selection and cell handover. The large deployment
of SCells will make mobility management in the network more complicated [22].
Due to the small scale of SCell coverage, the UE with high velocity will cross the
SCell in a short time, causing the UE to frequently switch between cells, which
increases the signaling overhead [23]. In addition, since the SCell has lower
transmission power and smaller coverage than the MCell, the traditional han-
dover process and parameters cannot be directly applied to MCell-SCell hybrid
networks [24]. The handover procedure between the SCell and MCell should
be modified. As for cell selection, when MCells and SCells are all in the open-
access mode, since UEs generally perform cell selection based on the reference
signal received power (RSRP) scheme and the transmission power gap between
SBSs and MBSs, UEs prefer to access to MBSs, which causes network load im-
balance. In summary, the deployment of SCells makes cell selection and cell
handover more complicated. It is necessary to conduct in-depth research on
these issues.
• Energy efficiency management: Energy efficiency has always been an important

indicator of network performance. Compared to traditional MCells, SCells have
lower energy consumption. The power consumption is only 15 W to maintain
the operation of a non-loaded SBS. However, with a surge of traffic, more and
more SCells will be deployed to weak coverage areas or hotspots, and the energy
consumption will greatly increase. It is necessary to study how to improve the
energy efficiency of the networks with SCells.
• Backhaul: SCell has a variety of backhaul methods, such as fiber, microwave,

DSL. Operators have to design suitable backhaul networks for different types of
SCells according to their corresponding requirements, so as to ensure the QoS
and reduce the cost of network construction.
2.3 Cell Association Criteria

One of the major issues in mobile cellular networks is how to provide the highest
possible quality of service (QoS) with limited resources. The process of allocating
resources to each UE and determining associated BS to maximize resource utilization
while minimizing interference is called user access or cell association. In the following
section, we will introduce four different cell association criteria.
2.3.1 Signal to Interference Plus Noise Power Ratio (SINR) based

Cell Association
Average received SINR is the best criterion to measure the channel quality for cell
selection. SINR refers to the received signal power over the sum of the interference
and noise power. This cell selection criterion can be represented as follows
iSINR = arg max γi , (2.1)

i
where i, iSINR and γi represent the cell index, the SINR of the i-th cell and the selected
cell index based on the max SINR criterion, respectively [25].
2.3.2 Reference Signal Received Power (RSRP) based Cell Associ-

ation
RSRP based cell association is specified in the LTE by the 3GPP organization in Re-
lease 12. When a typical UE accesses the network, each BS in its vicinity transmits a
reference signal to it; then the UE measures the strength of this reference signal to esti-
mate the distance of the BS. In fact, after using the inter-cell interference coordination
2.3. CELL ASSOCIATION CRITERIA 35
technology (ICIC) [26], RSRP based cell association is not optimal because it only re-
flects the signal strength received from each BS but does not reflect the actual channel
condition. However, it is still widely used because of its simplicity in the calculation.
This criterion is represented as follows
iRSRP = arg max si , (2.2)

i
where iRSRP and si represent the selected cell index based on the max RSRP criterion
and the RSRP of the i-th cell, respectively [25].
2.3.3 Reference Signal Received Quality (RSRQ) based Cell Asso-

ciation
Another criterion specified in LTE is RSRQ based cell association. It is a variant of
RSRP. RSRQ is defined by the RSRP over the total received power, i.e., received signal
strength indicator (RSSI) [25]. Under full load conditions, RSRQ is proportional to the
following value
RSRP S SINR
RSRQ = ∝ = (2.3)
RSSI S + I + N 1 + SINR
where S, I and N represent received signal power, interference power and Gaussian
white noise, respectively. The RSRQ based cell association criterion is represented as
follows
iRSRQ = arg max ρi , (2.4)

i
where iRSRQ and ρi represent the selected cell index based on the max RSRQ criterion
and the RSRQ of the i-th cell, respectively [25].
2.3.4 Cell Range Extension

In a HetNet, since there is a huge gap in the transmission power between SCells and
MCells, the UL and DL cell bounder are different. As can be seen in Fig. 2.3, MCell
DL coverage is larger than its UL coverage, and SCell UL coverage is larger than its
DL coverage, which results in the UE in the middle area of the two borders accessing
different BSs according to the UL and DL maximum SINR/RSRP/RSRQ criteria. The
Figure 2.3: The phenomenon of the Uplink and downlink imbalance.
phenomenon that the UL and DL optimal access cells are different is called the UL and
DL imbalance.
A possible solution to this problem is cell range extension (CRE) [7]. It adds a
positive cell selection offset to the reference signals of the SCells to increase their
coverage and offload more traffic from the MCells. Specifically, the three previous
criteria mentioned above can be extended to biased versions. For SINR based cell
association, the in equation (2.1) can be modified to
(
γMCell ,
γi = (2.5)
γSCell + αSINR .
That is, bias αSINR is added when calculating the SINR of MCells. Similarly, si and ρi
in equations (2.2) and (2.3) are modified to
(
sMCell ,
si = (2.6)
sSCell + αRSRP .
(
ρMCell ,
ρi = (2.7)
ρSCell + αRSRQ .
Fig. 2.4 shows the effect of cell coverage change after the offset is introduced.
As can be seen from the figure, the coverage of the MCell becomes smaller and the
coverage of the SCell becomes larger than before. Light-colored areas represent the
coverage area with bias. The UEs in these areas are offloaded to the micro base station.
CRE can bring some fairness to the UL, but it will cause some new problems at the
same time. For example, if the offset is set too large, the resource of MCells may not be
fully utilized. In addition, SCell edge UEs will suffer from high interference in the DL
2.3. CELL ASSOCIATION CRITERIA 37
Figure 2.4: The change of BS coverage after introducing a bias.
Table 2.2: Qualitative Comparison among Existing Cell Association Schemes for
Multi-tier Networks
RSRP RSRQ CRE
Objective Maximize received signal power Maximize SIR Balance traffic load
Applicability UL and DL UL and DL DL
Channel-aware X X X
Interference-aware × X X
Traffic load-aware × × X
Resource-aware × × ×
since the strongest SINR/RSRP/RSRQ is from MCells. Therefore, some techniques,

like ICIC [26], have been developed to combat this kind of interference. Only in this
way can increase the SCell resource utilization without influence the performance of
SCell edge UEs.
A qualitative comparison among these cell association schemes is given in Table
2.2 [27]. The specific key terms used in Table 2.2 are defined as follows. Channel-
aware refers to knowing the transmission power at the receiver and knowledge of the
instantaneous channel. Interference-aware refers to knowing the instantaneous inter-
ference at the receiver. Traffic load-aware refers to knowing the traffic load infor-
mation. Resource-aware refers to knowing the resource allocation information. The
schemes in the table are getting better and better from left to right, but the corre-
sponding required information is also increasing. Their purpose is to allocate UEs
as reasonably as possible to each BS for optimal performance. These schemes are
tractable and straightforward. However, to ensure optimal performance in multi-tier
HetNets, critical parameters such as offset values and transmission power need to be
optimized. More importantly, with the development of HetNets, some limitations of
these technologies are gradually revealed. Therefore, it is necessary to study new ac-
cess strategies to meet the requirements of the network structure.
Figure 2.5: A scenario of downlink and uplink decoupling.
2.4 Downlink and Uplink Decoupling

A better method proposed to solve the UL and DL imbalance problem and make better
use of the spectrum resources of SCells is DUDe, which allows UEs to connect to
different BSs in the UL and DL. DUDe brings in the benefits of having very high
offsets in the UL without the interference effects in the DL.
The DUDe strategy is shown in Fig. 2.5. The cell association criterion here is
RSRP. Since the MCell has a much larger transmission power than the SCell, it has
broader coverage, just like the blue cycle in the figure. According to traditional coupled
uplink/downlink access, UE2 and UE3 are associated with the MCell, and UE1 is
associated with the SCell. All the UEs are associated with the same BSs in the UL
and DL. For UE1 and UE3, they connect to the BS with the strongest RSRP in both
DL and UL. However, for UE2, since it is closer to the SCell than to the MCell, the
path loss will be lower if it connects to the SCell. If the DUDe strategy is adopted, the
DL connection will keep unchanged, but the UL connection will switch from UL1 to
UL2. In this way, UE2 is associated with the MCell in the DL and the SCell in the UL.
The connection pairs are all the best in both directions. In fact, DUDe not only pushes
more UEs to SCells in the UL, but it can also improve the SNR and energy efficiency.
If UEs transmit at a fixed power, a connection to a closer BS provides a higher SNR.
If the target SNR is fixed, UEs can transmit at a lower power via power control, then
the interference to neighboring BSs is reduced, and the UL capacity is increased.
Similarly, the UL and DL operating frequency bands can also be decoupled. As the
frequency of C-band is higher than existing systems, using the existing 4G spectrum
of sub-3 GHz to carry 5G NR UL services becomes a natural option. At Mobile World
Congress 2019, Huawei presented a ”5G Uplink and Downlink Decoupling” scheme,
where C-band is used for the DL and a sub-3 GHz band (for example, 1.8 GHz) for the
UL, thereby improving UL coverage, as shown in Fig. 2.6 [4]. Since the coverage of
2.5. MILLIMETER WAVE (MMWAVE) COMMUNICATIONS 39
Figure 2.6: Combining DL on C-band with UL on sub-3 GHz band.
BSs in the UL is widened, mobile operators can deploy fewer new sites in the 5G era.
2.5 Millimeter Wave (mmWave) Communications

Wireless communication has reached the bottleneck of insufficient traditional spectrum
resources. MmWave communications with abundant available spectrum resources may
be a potential choice for next-generation wireless communications. Millimeter-wave
refers to electromagnetic waves with wavelengths from 10 mm to 1 m and frequencies
from 30 GHz to 300 GHz (the industry has loosely considered any signal with a fre-
quency above 10 GHz to be a mmWave signal). However, the characteristics such as
high near-field path loss, severe oxygen and gaseous water molecules absorption have
posed significant challenges to mmWave communication. The difficulties in mmWave
communications are as follows [28].
• Range and directional communication: According to Friis’ transmission law,

it can be seen that the free space omnidirectional path-loss is proportional to
the square of the frequency [29]. Therefore, mmWave communication systems
require highly directional antennas and beamforming to compensate for severe
mmWave propagation path-loss [30, 31].
• Shadowing and rapid channel fluctuations: MmWave signals are incredibly vul-
nerable to shadowing. Shadowing can cause connection interruptions or poor
channel quality. For a given moving speed, the linear relationship between the
channel coherence time and the carrier frequency means that the channel coher-
ence time at the mmWave wave is minimal. For example, the Doppler spread at
60km/h at 60 GHz is approximately 3 kHz, that is, the channel will change in the
order of hundreds of microseconds, so how to quickly track the mmWave wave
channel change is an essential problem in mmWave communications [28, 29].
Luckily, beam steering [28] and dense deployment may solve these problems.
• Multiuser coordination: MmWave communication research and applications mainly

focus on point-to-point links (such as cellular backhaul [32]), as well as local
area network (LAN) and personal area network (PAN) systems limiting the num-
ber of users or having media access control (MAC) protocols that prohibit multi-
user simultaneous transmissions [33–36]. For spatial reuse with high spectral
efficiency, people need to study new mechanisms to coordinate the simultaneous
transmissions on multiple interfering links in mmWave networks.
• Processing power consumption: The power consumption is linear with the sam-
pling rate and exponentially with the number of bits per sample [37, 38]. For
low-power and low-cost devices, the high-resolution quantization at wide band-
widths with a large number of antennas is not practical. Especially in high-
definition (HD) video transmission, mmWave communications face even more
severe technical challenges.
• Oxygen and water molecules absorption: MmWave signals are mainly absorbed
by oxygen and water molecules during propagation. Oxygen molecules have dif-
ferent effects on mmWave signals in different frequency bands. Oxygen molecules
can attenuate signals by 15 dB/km in 57 to 64 GHz, but the loss in 28GHz,
38GHz and 73GHz are much lower [37, 39]. Humidity and rain fades are com-
mon problems for long-range mmWave communications but not an issue when
cell sizes are on the order of 200 m [37, 40].
Despite so many challenges above, the mmWave frequency spectrum is a candi-
date for future broadband cellular communication due to the global bandwidth short-
age. Moreover, mmWave communication has some advantages. Because of the highly
directional transmission and sensitivity to blockage, mmWave networks are noise-
limited rather than interference-limited [29,41]. In addition, the smaller wavelength of
mmWave signals also enables proportionally greater antenna gain for the same physi-
cal antenna size. Since mmWave cannot achieve universal coverage, especially in high
2.6. MULTI-CONNECTIVITY 41
Figure 2.7: Dual connectivity used for DUDe.
blockage areas, we consider a network with mmWave BSs coexisting with ultra-high
frequency (UHF) BSs in this thesis.
2.6 Multi-Connectivity
In heterogeneous wireless networks, since the frequency resources and coverage of a
single BS are limited, maintaining multiple connections is a better way to meet the
user’s capacity and coverage requirements, which is called multi-connectivity. In the
case of dual connectivity (DC) [42] (Fig. 2.9), each UE is allowed to simultaneously
utilize the spectrum from two cells transmitting over different frequencies via non-ideal
backhaul links (later releases may add support for infra-frequency band deployments).
In DC scenario, the information can be either split or transmitted twice. The two cells
are connected via a standard X2 interface. They operate separately, handling their
scheduling and control signaling (e.g., HARQ message), thereby significantly relaxing
the backhaul requirements [1]. This network architecture with functionality separa-
tion is estimated to save more than one-third of the current overall network power
consumption [43]. It has been introduced to Release 12 of the 3GPP [44].
Dual connectivity is expected to bring system enhancements as described below.
• Since dual connectivity simultaneously receives or sends multiple data streams,

and dynamically accesses the multiple cells with the best radio resources, it can
increase user throughput, especially for cell-edge users.
• It enhances mobile robustness. Dense small cell deployments lead to frequent

handovers for UEs when moving between SCells. However, if dual connectivity
is used, the signaling overhead will be greatly reduced as long as the UE remains
in the coverage of the macro cell.
• It alleviates the UL and DL imbalance problem in HetNets. Due to the differ-

ent transmission powers between MCells and SCells in the DL, UEs sometimes
prefer MCell to SCells although connecting to SCells may take advantage of
lower path loss. Dual connectivity increases the probability of UEs connecting
to SCells.
2.7 Device to Device Communication
2.7.1 Short-Range Communication Technologies

There are a number of different wireless technologies have been developed for very
short distances, such as Bluetooth, Wi-Fi Direct, ZigBee, etc. These technical param-
eters of these short-range wireless communication technologies are shown in Table
2.3 [45]. Among them, D2D communication technology has received great attention.
D2D communication refers to direct communication between two mobile users with-
out traversing the BS or core network. Compared to other short-range communication
technologies, D2D has a wider communication range, and it operates on a licensed
band, therefore can provide a guaranteed rate. D2D communication is more flexible
than traditional cellular communication. It can be applied to hotspots such as shopping
centers and stadiums to cover the dead zones and improve the network capacity. More-
over, it is a key enabler of IoT. Therefore, it is necessary to study D2D communications
in HetNets.
Table 2.3: Comparison of Short Range Technologies

Featured Standard Spectrum Coverage Max. data Uniformity
technology range rate of service
provision
D2D 3GPP Licensed 1000 m 1 Gbbps Yes
band
Wi-Fi 802.11 2.4 GHZ 200 m dB 250 Mbps No
direct
NFC ISO 13157 13.56 0.2m 424 kpbs No
MHZ
ZigBee 802.1504 868/915 10-100 m 250 kpbs No
MHZ, 2.4
GHZ
Bluetooth SIG 2.4 GHZ 10-100m 24 Mbps No
2.7. DEVICE TO DEVICE COMMUNICATION 43
2.7.2 Advantages of D2D communication

The specific advantages of D2D communication are as follows:
• Single-hop communication: D2D communication transforms two-hop commu-

nications (via BS) to single-hop communication, which does not require a rout-
ing protocol like a multi-hop link, so using a single-hop link has lower complex-
ity and latency.
• Short-range communication: D2D communication is typically used for short-

range data transmission, which allows the device to communicate at low trans-
mission power, thereby reducing power consumption and extending the battery
life. In addition, since the communication distance is relatively short, a higher
transmission rate and a lower delay can be obtained.
• Improve spectrum efficiency: Because of its low transmission power, D2D de-
vices can reuse the CUE spectrum resources without introducing to much in-
terference, thus enhance the spatial bandwidth efficiency through reasonable
resource allocation and interference management without lowering the perfor-
mance of MUE communication.
• Extend cell coverage: Cell edge UEs often encounter poor signal quality while
connecting to the BS. If an intermediate UE can act as a relay either between a
BS and a UE or between two UEs, the communication quality of the edge UE
can be improved and the coverage of cellular service is extended.
• Reduce BS load: Due to the boom of mobile devices, the load of BSs is increas-
ing dramatically. Introducing D2D communication into the network can separate
a part of the data exchange that does not need the BS to participate, such as local
advertising, proximity-based group gaming, content sharing, etc., and effectively
alleviate the load of BSs.
2.7.3 D2D Applications

With the development of the IoT, the applications of D2D are also increasing. Some
representative applications are shown in Figure 2.8 [46].
(1) Local service
Local message push: in this service, user data is directly transmitted between termi-
nals. By D2D communication, adjacent UEs can share data among themselves. Friends
Figure 2.8: D2D application scenarios.
can exchange photos or videos through their smartphones and team up to games. The
shops can send the latest promotional information to the surrounding customers.
Network traffic offloading: D2D communication can significantly offload a large
amount of data from the traditional cellular transmission. For example, while peo-
ple are visiting scenic spots, they often want the stories behind them. However, the
number of tourists is large, if every visitor in the scenic spot downloads the same con-
tent through the cellular network, it will cause a large waste of network resources. If
some local servers are deployed in these popular scenic spots to store the introduction
materials, when the tourists approach the spot, their mobile phones automatically re-
ceive data from the server in D2D mode, which can reduce the pressure of the cellular
network while improving the visitor experience.
(2) Internet of Things enhancement
The goal of developing mobile communications is to establish an extensive inter-
connected network that contains various types of terminals. This is the idea behind
developing IoT in cellular communications. New devices will have IoT features, such
as autonomous driving, automatically controlling vehicles and various devices. Be-
cause there will be many IoT terminals in the 5G network, the access load has become
a serious issue. D2D communication can effectively decrease the access load and la-
tency.
Internet of Vehicles: The Internet of Vehicles plays a vital role in the development
of unmanned technology. Driving at a high speed in complex road conditions requires
a short response time for the car, but a large amount of access to the internet of vehicles
causes a large delay and increases the probability of an accident. If D2D technology
is applied to the internet of vehicles, cars can easily complete a series of operations
such as autonomously discovering the surrounding vehicles, finding hazards, changing
lanes, etc., and truly ensure the safety of people.
Smart home: Instead of accessing all the devices with IoT features to the cellular
network, connecting them to an intelligent terminal by D2D first, and then connect the
terminal to the cellular network can effectively save the spectrum resources.
(3) Public safety scenarios
In massive emergency scenarios, for example, after a bomb explosion, people use
mobile phones to inform their families, network load surges in a short period of time,
so the communication requirements of many users cannot be satisfied. In this case,
D2D communication can be used to satisfy the basic voice call demands and push the
related information to people’s mobile phones.
(4) Emergency communications
When natural disasters such as earthquakes and mudslides occur, the cellular com-
munication system may be destroyed. D2D technology is applied to allow users to
establish a multi-hop self-organizing network to achieve information transmission be-
tween users and provide a solid support for disaster relief.
2.7.4 Key Technologies and Challenges of D2D

Although 3GPP has conducted research on D2D communication technology, D2D
communication still has many technical challenges and problems, such as peer dis-
covery, mode selection, security and interference management.
(1) Peer discovery
Looking at the demand of the D2D network, there should be an efficient method
for discovering peers before devices communicate with each other. From the user per-
spective, peer discovery techniques can be classified into two categories: restricted and
open. In the first case, devices cannot be discovered by the UEs without their permis-
sion. In the second case, devices can be discovered whenever they lie in the proximity
of other UEs. From the network perspective, peer discovery can be controlled lightly
Figure 2.9: Schematic representation of overlay inband, underlay inband, and outband
D2D.
or tightly by the BS [46]. Compared to light BS control, tight BS control is more ef-
ficient but requires more signaling overhead. Device discovery is the basis for D2D
communication and it is necessary to investigate an efficient peer discovery scheme to
establish D2D communication quickly.
(2) Mode selection
After the devices discover each other, it is necessary to determine whether D2D
communication is preferable compared to cellular communication. Sometimes, even if
the requirements of D2D communication are satisfied, the D2D communication mode
is not necessarily optimal in terms of performance. Therefore, it is necessary to de-
termine whether D2D communication can be selected through a mechanism to opti-
mize some performance objective, such as spectral efficiency, latency, or transmission
power. This process is called mode selection. D2D mode can be divided into two basic
modes: inband (on the cellular spectrum) and outband (on the unlicensed spectrum),
and the inband mode can be divided into underlay and overlay mode, which means the
D2D user shares the spectrum with the cellular user in an orthogonal or non-orthogonal
way. If DUEs work in underlay mode, they will cause interference to the CUEs. Fig.
2.9 illustrates the different kinds of D2D communications [47].
In-band Communication: it means that D2D communication uses the licensed

spectrum as cellular communication. The current cellular network has higher con-
trol over the licensed band and therefore in-band communication has higher reliability.
Some researchers believe that the unlicensed frequency band is uncontrollable due to
the large number of UEs, which makes it very difficult to guarantee the UE QoS. In-
band D2D communication can be further divided into underlay D2D and overlay D2D
according to the way in which the spectrum resources are used. The underlay D2D
communication link shares the same spectrum resources as the cellular link, while the
overlay D2D communication link uses dedicated spectrum resources for communica-
tion. Most of the current research is focused on underlay D2D communication. Next,
we will briefly introduce the advantages and disadvantages of in-band D2D communi-
cation.
In-band D2D communication has the following advantages: underlay D2D com-
munication can increase the spatial bandwidth efficiency of the system because it can
reuse the spectrum resource of CUEs. Because the BSs can effectively control the
licensed spectrum, the in-band D2D communication can perform effective QoS man-
agement. In-band D2D communication has the following disadvantages: for underlay
D2D communication, the reuse of spectrum resources may cause serious interference
problems; for overlay D2D communication, due to the use of dedicated spectrum re-
sources, it may cause waste of spectrum resources to some extent.
Out-band D2D communication: refers to D2D communication uses unlicensed

frequency bands. The main motivation for selecting out-band D2D communication is
that it does not interfere with cellular communication. UEs have to use an extra inter-
face for D2D communication using unlicensed bands. Only devices that have two kinds
of interfaces can perform D2D and cellular communications simultaneously. Out-band
D2D communication can be divided into the controlled mode when D2D control is
done by BSs and the autonomous mode when D2D control is done by users. One of
the most significant advantages of out-of-band D2D communication is that it does not
cause interference to the CUE, and thus does not have to be allocated dedicated cellu-
lar resources like in-band overlay D2D communication. However, there are also some
disadvantages, for example, the BS cannot effectively manage the interference of un-
licensed frequency bands. At present, research on out-of-band D2D communication is
relatively rare compared to in-band D2D communication, but some researchers have
begun to explore its advantages and consider whether it can be used as an alternative
to in-band D2D communication. The underlay mode DUEs share the same resources
as CUEs, thus introduces more interference. As shown in Fig. 2.10(a), when the UL
(a) Interference situation when share the uplink (b) Interference situation when share the down-
resources link resources
Figure 2.10: Interference situation in the cell when share the UL and DL resources.
spectrum resource is reused, the D2D receiver receives the interference caused by the
CUE using the same resource, if the distance between the CUE and the D2D receiver
is close, the interference will be high. Similarly, the communication between the D2D
pairs also affects the BSs. The closer the distance between the D2D transmitter and
the BS, the more the impact on the BS. When the UL spectrum resource is reused, as
shown in Fig. 2.10(b), the CUE will receive the signal sent by the base station, but will
also receive the interference caused by the D2D transmitter. Since the transmission
power of the BS is high, it will cause interference to nearby D2D users. In order to
further improve the spectrum utilization, multiple pairs of D2D users are allowed to
reuse the same cellular resources. Whether the UL resources or the downlink resources
are reused, interference will occur between D2D user pairs, and the situation becomes
more complicated. Since the pressure of data transmission in the DL is larger com-
pared with that in the UL, and BSs can better control the interference, UL spectrum
resources are preferentially selected to reuse.
(3) Security issues
Security is an important part of D2D communication. Because the data transmis-
sion in D2D communication is via other users’ devices, the transmitted data is prone
to many security risks. D2D links can be paralyzed by various attacks, such as eaves-
dropping, denial of service, IP spoofing, and so on. Because of these security issues
of D2D communication, it is necessary to ensure the transmission security of D2D
communication through strict authentication and key agreement mechanism. Closed
access is an effective measure to ensure security. In closed access, the UEs store a list
2.8. CELLULAR-ENABLED UNMANNED AERIAL VEHICULAR COMMUNICATION49
of reliable devices, such as colleagues, family members, etc.. Devices on the reliable
list are free to communicate with each other directly, and the devices out of this list can
communicate in cellular mode. For open access, each device can be used as a relay for
other devices, and the security cannot be guaranteed. How to ensure security in open
access mode is an important issue in current research.
(4) Interference management
Interference management is an important research challenge in D2D communica-
tion. As described before, DUEs in overlay mode may cause serious interference to
the CUEs. It is necessary to adopt interference management on the D2D communica-
tion to improve the system performance. At present, the interference problem in the
network is generally solved by resource allocation and power control.
2.8 Cellular-Enabled Unmanned Aerial Vehicular Com-

munication
With the maturity of unmanned aerial vehicle (UAV) services such as cargo delivery,
photography and surveillance, the global market for UAVs has substantially increased
over the past decade. The UAV industry has become a new driver of economic growth
as it taps into new business opportunities for the telecom operators [48]. For one thing,
UAVs are aerial terminals that can be served by cellular networks, which can be added
to its existing population of users [49, 50]. For another, UAVs can act as flying BSs to
boost throughput, coverage, and quality of service (QoS) of cellular networks [51–53].
The former case is considered in this thesis.
At present, most UAVs in the market depend on direct point-to-point communica-
tion with ground pilots or ground control stations (GCSs) over the unlicensed spec-
trum [54]. Although the unlicensed spectrum is free and can satisfy some personal
applications like visual LOS aerial photography, it is data rate and coverage are lim-
ited, unreliable, insecure, and vulnerable to interference. The surging number of UAVs
and ever-increasing requirements for UAV communications call for a more effective
and reliable communication strategy. A promising approach that has recently gained
popularity is utilizing cellular networks, i.e. cellular-enabled UAV communications.
Providing enhanced communication support for UAVs via LTE has recently been
approved by 3GPP, [55], while the feasibility of such approach was demonstrated in
[48, 56]. Although 4G networks can meet the requirements of some services, it cannot
support services requiring high data rate and/or extremely low latency. The major
challenges of 4G-enabled UAV communications are:

Limited UL capacity: The 4G bandwidth is limited and hard to meet the demand
for some data-hungry applications, such as high definition (HD) broadcast and aerial
surveillance.
High interference: Most of the channels between UAVs and BSs are LOS, and
the antennas on the UAVs are generally omnidirectional, in this case UAVs gener-
ate/receive UL/DL interference to/from more BSs than GUEs.
Limited coverage height: The coverage of 4G cells is designed for GUEs. The
flight height of UAVs is not within the main lobe of ground BS antennas, thus there are
blind areas in the airspace above 120 meters.
High latency: As computing resources are only deployed in the core network,
network congestion can cause high latency.
Fortunately, such 4G limitations have been overcome in 5G. With the use of the
C-band and mmWave band, the single carrier bandwidth of 5G networks can be 5-20
times higher than that of 4G networks [57]. Besides, MIMO and beamforming tech-
nologies can be applied to narrow the beamwidth and improve antenna gain, which
compensates for the high mmWave propagation loss, and reduces the intra-cell and
inter-cell interference, as well as the interference between the UAVs and GUEs1 . Fur-
thermore, MEC can be leveraged to reduce latency by integrating the processing and
storage modules within the MEC servers to ensure real-time performance. Thus, in this
thesis we focus on 5G-enabled UAV communications, and utilize the mmWave band
for UAV data transmission to reduce interference to GUEs and improve the network
data rate.
2.9 Reinforcement Learning

Reinforcement learning (RL) is a category of machine learning methods that are con-
cerned with how software agents ought to take actions in an environment so as to max-
imize the notion of cumulative reward. The idea of reinforcement learning algorithms
is very simple. Take the game as an example, if a certain strategy can be adopted in
the game to get a higher score, then further strengthen the strategy in order to continue
to achieve better results.
Some key terms that describe the elements of an RL problem are:
1 Although some may question the practicability of mmWave-enabled communication, the challenges
and solutions have been well addressed in [58, 59].

2.9. REINFORCEMENT LEARNING 51
Figure 2.11: Flappy Bird.
• Agent: It is an assumed entity that performs actions in an environment to gain

some reward.
• Environment (e): A scenario that an agent has to face.
• Reward (R): An immediate return given to an agent when he or she performs

specific action or task.
• State (s): State refers to the current situation returned by the environment.
• Policy (π): It is a strategy which applies by the agent to decide the next action
based on the current state.
• Value (V): It is expected long-term return with discount, as compared to the

short-term reward.
• Q value or action value (Q): Q value is quite similar to value. The only difference
between the two is that it takes an additional parameter as a current action.
Take Flappy Bird as an example, we need to click on the screen to fly the bird as far
as possible without hitting a pipe. The farther the bird flies, the higher the bonus. The
bird - Agent All kinds of water pipes - Environment The position of the bird - State
Click the screen to fly the bird - Action The farther the bird flies, the more points you
get - Reward
RL has a lot of applications, the most well-known one is for playing games. Al-
phaGo Master defeated a world champion in the game of go in 2016, then AlphaGo
Zero, a version created without using data from human games, exceeded AlphaGo
Master in 40 days. AlphaStar beat the top player Dario “ TLO ” Wunsch in StarCraft
II. Apart from this, RL can be applied to robotics for industrial automation, text sum-
marization engines, business strategy planning, etc..
RL algorithms can be divided into: model-based and model-free. In model-based

algorithms, you need to create a virtual model for each environment. The agent learns
to perform in that specific environment. However, if the model is inconsistent with the
real world, it will not perform well in the actual use scenario. By contrast, model-free
RL aims to acquire an effective behavior policy through trial and error interaction with
a black box environment [60], which is easier to implement and adjust to real scenarios.
Q-learning (QL) is a commonly used model free method. It updates Q values which
denotes value of doing action a in state s by the weighted average of the old value and
the new information. The value update rule is
temporal difference
z }| {
Qnew (st , at ) ← Q(st , at ) +α ( rt +γ max Q(st+1 , a) −Q(st , at )), (2.8)
| {z } |{z} a
old value reward | {z }
estimate of optimal future value
| {z }
new value (temporal difference target)
where γ ∈ [0, 1] is the discount factor, which makes rewards from the uncertain far
future less important than the ones in the near future that it can be fairly confident
about, and α ∈ [0, 1] is the learning rate determines to what extent newly acquired
information overrides old information. (2.8) is according to Bellman equation
Q(st , at ) = r + γ max Q(st+1 , at+1 ). (2.9)

at+1
It means the maximum return in each state is equivalent to the sum of the maximum
immediate reward and and the return (discounted by γ) obtained by following the opti-
mal policy thereafter until the end of the episode (i.e., the maximum expected reward
from the next state).
Q-learning only works in environments with discrete and finite state and action
spaces. Otherwise calculate and store the Q-value for each state-action combination
can be infeasible. To extend Q-learning, we can train a function approximator, for
example a neural network with parameters θ, to estimate the Q-values, i.e. Q(s, a) ≈
Q(s, a; θ). This is called deep Q-learning (DQL). To get the parameters θ, we need to
2.10. MOBILE EDGE COMPUTING 53
minimize the following loss function
0 0 0
L(θ) = E[(r + γ max
0
Q(s , a ; θ ) − Q(s, a; θ− ))2 ]. (2.10)
a
0 0
where s , a are the next action and state of s, a, θ− is a snapshot of the network pa-
rameters from a few iterations ago. Then, perform a greedy descent step on r +
0 0 0
γ maxa0 Q(s , a ; θ ) − Q(s, a; θ− ))2 with respect to θ. Q-Learning and DQL are off-
policy algorithms that learn about the greedy policy a = maxa Q(s, a); θ while using
a different behaviour policy for acting in the environment. This behaviour policy is
usually an ε-greedy policy, where a random action is taken with a probability ε to en-
sure good coverage of the state-action space, and the greedy action is taken with a
probability 1 − ε.
2.10 Mobile Edge Computing

Recent years have witnessed how MDs have become an indispensable part of daily
life. The development of mobile devices has triggered numerous new applications,
such as language recognition, map navigation, augmented reality, face recognition,
cloud gaming, and so on. These applications are computation-intensive and latency-
sensitive. Due to the limited battery power and computation resources of mobile de-
vices, computationally-intensive tasks and workloads are often offloaded to remote
computational servers, under the notion of cloud computing. To avoid long-distance
transmission latency from the MDs to cloud servers, MEC has been proposed as a sup-
plement to cloud computing, which enables cloud computing to be done at the edge
of a mobile network. Compared with cloud computing, edge computing servers are
deployed closer to users. MEC allows processes to take place in BSs, central offices,
and other aggregation points on the network. Fig. 2.12 shows the possible locations of
MEC in end-to-end network [61]. In this thesis, we consider MEC at MBSs and SBSs.
MEC has become a promising technique to satisfy the latency, capacity and security
requirements of many emerging services [62, 63]. At present, the smart factory, smart
city, live streaming/gaming and V2X are the most typical applications for MEC. Their
requirements for MEC are shown in Fig. 2.13 [61]. They are also the requirements for
typical 5G scenarios, i.e. the high bandwidth of eMBB, the ultra-high reliability and
low latency communication of URLLC, and the large connections of MIoT. Therefore,
MEC and 5G are mutually reinforcing. 5G network realizes the breakout of data traffic
Figure 2.12: The locations of MEC in end-to-end network.
Figure 2.13: Typical use cases and requirements of MEC.

2.11. OPTIMIZATION PROBLEM 55
through the flexible deployment of the user plane function (UPF) at the edge of the
network, and the flexibility of UPF deployment accelerated the development of edge
computing.
2.11 Optimization Problem

In this thesis, we are going to solve a series of resource allocation problems, they
are essentially optimization problems. A general optimization problem follows the
standard form below:
min f0 (x), (2.11)

x
s.t. fi (x) ≤ 0, i = 1, ..., m (2.11a)
hi (x) = 0, i = 1, ..., p (2.11b)
(2.11)
where x = [x1 , ..., xn ]T is the optimization variables, f0 (x) is the objective function,
fi (x) denote the inequality constraint functions, hi (x) denote the equality constraint
functions. The aim is to find the optimal solution x∗ which realises the smallest value
of f0 (x) while satisfying the constraints.
Since the late 1940s, a large effort has gone into developing algorithms for solving
various classes of optimization problems, analyzing their properties, and developing
good software implementations. However, we can only effectively solve a few classes
of them, for example, linear programs, least-squares problems, and convex optimiza-
tion problems. In fact, linear programs and least-squares problems are special cases of
the general convex optimization problem. The standard form of a convex optimization
problem is as follows:
min f0 (x), (2.12)

x
s.t. fk (x) ≤ 0, k = 1, ..., m (2.12a)
aTk x = bk , k = 1, ..., p (2.12b)
(2.12)
where the objective function f0 (x) and inequality constraint functions f1 , ..., fm are
convex, the equality constraints, characterized by ak ∈ Rn and bk ∈ R for k = 1, ..., p,
are affine.
Most problems in the world are non-convex, in this case, what we can do is trans-
forming them into convex optimization problems. There are very effective algorithms
that can reliably and efficiently solve large convex optimization problems with hun-
dreds or thousands of variables and constraints, and convex optimization problems
have very useful properties such as if the objective function is strictly convex, then the
problem has at most one optimal point, and the local minimum is also the global mini-
mum. Once you formulate a practical problem as a convex optimization problem, you
can solve it via many effective methods. Unconstrained minimization problems can
be solved by decent methods, such as the gradient descent method, steepest descent
method, Newton’s method, and so on. Equality constrained minimization problems
can be reduced to an equivalent unconstrained problem by eliminating the equality
constraints, after which descent methods can be applied to solve it. The other way
is to solve the dual problem which is unconstrained, and then recover the solution of
the original problem from the dual solution. Inequality constrained minimization prob-
lems can be solved by interior-point methods [64]. Of course, we can also optimization
toolbox solvers to solve the problems, such as CVX, MIDACO, etc.
2.12 Simulation Process

Cellular network simulations can be divided into two categories: link-level and system-
level. Link-level simulations are basically a software implementation of one or multi-
ple links between a BS and a UE, with a channel model to reflect the actual transmit
of the waveforms generated. The system-level simulation focuses on the performance
of the entire network to evaluate the impact of resource scheduling, power control,
routing, etc., on the system.
The process of system level simulation can be briefly decried as follows: first,
UEs are randomly distributed in one or more cells, and some basic parameters, such
as transmission power, bandwidth, are set for the BSs and UEs; then, the channel
fading information between all links is calculated according to the UE distribution in
the cell, and the UEs feed back the channel state information (CSI) of the channels
to the BSs, and BSs perform resource allocation and power control according to the
information and the corresponding algorithm. After multiple iterations, the advantages
and disadvantages of the algorithm are summarized according to information such as
system throughput, latency, fairness, and time complexity.
2.12. SIMULATION PROCESS 57
This thesis carries out system-level simulations using Matlab. The simulation is
mainly divided into three parts: network initialization, algorithm simulation and out-
put statistics. The initialization part is to set the coordinates of the UEs and BSs in
the network, and perform channel initialization and power initialization for the UEs,
and calculate the CSI information among all UEs in the whole system according to
the node coordinates and fading. The algorithm simulation part mainly verifies the
performance of the algorithm in the network. The output statistics part mainly calcu-
lates the throughput, latency and energy efficiency of the whole system, the capacity
between the links, etc., and the average of the results of multiple iterations is averaged
and visualized to figures or charts.
Chapter 3
Capacity Enhancement Decoupled

Access
3.1 Introduction
This chapter investigates the compatibility of mmWave communication and multi-
connectivity in DUDe. Increasing communication capacity has always been a major
goal from 1G to 5G. However, in 5G era, insufficient traditional spectrum resources
has become an bottleneck of wireless communications. In such a situation, mmWave
frequency spectrum is envisaged to offer an essential lifeline for future broadband cel-
lular systems in the quest to satisfy the anticipated explosive traffic demands. How-
ever, due to the directional transmission of mmWave, where the multi-user interference
may fall precipitously, mmWave systems are noise-limited [29, 41]. Nonetheless, the
smaller wavelength of mmWave signals enables proportionally greater antenna gain
for the same physical antenna array [29]. That said, the propagation characteristics,
such as high near-field path-loss, severe oxygen and water molecules absorption, pose
great challenges. Fortunately, recent research shows that these problems can be over-
come using highly directional antennas and beamforming techniques [30, 31], but as
mmWaves are sensitive to blockage, mmWave BSs alone can not achieve universal
coverage, using mmWave BSs with ultra-high frequency (UHF)1 BSs is an appealing
approach. UHF frequency bands can provide seamless coverage and high mobility,
while mmWave bands can support ultra-high data rates. A recent study in [66] showed
that to satisfy acceptable exposure limits at frequencies above 6 GHz, the maximum
1 UHF refers to the radio frequencies in the range between 300 MHz and 3GHz. Most frequencies
used in mobile 4G network are classified in the UHF band [65].
58
3.1. INTRODUCTION 59
transmission power in the UL should be several dBs below the power levels used in
current cellular technologies. Although the SNR can be maintained by mounting an-
tenna arrays on the BS, it is also possible to achieve this by shortening the distance
between BSs and UEs to reduce the path-loss. Nonetheless, mmWave signals tend to
experience lower SNRs, as the mmWave bandwidth is much wider than UHF band-
width and the transmission power per Hertz for mmWave is much lower than that of
the UHF signals. In this case, DUDe can be applied to shorten the BS-UE distance to
improve the UL SNR.
Since its inception, much research has been done on DUDe. For instance, the
authors in [2] used real data from Vodafone’s LTE network to evaluate the potential
throughput gain of DUDe. This technique was found to achieve 100%-200% improve-
ment in the 5th percentile UL throughput, and more so in the 50th percentile throughput
in a dense HetNet deployment. The analytical SINR coverage and joint UL-DL rate
of DUDe are obtained in [67]. The authors of [68] derived the association probability
based on stochastic geometry, and showed that as the density of SCells increases, a
large number of UEs choose to receive from an MCell in the DL, and transmit to an
SCell in the UL. The UL performance improvement brought by DUDe was investigated
in [69]. Their results show that DUDe can improve load balance, and is particularly
beneficial to ultra-dense networks. A number of studies have examined the application
of DUDe in mmWave-UHF hybrid networks including, e.g. [70], in which the authors
have derived the SINR and rate distributions from path-loss based and capacity-based
cell association perspectives. Also, in [71], O. W. Bhatti et al. analyzed the perfor-
mance of DUDe in such networks using real blockage data. However, those studies
are based on minimum path-loss (Min-PL) cell association scheme. According to this
association criteria, users may prefer UHF BSs to mmWave SBSs due to the high path
loss of mmWave, although their spectrum resources are more abundant.
Another way to improve the network data rate is allowing a user to be simulta-
neously served by multiple BSs [10, 11]. In the case of dual connectivity (DC) for
instance, each UE is allowed to simultaneously utilize the spectrum from two BSs
connected together via non-ideal backhaul links. In such scenarios, the information
can be either split or transmitted twice. This network architecture with functional-
ity separation is estimated to save more than one-third of the current overall network
power consumption [43]. However, as the transmission power of UEs is limited, adopt-
ing this in the UL could be counterproductive unless combined with DUDe to make it
more power efficient. [72] introduced DUDe to dual connectivity, but to the best of our
60 CHAPTER 3. CAPACITY ENHANCEMENT DECOUPLED ACCESS
knowledge, DUDe and dual connectivity in UHF/mmWave hybrid network have not
been considered before.
3.1.1 Main Contributions

This chapter investigates the feasibility of mmWave with DUDe in HetNets. In partic-
ular, we focus on decoupling the UL and DL serving BSs from the perspective of max-
imizing the total network capacity. To this end, we propose an efficient joint resource-
management and cell-association technique for hybrid HetNets, and compare its per-
formance for a wide range of metrics with benchmark single and dual connectivity
alternatives. Specifically, we investigate the performance of single and dual connec-
tivity techniques in both path-loss based and capacity-based approaches in three-tier
UHF-mmWave hybrid network, where mmWave SCells, UHF SCells and MCells co-
exist. For comparison, the special case where all SBSs are of the same type is also
considered. To evaluate the potentials of capacity-based cell association scheme, we
formulate the problem of joint cell association and network capacity maximization
as a mixed-integer nonlinear programming problem, which serves as an upper-bound
benchmark of the network capacity. Besides, the complexity and energy efficiency
analyses of the proposed technique are also provided.
3.1.2 Organization
The rest of this chapter is organized as follows. In Section 3.2, the assumptions con-
sidered in the system model are explained. In Section 3.3, the proposed joint resource-
management and cell-association technique with decoupling capability is presented.
In Section 3.4, the performance of path-loss based and capacity-based cell association
schemes are evaluated in the single and dual connectivity scenarios. Finally, Section
3.5 concludes the chapter.
3.2 System Model
3.2.1 Heterogeneous network model

We consider an orthogonal frequency-division multiple access (OFDMA) HetNet with
UHF MCells, UHF SCells, mmWave SCells and UEs following independent homoge-
neous Poisson Point Processes (HPPP) Φm , Φs1 , Φs2 and Φu with intensities λm , λs1 ,
3.2. SYSTEM MODEL 61
λs2 and λu , respectively. The ratio of the SCell density to the MCell density is denoted
by β, while the ratio of UHF SCells to all SCells is denoted by γ. The user traffic is
based on the full-buffer model.
In 2010, Jeffery G. Andrews et al. first modeled cellular networks based on stochas-
tic geometric theory and proposed an HPPP model [73]. It used a HPPP to model
the distribution of nodes in a cellular network and simulated the coverage and average
achievable rate of the cellular network.
3.2.2 Cell Association model

We consider macro base stations (MBSs) transmitting over UHF bands, and small base
stations (SBSs) transmitting over UHF or mmWave bands. It is also assumed that the
UEs can connect to any type of BSs and are capable of operating at UHF and mmWave
bands. Also, each UE can connect to one BS or two BSs over different sub-carriers. In
the DL, the UEs connect to the best one or two BSs according to the biased reference
signal received power (RSRP) criterion [25, 26]. If the UL and DL are coupled, the
UEs connect to the same BSs in both link directions; otherwise, the UL BSs will be
separately selected using the path-loss based association in [2], or the capacity-based
association proposed in this chapter. It is worth emphasizing that the UE can connect
up to 2 BSs on each link direction simultaneously (i.e. 4 BSs in total). While it is pos-
sible to allow each UE to be associated with more than two BSs for each link direction,
it was found that the resulting improvement in sum-rate is negligible. This is because
the potential capacity gain for the UEs connecting to more than two BSs is more often
outweighed by the amount of interference caused to adjacent BSs; ultimately leading
to an overall capacity reduction. Furthermore, complexity significantly increases, as
the more BSs a UE is connected to, the more overhead is generated, adding to the side-
information to be processed.
The core idea of capacity-based association is to adaptively switch between coupling

and decoupling the UL and DL on the basis of maximizing capacity at each link; simi-
larly for when adapting between single and dual connectivity. This association scheme
will be further explained in Chapter 4, while in this section, we will briefly describe
the Min-PL based association scheme.
In the Min-PL criterion, the UEs are connected to the BS with the lowest path-loss
[2, 72]. A typical UE is associated with a BS at x∗ ∈ Φl in the UL if and only if
−1
PuWl Ll −1 ≥ PuWk Lmin,k , ∀k, l ∈ {s1 , s2 , m}, (3.1)
where Pu is the transmission power of UE u, W is the UL cell bias value which is

positive referring to expanding the coverage of the cells, Ll is the UL path-loss of the
typical UE to the BS l, and Lmin,k = minx∈Φk Lk (x) is the minimum UL path-loss of the
typical UE to the kth tier BS.
3.2.3 Uplink and Downlink Interference Model
The system bandwidth is equally divided into N orthogonal resource blocks (RBs),
where each RB is divided into 12 resource elements (REs). Each RE refers to a sub-
carrier. The subcarrier spacing is 15 kHz for UHF and 60 kHz for mmWave. UEs in the
same cell occupy orthogonal RBs, so there is no intra-cell interference, but UEs in dif-
ferent cells interfere with each other. Moreover, there is no interference between UHF
and mmWave cells as their frequency bands are orthogonal. The DL SINR between
UE k and its serving BS i at RB n is given by
−1
Γi,n ~i,k,n Li,k,n
ηi,k,n = , j 6= i, (3.2)
η0 + ∑Bj=1 Γ j,n ~ j,k,n L−1
j,k,n
where Γi,n is the transmit signal power of the BS i at the RB n, ~i,k,n is Rayleigh fading
channel for UHF signals and Rician fading channel for mmWave signals. The K-factor
is set to be 7dB for LOS and 6dB for non-LOS (NLOS) [74]. The number of paths is
10. The adopted multipath channel model is the same as [75]. Li,k,n is the path-loss,
∑Bj=1 Γ j,n ~ j,k,n L−1
j,k,n is the interference power (for j 6= i), B is the number of BSs and
η0 is the additive white Gaussian noise (AWGN). On the other hand, the UL SINR is
given by
−1
Γk,n ~k,i,n Lk,i,n
ηk,i,n = −1
, l 6= k, (3.3)
η0 + ∑K
l=1 Γl,n ~l,i,n Ll,i,n
where Γk,n is the transmit signal power of UE k at RB n , K is the number of UEs and
Γl,n is the transmission power of other UEs.
3.2.4 Resource Allocation

Resource allocation is performed using the aperiodic channel quality indicator (CQI)
[76] reported from the UEs. In LTE and LTE-A systems, the CQI is often used for
packet scheduling (PS). Specifically, PS refers to selecting the scheduling time and
frequency for each UE. UEs report the CQI value for each RB to their serving BS. The
BSs utilize these CQI reports to select the preferred RBs for each UE. The CQI report
from a given UE includes information regarding the SINR for each physical RB from
the received pilot power and total interference every measurement period.
3.2.5 Propagation Model

The path-loss L(d) is modeled as

4π f
L(d) = 20 log + 10θ log(d) + χ, (3.4)
c
where d is the distance between the UEs and BSs, f is the operating frequency, c is the
light speed, θ is the path-loss exponent, and χ is the zero-mean log normal shadowing.
Line-of-sight (LOS) mmWave, none-line-of-sight (NLOS) mmWave and UHF links
have different values of α and χ. We adopt the same blockage model as in [77]. If the
distances between the users and mmWave SCells are less than the threshold µ = 200
m, then these links are assumed LOS with probability ω = 0.2, otherwise, these links
are assumed NLOS. The parameters µ and ω are environment-dependent.
It is assumed that the UHF BSs are equipped with omni-directional antennas, the
mmWave BSs are equipped with directional antenna arrays to compensate for the high
path-loss, and the UEs are equipped with omni-directional antennas [70, 77]. The an-
tenna gain at a mmWave BS is formulated by
(
GM , |θ| ≤ θb /2,
Gb (θ) = (3.5)
Gm , otherwise,
where θb = 10◦ is the beamwidth of the main lobe, GM = 18dBi and Gm = −2dBi
denote the gains of main lobe and side lobe, respectively. The mmWave UEs are
assumed to be in perfect alignment with their serving cells, while the beam directions
of interfering mmWave links are independently and uniformly distributed in [−π, π].
Moreover, the antenna gain of an interfering mmWave link is GM with a probability of
pM = θb /2π, and is GM with a probability of pm = 1 − pM .
3.2.6 Energy Efficiency Model

As DUDe will not affect DL cell association, and the BSs are generally assumed to
have an abundance of power supply, we only consider the energy efficiency (EE) in the
UL [78], which is defined as
∑u∈Φu τu
EE = , (3.6)
∑u∈Φu Pu
where τu is the data rate of UE u, Pu is the UL transmission power of UE u. Pu

(expressed in dBm) can be written as [79, 80]
Pu = min{Pmax , 10 log10 M + αL̃ + P0 + ∆mcs }, (3.7)
where Pmax is the maximum transmission power of UEs, P0 is the power baseline
value reflecting the noise level in the UL, ∆mcs is a parameter which depends on the
modulation and coding scheme chosen, M is the number of RBs assigned to the UE,
α ∈ {0, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1} is a compensation factor for L̃, which is the total
UL signal loss, including path-loss, shadowing, fast fading, etc.
3.3 Capacity-Based Cell Association

In this section we describe the proposed capacity maximization approach. Due to the
random nature of the user positions, some BSs may occasionally be highly loaded
while others are underutilized. Moreover, the high path-loss of mmWaves makes them
often less favorable to UEs than UHF BSs even though they offer much more band-
width. To increase the participation of mmWave SBSs, association schemes that max-
imize frequency reuse and capacity should be sought. To this end, we extend the
joint network capacity and QoS Maximization (NCQM) method presented in our pre-
vious work for conventional HetNets [81] to the case of hybrid-networks with DUDe.
Specifically, we refine the criteria and adapt it to encompass the new dimensions of
the considered network including dual connectivity and mmWave-BSs. The presented
scheme can be divided into two parts. The coupling/decoupling association (CDA)
part determines when to decouple the UL and DL, while the single/multi-BS associa-
tion (SMBA) part determines when to select dual connectivity over single connectivity
3.3. CAPACITY-BASED CELL ASSOCIATION 65
per user per link. Using this scheme, some UEs are offloaded to the BSs which can
provide more RBs rather than those with lower path-loss.
3.3.1 Coupling/Decoupling Association
In the UL, UEs are initially connected to the same BSs as those in the DL. It is worth
noting that one UE can connect to one or two BSs. Thus a association and interference
matrix Θ ∈ RB×K can be constructed as

 2, UE k is connected to BS i,

Θi,k = 1, if Pi,k > η0 + σ, (3.8)

0, otherwise,

where Pi,k is the UL interference power, σ is the interference threshold and η0 is the
noise power, Θ(i, k) represents the relation between UE k and BS i, as follows: a) UE
k is served by BS i (Θ(i, k) = 2); b)UE k interferes with BS i (Θ(i, k) = 1); and c) UE
k do not interfere with BS i (Θ(i, k) = 0). Since there are N RBs overall, the average
number of RBs that BS i can provide to each UE is given by

N
Ωi = , (3.9)
∑K
k=1 {Θ i,k = 1} + ∑K
k=1 {Θi,k = 2}
If UE k can get Ωi and Ω j RBs from its serving BS i and a neighboring BS j separately,
then a candidate pair matrix Λ ∈ RB×K can be constructed as
(
1, Ω j > Ωi , i 6= j
Λ j,k = (3.10)
0, otherwise.
If Λ( j, k) = 1, then UE k and BS j is a candidate link. With this in mind, the achievable

capacity for each Λ j,k is obtained as
C j,k = ∑ W log2 (1 + γk, j,n ), (3.11)

n∈Ω j,k
where W is the bandwidth of each RB, Ω j,k is the index set of RBs that BS can pro-
vide to UE k. If C j,k > Ci,k , calculate the updated network capacity Ctot ( j, k) with the
candidate association pair ( j, k), the way to calculate the network capacity is
B K
Ctot = ∑ ∑ ∑ W log2 (1 + γk,i,n )}. (3.12)
i=1 k=1 n∈Ωi,k
The candidate pair with the highest overall network capacity will replace the original
association pair (i, k). This algorithm is guaranteed to converge when it cannot find any
more pairings that would improve the overall capacity. The steps above are formulated
in a pseudo-code structure in Algorithm 1.
As the capacity for each candidate pair and the overall network capacity need to be
calculated many times, the time complexity of this association scheme can be higher
than that of Min-PL. According to [82], the time-complexity of the Shannon capac-
ity formula is unknown, therefore we assume it to be O(1). Assuming the number of
candidate pairs and deployed UEs are given by Ka and Kb respectively, and the num-
ber of candidate pairs satisfying C j,k > Ci,k is Kc , then the computational complexity
of this scheme is expressed as (Kc + 1)Kb O(1) + Ka O(1) = (Kc Kb + Kb + Ka )O(1).
Parameters Ka and Kc increase with the number of BSs and UEs.
Algorithm 1 Coupling/Decoupling Association (CDA)
0
1: Initialization: calculate Θ, Ωi , Ω j and current network data rate Ctot .
0
2: Set Ctot = Ctot ;
3: for k = 1 until K do
4: ω = (0, 0);
5: for i = 1 until B do
6: if Θi,k = 2 then
7: Calculate Ci,k ;
8: end if
9: for j = 1 until B do
10: if Θ j,k = 1 & Ω j > Ωi then
11: Λ j,k = 1 and calculate C j,k ;
12: if C j,k > Ci,k then
13: Calculate the updated network data rate Ctot ( j, k) with the candidate
association pair ( j, k);
14: if Ctot ( j, k) > Ctot then
15: Ctot = Ctot ( j, k);
16: ω = ( j, k);
Figure 3.1: Illustration of SMBA association scheme.
17: else
18: Reject the candidate association pair ( j, k);
19: end if
20: end if
21: end if
22: end for
23: end for
24: if ω 6= (0, 0) then
25: Replace the original association pair (i, k) with ω;
26: Recalculate Θ, Ωi , Ω j ;
0
28: end if
29: end for
3.3.2 Single/Multi-BS Association

The association scheme in this part is different from the dual connectivity association
scheme in [72] in the sense that UEs can connect to one more BS if higher capacity can
be achieved. Only a small portion of the UEs may be connected to two BSs, and thus
UEs can connect in a single or dual hybrid connectivity. Assuming UE u is connected
to BS j, and BS f is its neighbouring BS, then UE u may also connect to BS f if
conditions C1, C2 and C3 are all satisfied, which are stated as follows.
C1: The additional RBs from BS f do not overlap with the existing RBs provided
by BS j as follows
δ f ,u + δ j,u < N, (3.13)
where δ f ,u and δ f ,u are the numbers of RBs from BSs f and j separately. If BSs j and
f are in different frequency bands, i.e., UHF and mmWave, there is no need to check
this condition.
C2: The second connection does not interfere with any other BS (BS i, j and f in
Fig. 3.1) that covers UE u within its area, as follows
φ + δ f ,u + δ j,u < N, (3.14)
where φ is the number of RBs occupied by BS i or j. Similar to C1, BSs i, f , and j

may be in different frequency bands, and UEs transmit over different bands will not
interfere with each other.
C3: The second connection does not interfere with any UE (UE k in Fig. 3.1)
interfered by BS f , as follows
φ f + δ f ,u + λk < N, (3.15)
where φ f and λk are the numbers of RBs already occupied by BS f and UE k, sep-
arately. The achieved capacity for each candidate pair satisfying all the conditions is
calculated according to (3.11) and the priority of selection is given to the candidate
pair with the highest capacity. This process repeats until the overall capacity in (3.12)
cannot be increased any more, so this algorithm is also convergent. These steps are
formulated in a pseudo code structure in Algorithm 2. Assuming the number of can-
didate pairs that satisfy all the conditions is Ks , then the computational complexity can
be expressed as (Ks + 1)Kb O(1), where the parameter Ks increases with the number of
BSs and UEs.
The DUDe technique and our proposed schemes have some impact on the network
architecture and signaling overhead. For instance, a central unit is necessary to collect
the information from all BSs in the group (A group consists of one, or several, MBSs
and some SBSs.) and calculate the parameters needed in our algorithm, such as how
many UEs are within the coverage area of each BS in the UL, the number of allocated
resources per UE, and the path-loss of each link, so that the central unit can control the
BS handover. Besides, the UL-related control signaling needs to be transmitted from

the DL node, while the DL-related control signaling from the UE needs to be received
by the UL node, and forwarded to the DL node over the network infrastructure [1].
To construct the association and interference matrix Θ, each BS needs to send its local
association and interference map Θi to the central unit. If we assume there are B BSs
in the coverage of the central unit, BS i has Ki users, and the number of bits needed to
encode Θi,k is D, then the total number of bits required for all the BSs to forward Θi to
the central unit is ∑Bi=1 Ki D.
Algorithm 2 Single/Multi-BS Association (SMBA)

0
1: Initialization: calculate Θ, Ωi , Ω j and current network data rate Ctot .
0
3: for k = 1 until K do
4: ω = (0, 0);
5: for i = 1 until B do
6: for j = 1 until B do
7: if Θ j,k = 1 & C1 & C2 & C3 then
8: Calculate the updated network data rate Ctot ( j, k) with the candidate as-
sociation pair ( j, k) added;
9: if Ctot ( j, k) > Ctot then
10: Ctot = Ctot ( j, k);
11: ω = ( j, k);
12: else
13: Reject the candidate association pair ( j, k);
14: end if
15: end if
16: end for
17: end for
18: if ω 6= (0, 0) then
19: Add the original association pair ω;
20: Recalculate Θ, Ωi , Ω j ;
0
22: end if
23: end for
3.3.3 Joint Cell Association and Network Capacity Maximization
The capacity-based association schemes (CDA&SMBA) do not necessarily guaran-

tee global optimal network capacity. In order to quantify the upper-bound of the UL
network sum-rate, the problem of joint cell association and network capacity maxi-
mization (J-CA-NC-MAX) is formulated as a mixed integer nonlinear programming
problem. Particularly, the aim is to associate the K UEs with the B BSs, such that each
UE is associated with at least (most) ρ (ψ) BSs in the UL. Also, the sum of transmission
powers of the UEs associated with each BS must satisfy Pmax . The J-CA-NC-MAX
problem is formulated as follows.
J-CA-NC-MAX
B K
max Ctot = ∑∑ ∑ W log2 (1 + γk, j,n ), (3.16a)
j=1 k=1 n∈β j,k
K
s.t. ∑ I j,k Pj,k ≤ P, ∀ j ∈ {1, 2, . . . , B}, (3.16b)
k=1
B
∑ I j,k ≤ ψ, ∀k ∈ {1, 2, . . . , K}, (3.16c)
j=1
B
∑ I j,k ≥ ρ, ∀k ∈ {1, 2, . . . , K}, (3.16d)
j=1
0 ≤ Pj,k ≤ P, ∀k ∈ {1, 2, . . . , K}, ∀ j ∈ {1, 2, . . . , B}, (3.16e)
I j,k ∈ {0, 1}, ∀k ∈ {1, 2, . . . , K}, ∀ j ∈ {1, 2, . . . , B}, (3.16f)
where β j,k is the index of RBs occupied by the association pair (j, k), while I j,k is a
binary decision variable, defined as

1, if UE k is associated with BS j,
I j,k = (3.17)
0, otherwise.
In this work, it is assumed that each UE is associated with at least one BS and at most
two BSs (i.e. ρ = 1 and ψ = 2). Lastly, it should be noted that the J-CA-NC-MAX
problem is non-convex, and thus is computationally-intensive and time-consuming
even for moderate size of networks (i.e. not practical). However, it is considered in this
work as an upper-bound benchmark for the network capacity. The J-CA-NC-MAX is
3.4. PERFORMANCE ANALYSIS 71
solved via MIDACO, with tolerance set to 0.001 [83].
3.4 Performance Analysis

In this section, the performance of the proposed schemes in both single and dual con-
nectivity scenarios are evaluated and compared with the state-of-the-art alternatives.
The 5G NR frame structure supports both TDD and FDD transmissions, but for the
mmWave band, it only supports TDD at present [57]. Thus, TDD is applied in this
chapter. We assume half of the time is used for the UL and we do not consider the
overhead due to switching between transmission directions. We know that the network
traffic in the DL tends to be much larger than that in the UL, but we believe that with
the rise of new services and applications of 5G and IoT, the UL transmission is ex-
pected to increase [1]. Besides, DUDe increases the UL data rate without impacting
the DL data rate; thus, only the UL data rate is considered in this chapter. In the dual
connectivity scenario, we consider at most two uplink associations. For the Min-PL
scheme, UEs can connect to the first and second best serving BSs with the lowest path-
loss. In the case of the SMBA scheme, UEs can connect to one more BS if higher
capacity can be achieved from the decoupling. The list of all the association schemes
are shown in Table 3.1, while the simulation parameters are listed in Table 3.2. To
analyze the improvements, data rate and load situations in the MCells and SCells have
been selected as metrics. Monte Carlo simulation is utilized to evaluate the different
association and connectivity schemes.
3.4.1 Decoupling Access in Three-Tier HetNet

We consider three-tier UHF-mmWave hybrid networks in this section, where mmWave
SCells, UHF SCells and MCells coexist. Fig. 3.2 illustrates the average number of
UEs per cell in the UL for the different association schemes. The figure reveals that
the decoupled association schemes offload more UEs to underutilized SCells in the
Table 3.1: Cell Association Schemes

Conectivity Coupled association Decoupled association
Coupled Min-PL, CDA, SMBA,
Single
(biased RSRP) CDA&SMBA
Coupled-Dual
Dual Min-PL-Dual, CDA-Dual
(biased RSRP)
Table 3.2: Simulation Parameters

Parameters UE MCell UHF SCell MmWave SCell
Maximum transmission power 23 dBm 46 dBm 30 dBm 30 dBm
Downlink/Uplink bias N/A 0/0 dB 3/0 dB 5/0 dB
Spatial distribution HPPP
Spatial density 250/km2 5/km2 γβλm (1 − γ)βλm
Lognormal shadowing N/A µ = 0, σ = µ = 0, σ = LOS µ = 0, σ = 8.66 dB,
4 dB 4 dB NLOS µ = 0, σ = 9.02 dB
[29]
Path-loss exponent N/A 3 3 LOS 2.55, NLOS 5.76 [29]
Operating frequency N/A 2 GHz 2 GHz 28 GHz
Bandwidth N/A 20 MHz 20 MHz 200 MHz [84]
Subcarrier spacing N/A 15 kHz 15 kHz 60 kHz [84]
OFDM symbol period N/A 66.67 us 66.67 us 16.67 us [85]
Power control Transmit at the maximum power level
Noise density -174 dbm/Hz
Average number of UEs per cell in the UL
60
40
20
Min-PL-Dual
0 CDA-Dual
Coupled-Dual
MBS Min-PL
CDA&SMBA
UHF-SBS SMBA
mmWave-SBS CDA
Coupled
Figure 3.2: Average number of UEs per cell in the UL for different association schemes
in single and dual connectivity scenarios (β = 2, γ = 0.5).
UL direction, and thus potentially freeing more MCell resources for users out of SCell
coverage. Additionally, although the numbers of mmWave and UHF SCells are equal,
UHF MCell
Fraction of UEs associated with different cells
1.2
UHF SCell
mmWave SCell
2 3 4 5
1
0.8
0.6
0.4
0.2
0
Coupled Min-PL CDA SMBA CDA&SMBA Coupled-Dual Min-PL-Dual CDA-Dual
Association schemes
Figure 3.3: Ratio of UEs associated with different BSs at different ratios of SCell
density to MCell density and association schemes (γ = 0.5).
it is obvious that more UEs connect to UHF SCells. This is because the path-loss of
mmWave signals is higher, and there may be no LOS link between the BSs and UEs.
Even if the directional antenna array is applied, as the mmWave bandwidth is much
wider than UHF bandwidth, the transmission power per Hertz for mmWave is much
lower than UHF, and thus it will also decrease the SNR. Besides, compared with the
capacity-based (CDA, SMBA) association schemes, the receive signal power based
(Min-PL) association scheme provides higher probability for the UEs to connect to the
small BSs (SBSs) in both single and dual connectivity scenarios. The reason is that in
the CDA and SMBA schemes, UEs initially connect to the same BSs as those in the
DL.
Fig. 3.3 shows the fraction of UEs associated with MCells, UHF SCells and mmWave
SCells in the UL direction at various SCell to MCell densities. The numbers 2, 3, 4, 5
on the top of bars refer to the different ratios of SCells to MCells density. In general,
decoupling schemes make use of a bigger proportion of SCells, which in turn results in
an increase of SCell UEs. For the coupled association scheme and CDA scheme, the
ratio of association pairs with MBSs in dual connectivity is lower than that in single
connectivity. This is because the number of MBSs is much lower than that of SBSs
0.6
Min-PL
CDA
0.5 SMBA
CDA&SMBA
Min-PL-Dual
Fraction of decoupled UEs
CDA-Dual
0.4
0.3
0.2
0.1
0
1 1.5 2 2.5 3 3.5 4 4.5 5
Ratio of SCell to MCell density
Figure 3.4: Ratio of UEs decoupled in the UL and DL vs. ratio of SCell density to
MCell density (γ = 0.5).
and most UEs are associated with MCells in the DL, in the case when there is only
one MBS nearby, then the second association pair must be with an SBS. Accordingly,
the ratio of association pairs with UHF SBSs in dual connectivity is higher than that
in single connectivity, but the ratio of association pairs with the mmWave SBSs in
the dual connectivity scenario is almost the same as that in the single connectivity
scenario, which is because of the severity of the near-field path-loss of mmWaves. In
other words, the UEs prefer UHF SBSs to mmWave SBSs when both are available.
Fig. 3.4 shows the DUDe ratio at various SCell to MCell density ratios. When the
SCell density increases, there will be more overlapped regions, which provides more
flexibility for users to be handed-off to less loaded cells, and hence the number of
DUDe UEs increases. Moreover, the DUDe ratio becomes lower when each UE con-
nects to two BSs. For CDA and SMBA, as their initial association pairs are the same
as those in the UL, their DUDe ratios tend to be lower. In contrast, the UL associ-
ation pairs based on Min-PL are independent and experience a higher DUDe rate in
single and dual connectivity scenarios. It also explains why based on Min-PL more
UEs connect to SBSs in Fig. 3.3.
10 6
4
CDA&SMBA
3.5 SMBA
CDA
Min-PL
3 Coupled
Data Rate per UE (bps)
CDA-Dual
Min-PL-Dual
2.5 Coupled-Dual
1.5
0.5
0
10th 20th 50th
Percentile
(a) 10th, 20th, 50th percentile data rates
10 7
9
CDA&SMBA
8 SMBA
CDA
7 Min-PL
Coupled
CDA-Dual
6 Min-PL-Dual
Coupled-Dual
5
0
50th 80th 90th
Percentile
(b) 50th, 80th, 90th percentile data rates
Figure 3.5: 10th, 20th, 50th, 80th, 90th percentile UE UL data rates based on different
cell association schemes in single and dual connectivity scenarios (β = 5, γ = 0.5).
Figs. 3.5 (a) and (b) show the 10th, 20th, 50th, 80th, 90th percentile UE UL data
rates based on eight cell association schemes. A percentile is a measure used in statis-
tics indicating the value below which a given percentage of observations in a group
of observations fall [86]. For example, the 20th percentile is the value (or score) be-
low which 20% of the observations may be found. Because of the wide bandwidth of
mmWave, a few UEs associated with mmWave SBSs have extremely high data rates,
and so the 80th and 90th percentile data rates are extremely high. The data rate based
on the coupled scheme is the lowest in both single and dual connectivity scenarios.
This can be explained by the fact that higher coverage of SCells in the DUDe cases
results in a better distribution of UEs among the nodes, which provides better utiliza-
tion of the resources. Additionally, connecting to a nearer BS can yield a higher SINR,
and thus achieve higher throughput. Capacity-based cell association schemes (CDA,
SMBA, CDA&SMBA) achieve higher 50th, 80th and 90th percentile data rates than
the Min-PL scheme. It is because SCells serve fewer UEs in this case than with Min-
PL, though these UEs achieve higher data rates at the expense of the 10th and 20th
percentile per UE data rates. Furthermore, for Min-PL, CDA and coupled schemes,
dual-BS association can achieve higher 10th, 20th, 50th, 80th percentile data rates
than single-BS association.
Fig. 3.6 shows the impact of varying the SCell density within the MCell coverage
area on the UL network sum-rate. The densities of UEs and MCells remain constant
across the different cases. As the number of deployed SCells increases, the network
sum-rate grows. This is because adding more SCells means more UEs are offloaded
from the MCell to the SCells where they are granted more resources compared to
the resources offered by the congested MCells. Additionally, decoupled association
schemes provide better performance in comparison with the coupled scheme. Further-
more, capacity-based association schemes (CDA, SMBA, CDA&SMBA) have higher
network sum-rate than Min-PL which do not consider cell loads. It is worth noting
that dual connectivity does not necessarily improve the sum-rate, especially when the
number of BSs is small. An explanation for this could be that increasing the asso-
ciation pairs may increase interference and decrease the available RBs for the users
near the BSs. Determining whether to increase the number of connections from the
perspective of capacity gain seems more reasonable. Furthermore, CDA and Min-PL
have a tangible effect in the single connectivity scenario, but for dual connectivity,
its effect is limited. This may be because the frequency resource has been utilized
10 8
25
Coupled
Min-PL
CDA
20 SMBA
CDA&SMBA
Network sum-rate (bps)
Coupled-Dual
Min-PL-Dual
15 CDA-Dual
J-CA-NC-MAX
10
0
1 1.5 2 2.5 3 3.5 4 4.5 5
Figure 3.6: Network sum-rate in the UL vs. ratio of SCell density to MCell density in
single and dual connectivity scenarios (γ = 0.5).
efficiently in dual and hybrid connectivity. On the other hand, it is evident that the
J-CA-NC-MAX scheme serves as an upper-bound to the network sum-rate due to the
following reasons. First, the sum-rates of CDA and SMBA are dependent upon the
initial association pairs, which are chosen to be the same as those in the UL, but in
fact, they can also be set by the Min-PL scheme or randomly. In turn, different ini-
tial association pairs can lead to different final sum-rates. Second, the higher network
sum-rate achieved by the J-CA-NC-MAX scheme is at the expense of some UEs’ data
rates being lower, which impairs the fairness of the network UEs. However, for the
CDA and SMBA schemes, the data rate of each UE will not be lower than the coupled
association scheme.
Fig. 3.7 shows the UL network sum-rate when the UHF SCells make up different pro-
portions of all the SCells. It can be seen that the different dual-BS association schemes
follow a similar trend. Specifically, the sum-rate increases first and then decreases
under different ratios of UHF SCell to total SCell densities. As might be anticipated,
wireless networks with both UHF and mmWave SCells have better coverage than those
with mmWave SCells. However, for the other association schemes, the network sum-
rate decreases when the number of mmWave BSs decreases. This is because of the
10 9
4.5
Coupled
4 Min-PL
CDA
SMBA
3.5
CDA&SMBA
Coupled-Dual
3 Min-PL-Dual
CDA-Dual
2.5 J-CA-NC-MAX
1.5
0.5
0
0 0.2 0.4 0.6 0.8 1
Ratio of UHF SCell to SCell density
Figure 3.7: Network sum-rate in the UL vs. ratio of UHF SCell density to SCell density
in single and dual connectivity scenarios (β = 5).
decrease in the bandwidth available. Moreover, CDA can achieve higher sum-rate in
the single and dual hybrid connectivity scenario than in the dual connectivity scenario,
and the advantage of decoupled association is less effective in the dual connectivity.
This is due to the fact that the UEs distribution among the BSs is more balanced in the
dual connectivity scenario.
Fig. 3.8 shows the UL energy efficiency when the UHF SCells make up different pro-
portions of all the SCells. It can be seen that for the same association scheme, single
connectivity is more energy efficient than dual connectivity, because the path-loss be-
tween a UE and its second serving BS is relatively higher. Moreover, in most cases,
decoupled association schemes have higher energy efficiency than the coupled asso-
ciation scheme, since DUDe shortens the distance between UEs and BSs. The CDA
scheme reconciles the RB utilization and interference, and thus, it has higher energy
efficiency than the Min-PL scheme in the single and dual connectivity scenarios. For
single connectivity and single-dual hybrid connectivity, the energy efficiency increases
with the number of mmWave SCells. It is because there is almost no interference
among mmWave UEs and when the density of mmWave SBSs is high, the directional
0.9
Normalized energy efficiency 0.8
0.7
0.6
0.5
0.4 Coupled
Min-PL
0.3 CDA
SMBA
0.2 CDA&SMBA
Coupled-Dual
0.1 Min-PL-Dual
CDA-Dual
0
0 0.2 0.4 0.6 0.8 1
Ratio of UHF SCell to SCell density
Figure 3.8: Normalized network energy efficiency when the UHF SCells make up
different proportions of all the SCells (β = 5).
antenna can compensate for the high path-loss of mmWave signals, and thus, the Het-
Net with more mmWave SBSs can achieve higher sum-rate. By contrast, the energy
efficiency of the dual connectivity schemes increases first and then decreases. This is
because mmWave signals are sensitive to blockage, and since the distance between a
UE and its LOS mmWave serving BS can be long, deploying more UHF SBSs can
shorten the UE-BS distance, but at the same time aggravate the co-channel interfer-
ence.
Fig. 3.9 shows the number of handovers (switching among the serving BSs) for the
different schemes, when there are 40 UEs, 1 MBS and at most 5 SBSs. The Min-
PL and coupled association schemes can complete the BS association in one iteration.
For CDA and SMBA, however, it takes a number of iterations to converge, though in
the meantime, the UEs can connect to their DL serving BSs in the UL first, and in
each iteration, only one association pair is changed without affecting the other existing
associations. It can be seen that the number of handovers increases with the number of
BSs. The CDA scheme requires more iterations to converge in the single connectivity
scenario than in the dual connectivity scenario because the UEs distribution among the
BSs is more balanced in the dual connectivity scenario.
15
CDA
SMBA
CDA&SMBA
CDA-Dual
Number of handovers
10
0
1 1.5 2 2.5 3 3.5 4 4.5 5
Figure 3.9: Number of handovers in the UL vs. ratio of SCell density to MCell density
in single and dual connectivity scenarios (γ = 0.5).
3.4.2 Decoupling Access in Two-Tier HetNet
In order to promote system integration, we also consider the special case where all
SBSs are of the same type, i.e., γ = 1 or γ = 0. Although many of the results are sim-
ilar to those in the three-tier HetNet, there are some subtle differences. For HetNets
consisting of UHF MCells and mmWave SCells (Fig. 3.10 (a)), the ratio of mmWave
SCell UE in the dual connectivity scenario is higher than that in the single connec-
tivity scenario. Due to the high path-loss and blockage of mmWave, the coverage of
mmWave SCell is smaller than that of UHF MCells in the UL. However, if there is
only one MBS nearby then the second association pair must be with a mmWave SCell.
Additionally, by comparing Fig. 10 (a) with Fig. 10(b), it is evident that the ratio of
mmWave SCell is doubled, but the mmWave SCell UEs increase more than twice. This
also verifies that the UEs prefer UHF BSs to mmWave SBSs when both are available.
Compared to Fig. 3.11 (b), it is obvious that the sum-rate of the HetNets with only
UHF SCells (Fig. 3.11 (a)) is much lower than those with mmWave SCells, and this
is because the mmWave bandwidth is much wider than UHF bandwidth. Moreover,
the sum-rate based on Min-PL, CDA and coupled schemes increases more slowly with
Fraction of UEs associated with mmWave SCells

Coupled
0.9 Min-PL
CDA
0.8 SMBA
CDA&SMBA
0.7 Coupled-Dual
Min-PL-Dual
0.6 CDA-Dual
0.5
0.4
0.3
0.2
0.1
0
2 2.5 3 3.5 4 4.5 5
(a) γ = 0
0.4
Fraction of UEs associated with mmWave SCells
Coupled
0.35 Min-PL
CDA
SMBA
0.3
CDA&SMBA
Coupled-Dual
0.25 Min-PL-Dual
CDA-Dual
0.2
0.15
0.1
0.05
0
2 2.5 3 3.5 4 4.5 5
(b) γ = 0.5
Figure 3.10: Ratio of UEs associated with mmWave SCells in the UL vs. ratio of SCell
density to MCell density with different ratios of mmWave SCells.
10 8
1.8
1.6
1.4
1.2
Coupled
1 Min-PL
CDA
SMBA
CDA&SMBA
0.8 Coupled-Dual
Min-PL-Dual
CDA-Dual
0.6
1 1.5 2 2.5 3 3.5 4 4.5 5
(a) γ = 1
10 8
18
Coupled
16
Min-PL
CDA
14 SMBA
CDA&SMBA
12 Coupled-Dual
Min-PL-Dual
CDA-Dual
10
0
1 1.5 2 2.5 3 3.5 4 4.5 5
(b) γ = 0.5
Figure 3.11: Network data rate in the UL vs. ratio of SCell to MCell density with
different ratios of mmWave SCells.
UHF MCell
Fraction of UEs associated with different cells
1.2
UHF SCell
mmWave SCell
2 3 4 5
1
0.8
0.6
0.4
0.2
0
Coupled Min-PL CDA SMBA CDA&SMBA Coupled-Dual Min-PL-Dual CDA-Dual
Association schemes
Figure 3.12: Ratio of UEs associated with different BSs at different ratios of SCell
density to MCell density and association schemes (less density scenario, γ = 0.5).
the increase in the SCell to MCell density in the dual connectivity scenario than in the
single connectivity scenario, which is different from the result in Fig. 3.11 (b). This
is because increasing the number of BSs transmitting over UHF bands increases not
only resource utilization but also collaterally interference, and the dual connectivity
can make the interference even higher. Moreover, the sum-rate based on the coupled
scheme is higher than the Min-PL scheme in the dual connectivity scenario. This is
because a few UEs near SBSs acquire rich resources and thus achieve a very high data
rate with the coupled scheme.
3.4.3 Reduced Density Scenario

The BS and UE densities are reduced by half in Fig. 3.12, but the number of BSs
and UEs is kept constant. Compared with Fig. 3.3, the number of UEs associated
with the mmWave BSs decreases while the number of UEs associated with UHF BSs
increases. When the BS density decreases, the average distance between UEs and
BSs becomes longer, making it more difficult for UEs to access the mmWave BSs.
Besides, compared with Fig. 3.6, the network sum-rate in Fig. 3.13 dropped due to
the decreased BS density. Furthermore, the network sum-rate for the dual connectivity
10 8
18
16
14
12
10
8
Coupled
Min-PL
6 CDA
SMBA
4 CDA&SMBA
Coupled-Dual
Min-PL-Dual
2 CDA-Dual
0
1 1.5 2 2.5 3 3.5 4 4.5 5
Figure 3.13: Network sum-rate in the UL vs. ratio of SCell density to MCell density
in single and dual connectivity scenarios (less density scenario, γ = 0.5).
association schemes is severely degraded. It is because the distance to the second best
BS becomes longer, thus connecting to a far away BS is not energy efficient and will
decrease the available RBs for the UEs near the BSs. In these circumstances, the sum-
rate of the CDA and SMBA schemes in the dual connectivity scenario is even lower
than those in the single connectivity scenario.
3.5 Summary
This chapter proposed a DUDe based resource allocation and multi-BS association
technique to improve the performance of hybrid HetNets. Both capacity maximiza-
tion and minimum path-loss based approaches were considered using single and dual
connectivity. The results provide an insight into the performance of DUDe combined
with mmWave and dual connectivity techniques for various performance metrics of
interest. It was shown that capacity-based cell association schemes (CDA, SMBA,
CDA&SMBA) can achieve higher data rates than path-loss based cell association.
CDA scheme reconciles the RB utilization and interference, thus has higher energy
3.5. SUMMARY 85
efficiency than the Min-PL scheme in the single and dual connectivity scenarios. More-
over, it has been demonstrated that dual connectivity does not necessarily improve the
sum-rate, especially when the density of UHF BSs is low. Determining whether to
increase the number of connections from the perspective of capacity gain seems more
reasonable. The advantage of decoupled association is less effective in dual connectiv-
ity, because the UEs distribution among the BSs is more balanced in the dual connec-
tivity scenario than in the single connectivity scenario. Finally, the available mmWave
bandwidth is much wider than UHF bandwidth, and so the network sum-rate mostly
depends on the density of mmWave BSs, but UHF SBSs and MBSs are still important
to provide umbrella coverage to guarantee a consistent service.
Chapter 4
Interference Management for D2D

based on DUDe
4.1 Introduction
This chapter analyzes the application of DUDe to D2D-underlay HetNets. Most of the
current research focuses on in-band underlay D2D, in which D2D re-uses the licensed
spectrum; hence, the spectrum resources are simultaneously used by the cellular and
D2D UEs (DUEs). This mode is advantageous due to the reuse of the spectral resource
of the cellular UEs (CUEs). Besides, as the BSs can effectively control the licensed
spectrum usage, in-band D2D can be effective in QoS management. Many studies
on D2D assume homogeneous networks or only consider one MCell [87–89]. This is
the most common scenario for D2D communication and has been extensively studied.
At present, mode selection, resource allocation, and power control are mainly used to
manage the interference in macrocell and D2D hybrid networks.
Mode Selection: the DUE in the coverage of the MCell can select the cellular mode
or the D2D mode for communication. The cellular mode means that data transmission
between UEs is through the MBS, and the D2D mode is direct communication between
UEs without MBS forwarding. The D2D mode can be divided into two basic modes:
underlay and overlay, which means that DUEs share the spectrum with CUEs in an
orthogonal or non-orthogonal way. If DUEs work in underlay mode, it will cause
interference to the CUEs. A large and growing body of literature has investigated the
mode selection for D2D communication.
The paper [90] discussed the mode selection for D2D communication underlay-
ing a cellular network in a single cell scenario, and a mode selection algorithm based
86
on DUE throughput was proposed. By calculating the throughput when DUE uses
cellular, underlay and overlay mode respectively, the mode that maximizes the DUE
throughput is selected for communication. Then, based on the learning from the sin-
gle cell studies, this mode selection algorithm was extended to a multi-cell scenario.
Although this mode selection algorithm can greatly improve system throughput, the
computational complexity is high. Especially when there are a large number of DUEs
in the network, this algorithm becomes impractical.
In [91], authors formulated a stochastic optimization problem that aims to max-
imize the average sum-rate of the network, while satisfying the QoS requirement of
each UE. By solving the problem, they developed an optimal opportunistic subchannel
scheduling algorithm, which performs both subchannel scheduling and transmission
mode selection opportunistically. However, this paper only considered the cellular
mode and the underlay mode, and underlay mode was not studied.
In [92], relay nodes are introduced into the macro cellular network, and a distance-
dependent mode selection algorithm is proposed, which aims at selecting the optimal
transmission mode with the overall maximized capacity, and the QoS of CUEs is sat-
isfied at the cost of nearly the least information for decision-making. It needs merely
a few parameters instead of traversal analysis in [90], which is much more significant
in practical application.
From the papers above, it can be seen that a reasonable mode selection mechanism
can bring great gain to the network. Besides, from a realistic point of view, we should
focus more on the underlay mode, since the underlay mode can only be applied when
free spectrum resource is available. Therefore, when introducing D2D communication
into an MCell-SCell hybrid network, we should put emphasis on the underlay mode.
Resource Allocation: it is an important way to control the interference, espe-
cially when DUEs reuse the CUE resources. There have been a lot of studies on
resource allocation in the MCell-D2D hybrid network. They are mainly based on
game theory [93–96], mixed integer nonlinear programming [97], interference avoid-
ance [98, 99] and fractional frequency reuse [100]. Besides, the distance between the
transmitter and receiver of a D2D pair is often shorter than that of an MBS-MUE pair,
it is likely that the D2D link has a higher channel gain than the MBS-MUE link. There-
fore, when the network sum rate is optimized, the MUE rate could be sacrificed [101].
However, in general, the priority of the MUE QoS is higher than that of the DUE, so
the requirements of the MUE QoS must be satisfied when optimizing the network sum
rate.
88CHAPTER 4. INTERFERENCE MANAGEMENT FOR D2D BASED ON DUDE
Through the analysis to the documents, we find that most of the current litera-
ture pays particular attention to the resource allocation in D2D underlay mode. These
papers aim at different scenarios, adopt different optimization indicators, and utilize
different methods to reduce the interference when DUEs reuse the resources of CUEs.
There have been a lot of studies on interference management in HetNets composed of
MCells, SCells, and D2D pairs.
Power Control: it refers to reducing the interference in the system by adjusting
the transmission power of nodes in the network. At present, there are two categories
of power control schemes in D2D communication: static power control and [102, 103]
,dynamic power control [104–107].
Static power control means that the transmission power of UEs are determined
and fixed the D2D links are established. In [102], the SINR thresholds of MUEs and
the DUEs are predefined, and then the transmission powers of the UEs are calculated
. In [103], a distance-dependent resource allocation algorithm is proposed. In this
algorithm, the transmission power of DUEs is fixed, and MUEs use open loop power
control method to determine the transmission power. Since the static power control
scheme does not have to adjust the UE transmission power according to the network
status, its signalling overhead is relatively small, but because it is not flexible enough,
the network performance cannot be fully improved.
Dynamic power control means that the MBSs or UEs can dynamically adjust their
transmission power according to their channel states [104–107]. In [105], a close
loop power control scheme is proposed, which uses open loop fraction power con-
trol scheme to calculate the initial transmission power of the DUE, compares the SINR
at the receiver with a predefined target SINR, then adjusts the transmission power of
the DUE to satisfy target SINR, but this method cannot guarantee the QoS of the MUE.
In [106], authors solve the DUE power allocation problem through a subgradient-based
algorithm by applying the Lagrangian dual theory. This algorithm can find an optimal
DUE transmission power, but since it pursues DUE rate maximization but only guaran-
tees the minimum rate of the MUE, the fairness between the MUE and the DUE cannot
be guaranteed. In [107], a centralized and a distributed power control algorithms are
proposed. The goal of centralized power control is twofold: ensure that the CUEs have
sufficient coverage probability by limiting the interference created by underlaid D2D
users while scheduling as many D2D links as possible. For the distributed power con-
trol method, the optimal on-off power control strategy is proposed, which maximizes
the sum-rate of the D2D links. Compared with the static power control scheme, the
dynamic power control scheme can better improve the system performance but has
more signalling overhead and higher computational complexity.
However, as HetNets are becoming the mainstream, it is necessary to optimize
the performance of D2D in HetNets. Different from the two-tier MCell-D2D net-
works, D2D-underlay HetNets are composed of MCell, SCell and D2D pairs. Many
of the research so far considers MCells and SCells that use different frequency bands
[108–110]; so there is no interference between these cell types, resulting in a two-
tier network scenario. There are relatively few studies that consider both SUEs and
DUEs reusing the spectrum of MCells, and they can be further divided into two cat-
egories. Some researchers assume the resource of one MUE cannot be reused by
multiple D2D/SUE links simultaneously within the MCell. For example, in [111],
a coalition formation algorithm and a constrained deferred acceptance algorithm were
devised to solve the subchannel allocation problem, where the subchannel of a MUE
can be reused by at most one SUE/DUE. In order to optimize spatial reuse gain and
enable as many as possible D2D terminals to access the network, a few papers have
focused on the scenario where the resource of one MUE may be reused by multiple
D2D/SUE links simultaneously within the MCell, as long as the interference is low
enough. For instance, in [112], the authors devised a dynamic programming approach
for efficiently matching communicating endpoints and assigning spectrum resources.
In [113], regional restrictions on MBSs, SBSs, and D2D pairs were conducted to solve
the resource allocation problem. In [114], mode selection, resource allocation and
power control of DUEs in HetNets were studied, but only considered a relatively sim-
ple scenario with only one MUE, one femto-cell UE, and one D2D pair.
Overall, most of the current research focuses on in-band underlay D2D, in which
D2D re-uses the licensed spectrum; hence, the spectrum resources are simultaneously
used by the cellular and D2D UEs. This mode is advantageous due to the reuse of the
spectral resource of the cellular users (CUEs). Besides, as the BSs can effectively con-
trol the licensed spectrum usage, in-band D2D can be effective in QoS management.
However, in an in-band underlay D2D scenario, interference management is a major
challenge, as there exists not only inter-tier interference among MCell UEs (MUEs),
SCell UEs (SUEs) and D2D UEs (DUEs), but also intra-tier interference among UEs
associated with different BSs sharing the same subchannel. Apart from mode selec-
tion, resource allocation and power control which been mentioned in many published
studies, DUDe is a new promising technique to solve this problem.
DUDe allows the UEs to connect to a geometrically closer SCell in the UL, rather
than the same one in the DL, which shortens the distance between UEs and their serv-
ing BSs in the UL, reduces path-loss, and lowers transmission power and interference
to neighboring BSs. Besides, DUDe pushes more UEs to under-utilized SCells in the
UL, which allows more efficient resource utilization of SCells, and higher data rates.
At the same time, DUDe can facilitate a more benign environment for D2D receivers
by lowering the CUE UL transmission power, which results in less interference, and
enables more D2D transmissions. Although there have been some research on DUDe,
few attention has been paid to the performance of DUDe in D2D scenarios. Specifi-
cally, the authors in [115] focused on how to calculate the total power saved by UEs
and the extra area that can be utilized to enable D2D pairs because of DUDe; however,
neither subchannel allocation nor transmission power optimization have been taken
into consideration. The authors in [116] proposed a fractional frequency reuse scheme
based on DUDe cell association, where inner and outer subregions of MCells are pre-
assigned to different sub-bands, and MCell-edge UEs are associated with SCells in the
UL to mitigate the interference. However, fractional frequency reuse scheme reduces
the available resources in each subregions. Thus, the quality of services is degraded,
especially when the number of UEs at a certain subregion is extremely high, as the BS
may not able to accommodate all the UEs. Furthermore, the UEs adjust their trans-
mission powers via channel inversion power control, which cannot optimize the data
rate.
This chapter fills a gap in the literature on DUDe-based D2D communications in Het-
Nets. In this work, we consider a HetNet scenario which is closer to reality, where
the resource of one MUE could be reused by multiple D2D/SUE links simultaneously
within the MCell as long as the interference is kept sufficiently low. In contrast to other
research considering UEs connecting to the same BS in the UL and DL, we investigate
the performance of DUDe in D2D-underlay HetNets. In turn, a joint cell-association,
subchannel allocation, and power control (J-CA-SCA-PC) problem for UL network
sum-rate maximization is formulated, subject to maximum transmission power and
minimum rate constraints. However, J-CA-SCA-PC happens to be non-convex and
NP-hard, and thus is computationally-expensive. Hence, this work decouples problem
J-CA-SCA-PC, and solves it via a low-complexity and near-optimal scheme. Specifi-
cally, the main contributions of this work can be summarized as follows:
• Formulated a joint cell-association, subchannel allocation and power control

problem for network sum-rate maximization in D2D-underlay HetNets, subject
to maximum transmission power and minimum rate constraints.
• Proposed a DUDe cell association scheme for D2D underlay HetNets. Partic-
ularly, as we consider the scenario where the resource of one MUE could be
reused by multiple D2D/SUE links simultaneously, it is necessary to manage
the interference carefully. In turn, UE clustering and subchannel allocation is
achieved via a greedy coloring scheme and a modified Munkres algorithm of
low-order polynomial time-complexity [117]. Furthermore, a difference of con-
vex (D.C.) functions based power allocation algorithm is devised with proven
complexity and convergence to maximize the network sum-rate, while satisfy-
ing the UE transmission power and data rate constraints.
• Compared the proposed scheme to various coupled/decoupled cell-association,

subchannel allocation, and power control schemes. In addition, it should be
noted that decoupling cell-association, subchannel allocation, and power con-
trol does not necessarily guarantee global optimality, in comparison to the for-
mulated J-CA-SCA-PC problem. Thus, the proposed scheme is also compared
with the upper bound, obtained by solving problem J-CA-SCA-PA via a global
optimization package. Numerical results show that the proposed scheme effi-
ciently yields near-optimal network sum-rate in comparison to the J-CA-SCA-
PA scheme, while posing a reasonable tradeoff between complexity and opti-
mality. Moreover, it has been shown to achieve about 20% higher network sum-
rate, 30% higher CUE sum-rate, and 15% higher DUE sum-rate than its coupled
counterparts.
As the outset, it may seem rather intuitive that DUDe can improve the performance
of D2D communications compared to the coupled schemes. However, to the best of
our knowledge, no prior work has devised a low-complexity but near-optimal joint
decoupled cell-association, subchannel allocation, and power control scheme for sum-
rate maximization in D2D-underlay HetNets, or performed extensive comparisons with
state-of-the-art and upper-bound benchmark schemes.
4.1.2 Organization
The rest of this chapter is organized as follows. The system model is presented in
Section 4.2, while the proposed scheme is presented in Section 4.3. The performance
of our algorithm and state-of-art alternatives are evaluated in Section 4.4. The conclu-
sions are drawn in Section 4.5.
4.2 System Model

We consider an orthogonal frequency-division multiple access (OFDMA)-based Het-
Net composed of a MCell, S SCells, C CUEs and D D2D pairs. The positions of SCells,
CUEs and D2D transmitters follow uniform distribution with densities of λs , λc and
λd , respectively. Assuming all UEs have low mobility, the channel state information
(CSI) between UEs and BSs, as well as between D2D pairs remains stationary [118]1 .
The UEs’ traffic is based on the full-buffer model. Furthermore, we assume in-band
underlay D2D communication, where the D2D pairs and SUEs share N UL subchan-
nels with the MUEs, and each MUE’s resource can be reused by multiple CUEs and
DUEs2 . Let IC = {1, 2, ...,C}, ID = {C + 1,C + 2, ...,C + D} and N = {1, 2, ..., N} in-
dicate the index sets of CUEs, DUEs and subchannels, respectively. For convenience,
let I = {1, 2, ...,C,C + 1,C + 2, ...,C + D} denote the index set of all UEs (i.e. CUEs
and DUEs). The MBSs and SBSs are all closed-access, meaning that each cell can
only schedule the UEs that belong to it. We assume that frequency division duplexing
(FDD) is applied in the HetNet, and all BSs and D2D pairs transmit over ultra-high fre-
quency (UHF) bands [1]. A UE within the coverage of both an MBS and an SBS could
be served by either a MBS or a SBS. Lastly, we consider unicast data transmissions,
where UEs can transmit data to only one endpoint at a time.
4.2.1 Transmission Data Rate

As DUDe only increases the UL data rate, we only consider UL transmission here,
while accounting for interference between all CUEs and DUEs, as shown in Fig. 4.1.
1A CUE may be associated with a MBS (i.e. MUE) or a SBS (i.e. SUE). Moreover, UEs is used to
collectively refer to CUEs and DUEs.
2 We assume each UE occupies one subchannel, but it can be extended to multiple subchannels by
assuming that virtual UEs exist at the positions of actual UEs. If a UE is not assigned any subchannel,
then its communication is blocked.
Figure 4.1: Interference between MUEs, SUEs and DUEs
Particularly, UEs in different colors are assigned different subchannels; only UEs oc-
cupying the same color cause interference to each other. Now, let M = {0, 1, . . . , M}
denote the set of BSs, where BS m = 0 corresponds to the MBS, and the rest (i.e.
m = 1, . . . , M) are SBSs. For convenience, let χi,n (for i ∈ IC ) be a binary decision
variable, defined as

1, if CUE i is associated with BS m,
χi,m = (4.1)
0, otherwise,
where it should be noted that each CUE i ∈ IC can be associated with one BS (i.e.
∑m∈M χi,m = 1). Also, let ρi,n (for i ∈ IC ) and ρk,n (for k ∈ ID ) be defined as

1, if subchannel n is allocated to CUE i,
ρi,n = (4.2)
0, otherwise,
and 
1, if subchannel n is allocated to DUE k,
ρk,n = (4.3)
0, otherwise,
respectively. We assume that each CUE i ∈ IC and DUE k ∈ ID can be allocated to one
subchannel n ∈ N . That is, ∑n∈N ρi,n = 1, and ∑n∈N ρk,n = 1, ∀i ∈ IC , and ∀k ∈ ID ,
respectively. Thus, the data rate of CUE i ∈ IC is

!
ρi,n pi hi,m,n
RCi = B log2 1 + ∑ ∑ χi,m , (4.4)
m∈M n∈N
Ii,m,n + σ2
where B is the subchannel bandwidth. Moreover, σ2 = N0 B is the variance of the ad-

ditive white Gaussian noise (AWGN), N0 is the noise spectral density, and pi is the
transmission power of CUE i. Also, hi,m,n is the fading channel gain between trans-
mitter i and receiver m on subchannel n, including path-loss, where Rayleigh fading
channels are assumed in this work. The interference Ii,m,n received at BS m on sub-
channel n for CUE i is
Ii,m,n = ∑ ρ j,n p j h j,m,n + ∑ ρk,n pk hk,m,n, (4.5)

j∈IC , j6=i k∈ID
Similarly, the transmission data rate of DUE k ∈ ID on subchannel n ∈ N is written as

!
ρk,n pk hk,k,n
RD
k = B log2 1+ ∑ , (4.6)
n∈N
Ik,n + σ2
where hk,k,n is the channel gain between the D2D transmitter k and its receiver k on
subchannel n. The interference received at D2D receiver k on subchannel n is given by
Ik,n = ∑ ρi,n pihi,k,n + ∑ ρl,n pl hl,k,n . (4.7)

i∈IC l∈ID ,l6=k
4.2.2 Path-loss and Energy-Efficiency Models
The path-loss L(d) is modeled as

4π f
L(d) = 20 log + 10θ log(d) + ψ, (4.8)
c
where f is the operating frequency, c is the speed of light, θ is the path-loss exponent,
d is the distance between the transmitter and receiver, and ψ is the zero-mean log
normal shadowing. It is assumed that the UHF BSs are equipped with omni-directional
antennas.
4.3. RESOURCE ALLOCATION IN D2D-UNDERLAY HETNETS WITH DUDE95
The network energy-efficiency (EE) in the UL is determined as [78]
∑i∈IC RCi + ∑k∈ID RD

k
EE = . (4.9)
∑i∈IC pi + ∑k∈ID pk
4.3 Resource Allocation in D2D-Underlay HetNets with

DUDe
4.3.1 Problem Formulation

In this work, our focus is to maximize the network sum-rate with constraints on the
minimum data rate and maximum transmission power per UE. Therefore, the joint
cell-association, subchannel allocation, and power control (J-CA-SCA-PC) problem
can be formulated as3
J-CA-SCA-PC: (4.10)
max R , ∑ RCi + ∑ RDk (4.10a)

χ ,ρ
ρ, p i∈IC k∈ID
s.t. RCi ≥ RCmin , ∀i ∈ IC , (4.10b)

RD D
k ≥ Rmin , ∀k ∈ ID , (4.10c)
0 ≤ pi ≤ pCmax , ∀i ∈ IC , (4.10d)
0 ≤ pk ≤ pD
max , ∀k ∈ ID , (4.10e)
∑ χi,m = 1, ∀i ∈ IC , (4.10f)
m∈M
∑ ρi,n = 1, ∀i ∈ IC , (4.10g)
n∈N
∑ ρk,n = 1, ∀k ∈ ID , (4.10h)
n∈N
ρi,n , ρk,n ∈ {0, 1}, ∀i ∈ IC , ∀k ∈ ID , ∀n ∈ N , (4.10i)

χi,m ∈ {0, 1}, ∀i ∈ IC , ∀m ∈ M , (4.10j)
where RCmin and RD

min are the minimum data rate values per CUE and DUE, respectively.
Moreover, pCmax and pD max are the maximum transmission power per CUE and DUE,
3 Theuse of “joint” generally refers to the overall problem we are considering, which is solved by
decoupling it into three sub-problems, i.e. cell-association, subchannel allocation, and power control.
respectively. Additionally, χ, ρ and p are the network cell-association, subchannel

allocation, and transmission power matrices of all CUEs and DUEs, respectively. In
problem J-CA-SCA-PC, (4.10a) is the objective function, representing the network
sum-rate. Constraints (4.10b) and (4.10c) correspond to the minimum CUE and DUE
data rates, respectively. Constraints (4.10d) and (4.10e) represent the upper and lower
bounds of the CUE and DUE transmission powers, respectively. Constraint (4.10f)
ensures that each CUE is associated with one BS, while Constraints (4.10g) and (4.10h)
respectively ensure that each CUE and DUE is assigned one subchannel. The last two
constraints define the range of values the binary decision variables take.
Remark 1. Problem J-CA-SCA-PC is a mixed-integer non-linear programming (MINLP)

problem. More importantly, it is non-convex (and NP-hard [112, 119, 120]) due to the
binary decision variables for subchannel allocation, and also the existence of the in-
terference among the CUEs and D2D pairs sharing the same subchannel. A possible
approach is to exhaustively search for every possible BS, subchannel and UEs combi-
nation, and compute the corresponding optimal power allocation for each combina-
tion (which is still non-convex). However, this approach is impractical, and excessively
complex.
Alternatively, we decouple problem J-CA-SCA-PC into three sub-problems: (1)

cell-association, (2) subchannel allocation, and (3) power control. Firstly, we fix the
UL transmission powers, and decide on the serving BS for each UE. After that, we
allocate a suitable subchannel to each UE, while fixing their transmission powers and
serving BSs. Finally, the transmission powers are optimized based on the obtained
cell-association and subchannel allocation to maximize the network sum-rate. This
approach is widely used in the literature, such as [89, 110].
4.3.2 Cell-Association
As for traditional UL and DL coupled cell-association, CUEs connect to the BS with
the maximum biased DL RSRP [25, 26] in the UL and DL. However, we decouple the
UL and DL in this chapter, and allow the CUEs to be associated with a BS according to
the minimum path-loss (Min-PL) criterion in the UL [2]. In the Min-PL criterion, the
CUEs are connected to the BS with the lowest path-loss [2]. Particularly, CUE i ∈ IC
is associated with BS m ∈ M in the UL if
−1
pi,mWm Li,m −1 ≥ pi,m0 Wm0 Li,m0, ∀m, m0 ∈ M , (4.11)
where W is the UL cell bias value, which is positive referring to expanding the coverage
of the cells, Li,m is the UL path-loss between CUE i and BS m. Moreover, pi,m and pi,m0
(expressed in dBm) are UL transmission powers of UE i, which are related to the UE-
BS distances, and set according to the fractional power control (FPC) scheme in [79],
as
pi = min{pmax , 10 log10 M + αL + p0 }, (4.12)
where pmax is the maximum transmission power per UE, and M is the number of sub-
channels assigned to the UE. Moreover, p0 is a target received power at BSs, and
0 < α < 1 is a compensation factor for UL path-loss L. Particularly, α allows the UE
to partially compensate the path-loss, as the higher the path-loss the lower the received
power. In turn, increasing α increases the received power at the cell-edge UEs, which
also increases inter-cell interference [121]. In this chapter, α = 0.6, as it balances the
cell-edge and cell-centre UEs’ data rates.
4.3.3 Resource Block Allocation
Once cell-association is obtained, and in order to allow as many CUEs/DUEs as possi-

ble to access to the network, we assume the SUEs and DUEs can reuse the subchannels
of MUEs, and one subchannel can be reused by multiple UEs simultaneously in differ-
ent cells, which makes it necessary to control inter-tier and co-tier interference4 . Now,
we first assign an unoccupied subchannel to each MUE, and then cluster all the SUEs
and DUEs using the graph coloring theory to control the interference5 . The UEs whose
mutual interference is under a certain threshold are clustered into one group, and the
UEs in the same group are allocated the same subchannel.
We define G = (V, E) as a graph corresponding to the interference in the network,
where V = Vi (for 1 ≤ i ≤ C + D), and E = ei, j (for i 6= j and 1 ≤ i, j ≤ C + D) represent
the vertices and edges of the graph, respectively. Specifically, a vertex represents a UE
requesting a subchannel. An edge connects two different vertices, which means these
two UEs cannot be allocated to the same subchannel due to strong mutual interference.
The criteria for judging whether there is strong interference between UE i and UE j
4 Resource block allocation is based on the aperiodic channel quality indicator (CQI) reported from
the UEs [76].
5 We assume a control node in the network knows the whole interference map.
(i 6= j) on subchannel n are
pi Li,m γm
> , (4.13)
p j L j,m φ
and
p j L j,l γl
> , (4.14)
pi Li,l φ
where i, j ∈ I , the receiver of UE i is BS/UE m, while the receiver of UE j is BS/UE

l. Moreover, Li,m denotes the path-loss between transmitter i and receiver m, while γi
and γ j are the signal-to-interference (SIR) thresholds at the serving BS/UE of UEs i
and j. Also, φ is a parameter proportional to the density of the UEs, which is used to
adjust the SIR thresholds as the potential interference level after subchannel allocation
is proportional to the number of UEs. More specifically, the higher number of UEs in
the network is, the more UEs will reuse a subchannel, and thus, the SIR threshold de-
termining if two UEs can reuse a subchannel should be lower. If either (4.13) or (4.14)
is not satisfied, these two UEs cannot share the same subchannel. We use different
colors to represent different subchannels, and assign a color to each vertex. Notably,
two vertices on the same edge cannot use the same color.
The greedy coloring algorithm is utilized to solve this graph coloring problem,
whose time-complexity is O (|V | + |E|) [122]. Particularly, the vertices are ranked
in descending order by their degrees6 . The colors are represented by the numbers
{0, 1, 2, . . . }, and each vertex is given the color with the smallest number that is not
already used by one of its neighbors. Then, the vertices with more edges incident to
them are served first, and so on. The vertices in the same color represent the UEs in the
same group. Intuitively, the UEs in a group are located far apart, so as to ensure that
the mutual interference of the UEs within each group is kept under a certain threshold.
It should also be noted that the higher the SIR threshold, the more the clusters, and the
fewer the UEs in each group; however, the number of groups cannot exceed that of the
subchannels.
6 Ingraph theory, the degree (or valency) of a vertex of a graph is the number of edges that are
incident to the vertex.
Figure 4.2: Weighted bipartite graph matching.
After clustering the SUEs and DUEs into G groups, the next step is to allocate one
subchannel to each group. Particularly, the subchannel allocation can be modeled as
a weighted bipartite graph matching, as shown in Fig. 4.2, where the two vertex sets
G = {1, 2, ..., G} and N = {1, 2, ..., N} represent the grouped user equipment (GUE)
(i.e. SUEs and/or DUEs) set and the subchannel set, respectively. Moreover, let wg,n
be a weight representing the sum-rate gain when subchannel n is allocated to GUE
g ∈ G , as  
wg,n =  ∑ RCi,n + ∑g RDk,n − Rn, (4.15)

g
i∈IC k∈ID
g g
where IC and ID refer to the CUE and DUE index sets of the gth group, whereas RCi,n
g g
and RDk,n represent the data rates of CUE i ∈ IC and DUE k ∈ ID over subchannel n,
respectively. Moreover, Rn is the data rate on subchannel n when group g has not
occupied it yet, which could be zero.
To solve this problem, the Munkres algorithm is utilized to effciently find the op-
timal one-to-one UE group-subchannel assignment [123]. Particularly, it optimally
performs subchannel assignment for the clustered UE groups, such that the network
sum-rate is maximized, while ensuring low complexity. In turn, a G × N capacity gain
matrix W = [wg,n ] is constructed. However, as the number of groups could be less
than the number of subchannels, the traditional Munkres algorithm for square matrices
is not suitable, so we utilize the modified Munkres algorithm to solve the assignment
problem for rectangular matrices [117, 124]. Specifically, given the G × N cost matrix
0
W = [−wg,n ] of real members, and find a set of k independents k = min(G, N), which
are in different rows and columns, such that the sum of the elements is minimum7 .
7 The Munkres algorithm is initially used to solve the minimal assignment problem, i.e. find
The modified Munkres algorithm is outlined in Algorithm 3, which has a low-order

polynomial run-time with a worst-case complexity of O (N 3 ) [125], and the proof of
convergence of Algorithm 3 can be found in [126].
4.3.4 Power Control
After subchannel allocation, the final step is to maximize the network sum-rate while
satisfying the minimum UE data rate and maximum transmission power constraints. In
OFDMA networks, due to the orthogonality of subchannels, the interference between
two different subchannels can be ignored. Thus, maximizing the network sum-rate is
equivalent to maximizing the sum-rate over each subchannel. This network sum-rate
maximizing power control (PC) problem can be formulated as
PC: (4.16)
max R (p) = ∑ Rn (p) (4.16a)

p n∈N
 
= ∑  ∑n RCi,n (p) + ∑n RDk,n (p)

n∈N i∈IC i∈ID
s.t. RCi,n (p) ≥ RCmin , ∀i ∈ ICn , ∀n ∈ N , (4.16b)

RD D n
k,n (p) ≥ Rmin , ∀k ∈ ID , ∀n ∈ N , (4.16c)
0 ≤ pi ≤ pCmax , ∀i ∈ ICn , (4.16d)
0 ≤ pk ≤ pD
max , ∀k ∈ IDn , (4.16e)
where ICn and IDn are the index sets of the CUEs and DUEs allocated to subchannel n,
respectively. Moreover, RCi,n (p) is the data rate of CUE i on subchannel n, which can
be expressed as
RCi,n (p)
!
pi hi,i,n (4.17)
= B log2 1+ ,
∑ j∈ICn , j6=i p j h j,i,n +∑l∈IDn pl hl,i,n +σ2
a maximum cardinality matching E of minimum cost, where the cost of matching E is given by
c (E ) = ∑e∈E w(e). As for the maximal assignment, the best way is to replace the values wg,n with
−wg,n .
Algorithm 3 Modified Munkres’ Algorithm

0
1: Initialization: W
2: Rotate the matrix if necessary so that the number of columns exceeds the number of rows
and let k = min(G, N);
3: Find the smallest element of each row of the matrix and subtract it from every element in
its row;
4: Find a zero in the resulting matrix;
5: if there is no 0∗ in its row or column then
6: Star it;
7: Repeat for each element in the matrix;
8: end if
9: while (1) do
10: Cover each column containing a 0∗ ;
11: if k columns are covered then
12: break;
13: else
14: while (1) do
15: if there exists an uncovered zero then
16: while (1) do
17: Prime it;
0
18: if (there is no 0∗ in the row containing this 0 ) then
19: Construct a series of alternating primed and starred zeros as follows,
0
20: while there exists a 0 that has 0∗ in its column do
0
21: Let Z0 represent the uncovered 0 found in Step 15, Z1 denote the 0∗ in
0
the column of Z0 (if any), and Z2 denote the 0 in the row of Z1 ;
22: end while
23: Unstar each 0∗ of the series;
0
24: Star each 0 of the series;
25: Erase all primes and uncover every line in the matrix;
26: Go to step 9;
27: else
28: Cover this row and uncover the column containing the 0∗ ;
29: break;
30: end if
31: end while
32: else
33: Add the smallest uncovered value to every element of each covered row, and
subtract it from every element of each uncovered column;
34: break;
35: end if
36: end while
37: end if
38: end while
39: Done: if wg,n is a 0∗ , then the element associated with row g is allocated to that associated
with column n.
where hi,i,n is the channel gain between CUE i and its serving BS i on subchannel
n, while h j,i,n is the channel gain between CUE j ( j not necessarily equals i) and
the serving BS of CUE i on subchannel n. Moreover, the data rate of DUE k over
subchannel n is written as
RD
k,n (p)
!
pk hk,k,n (4.18)
= B log2 1+ .
∑ j∈ICn p j h j,k,n +∑l∈IDn ,l6=k pl hl,k,n +σ2
The power allocation problem on each subchannel is non-convex due to the in-
terference terms in the rate function of each CUE and DUE, and thus is difficult to
optimally solve. However, note that the sum-rate function Rn (p) can be re-written as
Rn (p) = fn (p) − gn (p) , (4.19)
where
 
fn (p) = B ∑ log2  ∑ p j h j,i,n + ∑ pl hl,i,n +σ2

i∈ICn j∈ICn l∈IDn
  (4.20)
+ B ∑ log2  ∑ p j h j,k,n +∑ pl hl,k,n +σ2,
k∈IDn j∈ICn l∈IDn
and
 
gn (p) = B ∑ log2  ∑n p j h j,i,n + ∑ pl hl,i,n +σ2

i∈ICn j∈IC , j6=i l∈IDn
  (4.21)
+B ∑n log2 ∑n p j h j,k,n + ∑ pl hl,k,n +σ2.
i∈ID j∈IC n l∈ID ,l6=k
It can straightforwardly be verified that fn (p) and gn (p) are concave functions in
p [127], and thus the sum-rate function Rn (p) is a D.C. function. As gn (p) is slowly
sensitive to a change in p, it can be well-approximated by its first order Taylor expan-
sion at a fairly large neighborhood of p(`) in each iteration `, as

gn (p) ≈ gn p(`) + ∇gTn p(`) p − p(`) , (4.22)

where ∇gTn p(`) is the gradient of gn p(`) at p(`) , and (·)T denotes the transpose
operator [128]. In turn, the sum-rate function in (4.19) can be approximated as

R̄n (p) ≈ fn (p) − gn p(`) + ∇gTn p(`) p − p(`) , (4.23)
which can be verified to be concave in p, since it is a difference between a concave

function and a linear function [129]. Additionally, note that Constraint (4.16b) can be
written in linear form (∀i ∈ ICn ) as
 
pi hi,i,n ≥ R̄Cmin  ∑n p j h j,i,n + ∑ pl hl,i,n +σ2  , (4.24)

j∈IC , j6=i l∈IDn
C
where R̄Cmin , 2Rmin /B − 1. Similarly, Constraint (4.16c) can be re-expressed (∀k ∈ IDn )
as
 
pk hk,k,n ≥ R̄D
min
 ∑ p j h j,k,n + ∑ pl hl,k,n +σ2  , (4.25)
j∈ICn n
l∈ID ,l6=k
DRmin /B − 1. Based on the above, the D.C. programming-based power con-

and R̄Dmin , 2
trol problem for sum-rate maximization over each subchannel n ∈ N can be expressed
as
DC-PC (n, `)
`): (4.26)

(`) T (`) (`)
max fn (p) − gn p + ∇gn p p−p (4.26a)
p
s.t. (4.24), (4.25), (4.16d), and (4.16e), (4.26b)
which can be verified to be a concave maximization problem with a linear constraints

set. Thus, DC-PC (n, `) can be solved efficiently via any standard optimization pack-
age [129]. Now, to efficiently solve problem DC-PC (n, `) in each iteration `, the
Frank-Wolfe iterative algorithm is employed [130]. Specifically,
n oa feasible solution
p(0) is initialized, and then a sequence of improved solutions p(l) is generated until
convergence, as outlined in Algorithm 4.
Algorithm 4 D.C. Programming-Based Power Control

1: Initialization: Set error tolerance ε ∈ (0, 1), initialize iteration index ` = 0, select a feasi-
ble p(0) , and calculate Rn p(0) .

2: repeat
3: Set ` = ` + 1;
4: Solve problem DC-PC (n, `) to obtain the solution p∗ ;
Set p(l) = p∗ , and calculate R p(l) ;

5:
6: until Rn p(`) − Rn p(`−1) ≤ ε or p(`) − p(`−1) ≤ ε.

7: Output: p? = p(`) .

The computational-complexity of Algorithm 4 is O ICn + |IDn | 3 [128], and
since there is a total of N subchannels, then the overall computational-complexity of
3
solving problem PC is O N ICn + |IDn | . As for convergence, note that the solu-
tion p(`+1) obtained in the (` + 1)th is based on the solution of the previous iteration
(i.e. p(`) ). Also, since DC-PC (n, `) is a concave maximization problem, then the
following inequality holds [128]

fn p(`+1) − gn p(`+1) ≥

fn p(`) − gn p(`) + ∇gTn p(`) p(`+1) − p(`) ≥ (4.27)

fn p(`) − gn p(`) ,

which implies that the sum-rate function Rn p(`) improves at each iteration `. More-
over, since the constraints set is linear, and hence
n convex
o (i.e. closed and compact)
[129], then by Cauchy theorem, the sequence p (`) always converges [131]. This

happens when Rn p(`) − Rn p(`−1) ≤ ε or p(`) − p(`−1) ≤ ε, as there is no im-
provement, for some ε > 0.
Remark 2. It should be noted that decoupling cell-association, subchannel alloca-

tion, and power control does not necessarily guarantee global optimality, in compar-
ison to the formulated J-CA-SCA-PC problem. However, it will shortly be demon-
strated that decoupling cell-association in the DL and UL, while Min-PL for UL cell-
association, graph-coloring and Munkres algorithm for subchannel allocation, and
D.C. programming-based power allocation poses a reasonable tradeoff between com-
plexity and optimality.
4.4. PERFORMANCE EVALUATION 105
Table 4.1: Resource Allocation Schemes

Cell- Subchannel Al-
Scheme Power Control
Association location
Fractional power
C-G-FPC Biased RSRP Greedy
control
Fractional power
D-G-FPC Min-PL Greedy
control
Fractional power
D-R-FPC Min-PL Random
control
Greedy coloring Fractional power
D-GCM-FPC Min-PL
& Munkres control
D.C.
C-G-DCP Biased RSRP Greedy
programming
D.C.
D-G-DCP Min-PL Greedy
programming
D.C.
D-R-DCP Min-PL Random
programming
Greedy coloring D.C. program-
D-GCM-DCP Min-PL
& Munkres ming
C-GCM-DCP Biased RSRP Greedy coloring D.C. program-
& Munkres ming
4.4 Performance Evaluation

In this section, the performance of the proposed (D) decoupled cell-association is eval-
uated and compared to the coupled (C) cell-association scheme for different combina-
tions of subchannel allocation and power control schemes. Particularly, the proposed
subchannel allocation scheme—based on greedy coloring and the Munkres (GCM)
algorithm—is compared with the Random (R), and Greedy (G) subchannel allocation
schemes8 . Moreover, the proposed D.C. programming-based power control (DCP) is
compared with the FPC scheme [79]. The different schemes are listed in Table 4.19 .
8 In the Random scheme, subchannels are randomly allocated to the UEs, while in the Greedy
scheme, the subchannel with the highest SINR is allocated to each UE [132].
9 The initial feasible transmission power values of the DCP scheme are set according to the FPC
scheme. Particularly, the minimum data rate per CUE/DUE is 500 kbps, in alignment with the maximum
transmission power values given in Table II. Specifically, this is the minimum data rate that all the UEs
can achieve according to the FPC scheme. When maximizing the network sum-rate, the data rate of
some UEs may be sacrificed to reduce the interference to other UEs; however, the minimum data rate
constraint is used to guarantee the data rate of each UE will not drop below the minimum requirement.
In practice, and in the case that the channel conditions are severe and/or the available transmission power
is insufficient to satisfy the minimum rate constraints, some CUEs/DUEs may experience an outage.
106
8.5
J-CA-SCA-PC
C-G-FPC
8 D-G-FPC
D-GCM-FPC
D-R-FPC
7.5
C-G-DCP
D-G-DCP
D-GCM-DCP
7 D-R-DCP
6.5
5.5
5
2 2.5 3 3.5 4 4.5 5 5.5 6
Figure 4.3: Network sum-rate vs. ratio of SCell to MCell density.
As per Remark 1, obtaining the global optimal solution of problem J-CA-SCA-

PC may be possible in theory; but may be practically infeasible due to the excessive
computational complexity. Nevertheless, to highlight the near-optimality of our pro-
posed algorithmic designs, we simulate a small HetNet with one MBS, 2-6 SBSs, 3
CUEs and 3 D2D pairs, which are randomly and uniformly deployed in a 450m×450m
square network area. The number of subchannels is assumed to be 2. Clearly, there
is an exponential number of possible combinations for the cell-association, and sub-
channel allocation of the UEs. On top of that, the power control is non-convex for
all the UEs sharing the same subchannel. In turn, problem J-CA-SCA-PC is imple-
mented in a commercial global optimization package, known as MIDACO [83, 133]10 .
As shown in Fig. 4.3, the network sum-rate of the proposed D-GCM-DCP scheme is
very close to that of the J-CA-SCA-PA scheme (as per Remark 2), especially when
the possible BS-subchannel-UE combination number is low. More importantly, our
proposed scheme is superior to the other cell-association, subchannel allocation and
power control schemes.
Remark 3. From this point onwards, and due to the excessive computational-complexity
10 Note that the J-CA-SCA-PC scheme serves as an upper-bound benchmark scheme for network
sum-rate maximization. Moreover, in MIDACO, the tolerance has been set to 10−3 , which implies that
the global optimal solution is accurate up to three decimal places.

Parameters CUE DUE MCell SCell
Minimum data rate per UE 500 kbps 500 kbps N/A N/A
Spatial density 150/km2 150/km2 5/km2 25/km2
D2D maximum distance 30 m
Spatial distribution HPPP
Lognormal shadowing µ = 0, σ = 4 dB
Path-loss exponent 3
Operating frequency 2 GHz
UL bandwidth 5 MHz
Subcarrier spacing 15 kHz
Number of subchannels 25
Noise spectral density -174 dbm/Hz
of the J-CA-SCA-PC scheme, it is omitted from the following simulation results, which
are based on a larger HetNet, with the full simulation parameters as given in Table
4.2. Particularly, the simulation parameters are applicable to both LTE and 5G cellu-
lar networks [57].
Fig. 4.4 shows the relationship between the number of SCells and the network
sum-rate. It can be seen that decoupled association schemes achieve at around 20%
higher sum-rates than the coupled association ones. This is because DUDe short-
ens the distance between the UEs and their serving BSs, which in turn reduces the
path-loss. As FPC is utilized, uplink transmission power and interference are also
reduced accordingly. Similarly, deploying more SCells offloads more UEs from the
MCell to the SCells, which shortens the distance between the UEs and BSs, and
as a result, the network sum-rate increases with the number of SCells. This fig-
ure also shows that our subchannel allocation algorithm—based on greedy coloring
and Munkres’ algorithm—outperforms the Greedy and Random subchannel allocation
schemes. Moreover, the proposed power allocation scheme based on D.C. program-
ming also improves the network sum-rate in comparison to FPC. However, the per-
formance gain of optimizing subchannel allocation and power control is less remark-
able in comparison to the decoupled access. Specifically, the C-GCM-DCP scheme
achieves 7% higher network sum-rate than the C-G-FPC scheme, while the D-GCM-
DCP achieves 20% higher sum-rate. This may be because DUDe has reduced the
interference significantly that there is not much room left for further improvement via
subchannel allocation and power control; although we used a near-optimal subchannel

allocation and power control scheme.
108
1.25
C-G-FPC
D-G-FPC
1.2 D-GCM-FPC
D-R-FPC
C-G-DCP
1.15
D-G-DCP
D-GCM-DCP
D-R-DCP
1.1 C-GCM-DCP
1.05
0.95
0.9
2 2.5 3 3.5 4 4.5 5 5.5 6
Figure 4.4: Network sum-rate vs. ratio of SCell to MCell density.
Fig. 4.5 shows the relationship between the number of SCells and the CUE sum-
rate. The D-GCM-FPC and D-GCM-DCP schemes can achieve around 30% higher
CUE sum-rate than C-G-FPC and C-G-DPC schemes. It is interesting to find that our
power allocation scheme reduces the CUE sum-rate, which lies in the fact that the
distance between a D2D transmitter and receiver is often shorter than that between a
CUE and its serving BS (i.e. the D2D link has higher SINR). Thus, to maximize the
network sum-rate, when the D2D link and CUE link share the same subchannel, more
power is allocated to the D2D link and the CUE link suffers more interference, and
thus, the impact of the D.C. programming-based power allocation on the CUE sum-rate
is negative. However, the distance gap decreases when the number of SCells increases,
and so the impact of the D.C. power allocation scheme also decreases, and the CUE
sum-rate gap of the schemes with and without D.C. power allocation decreases.
Fig. 4.6 shows the 10th , 20th , 50th , 80th , and 90th percentile data rates per UE.
It can be seen that the D-GCM-FPC and D-GCM-DCP schemes can achieve about
twice higher 10th percentile data rate than the C-G-FPC and C-G-DPC schemes. The
D.C. power allocation increases the 80th , and 90th percentile data rates, but decreases
107
5.5
C-G-FPC
D-G-FPC
D-GCM-FPC
5 D-R-FPC
C-G-DCP
D-G-DCP
CUE Sum-Rate (bps)
D-GCM-DCP
4.5 D-R-DCP
3.5
3
2 2.5 3 3.5 4 4.5 5 5.5 6
Figure 4.5: CUE sum-rate vs. ratio of SCell to MCell density.
10 6
3
C-G-FPC
D-G-FPC
2.5 D-GCM-FPC
D-R-FPC
C-G-DCP
2 D-G-DCP
D-GCM-DCP
D-R-DCP
1.5
0.5
0
0 10th 20th 50th 80th 90th
Percentile
Figure 4.6: 10th , 20th , 50th , 80th , and 90th percentile UE data rates based on different
schemes.
107
7.4
C-G-DCP
D-G-DCP
7.2 D-CGM-DCP
R-R-DCP
7
D2D Sum-Rate (bps)
6.8
6.6
6.4
6.2
6
0.025 0.05 0.075 0.1 0.125
Maximum transmission power of DUEs (W)
Figure 4.7: DUE sum-rate vs. maximum transmission power of DUEs.
the 10th and 20th percentile per UE data rates. In other words, it improves the data
rates of the links with high channel quality at the expense of those with low channel
quality. Furthermore, the coupled association schemes have the lowest data rate and
our decoupled GCM algorithms have the highest data rate irrespective of the power
allocation scheme.
Fig. 4.7 shows the relationship between the maximum transmission power of DUEs
and the D2D sum-rate. According to the FPC scheme, as long as the maximum trans-
mission power is sufficiently high, the transmission power of each UE is only affected
by their path-loss, and so we only consider the performance of the schemes with the aid
of power optimization based on D.C. programming. It can be seen that the D2D sum-
rate increases with the maximum transmission power, which means that the sum-rate
improvement mainly depends on the transmission power increase. However, when the
transmission power becomes higher, its effect on the data rate becomes less, because
interference becomes more dominant. The DUEs sum-rate of the proposed scheme is
about 15% higher than the coupled scheme.
Fig. 4.8 shows the relationship between the number of UEs and the normalized
energy-efficiency. Although with the aid of D.C. power allocation, UEs can achieve
higher data rate (as shown in Fig. ??), their energy-efficiency reduces. This is because
0.9
Normalized Energy-Efficiency 0.8
0.7
C-G-FPC
D-G-FPC
0.6
D-GCM-FPC
D-R-FPC
0.5
C-G-DCP
D-G-DCP
0.4
D-GCM-DCP
D-R-DCP
0.3
0.2
0.1
0
40 50 60 70 80
Number of UEs
Figure 4.8: Normalized energy-efficiency vs. number of UEs.
the D.C. power allocation scheme will allocate as much power as possible to UEs, so
as to improve their data rates, but the interference grows with the UE transmission
power, especially when the transmission power and UE density are high. Besides, it
is surprising to find that the energy-efficiency of the schemes with D.C. power opti-
mization does not decrease when the number of UEs increases. A possible explanation
for this is that in high density scenarios, interference bottleneck happens before the
transmission power bottleneck, such that the transmission power of the UEs decreases
with the increase of UEs.
Lastly, and to shed light on the convergence of the proposed scheme, Fig. 4.9 il-
lustrates the average number of iterations of Algorithms 1 and 2. It can be seen that
Algorithm 1 increases slightly with the increase in the number of CUEs and DUEs11 .
This implies that the devised greedy coloring and modified Munkres algorithm is ef-
ficient when assigning the UEs to the subchannels. As for Algorithm 2 (over all 25
subchannels), one can see that with the increase the number of CUEs and DUEs the
average number of iterations increases relatively higher than Algorithm 1. This is
due to the relatively higher complexity of Algorithm 2 in comparison to Algorithm
11 The number of DUEs make up half of the total number of CUEs and DUEs.
160
Algorithm 1
Algorithm 2
140
Average Number of Iterations
120
100
80
60
40
30 35 40 45 50 55 60 65 70
Total Number of CUEs and DUEs
Figure 4.9: Average number of iterations of Algorithms 1 and 2.
3
n n in comparison to O N 3 ). More importantly, the com-

1 (i.e. O N IC + |ID |

plexity over each subchannel is O ICn + |IDn | 3 , which implies that if Algorithm
2 can be executed in parallel over all subchannels, then the complexity can be greatly
reduced. This in turn proves the efficient execution of the proposed scheme to obtain a
near-optimal network sum-rate.
4.5 Summary
A joint DL/UL decoupled cell-association, subchannel allocation and power control
scheme for D2D-underlay HetNets has been proposed, and its performance has been
compared with state-of-the-art alternatives. In the proposed scheme, the Min-PL based
DUDe has been used for cell-association in the UL. Moreover, a greedy coloring and
a modified Munkres algorithm have been employed to solve the subchannel allocation
problem. After that, D.C. programming is utilized to maximize the network sum-rate
while meeting UE maximum transmission power and minimum date rate constraints.
Numerical results show that the DUDe association schemes achieve higher data rate
than their coupled counterparts. The proposed subchannel allocation scheme has been
shown to outperform the Greedy and Random subchannel allocation schemes, and
4.5. SUMMARY 113
power allocation using D.C. programming dramatically improves the network sum-
rate. It was also found that power optimization based on D.C. programming tends to
allocate more power to UEs with higher channel quality at the expense of the UEs with
lower channel quality. However, care must be taken as energy-efficiency may decrease
if the power optimization stage continues to allocate power to the UEs even when the
data rate improvement is marginal. Overall, the results showed that the proposed joint
scheme can achieve about 20% higher network sum-rate, 30% higher CUE sum-rate,
and 15% higher DUE sum-rate than its coupled counterparts.
Chapter 5
DUDe Cellular-Enabled UAV

Communications
5.1 Introduction
This chapter investigates the performances of applying DUDe to cellular-enabled UAVs
communications. Research in cellular-enabled UAV communication is still in its in-
fancy. The authors in [49] designed a cellular-enabled UAV communication system
from the perspective of connectivity-constrained trajectory optimization, where each
UAVs’ starting point and destination are fixed and a target quality-of-connectivity with
the cellular network should be guaranteed. The average secrecy rates of the UAV-
to-ground (U2G) and ground-to-UAV transmissions are maximized in [50] by jointly
optimizing the UAV’s transmission power and the trajectory. Although the formulated
problems are non-convex, the authors proposed iterative algorithms to solve them by
applying successive convex optimization and the block coordinate descent methods.
The authors in [134] proposed a cooperative non-orthogonal multiple access (NOMA)
scheme to mitigate the severe UL interference due to the LOS U2G channels. Specif-
ically, some BSs with better channel conditions are selected to decode the UAV’s sig-
nals first, and then forward the decoded signals to their backhaul-connected BSs for
interference cancellation. The throughput of ground passive receivers is maximized by
cancelling the highest received jamming power at each passive receiver via successive
interference cancellation in [135]. The hovering position for the UAV and the power
allocation scheme to meet the data rate constraints of the users is derived in [136].
Generally speaking, most of the existing research focuses on trajectory design and re-
source allocation to minimize the interference between UAVs and GUEs [137–140].
114
However, they ignored an important character of UAV communications, which is the

UL and DL data rate imbalance. UAV uplinks are dominated by data transmission,
requiring high data rates up to hundreds of Mbps. Contrarily, the UAV downlinks are
dominated by control and non-payload communication (CNPC) communications with
low data rates, which consists of command and control (C&C) data, air traffic con-
trol (ATC) relay data, and sense and avoid (S&A) data [48]. They are crucial to the
UAV operation requiring low latency, ultra-reliability, and high security, but could not
be guaranteed due to the interference from cellular networks [13]. To overcome this
problem, we propose a DL-UL decoupled (DUDe) access scheme, which prevents the
UAVs-GUEs interference via decoupled BS association and frequency band allocation,
and can solve the UL/DL imbalance problem. In this sense, our decoupled scheme can
be viewed as an add-on functionality to any existing interference reduction technique.
Conventional DUDe was originally proposed in [2] to allow UEs to be served by
different BSs on the UL and DL directions in contrast to legacy 1G-4G systems where
the same BS must serve the user on both directions. DUDe not only shortens the
distance between the UEs and serving BSs in the UL—which can improve commu-
nication energy-efficiency (EE)—but also makes better use of the spectral resources
of small BSs. References [69] show that DUDe can improve load balancing, and is
particularly beneficial to ultra-dense networks. Current researches focus on decouple
the UL and DL serving BSs, inspired by this, and in comparison to existing work, we
consider decoupling the DL/UL not only on the basis of BSs but also on the basis of
operating frequencies. By utilizing different frequency bands for LOS links and NLOS
links, as well as CNPC links and data links, the interference between UAVs and GUEs
can be eliminated from the very beginning.

The main contributions of this chapter can be summarised as follows:
• We propose a novel DUDe based access scheme for cellular-enabled UAV com-
munications, in which both dimensions of serving BSs and operating frequencies
are utilised. The proposed DUDe eliminates interference between the UAVs and
GUE, as well as part of the inter-cell interference among GUEs, leading to: (1)
interference-limited GUE communications, and (2) noise-limited UAV and LOS
GUE communications, which is fundamentally different from the the conven-
tional scenario where UAVs and GUEs share the same resources. To the best of
116 CHAPTER 5. DUDE CELLULAR-ENABLED UAV COMMUNICATIONS
our knowledge, the proposed DUDe scheme has not been introduced to cellular-
enabled UAV systems before.
• We devise an optimal low-complexity power allocation algorithm, based on frac-

tional programming and successive convex approximation, to maximise the UL
communication EE in tandem with the proposed DUDe scheme.
• Finally, the performance of the proposed DUDe access scheme and power alloca-
tion algorithm are compared with benchmark schemes in terms of EE, sum-rate
and data rate per UE. It is demonstrated that the proposed DUDe can achieve
several times higher sum-rates in both UL and DL directions than the coupled
benchmark counterparts, while the optimal power allocation algorithm can pro-
vide 15% higher GUE sum-rate and twice higher GUE EE than conventional
fractional power control.
5.1.2 Organization
In the rest of the chapter, Section 5.2 presents the system model. Section 5.3 proposes
the DUDe access scheme for cellular-enabled UAV networks. Section 5.4 utilizes frac-
tional programming to optimize the EE of the proposed scheme. Section 5.5 evaluates
the performance of the proposed scheme, and compares it with coupled benchmarks.
Finally, Section 5.6 presents the conclusions.
5.2 System Model
5.2.1 Network Model

In this work, an OFDMA HetNet composed of MCells, SCells, UAVs and GUEs is
considered. The positions of MCells, SCells and GUEs follow uniform distribution
with densities of λm , λs and λg , respectively. The horizontal locations of the UAVs
also follow uniform distribution with intensity λu , and their altitudes are randomly and
uniformly distributed in the altitude of 50-200 m1 . Lastly, we assume a full-buffer UE
traffic model2
1 It is assumed that the UAVs are independent users, and the resource allocation problem is considered
from the network operator’s perspective. Thus, the positions of the UAVs are assumed to be fixed, and
the trajectory or/and position optimization is assumed to be outside the HetNet control. Nevertheless,
interested readers in scenarios where trajectory optimisation is possible are kindly referred to [138–141].
2 UE is collectively used to refer to a GUE or a UAV.
Table 5.1: Antenna parameters

Frequency Band UHF mmWave mmWave
Number of antenna elements 1 4 16
Half-power beamwidth (degree) θb 360 49.6 24.8
Main-lobe gain (dBi) GM 0 6 12
Side-lobe gain (dBi) Gm 0 -0.8839 -1.1092
5.2.2 Propagation Model

The path-loss L(d) can be modeled as

4πd0 f d
L(d) = 20 log + 10φ log + χ, (5.1)
c d0
where d0 denotes close-in reference distance, f is the operating frequency, c is the

speed of light, d represents UE-BS distance, φ is path-loss exponent, and χ is the log-
normal shadowing.
5.2.3 Antenna Elements

In this work, the BSs and both types of UEs are assumed to support UHF and mmWave
bands. Each BS and UE are assumed to have one UHF omnidirectional antenna. On the
other hand uniform planar square arrays (UPA) with half-wavelength antenna element
spacing are assumed for mmWave transmissions to compensate for the high path-loss.
The antenna parameters are shown in Table 5.1 [142–144]. Moreover, the antenna gain
at the mmWave BS is
(
GM , |θ| ≤ θb /2,
Gb (θ) = (5.2)
Gm , otherwise,
where θb is the mainlobe beamwidth, GM is the mainlobe gain, and Gm is the sidelobe
gain. We assume the mmWave UEs are in perfect alignment [144] with their serving
BSs. The antenna gain of a UE-BS link is the product of the antenna gains at the
receiver and the transmitter.
5.2.4 Blockage Model

For GUE communications, the generalized blockage ball model is considered [77], as
it is widely used in many studies [70, 145]. If the UE-BS distance is less than µ =
200 m, this link is assumed to be LOS with probability ω = 0.2, otherwise, this link
is assumed to be NLOS. For UAV communications, the blockage model in [146] is
applied, where the LOS probability is
1
P(LOS, θ) = , (5.3)
1 + exp(−b(θ − a))
with θ being the elevation angle of the UAV at the BS antenna, and a and b are S-curve
parameters related to the environment.
5.2.5 Resource Block Allocation
Resource block (RB) allocation is performed using the aperiodic channel quality indi-
cator (CQI) [76] reported from the UEs. In LTE and LTE-A systems, the CQI is often
used for packet scheduling (PS). Specifically, PS refers to selecting the scheduling time
and frequency for each UE. UEs report the CQI value for each RB to their serving BS.
The BSs utilize these CQI reports to select the preferred RBs for each UE. The CQI
report from a given UE includes information regarding the SINR for each physical RB
from the received pilot power and total interference every measurement period. The
greedy RB allocation algorithm in [132] is adopted in this chapter. We assume the
RBs of each BS are equally assigned to its UEs, which can be easily extended to the
unequal case. For each UE, all the RBs that are not yet assigned are sorted by SINR
(from the viewpoint of the UE), and those with high SINR are preferentially assigned
to the UE.
5.2.6 Transmission Data Rate
The transmission data rate between transmitter i ∈ I = {1, 2, ..., I} and its receiver
m ∈ M = {0, 1, 2, ..., M} on RB n ∈ N = {0, 1, 2, ..., N} is given by
Pi,n Gi,m |hi,m,n |2

Ri,m,n = B log2 1 + , (5.4)
∑ j∈I , j6=i Pj,n G j,m |h j,m,n |2 + σ2
where B is the RB bandwidth, Gi,m is the antenna gain, σ2 = N0 B is the variance of the
AWGN, N0 is the noise spectral density, Pi,n is the transmission power of transmitter
i on RB n, and hi,m,n is the channel gain between transmitter i and receiver m on RB
n, including path-loss and fading. As we consider both UL and DL in this chapter, a
5.3. DUDE ACCESS IN CELLULAR NETWORKS SERVING UAVS AND GUES119
transmitter can be a UE or BS3 . In this chapter, one transmitter can occupy multiple
RBs. Moreover, the total data rate of transmitter i is the sum of its data rate on each
allocated RB. We assume different fading models for mmWave and UHF communica-
tions; specifically, Nakagami−m fading [147] and Rayleigh fading [148], respectively.
5.3 DUDe Access in Cellular Networks Serving UAVs

and GUEs
Table 5.2: Transmission Parameters of the Different Schemes

Scheme DUDe Coupled (UHF) Coupled (mmWave)
Direction DL UL DL UL DL UL
Frequency mmWave UHF mmWave UHF UHF mmWave
Bandwidth 200 MHz 20 MHz 200 MHz 20 MHz 20 MHz 200 MHz
Time 50% 50% 50% 50% 50% 50%
UE GUE GUE UAV GUE GUE UAV+GUE GUE UAV+GUE
Association Biased RSRP Min-PL Biased RSRP Biased RSRP
Figure 5.1: Cell Association and Interference map of the proposed scheme.
In conventional cellular networks, the UL and DL are coupled, and UEs connect to
the BS with the highest biased reference signal received power (RSRP) [25] in the DL,
and transmit to the same BS in the UL. The operating frequency is also the same in
the two link directions. This is known as coupled uplink/downlink access (CUDA) [1].
In homogeneous networks, the CUDA mode is optimal, because the best serving BS
in the UL and DL are the same. However, with the development of 4G and 5G net-
works, more low-cost SCells are deployed, which makes cellular networks heteroge-
neous. Due to the transmission power gap between SCells and MCells, MCells usually
3 It should be noted that if RB n is not allocated to UE i, then Pi,n is set to zero.
have wider coverage than SCells in the DL, hence more UEs are associated with them.
However, this is not the case in the UL. Most UEs are battery-powered with similar
transmission power, as a result, their coverage is also similar. The CUDA scheme
makes the UL channel quality of the MCell edge UEs terrible. As the operating fre-
quencies keep getting higher, the situation becomes even worse. This is called UL and
DL imbalance [3]. For the sake of reducing the path-loss, a better choice is to allow
MCell edge UEs to connect to a geometrically closer SCell rather than the coupled
MCell in the UL, which is the essence of the DUDe technique [2]. Furthermore, since
cellular GUEs require higher data rate in the DL than in the UL, allowing them to
transmit over high-frequency bands in the DL and low-frequency bands in the UL is
another way to improve the UL coverage while guaranteeing a high throughput in the
DL, which is akin to decoupling in terms of the operating frequency.
Likewise, UAV UL and DL communication also have different requirements as ex-
plained in Section I. Therefore, splitting the UAV data links from the control links is
intuitive. The GCSs operating on the L/C bands can be utilized for CNPC links, which
can provide a wide coverage, and thus avoid frequently switching of serving BSs dur-
ing the flight. There is approximately 17 MHz (960-977 MHz) at the L-band, and 61
MHz (5.03-5.091 GHz) at the C-band presently allocated for UAV CNPC links [149].
Transmitting on a dedicated band can also avoid interference and guarantee the relia-
bility of control links. Meanwhile, cellular BSs can be utilized for data transmission.
Instead of choosing the serving cellular BSs according to the DL biased RSRP, using
a minimum path-loss (min-PL) scheme is a better choice, where UEs are connected to
the BS with the lowest path-loss [2] in the UL. Particularly, UE i ∈ I is associated with
BS m ∈ M in the UL if
−1
pi,mWm Li,m −1 ≥ pi,m0 Wm0 Li,m0, ∀m, m0 ∈ M , (5.5)
where W is the UL cell bias value, which is positive and refers to expanding the cov-
erage of the cells. Moreover, Li,m is the UL path-loss between UE i and BS m. Due
to the high capacity requirement of the UAV UL data transmission as well as the LOS
dominant characteristic, mmWave bands can be utilized for UAV data links. With
the development of directional antennas and beam-forming technologies, the received
signal strength of mmWave communications can be guaranteed and inter-cell interfer-
ence can be significantly reduced. Correspondingly, to reduce the interference between
UAVs and GUEs in the UL, UHF bands are utilized for the GUE UL transmission (The
mmWave band is utilized for the UL-dominant UAV data transmission). Thus, the UL
5.4. ENERGY-EFFICIENCY MAXIMIZATION 121
coverage can also be guaranteed. As for GUE DL transmission, UHF bands are utilized
to provide umbrella coverage. Additionally, in order to improve the network capacity,
mmWave bands are used in the DL LOS links. Besides, the serving BSs of GUEs are
also decoupled, which are decided according to the biased RSRP scheme in the DL
and min-PL scheme in the UL. The whole band allocation and cell association strategy
of the proposed DUDe scheme is given in Table 5.2.
Since the current 5G NR frame structure only supports time division duplexing
(TDD) mode for mmWave transmission [57], for consistency, TDD is utilized for both
UHF and mmWave transmission in this chapter. According to the LTE standard, the
DL/UL configuration is semi-static, but in the actual LTE-TDD deployment, cells op-
erating over the same frequency are set for the same DL/UL configuration. 5G NR
standard supports the dynamic and flexible setting of the DL/UL configuration in a
TDD mode, and adjacent cells at the same frequency can have similar or different
DL/UL configuration. However, for simplicity, the static consistent DL/UL configura-
tion is considered in this chapter so that cross-link interference [150] is avoided. The
cell association and interference map of the proposed scheme is shown in Fig. 5.1,
where the dotted lines refer to interference.
5.4 Energy-Efficiency Maximization

The EE of the GUEs and UAVs under the proposed scheme is optimized in this sec-
tion. As per Fig. 5.1, the intra-cell interference can be ignored since all sub-carriers
are orthogonal to each other in OFDM, but because of frequency reuse, inter-cell in-
terference still exists. In the DL, the GUEs transmitting over the UHF band cause
interference to each other (i.e. an interference-limited scenario). On the other hand,
GUEs transmitting over the mmWave band also suffer from interference, but due to
beamforming and the fact that mmWave signals are sensitive to blockage, the inter-
ference level is negligibly small (i.e. a noise-limited scenario [29, 41]). Moreover,
the larger bandwidth of the mmWave signal necessitates a wider band receiving filter
hence more noise passes through with the desired signal. In such a case, the noise
power becomes more dominant than the interference and thus the SNR can be used
as an approximation for the SINR; hence mmWave communication networks can be
modeled as noise-limited systems [151, 152]. In the UL, all UAVs transmit over the
mmWave band, while all GUEs transmit over the UHF band, and the interference rela-
tionship is quite similar to that in the DL. Due to the DUDe structure, the interference
between UAVs and GUEs is eliminated, which is quite different from the assumptions
in most research. Taking the high mobility of UAVs into consideration, we propose a
practical low complexity power allocation scheme which can optimize EE in the UL
direction, while the DL scenario can be easily analogized. Specifically, we consider
EE maximization at the GUEs (UHF band) and the UAVs (mmWave band) separately,
subject to a minimum data rate and maximum transmission power requirements per
UE.
5.4.1 GUEs Energy-Efficiency Maximization
The transmission data rate of GUE i to its serving BS m over RB n is written as
γG
 
i,n
z }| {
2

P |h
i,n i,m,n |
RG
 
i,m,n (P) = B log2 1 +
 , (5.6)
2 2
∑ Pj,n |h j,m,n | + σ 
 G

j∈I , j6=i
where γG G
i,n is the received SINR at the BS, I is the index set of GUEs over the UHF
band, and P is the power allocation matrix of all GUEs over all RBs. Thus, the total
rate of GUE i in the UL is obtained as
RG
i (P) = ∑ RG
i,m,n (P) . (5.7)
n∈NUHF
Similarly, the total power consumption of GUE i is
PG
i = ∑ Pi,n , (5.8)
n∈NUHF
where NUHF is the index set of UHF RBs. The transmitter is considered to be OFF
when the UE is not allowed to transmit or during periods when the UE is not transmit-
ting a sub-frame [153]. The GUEs energy-efficiency maximization (GUEs-EE-MAX)
problem can be formulated as
GUEs-EE-MAX: (5.9)
∑ G RG (P)
max EEG = i∈I Gi (5.9a)
P ∑i∈I G Pi (P)
s.t. RG G G
i (P) ≥ Rmin , ∀i ∈ I , (5.9b)
PG G
i ≤ Pmax , ∀i ∈ I G , (5.9c)
Pi,n ≥ 0, ∀i ∈ I G , ∀n ∈ NUHF . (5.9d)
where RG G
min and Pmax are the minimum rate and maximum transmission power per
GUE, respectively. In problem GUEs-EE-MAX, (5.9a) defines the EE objective func-
tion as the ratio of the GUEs sum-rate to the total power consumption. Constraint
(5.9b) ensures that the total rate of each GUE satisfies the minimum rate requirement,
while Constraint (5.9c) defines the maximum value of the total transmission power
of each GUE. We assume that all the UEs have the same minimum rate and maxi-
mum transmission power requirements, but it can be easily extended to the scenario
where different UEs have different requirements. Lastly, Constraint (5.9d) ensures the
non-negativity of the transmission power of each GUE over each RB. The feasibility
of problem GUEs-EE-MAX depends on the values of RG G G
min and Pmax , where Rmin is
used to guarantee the data rate of each UE will not drop below the minimum require-
ment, when energy-efficiency is maximized. On the other hand, if RG min is too high,
the power allocation scheme may not satisfy the minimum data rate of each UE, in
which case problem GUEs-EE-MAX becomes infeasible. Thus, in our work, RG min
is set according to the data rate that can be achieved via the fractional power control
(FPC) scheme. The FPC scheme is applied in 4G and 5G cellular networks [79, 154],
where the transmission power of GUE i on RB n is given by
1
Pi,n = min{Pmax , 10 log10 M + wL + P0 }, (5.10)
M
in which Pmax is the maximum transmission power per UE, M is the number of allo-
cated RBs per UE, P0 is the target received power, and w ∈ {0, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1}
is the compensation factor for path-loss L. Based on (5.10), the initial transmission
G . Thus, problem GUEs-EE-MAX is always feasible in our
power will not exceed Pmax
work.
Remark 4. The rate function of each GUE is non-convex due to the interference terms,
and thus the EEG function is not necessarily convex [155]. Moreover, the constraints
set is not convex due to the minimum rate constraint.
To efficiently solve problem GUEs-EE-MAX, consider the following lower-bound

approximation [156]
log2 (1 + γ) ≥ α log2 γ + β, (5.11)
where γ, γ̄ ≥ 0, and the bound is tight for γ = γ̄. Moreover,
γ̄
α= , and β = log2 (1 + γ̄) − α log2 γ̄. (5.12)
1 + γ̄
Thus, the lower-bounded objective function is obtained as in (5.13). Then, by adopting

G
the transformation Qi,n = log2 Pi,n , the lower-bounded EE function EE (P) is rewritten
as in (5.14).

∑i∈I G RG ∑i∈I G ∑n∈NUHF B log2 1 + γG ∑ ∑ B α log γG +β
i (P) i,n i∈I G n∈NUHF i,n 2 i,n i,n G
G
= ≥ , EE (P)
∑i∈I G Pi (P) ∑i∈I G ∑n∈NUHF Pi,n ∑i∈I G ∑n∈NUHF Pi,n
(5.13)

2Qi,n |hi,m,n |2
∑i∈I G ∑n∈NUHF B αi,n log2 Q + βi,n
G ∑ j∈I G , j6=i 2 j,n |h j,m,n |2 +σ2
EE (Q) =
∑i∈I G ∑n∈NUHF 2Qi,n
G
Ri (Q)
z }| !! !{
∑i∈I G ∑ B αi,n Qi,n + log2 |hi,m,n |2 − log2 ∑ 2Q j,n |h j,m,n |2 + σ2 + βi,n
n∈NUHF j∈I G , j6=i
= !
∑i∈I G ∑ 2Qi,n
n∈NUHF
| {z }
G
Pi (Q)
G
R (Q)
, G
P (Q)
(5.14)
G
Remark 5. For each rate function Ri (Q) in (5.14), the negative log-sum-exp term is
G
concave in Q [155], and thus Ri (Q) is concave, ∀i ∈ I G . More importantly, the sum
of the concave functions is also concave [157], and hence, the lower-bounded rate
G
function R (Q) is concave in Q.
G
Remark 6. In (5.14), the total transmission power Pi (Q) of each GUE i ∈ I G is
G
convex in Q, while implies that P (Q) is also convex in Q.
Based on all the above, the transformed GUEs EE maximization problem is ex-
pressed as
T-GUEs-EE-MAX: (5.15)
G
G ∑i∈I G Ri (Q)
max EE = G
(5.15a)
Q ∑i∈I G Pi (Q)
G
s.t. Ri (Q) ≥ RG
min , ∀i ∈ I G , (5.15b)
∑ 2Qi,n ≤ Pmax
G
, ∀i ∈ I G . (5.15c)
n∈NUHF
Remark 7. The constraints set of problem T-GUEs-EE-MAX can be verified to be

convex in Q.
G
Based on Remarks 5, 6 and 7, the transformed lower-bounded EE function EE (Q)
is a ratio of a concave function to a convex function in Q, with a convex constraints
set. Particularly, problem T-GUEs-EE-MAX takes the form of a concave-convex
fractional programming problem [155], and thus can be globally optimally solved via
the Dinkelbach’s algorithm [158]. Particularly, define the auxiliary function
n o
G G
FG (λ) , max R (Q) − λP (Q) , (5.16)
Q
for λ ≥ 0, which is the unique maximizer of FG (λ), for fixed values of αi,n and βi,n ,
∀i ∈ I G and ∀n ∈ NUHF [159]. Hence, problem T-GUEs-EE-MAX is solved via the
algorithm outlined in Algorithm 6.
Algorithm 5 Dinkelbach’s Algorithm

1: Initialization: Set error tolerance ε ∈ (0, 1), iteration index ` = 0, and λ(0) = 0.
2: repeat n o
G G
3: Compute Q(`) = arg max R (Q) − λ(`) P (Q) ;
Q
(`) G G
G = R Q(`) − λ(`) P Q(`) ;

4: Calculate F λ
G
R (Q(`) )
5: Update λ(`+1) = G (`) ;
P (Q )
6: Set ` = ` + 1;
until FG λ(`) ≤ ε.

7:
8: Output: Q̂ , Q? .
Algorithm 6 Solution of Problem T-GUEs-EE-MAX

1: Initialization: Set error tolerance ε ∈ (0, 1), iteration index ` = 0, αi,n = 1, βi,n = 0,
c G P̂(0) .
∀i ∈ I G and ∀n ∈ NUHF , select a feasible P̂(0) , and calculate EE

2: repeat
3: Set ` = ` + 1;
(`) (`)
4: Update αi,n and βi,n by (10), ∀i ∈ I G , and ∀n ∈ NUHF ;
5: Compute Q̂(`) by solving problem T-GUEs-EE-MAX via Algorithm 5;
(`)
6: Set P̂(`) = 2Q̂
c G P̂(`) ;

7: Evaluate EE
c G P̂(`) − EE c G P̂(`−1) ≤ ε.

8: until EE
9: Output: P?
Lemma 1. Algorithm 5 monotonically converges to the global optimal solution Q̂ of

problem T-GUEs-EE-MAX4 .
Proof: Refer to Appendix A.1.

After that, the obtained solution is transformed into its original form as P̂ = 2Q̂ ,
and hence the value of the original objective function in (5.9a) is calculated as
∑i∈I G RG

G
c P̂ = i P̂
EE , (5.17)
∑i∈I G PG
i P̂
which is based on fixed values of αi,n and βi,n . Additionally, Q̂ is obtained from Al-
gorithm 5 based on a lower-bounded objective function. Hence, to obtain the solution
of problem GUEs-EE-MAX, αi,n and βi,n must be repeatedly updated ∀i ∈ I G and
∀n ∈ NUHF , while solving problem T-GUEs-EE-MAX via Algorithm 5, as shown in
Algorithm 6.
Lemma 2. Algorithm 6 converges in a finite number of iterations to the global optimal

solution P? of problem GUEs-EE-MAX.
Proof: Refer to Appendix A.2.
5.4.2 UAVs Energy-Efficiency Maximization

Different from GUE communications, UAV communications can be seen as noise-
limited [29, 41]. Thus, the transmission data rate between UAV i ∈ I A and its serving
4 Algorithm 5 has a linear convergence rate and polynomial-time complexity [160].
BS m on RB n is given by5
Pi,n Gi,m |hi,m,n |2

RAi,m,n (Pi,n ) = B log2 1+ , (5.18)
σ2
where I A is the index set of UAVs transmitting over the mmWave band. In turn, the
total data rate of UAV i in the UL is determined as
RAi (P) = ∑ RAi,m,n (Pi,n ) , (5.19)

n∈NmmW
where NmmW is the index set of mmWave RBs.
Remark 8. The data rate function RAi,n (Pi,n ) is concave in Pi,n , and hence the sum-rate
function RAi is a sum of concave functions, and thus is also concave in P [157].
The total power consumption of UAV i is given by
PAi = ∑ Pi,n . (5.20)

n∈NmmW
Hence, the UAVs energy-efficiency maximization (UAVs-EE-MAX) problem is ex-

pressed as
UAVs-EE-MAX: (5.21)
RA (P)
max EEA = A (5.21a)
P P (P)
s.t. RAi (P) ≥ RAmin , ∀i ∈ I A , (5.21b)
PAi ≤ Pmax
A
, ∀i ∈ I A , (5.21c)
Pi,n ≥ 0, ∀i ∈ I G , ∀n ∈ NmmW . (5.21d)
where RAmin and Pmax

A are the minimum rate and maximum transmission power per UAV,
respectively6 .
5 Recall that all UAVs transmit over the mmWave band, and beamforming is applied, and thus inter-
cell interference can be neglected [29, 41].
6 In a similar manner to problem GUEs-EE-MAX, the feasibility of problem UAVs-EE-MAX can
be guaranteed. In the simulations, this has been achieved by setting the initial transmission power of
each UE as that of the FPC scheme.
Remark 9. The objective function EEA (P) is a ratio of a concave function RA (P) to
a linear function in P.
Based on Remark 9, problem UAVs-EE-MAX takes the form of a concave-linear

fractional programming problem [155]. More importantly, each stationary point of
the objective function EEA (P) is the global maximizer [155]. Therefore, define the
auxiliary function
n o
FA (λ) , max RA (P) − λPA (P) . (5.22)
P
Hence, problem UAVs-EE-MAX can be globally optimally solved via Algorithm 5,
G
but by replacing R (Q), P (Q), and FG (λ) by R (P), P (P), and FA (λ), respectively7 .
5.4.3 Computational Complexity
The computational-complexity of the Dinkelbach-based

algorithm (i.e. Algorithm 5)
with stopping criteria ε is O ε12 log(K) , where K is the number of terminals (i.e. K =
|I G |) [158]. In Algorithm 6, a concave optimization problem is solved in each iteration
via Algorithm 5, and hence Algorithm 6 has polynomial time-complexity, then the
2
1
overall complexity of Algorithm 6 is O ε2
log(|I G |) . Since problem UAVs-EE-

1 A
MAX is solved via Algorithm 5, its computational complexity is O ε2 log(|I |) . For
comparison, Table 5.3 gives the complexities of some other optimal power allocation
schemes in similar scenarios. It is evident that the complexity of our proposed scheme
is relatively lower than the other schemes.
Table 5.3: Complexity of Different Schemes

Paper Complexity
[161] O (MN), where M is the number of UEs, N is the number of time-slots.
[162] O (FKM), where F is the number of resource blocks, K is the number
of vehicle to vehicle links, M is the number of vehicle to UAV links.
[163] O (NMK), where N is the number of smart devices, M is the number of
UAVs and K is the number of subchannels.
7 The proof of Lemma 1 of problem T-GUEs-EE-MAX is tenable to problem UAVs-EE-MAX, and

thus is not provided to avoid unnecessary repetition.

In this section, the performance of the GUEs and UAVs under the proposed DUDe ac-
cess scheme is evaluated and compared with benchmark schemes in terms of energy-
efficiency, sum-rate, and data rate per UE. In the simulations, the total DL and UL
transmission time is assumed to be the same. Considering the current UAV regula-
tions [13], we set the altitude of UAVs to be 50-200 meters8 . As the data rate of the
UAV CNPC links is limited, and will not cause any interference to other links, it is
ignored in the simulations for simplicity. As for benchmarks, we consider both UAVs
and DUEs transmitting over the UHF or mmWave band, with access according to the
CUDA (coupled) scheme. In the UL, fractional power control (FPC) is applied. In the
DL, the transmission power is uniformly distributed over the whole bandwidth. The
transmission parameters of the different schemes are given in Table 5.4.

Parameters GUE UAV MCell SCell
Downlink/Uplink bias N/A N/A 0/0 dB 3/0 dB
Spatial distribution Uniform distribution
Altitude of UAVs 50-200 m
S-curve parameters a = 9.6, b = 0.28 [144, 146]
Blockage ball model parameters µ = 200 m, ω = 0.2 [77]
Spatial density 250 per km2 50 per km2 5 per km2 20 per km2
Path-loss exponent UHF: GUE-UAV 2, GUE-BS 3, d0 = 1 m [148]; mmWave: LOS 2.55, NLOS 5.76, d0 = 5 m [29, 164]
Lognormal shadowing UHF: µ = 0, σ = 4 dB [148]; mmWave: LOS µ = 0, σ = 8.66 dB, NLOS µ = 0, σ = 9.02 dB [29]
Nakagami-m parameters mL = 3, mN = 2 [144, 147]
Operating frequency 2 GHz & 28 28 GHz 2 GHz & 2 GHz & 28
GHz 28 GHz GHz
Bandwidth UHF: 20 MHz; mmWave: 200 MHz [84]
Subcarrier spacing UHF: 15 kHz; mmWave: 60 kHz [84]
Power control FPC with p0 = −85 dBm, and α = 0.8 [55], fixed power consumption ρG = ρA = −50 dBm [153]
Noise spectral density -174 dbm/Hz
5.5.1 Proposed DUDe Access Scheme Simulations

Fig. 5.2 depicts the 10th , 20th , 50th , 80th , and 90th percentile data rate per UE in
the DL and UL. A percentile indicates the value below which a given percentage of
observations falls [86]. For example, the 10th percentile is the value (or score) below
which 10% of the observations may be found. The 10th percentile data rate of mmWave
UEs in the UL is around 1.6 × 108 bps, which means, 10% of mmWave UEs’ UL data
rates are below 1.6 × 108 bps. It can be seen that the decoupled scheme can achieve
several times higher data rates in both the UL and DL than its coupled counterparts.
8 3GPP identified that cellular networks should cater for UAVs flying between ground level and 300
meters [13, 55].
DL mmWave UE Data Rate (bps) UL mmWave UE Data Rate (bps)

8 6
10 10
UL UHF UE Data Rate (bps)

6 6
Decoupled Decoupled
Coupled (mmWave) Coupled (UHF)
4 4
2 2
0 0
10th 20th 50th 80th 90th 10th 20th 50th 80th 90th
Percentile Percentile
10 8 10 6
DL UHF UE Data Rate (bps)

3
Decoupled Decoupled
10
Coupled (mmWave) Coupled (UHF)
2
5
1
0 0
10th 20th 50th 80th 90th 10th 20th 50th 80th 90th
Percentile Percentile
Figure 5.2: 10th , 20th , 50th , 80th , and 90th percentile data rate per mmWave link and
UHF link in the UL and DL.
This is partly because the decoupled scheme occupies wider bandwidth (as shown
in Table 5.2), but it is worth noting that the bandwidth of the decoupled scheme is
only 10% wider than the mmWave coupled scheme. The huge rate gap between the
decoupled and coupled mmWave schemes is due to the DUDe access. According to the
DUDe scheme, the mmWave band is monopolized by UAVs in the UL and LOS GUEs
in the DL. For comparison, according to the mmWave coupled scheme, the mmWave
band is shared by the GUEs and UAVs in the UL, and NLOS and LOS GUEs in the
DL. In this case, the DUDe scheme can make better use of the mmWave band and
provide higher data rates. Besides, DUDe also shortens the UE-BS distance in the UL,
which increases the received signal strength and thus improves the SINR. It can also
be seen that, due to the wider bandwidth, the mmWave band can provide high data
rate for LOS links. Although mmWave signals suffer from high path-loss, it can be
compensated for by beamforming. As such, the interference among the UAVs is very
low and can be seen as noise-limited.
Fig. 5.3 illustrates the impact of the UAV altitude on the UL sum-rate and EE,
where—for instance—50/80 refers to the altitude range 50-80 m. As for the coupled
mmWave and UHF schemes, the sum-rate and EE of UAVs decreases when the alti-
tude increases, which is due to the increase in path-loss. However, for the decoupled
10 9
2.5 1
2 0.8
UL UAV Sum-Rate (bps)
UL UAV Normalized EE
Decoupled
Coupled (UHF)
1.5 0.6
Coupled (mmWave)
Decoupled
1 0.4 Coupled (UHF)
Coupled (mmWave)
0.5 0.2
0 0
50/80 80/110 110/140 140/170 170/200 50/80 80/110 110/140 140/170 170/200
UAV Altitude (m) UAV Altitude (m)
Figure 5.3: UL UAV sum-rate and normalized EE vs. UAV altitude.
scheme, the sum-rate and EE slightly increase at first as the LOS probability increases
with altitude. Although the path-loss also increases, since many UAVs are associ-
ated with nearby SCells (while more UAVs are associated with MCells in the coupled
schemes), and thus the impact of path-loss does not surpass the impact of LOS prob-
ability. It can also be seen that the EE of the coupled UHF scheme is higher than that
of the coupled mmWave scheme and the decoupled mmWave/UHF hybrid scheme.
This is because, in contrast to the coupled benchmark which uses UHF, our decou-
pled scheme uses mmWave frequency for the UAVs, which has higher sensitivity to
path-loss as opposed to UHF frequencies. It is also important highlight that while the
coupled technique provides 10 times better EE in this case (as shown in Fig. 5.3 (b),
our decoupled technique provides over 100 times better sum-rate, as can be seen from
Fig. 5.3 (a). To further elaborate, when comparing the GUE performances, as shown
in Fig. 5, in which case both our decoupled technique and the coupled counterparts
use UHF frequencies, our proposed technique is superior in terms of sum-rate and EE.
Fig. 5.4 shows the relationship between the number of UAVs/GUEs and UL and
DL sum-rates, where 2/10 refers to 2 UAVs and 10 GUEs, and there are 1 MBS and
4 SBSs here. The UL sum-rate of the coupled mmWave scheme is the highest at first,
then it is exceeded by the sum-rate of the decoupled scheme. This is because, for the
decoupled scheme, only the UAVs transmit over the mmWave band in the UL, while
the GUEs transmit over UHF band. Moreover, when the number of BSs is greater than
the number of UAVs, the mmWave resource cannot be not fully utilized; whereas for
the coupled mmWave scheme, both GUEs and UAVs transmit over the mmWave band
10 9 10 9
3 2.5
2.5
2
UL Sum-Rate (bps)
DL Sum-Rate (bps)
2
1.5
1.5
1
Decoupled Decoupled
1 Coupled (UHF) Coupled (UHF)
Coupled (mmWave) Coupled (mmWave)
0.5
0.5
0
2/10 4/20 6/30 8/40 10/50 2/10 4/20 6/30 8/40 10/50
Number of UAVs/GUEs Number of UAVs/GUEs
Figure 5.4: Network sum-rate in the UL and DL vs. number of UAVs/GUEs.
hence can achieve better spectral utilisation. This is corroborated by the results when
the number of UAVs for the decoupled case exceeds the number of base-stations it
becomes superior to the coupled counterpart. As for the DL, although there are more
UEs, most of them are associated with MCells, and some BSs are not activated at first,
and so the DL network sum-rate increases with the increase in the number of UEs.
5.5.2 Energy-Efficiency Maximization

In this subsection, the performance of the proposed optimal (OPT) EE-maximizing
power allocation is compared to that of FPC. Fig. 5.5 compares the UL GUE sum-
rate and normalized EE of the proposed decoupled access scheme to the coupled UHF
scheme, both of which use the UHF band for GUE UL transmission. It can be seen
that the proposed OPT power allocation scheme can achieve around 15% higher GUE
sum-rate and twice higher EE than the FPC scheme. Furthermore, the performance of
the coupled schemes is worse than that of the corresponding decoupled scheme. This
is because in the decoupled schemes, the UHF band is monopolized by GUEs in the
UL, while in the coupled UHF scheme, it is shared by the GUEs and UAVs. Moreover,
DUDe shortens the UE-BS distance, and hence improves the SINR.
Fig. 5.6 is similar to Fig. 5.5 except it compares the UL UAV performance. It
can be seen that the OPT power allocation scheme achieves much higher EE than the
FPC scheme. However, since the UAVs utilize the mmWave band in the decoupled
schemes, the transmission is noise-limited rather than interference-limited, and the
10 7
13 1
Decoupled-FPC Decoupled-FPC
Coupled (UHF)-FPC Coupled (UHF)-FPC
12 Decoupled-FP-OPT
0.9 Decoupled-FP-OPT
Coupled (UHF)-OPT Coupled (UHF)-OPT
11 0.8
UL GUE Sum-Rate (bps)
UL GUE normalized EE
10 0.7
9 0.6
8 0.5
7 0.4
6 0.3
5 0.2
2 3 4 5 2 3 4 5
Ratio of SCell to MCell density Ratio of SCell to MCell density
Figure 5.5: UL GUE sum-rate and normalized EE vs. ratio of SCell density to MCell
density.
9
10
2.5 1
Decoupled-FPC Decoupled-FPC
Coupled (mmWave)-FPC 0.9 Coupled (mmWave)-FPC
Decoupled-FP-OPT Decoupled-FP-OPT
2 Coupled (mmWave)-OPT 0.8 Coupled (mmWave)-OPT
0.7
UL UAV Sum-Rate (bps)
UL UAV normalized EE
1.5 0.6
0.5
1 0.4
0.3
0.5 0.2
0.1
0 0
2 3 4 5 2 3 4 5
Ratio of SCell to MCell density Ratio of SCell to MCell density
Figure 5.6: UL UAV sum-rate and normalized EE vs. ratio of SCell density to MCell
density.
optimal transmission power that maximizes the UAVs’ EE is much lower than that of
the GUEs transmission. The OPT power allocation scheme cannot attain high sum-
rate and EE simultaneously. Nonetheless, similar to Fig. 5, the performance of the
decoupled schemes is superior to the corresponding coupled schemes.
Fig. 5.7 illustrates the 10th , 20th , 50th , 80th , and 90th percentile data rate per GUE
and UAV in the UL, where the ratio of SCell density to MCell density is 4. The
UAV data rate shows a relatively average distribution; however, the GUE data rate
distribution is more uneven. This is because the OPT power allocation scheme tends
to allocate more power to those UEs with higher channel condition and/or lower path-
loss. Due to the minimum data rate constraint, the two power allocation schemes
achieve similar 10th percentile data rate per UAV, while the 90th percentile data rate of
the OPT power allocation scheme is lower than the FPC scheme. This validates the
conclusions in Fig. 5.6 in another way.
UL UAV Data Rate (bps)
10 8
6
Decoupled-FPC
Decoupled-OPT
4
0
10th 20th 50th 80th 90th
Percentile
UL GUE Data Rate (bps)
10 6
6
Decoupled-FPC
Decoupled-OPT
4
0
Percentile
Figure 5.7: 10th , 20th , 50th , 80th , and 90th percentile data rate per UAV and GUE in
the UL.
5.6 Summary
A new access scheme in heterogeneous cellular networks with UAVs has been pro-
posed in this chapter. Particularly, to meet the different requirements of UAV data links
and CNPC links, as well as GUE uplinks and downlinks, we separate the UAV CNPC
links from the high-capacity data communication links, and separate the GUE uplinks
5.6. SUMMARY 135
from the downlinks. All the CNPC signals are connected to the GCSs operating in the
L/C band, while the data signals are communicated to cellular BSs via the mmWave
band. To further reduce interference between the UAVs and GUEs, the GUE down-
links and uplinks are also decoupled in terms of operating frequency and serving BS.
The results demonstrated that the proposed access scheme can achieve several times
higher sum-rates in both UL and DL than its coupled counterparts. Furthermore, an
optimal power allocation algorithm has been devised to optimize the EE of the DUDe
scheme, and shown to achieve 15% higher GUE sum-rate and twice higher GUE EE
than the conventional fractional power control scheme. Finally, it was shown that for
noise-limited UAV communications, there is a trade-off between sum-rate and EE as
simultaneously maximizing both is unattainable.
Chapter 6
Q-Learning and DQL based Power

Control for UAVs
6.1 Introduction
This chapter solve the same EE optimization problem as in Section 5.4, but use re-
inforcement learning methods. Due to the rapid changes in the UAV wireless envi-
ronment, such as air-to-ground channels, spatial and time variations of non-stationary
signal behaviour, and detection of UAVs via UAVs-enabled protocols, conventional
optimization methods with ideal assumptions (e.g. perfect CSI) may not work in prac-
tical and real-time applications. Hence, it is necessary to augment classical algorithms
and solutions with artificial intelligence (AI) and machine learning (ML)-based tech-
niques [165]. Moreover, when both UAVs and GUEs transmission resources are opti-
mized, the network access schemes can be designed to overcome the classical meth-
ods’ excessive overhead and delays, while incorporating ML techniques to achieve an
acceptable/sub-optimal EE solution in rapidly changing wireless environments.
Reinforcement learning (RL) algorithms are among the most promising ML tech-
niques to use in radio resource management (RRM) for UAV-enabled cellular com-
munications [166]. This is due to the nature of RL, which is based on maximizing a
reward function by exploring the action(s) domain—via trial-and-error interactions—
to allow the learner to discover the best choices based on the received rewards [167].
In turn, RL has become a base for resource allocation in wireless networks, due to its
simplicity and ability to provide reliable and efficient learning through interaction with
the network. Q-learning (QL) is a model-free RL approach, which is based on finite
136
states and actions to obtain acceptable/near-optimal solutions with low computational-

complexity [168]. However, in QL, the sizes of the state and action spaces grow ex-
ponentially for each additional unknown network feature and/or parameter, leading
to the curse of dimensionality, especially in the training phase. Alternatively, deep
Q-learning (DQL) has been proposed, which utilizes a deep neural network (DNN),
called deep Q-network (DQN) along with other techniques (e.g. replay memory) to
perform a stable and efficient training, and reliably estimate the Q-function [169]. Par-
ticularly, DQL is based on quickly performing predictions using only a small num-
ber of simple operations to obtain an output, which greatly reduces execution time.
Consequently, deep RL approaches have found numerous applications in cellular net-
works [170–172]. Add to this the 3GPP technical requirements for the enhancement
of UAVs [173], which only proposes AI/ML to control the UAVs, but did not discuss
how AI/ML can be used in the scheduling and resource allocation. This motivates us to
investigate the potentials of AI/ML techniques in cellular-enabled UAVs by applying
QL and DQL as resource allocation tools.
6.1.1 Related Works

Recently, a number of research works have proposed learning-based resource alloca-
tion for cellular-enabled UAV networks [171]. For instance, in [174], an interference
management scheme is proposed with the aim of achieving a tradeoff between max-
imizing EE and minimizing wireless latency and interference to the ground network.
Specifically, a DQL algorithm based on echo state network (ESN) cells is devised to
allow each UAV to map each observation of the network state to an action, and hence
learn its optimal path, transmission power and cell association. The proposed algo-
rithm has been shown to minimize the interference to the GUEs and the transmission
delay of the UAVs. A 3D energy-efficient and fair UAV scheduling scheme based on
deep RL (DRL) is proposed in [175] to allow the UAVs to hover around and serve the
users, and also recharge their batteries. The proposed algorithm has been shown to
outperform existing scheduling algorithms in terms of coverage, energy-efficiency and
fairness. In [176], a novel DRL-based control algorithm is devised for energy-efficient
coverage and connectivity, which is demonstrated to outperform baseline schemes in
terms of coverage, fairness and energy consumption. In [177] and [178], the proposed
QL- and DRL-based methods have been applied to UAV-BSs with energy constraints
to achieve energy-efficiency and coverage fairness to the GUEs, while reducing the
collision incidents and co-channel interference (CCI). To the best of our knowledge,
138CHAPTER 6. Q-LEARNING AND DQL BASED POWER CONTROL FOR UAVS
none of the prior works in the literature have considered decoupled access in cellular-
enabled UAVs with RL-based power control for energy-efficiency maximization.
The main contributions of this chapter can be summarized as follows:
• A novel and simple QL algorithm is proposed for EE-maximizing power control,

while alleviating the excessive computational delays of the classical fractional
programming and successive convex approximation solutions. This algorithm
has outperformed the benchmark schemes in terms of EE.
• A novel DQL algorithm is proposed to optimize the EE and overcome the large
state-action matrix in the QL algorithm. Although the DQL performance is
slightly worse than the QL, it outperforms the conventional fractional power
control (FPC) scheme.
• The performance of the proposed DUDe QL and DQL power control schemes
are compared with state-of-art alternatives in terms of EE, sum-rate and data
rate per UE. It is demonstrated that the proposed DUDe can achieve several
times higher sum-rates and EE than the coupled benchmark counterparts. The
QL (DQL) power control scheme improves EE by around 100% (70%) for the
UHF band, and by around 160% (130%) for the mmWave band, in comparison
to conventional FPC scheme.
6.1.3 Organization
The rest of this chapter is organized as follows. Section 6.2 outlines the QL and DQL
algorithms for EE maximization, while Section 6.3 discusses their implementation.
Section 6.4 evaluates the performance of the proposed QL and DQL power control
schemes, and compares them with several benchmarks. Finally, Section 6.5 draws the
conclusions.
The System Model in this chapter is the same as in Section 5.2, and the EE opti-
mization problem is based on the DUDe access scheme proposed in Section 5.3.
6.2. RL-BASED OPTIMIZATION OF ENERGY-EFFICIENCY 139
6.2 RL-Based Optimization of Energy-Efficiency
6.2.1 Q-Learning (QL)
In this work, the power control is centralized, and the QL agent is assumed to be
located at the MBS, where the learning process is modeled as a Markov decision pro-
cess (MDP). Now, let S be the set of possible transmission power states over the as-
signed RBs, and A (s) be the discrete set of actions in terms of the transmission powers
over the assigned RBs in state s. Assuming discrete
time-steps t resembling the train-
ing rounds, the QL agent takes a(t) ∈ A s(t) based on some policy φ. Particularly,
φ (s, a) represents
the probability of taking action vector a in state s1 . By applying
a(t) ∈ A s(t) and transitioning from state s(t) to s(t+1) , a reward r(t+1) , s(t) , a(t)
is given to characterize the benefit from taking action vector a(t) in state s(t) . The well-
known QL algorithm aims to find the optimal policy φ∗ that maximizes an expected
reward function. Thus, let the future cumulative discounted reward at time-step t be
given by [179]
∞
R (t) = ∑ δτr(t+τ+1), (6.1)
τ=0
where δ ∈ [0, 1) is the discount factor for future rewards. Also, let the Q-function
associated with policy φ as the expected reward when a is taken in state s, as
h i
Qφ (s, a) = E R (t) |s(t) = s, a(t) = a , (6.2)
which satisfies the Bellman optimality equation as [180]
Qφ (s, a) =
!
(6.3)
a
s0 , a Qφ s0 , a
0 0

R (s, a) + δ ∑ Ps,s0 ∑ 0φ ,
s0 ∈S 0
a ∈A (s )
h i
with R (s, a) = E r (t+1) (t) (t)
|s = s, a = a being the expected reward of (s, a) ∈ S × A .

Moreover, Ps,sa = Pr s(t+1) = s0 |s(t) = s, a(t) = a represents the transition probability
0
from state s to state s0 upon applying a. In turn, the optimal Q-function associated with
φ∗ is obtained as
1 It
should be noted that both s and a are of dimension 1 × N, where N is the number of the reused
RBs in the system.
a
s0 , a0 .

Qφ∗ (s, a) = R (s, a) + δ ∑ Ps,s0 max Qφ∗ (6.4)
0
s ∈S a0
The QL agent assigned a Q matrix for each UE in the network, denoted Q(s, a), which
serves as a lookup table for each action-value combination. Moreover, the QL algo-
rithm updates each entry in the Q matrix in each time-step t as

Q s(t) , a(t) ← (1 − η)Q s(t) , a(t)
(6.5)
+ η r(t+1) + δ max Q s(t+1) , a ,
a
where 0 < η ≤ 1 is the learning rate to control the speed of reaching a solution. To
avoid being stuck at non-optimal policies and to deal with the exploitation versus ex-
ploration tradeoff issue [181], the ε-greedy policy is used for each time-step t, which
implies that the QL agent takes action a∗ that maximizes the Q-function with proba-
bility 1 − ε + |A ε(s)| for exploitation, and a random action with probability ε + |A ε(s)| for
exploration [181].
Algorithm 7 Q-Learning
1: Initialization: Q(s, a) with zero values, δ, η, and ε.
2: for each time-step t do
3: For the current state s(t) , pick the action vector a(t) using the ε-greedy policy, as

arg max Q s(t+1) , a , with prob. 1−ε+ A sε(t) ,
a(t) ← a | ( )|
a random action vector, with prob. ε + ε
.
|A (s(t) )|
Perform action a(t) , obtain reward r(t+1) = r s(t) , a(t) and observe the new state s(t+1) .

4:
Update Q s(t) , a(t) as

5:

Q s(t) , a(t) ← (1 − η)Q s(t) , a(t)

+ η r(t+1) + δ max Q s(t+1) , a .
a
6: Set t ← t + 1 and current state s(t) ← s(t+1) .

7: end for
8: Output: State s and action a vectors.
In this work, the reward function r(s, a) takes one of the predefined values v1 >
v2 > v3 , described as

v , (EE ≥ ζmin ) ∩ (Ri ≥ Rmin ) ∩ (Pi ≤ Pmax ),
 1


r(s, a) = v2 , (EE ≥ ζmin ) ∪ (Ri ≥ Rmin ) ∩ (Pi ≤ Pmax ), (6.6)



v , otherwise,
3
where EE is the energy-efficiency value, Ri is the data rate of UE i, and Pi the total
transmission power of that UE over all used RBs, with thresholds ζmin , Rmin , and Pmax ,
respectively. Every time an the action vector a is selected, the MBS calculates the
rewards, and measures how well the action vector contributes to the maximization of
the UEs energy-efficiency, while ensuring the minimum date rate and maximum trans-
mission power constraints are satisfied. The QL algorithm is outlined in Algorithm
7, which summarizes the process of evaluating the Q values, and obtaining the GUEs
and UAVs allocated RBs states, and transmission power action vectors. Each UE in-
dependently learns its own policy, treating other agents as part of the environment, i.e.
independent Q-learning [182].
The QL algorithm is guaranteed to converge when all actions are repeatedly sam-
pled and the rewards are bounded [168, 183]. More importantly, the QL algorithm has
two serious issues: (1) the amount of memory need to store and update the Q(s, a) ma-
trix grows exponentially as the number of states and actions increases, and (2) some
states may rarely be visited, which excessively increases the time needed to explore all
state-action combinations to obtain a good estimate of Q(s, a), which is impractical.
Independent Q-learning makes the environment become non-stationary from the point
of view of each UE. Since it contains other UEs who are themselves learning, the con-
vergence can not be guaranteed. Fortunately, substantial empirical evidence has shown
that independent Q-learning often works well in practice [184].
6.2.2 Deep Q-Learning (DQL)

As for DQL, and as mentioned earlier, a DNN called DQN is utilized to estimate the Q-
function instead of the Q(s, a) matrix in the QL algorithm. In this work, a multi-layer
deep forward neural network is utilized to replace the classical state-action matrix and
find the optimal policy. This is achieved by exploiting correlations in the space of the
input raw data and identifying the important features that distinguish such input [185].
Moreover, an experience-replay mechanism is used to store the reciprocal experience
and randomly pick a group of samples from the stored experience to train the DQL
instead of the direct successive samples of the QL algorithm. Furthermore, a second
neural network is added to provide the target Q-values. These values will be used to
calculate the loss value for each action at DQL training round [171].
Now, let the DQN be denoted Q (s, a;θθ), where θ is a real-valued vector completely
characterizing the function Q (s, a;θθ), such that Q (s, a;θθ) ≈ Qφ∗ (s, a). In turn, the
search for the best Q-function translates to finding the best θ of finite dimensions via
training.In particular, the DQLagent gathers experiences and forms a data set D in the
form of s(t) , a(t) , r(t+1) , s(t+1) by collecting experiences until-step t. To this end, two
(t) (t)
DQNs are defined, namely the target DQN with θ target , and the train DQN with θ train .
(t) (t)
Moreover, θ target is updated to become equivalent to θ train over a specific number of
time-steps [179]. In each time-step t, the DQN is trained by minimizing a least squares
loss function (i.e. a gradient-descent) based a random mini-batch from D , which is
expressed as [186]
h i2
L θ (t) (t) (t) (t) (t)
train = E y − Q s , a ;θθtrain , (6.7)
where y(t) is the target value function, given by

(t) (t) (t) (t+1) (t)
y = r s ,a + δ max Q s , a;θθtarget . (6.8)
a
Due to the possible instability (or divergence) of the DQL, the aperiodic store expe-
rience is used to improve the learning stability of the DQL [187]. In addition, ε is
updated using the decay rate υ as ε = ε(1 − υ), while slowly smoothing the target
parameters in every training round with ξ, as
(t) (t−1) (t)

θ train = ξθθtrain + (1 − ξ)θθtrain
(t) (t−1) (t)
(6.9)
θ target = ξθθtarget + (1 − ξ)θθtarget ,
ultimately reducing the correlations between the target and estimated Q-values, and
thus stabilizing the DQL algorithm.
For DQL, the reward function r(s, a) takes one of the predefined values and v1 > v2 ,
as 
v , (EE ≥ ζ ) ∩ (R ≥ R ) ∩ (P ≤ P ),
1 min i min i max
r(s, a) = (6.10)
v , otherwise,
2
which maintains the minimum capacity and maximum transmission power for each
UE, while maximizing the EE. The DQL algorithm is summarized in Algorithm 8,
which is guaranteed to converge efficiently [179, 188].
Algorithm 8 Deep Q-Learning

1: Initialization: Experience memory D , δ, ε, υ, and ξ. Also, initialize training parameters
θtrain , and target parameters as θtarget = θtrain .
2: for each time-step t do
3: For the current state s(t) , pick the action vector a(t) using the ε-greedy policy, as

arg max Q s(t+1) , a;θθ(t)

ε
target , with prob. 1−ε+ |A (s(t) )| ,
(t) a
a ←
a random action vector, with prob. ε + A sε(t) .
| ( )|
Perform action a(t) , obtain reward r(t+1) = r s(t) , a(t) and observe the new state s(t+1) .

4:
Store s(t) , a(t) , r(t+1) , s(t+1) in experiences memory D .

5:
6: Pick a random mini-batch of from D .
7: Determine the target value function y(t) as

(t)
y(t) = r s(t) , a(t) + δ max Q s(t+1) , a;θθtarget .
a
(t)
8: Update parameters θ train by minimizing the loss function
h i2
L θ (t)
train = E y (t)
− Q s (t) (t) (t)
, a ;θ
θ train .
(t) (t)
9: Update the target parameters θ train and θ target using ξ as
(t) (t−1) (t)

θ train = ξθθtrain + (1 − ξ)θθtrain
(t) (t−1) (t)
θ target = ξθθtarget + (1 − ξ)θθtarget .
10: Set t ← t + 1 and current state s(t) ← s(t+1) .

11: Update ε = ε(1 − υ).
12: end for
13: Output: State s and action a vectors.
Since a multi-layer deep neural network is utilized in this work, Fig. 6.1 illustrates
the operation of the proposed power control scheme using DQL. The state and action
vectors each have 1×(N ×I) elements to describe each possible state and action, where
N is the number of the reused RBs in each frequency band, and I is the total number of
UEs in the UHF or mmWave band. Both s and a represent inputs to the DNN, while
the output is the estimate of expected long-term reward based on a given status s of the
DQL. The input layers for both s and a are followed by multiple deep layers; starting
with fully connected layer, described by y1 = ws · s + bs , where the input vector s is
weighted by vector ws and bs is the bias vector. The next layer is the Rectified Linear
Unit (ReLU) used to suppress any negative output value of the previous fully connected
layers to zero, and the output is y2 = max(y1 , 0), then another fully connected layer is
Figure 6.1: Proposed power control scheme using DQL
applied. In order to update s by adding the actions a, the Add layer has been used to
obtain the output. Finally, to remove any negative power, a ReLU layer has been used.
Lastly, a fully connected layer with a single output is used to provide the state-action
function Q(s, a), as illustrated in Fig. 6.2.
6.3 Implementation of QL and DQL

This section discusses the implementation of the QL and DQL algorithms. It should
be noted that since the UHF and mmWave UEs do not interfere with each other, the
QL/DQL algorithm is executed for the UE over each band separately to obtain the
transmission power values for GUEs and UAVs EE maximization.
6.3. IMPLEMENTATION OF QL AND DQL 145
Actions
Wireless Environment
Reward Observations
RL Algorithm
Updated policy
Observations Actions
Inputs Inputs
Critic(a) Fully
Connected
ReLU
Layer
Critic(b) Fully Fully

Connected Connected
Add Layer
ReLU Layer
Fully connected
Output Q(s,a)
Policy
DQL Agent
New actions
Figure 6.2: DQL Model
Now, the state vector s contains the power value of the user RBs, say UE i, starting
with a low power value (e.g. Pi,n = 10−6 W) up to the maximum transmission power
value Pmax = 0.1 W, but each UAV or GUE’s maximum transmission power is 0.2 W,
since we assume one user can occupy multiple RBs. Each action in the action vector
a involves multiplying the RB power by one of three values in {0.1, 1, 10} as per the
ε-greedy policy, which facilitates the exploration and exploitation to maximize the Q-
function. Upon applying an action vector, the states vector is updated to values in the
range 10−6 , 0.1 W, which designates the transmission power values of either UHF or

mmWave UEs.
In this work, the final or exit state is not known a priori, since the optimal value for
the EE is not known before executing the QL (or DQL) algorithm. However, conver-
gence to a certain Q value or reaching the maximum number of iterations terminates
the search process, and the final obtained Q value resembles the best EE value. In turn,
the learning rate η, discount factor δ, and maximum number of iterations Tmax govern
the convergence speed and accuracy of the obtained EE solution. Particularly, a higher
learning rate η allows better solution exploration, and a value of δ → 1 puts more em-
phasis on long-term higher rewards. Also, the higher Tmax is, the better the exploration
and exploitation, which guarantees the optimal states for UEs transmission powers.
Hence, the values of η, δ and Tmax pose a trade-off between accuracy of the obtained
solution and speed of convergence. To highlight this, Table 6.1 summarizes the pa-
rameters for two scenarios, which will be considered in the performance evaluation in
Section 6.7.
Table 6.1: QL Parameters

Parameters Scenario 1 Scenario 2
−6 −5 −4 −3 −2 −1
Possible States of One RB s ∈ 10 , 10 , 10 , 10 , 10 , 10 , W
Possible Actions of One RB a ∈ {0.1, 1, 10} W
v1 = 10
v2 = 1
v3 = −10
Reward Function Values ζG
min = ζ A = 4 MBits/J
min
RGmin = 0.4 MBits/s
A
Rmin = 4 MBits/s
G = PA = 0.2 W
Pmax max
Discount Factor δ 0.1
ε-Greedy Parameter 0.333
Learning Rate η 0.1 0.01
Tmax 10,000 25,000
In a similar manner to the QL algorithm, the DQL is evaluated based on two sce-
narios, as shown in Table 6.2.

In this section, the performance of the coupled UHF and coupled mmWave with FPC
are compared to the DUDe access scheme in terms of sum-rate, energy-efficiency, and
data rate per UE. Specifically, the performance of the QL and DQL power control
schemes based on DUDe access are evaluated and compared with the optimal scheme
proposed in last chapter and fractional power control schemes used in 4G and 5G
networks, namely Decoupled-Optimal and Decoupled-FPC, respectively2 . Since the
2 The
optimal EE-maximizing power control schemes are based on the solutions of problems GUEs-
EE-MAX and UAVs-EE-MAX, as discussed in subsections 5.4.1, and 5.4.2, respectively.
Table 6.2: DQL Parameters

Parameters Scenario 1 Scenario 2
States Input Size: 5 Users × 3 RBs = 15 Neurons
Input Layer Output: 24 Neurons—Normalization: None
Actions Input Size: 15 Neurons—Output: 50 Neurons
Input Layer Normalization: None
States Critic(a) Fully Input Size: 50 Neurons—Output: 50 Neurons
Connected Layer Normalization: None
States Critic(b) Fully Input Size: 50 Neurons—Output: 50 Neurons
Action Critic Fully Input Size: 50 Neurons—Output: 50 Neurons
ReLU Layers for
f (x) = max(0, x)
Critic & Action Paths
Add layer Adding neurons element wise
Fully Connected Input Size: 50 Neurons—Output: 1 Neuron
Output Q(s, a) Normalization: None
s ∈ 10−6 , 10−5 , 10−4 , 10−3 , 10−2 , 10−1 , W

Possible States
Possible Actions of One RB a ∈ {0.1, 1, 10} W
v1 = 10
v2 = −10
Reward Function ζGmin = ζ A = 4 MBits/J
min
Values RGmin = 0.4 MBits/s
A
Rmin = 4 MBits/s
G = PA = 0.2 W
Pmax max
Decay Rate υ 0.005
Smoothing Factor ξ 0.001
Discount Factor δ 0.1
ε-Greedy Parameter 0.333
Learning Rate η 0.1 0.01
Tmax 10,000 25,000
UAV CNPC links require low data rate, and will not interfere with other links, they are
not considered in the simulations, for simplicity. Table 6.3 summarizes the simulated
transmission parameters.
Fig. 6.3 illustrates the 10th , 30th , 50th , 70th , and 90th percentile data rate per GUE
and UAV in the UL, where the ratio of SBS to the MBS is 4. In Fig. 6.3 (a), the UEs
data rates shows that due to EE optimization the 70% and 90% percentile data rates in
Decoupled-Optimal, QL and DQL are lower than the decoupled FPC. Although the
minimum rates are respectively 4 × 106 and 4 × 105 bps for the GUEs and UAVs, some

Parameters GUE UAV MBS SBS
DL/UL Bias N/A N/A 0/0 dB 3/0 dB
Spatial Distribution Uniform Distribution
Altitude of UAVs 50-200 m [13, 55]
S-curve Parameters a = 9.6, b = 0.28 [144, 146]
Blockage Ball Model Parameters µ = 200 m, ω = 0.2 [77]
Spatial Density 30 per km2 30 per km2 5 per km2 20 per km2
Path-Loss Exponent UHF: GUE-UAV 2, GUE-BS 3, d0 = 1 m [148]; mmWave: LOS 2.55, NLOS 5.76, d0 = 5 m [29, 164]
Lognormal Shadowing UHF: µ = 0, σ = 4 dB [148]; mmWave: LOS µ = 0, σ = 8.66 dB, NLOS µ = 0, σ = 9.02 dB [29]
Nakagami-m Parameters mL = 3, mN = 2 [144, 147]
Operating Frequency 2 GHz & 28 28 GHz 2 GHz & 2 GHz & 28
GHz 28 GHz GHz
Bandwidth UHF: 1.2 MHz; mmWave: 4.8 MHz
Subcarrier Spacing UHF: 15 kHz; mmWave: 60 kHz [84]
Power Control FPC with P0 = −85 dBm, and α = 0.8 [55]
Noise Spectral Density -174 dBm/Hz
Figure 6.3: 10th , 30th , 50th , 70th , and 90th percentile data rate per user in the UL: (a)
mmWave band and (b) UHF band - SBS to MBS ratio = 4
of the UEs under the QL and DQL schemes are below the thresholds. This is because
the EE thresholds (i.e. ζG A
min and ζmin ) appear as soft thresholds (as per (6.6) and (6.10)),
which leads to a tradeoff between the data rate and EE. Additionally, Fig. 6.3 (a) shows
that both QL and DQL algorithms improve their learning policies and assign power to
UEs to increase their data rates when the number of training iterations is increased and
their learning rates are decreased. This can be verified by comparing Scenarios 1 and
2 for the Decoupled-QL and Decoupled-DQL schemes, and this is due to the fact
that more states are visited in search for the best state. More importantly, this implies
that improving the learning increases the rate of the UEs for EE-maximization in the
mmWave band since it is noise-limited. Contrarily, Fig. 6.3 (b) depicts an opposite
pattern, since the UHF band is interference-limited, and thus, Scenario 2 decreases
the rate of the UEs to improve the EE, in comparison to Scenario 1.
Figure 6.4: UL GUE sum-rate vs. ratio of SBS to MBS.
In Fig. 6.4, it can be seen that the GUEs sum-rate of the decoupled schemes are
at least 60% higher than the Coupled-UHF scheme. This is because the decoupled
schemes have wider bandwidth for GUE UL communications (as shown in Table 5.2),
shorten the distances between the GUEs and BSs, and eliminate the interference be-
tween UAVs and GUEs. The sum-rate of the Coupled-mmW scheme is the high-
est, since it utilizes the mmWave band for GUE UL communication, while the other
schemes utilize the UHF band, and the mmWave bandwidth is wider than the UHF
bandwidth. The sum-rate of the Decoupled-Optimal scheme remains relatively con-
stant with the increase in the SBS to MBS ratio, since it mainly aims to achieve the
optimal energy-efficiency, as will shown in Figs. 6.6 and 6.7. In comparison to the
Decoupled-Optimal scheme, both the Decoupled-QL and Decoupled-DQL schemes
yield higher data rates at the expense of higher transmission power, which will translate
to lower EE values. To see this, for both schemes, Scenario 2 yields lower sum-rate
than Scenario 1, as increasing the training iterations and reducing the learning rate
lower the sum-rate to improve the EE by carefully selecting the transmission power.
Figure 6.5: UL UAV sum-rate vs. ratio of SBS to MBS.
Similarly, in Fig. 6.5, the UAV sum-rates of the Coupled-UHF and Coupled-
mmW schemes are much lower than the decoupled schemes, since the UAVs under
those two schemes are allocated narrower bandwidth and suffer from higher path-loss.
Besides, the UAVs under the Coupled-UHF scheme also suffer from ICI, while the
UAVs under the other schemes are allocated the mmWave band, and thus, their ICI is
minimal. In comparison to the Decoupled-Optimal scheme, both the Decoupled-
DQL and Decoupled-QL tend to explore if increasing the sum-rates may help in
achieving better EE for the UAV UL transmissions and this is appear as a small in-
crease in the sum-rate when the SBS to MBS ratio increases. Adding to this, Scenario
2 improves the sum-rates for both the Decoupled-DQL and Decoupled-QL in com-
parison to Scenario 1.
As for EE, as shown in Fig. 6.6 and Fig. 6.7, the decoupled schemes can achieve up
to several times higher EE than the coupled schemes, as they prevent the interference
between UAVs and GUEs, reduce the interference among GUEs, and shorten the UE-
BS distances. Also, Figs. 6.6 and 6.7 demonstrate that the EE improvement for the
Decoupled-Optimal, Decoupled-QL and Decoupled-DQL schemes as the SBS to
Figure 6.6: UL mmWave UEs normalized EE vs. ratio of SBS to MBS.
Figure 6.7: UL UHF UEs normalized EE vs. ratio of SBS to MBS.
MBS ratio increases. This is attributed to the decrease in the number of UEs associated
with the same SBS or MBS, and the decrease in UE-BS distances. In addition, the
Decoupled-QL and Decoupled-DQL schemes yield an improvement in the EE as the
Figure 6.8: Network UL sum-rate vs. ratio of SBS to MBS.
training iterations increase and the learning rate decreases, which can be verified by
comparing Scenario 1 and Scenario 2 for both schemes.
Fig. 6.8 illustrates the total UL sum-rate, where one can see that the sum-rate
of the Decoupled-QL and Decoupled-DQL schemes improves as the ratio of SBS
to MBS increases. Also, the sum-rate for the Decoupled-QL (Scenario 2) scheme
shows has a minor improvement over the Decoupled-QL (Scenario 1) scheme, while
Decoupled-DQL (Scenario 2) shows around 5 Mbps improvement over Decoupled-
DQL (Scenario 1). Lastly the QL and DQL algorithms are limited by the maximum
number of iterations in search for the best trade-off between sum rate, minimum user
rate, and the EE, which also control how long it takes to run the optimization process.
6.5 Summary
In this chapter, two power control schemes based on QL and DQL have been proposed
to improve the EE of an cellular-enabled UAV network in DUDe scenario. Then, the
proposed QL and DQL schemes have been compared with the FPC scheme (which is
applied in 4G and 5G networks), and the optimal EE-maximizing benchmark power al-
location scheme proposed in last chapter. The results revealed that the proposed DUDe
access schemes can achieve several times higher sum-rates and EE than their coupled
6.5. SUMMARY 153
counterparts. Moreover, it is shown that although RL methods can achieve optimal

results in theory, in practical scenarios with predominant dynamic environments, and
limited time to execute the optimization process, QL and DQL with limited number
of iterations may only achieve a near-optimal EE performance. Nonetheless, the pro-
posed QL (DQL) algorithm has been shown to achieve better EE performance than
the baseline FPC scheme by around 100% (70%) for UHF band, and by around 160%
(130%) for the mmWave band, in comparison to conventional FPC scheme.
Chapter 7
DUDe Access in Mobile Edge

Computing
7.1 Introduction
This chapter analyzes the application of DUDe on MEC. Different from cloud com-
puting, MEC is limited by real-time delay constraints and limited computing resources
[189]. If too many tasks are offloaded to one MEC server, then this may lead to severe
congestion and latency. If an MD’s task is offloaded to a BS that is too far away, then
data transmission may be drastically impaired due to excessive path-loss, which ad-
versely affects communication resource utilization. Thus, the main problems of MEC
resource allocation can be divided into computation offloading decision (whether to of-
fload or not), user association/offloading node selection (where to offload), subchannel
assignment, and transmission power control (interference management). Some exist-
ing research works focus on MEC with multi-user single-server scenarios, while others
consider MEC with homogeneous multi-server scenarios. However, only a few studies
have been conducted to investigate the heterogeneous servers scenario, where servers
are located at the SBS, MBS or cloud, and have different computing and communica-
tion capabilities, hence the focus of this chapter.
Generally speaking, the optimization objectives in most works can be divided into
three categories:
(1) Energy consumption minimization: To improve the battery lifetime, the energy
consumption of MDs and/or BSs is optimized, subject to constraints on latency.
(2) Latency minimization: For those applications with stringent latency constraints,
154
Table 7.1: Summary of MEC Scenarios and Objectives

MD Edge Servers Objectives
Ref.
Energy &
Single Multiple Single Multiple Energy Latency
Latency
[190] X X X
[191] X X X
[192] X X X
[193] X X X
[194] X X X
[195] X X X
[196] X X X
[197] X X X
[198] X X X
[199] X X X
it is necessary to shorten the latency while satisfying the transmission power

and/or energy consumption constraints.
(3) Joint energy consumption and latency minimization: The weighted sum of en-
ergy consumption and latency is minimized.
A summary of different MEC scenarios and objectives studied in the literature is given
in Table 7.1, where it should be noted that [190, 191] pertain to the heterogeneous
servers scenario, while the rest are for homogeneous servers.
In MEC-based systems, computational tasks offloading is potentially limited by the
type of MD-BS association, and almost all the previous works consider offloading a
MD’s computation task to the MEC servers available at its serving BS. However, with
the deployment of more and more low-cost SBSs, the traditional homogeneous net-
works become HetNets. Due to the transmission power disparity of SBSs and MBSs,
the DL coverage of MBSs is usually much greater than that of SBSs. Consequently,
more MDs are associated with MBSs in the DL. Offloading too many MDs to the
MBSs may cause severe network congestion and latency. Furthermore, unlike BSs,
MDs have similar transmission powers and transmission coverage. If the associated
cell in the UL is the same as that in the DL, the link quality of the MCell edge MDs
will be poor, and may cause high UL transmission latency. Although cell range ex-
tension can offload more MDs to the SBSs [7], it may impair the DL transmission
performance. For the sake of balanced offloading and reduced path-loss, it would be
a better choice for some MDs to connect to a geometrically closer SCell in the UL
156 CHAPTER 7. DUDE ACCESS IN MOBILE EDGE COMPUTING
rather than the same BS in the DL, i.e. DUDe access [2]. Moreover, since different
types of BSs have different computation capabilities, choosing the serving BS from the
computation resource perspective is also necessary.
Figure 7.1: A DUDe MEC model.
As shown in Fig. 7.1, according to the DUDe access, a MD’s UL and DL serving
BSs may not necessarily be the same. Accordingly, the MD’s tasks can be executed
at the MEC server available at the UL and/or DL serving BS. In such a case, the
BSs should be able to communicate with each other, which is possible for 4G and
beyond cellular networks, since the BSs are inter-connected via traditional backhaul
connections (e.g. via the X2 interface, in accordance with 3GPP LTE terminology) [1].
Despite the extra backhaul delay, the DUDe MEC scheme is also capable of providing
a fairly lower offloading latency as compared to the conventional offloading scheme
with coupled association [200]. Hence, it is necessary to study optimal or near-optimal
resource allocation algorithms for DUDe MEC systems. To the best of our knowledge,
only one paper applied DUDe to MEC-based networks [200], which presents a DUDe
MEC framework and compares it with the CUDA scheme; however, it does not propose
any offloading or resource allocation algorithms.

The main contributions of this chapter can be summarized as follows:
• A DUDe joint BS association and subchannel allocation algorithm is proposed.

Different from most studies that consider BS association and subchannel alloca-
tion separately, this work jointly performs BS association and subchannel allo-
cation by utilizing the student-project allocation (SPA) matching model [201].
Furthermore, the constraint that one MD must connect to the same BS in the
UL and DL is broken, and the UL serving BS is chosen according to both the
communication and computation capacity of BSs.
• Given the obtained BS association and subchannel allocation, an optimal power

allocation algorithm is devised to minimize the sum-latency of the network MDs.
The formulated problem is a sum-of-ratios problem, which is non-convex and
NP-hard. To efficiently and optimally tackle it, the proposed algorithm tightly
approximates the problem as a convex optimization problem, and successively
solves it until convergence to the global optimal power allocation solution.
• Finally, the performance of the proposed algorithms are compared with bench-
mark schemes in terms of latency, energy-efficiency (EE) and data rate. Specifi-
cally, the simulation results demonstrate that the latency of the proposed DUDe
access and power control scheme is much lower than that of the CUDA scheme
by up to 60%, and its EE and data rate are also increased.
7.1.2 Organization
The rest of the chapter is organized as follows. Section 7.2 presents the system model
adopted in this chapter. Section 7.3 proposes the DUDe joint BS association and sub-
channel allocation scheme, which considers the communication and computation ca-
pacity of the different types of BSs. Section 7.4 proposes an optimal power allocation
scheme to minimize the network sum-latency. Section 7.5 evaluates the performance
of the proposed scheme and compares it with coupled benchmarks. Finally, Section
7.6 gives a summary and critique of the proposed algorithms as well as the findings.
7.2 System Model
7.2.1 Network Model

In this work, a two-tier OFDMA HetNet—composed of one MBS and several SBSs—
is considered, and the Rayleigh fading channel model is adopted. The positions of
the MBS and the SBSs are uniformly distributed, where the density of the SBSs is λs .
The MDs’ locations also follow a uniform distribution with a density of λu . Let K =
{1, . . . , k, . . . , K} and N = {1, 2, ..., N} be the index sets of the MDs and subchannels,
respectively. Also, let M = {0, 1, . . . , M} denote the set of M + 1 BSs, where BS m = 0
corresponds to the MBS, and the rest (i.e. m = 1, . . . , M) are SBSs. Both tiers operate
on the same frequency band, and are based on FDD. Moreover, an MD can only occupy
one subchannel, and MDs in different cells can reuse the same subchannel. As such,
there is inter-cell interference, but no intra-cell interference.
7.2.2 Transmission Model

The transmit powers of all SBSs are assumed to be identical, but the transmit power of
the MBS is greater than the SBSs. In the DL, the maximum transmit power is equally
allocated over the whole bandwidth; while in the UL, the initial transmit power of MD
k is set according to the fractional power control (FPC) as [202, 203], which is applied
in 4G and 5G
Pkul = min{Pmax , wPL + P0 }, (7.1)
where Pmax is the maximum transmission power of each MD, P0 is the target received
power, and w ∈ {0, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1} is the compensation factor for path-loss
PL. The path-loss PL(d) can be modeled as

4πd0 f d
PL(d) = 20 log + 10φ log + χ, (7.2)
c d0
where d0 denotes the close-in reference distance, f is the operating frequency, c is the
speed of light, d represents the MD-BS distance, φ is the path-loss exponent, and χ is
the log-normal shadowing.
7.2.3 MEC Model

In this work, full offloading is adopted, while assuming that there is only one compu-
tation task to be processed for each MD during a computation offloading period. As
the UL serving BS may be different from the DL serving BS, the tasks can be executed
at the cloudlet attached to the UL serving BS or the DL serving BS or both of them.
However, for simplicity, it is assumed that all the tasks are executed at the UL cloudlets
in this chapter. The MBS’s cloudlet is assumed to have higher computation capacity
than the SBSs’ cloudlets, which are denoted F M and F S (in CPU cycles per second),
respectively1 . Also, each SBS is connected to the MBS via a backhaul link with fi-
nite capacity2 Cbh . Let the computational task of each MD k ∈ K = {1, 2, ..., K} be
defined as a tuple Tk , (BCk , BIk , BO C
k ), where Bk is the required number of CPU cycles
to complete MD k0 s task, BIk is the number of bits of the offloaded task, and BO k is the
number of output bits representing the result of the task execution. For simplicity, the
output bits and the computation data size are assumed to be proportional to the input
task size. Particularly, let B0k = αk BIk , where αk is the proportion of output to input
bits [189]. Moreover, BCk = βk BIk , where βk is the number of CPU cycles per bit, and
depends on the type of executed task. To offload a task, MD k first transmits the BIk
input bits to its UL serving BS, and the cloudlet of which executes the BCk CPU cycles.
Finally, the BO 3
k output bits are transmitted back to the MD .
7.2.4 Latency Model

The total offloading latency for a typical MD k is given by
Lk = Lkul + Lkexe + Lkbh + Lkdl , (7.3)
where Lkexe is the time delay for the cloudlet to execute MD k’s task, Lkul is the UL
transmission latency, and Lkdl is the DL transmission latency4 . Moreover, there are two
cases for the backhaul latency Lkbh , which are defined as
• UL serving BS 6= DL serving BS: Lkbh = BO bh
k /C ;
• UL serving BS = DL serving BS: Lkbh = 0.
7.2.5 Communication Model

Let χk,m be a binary decision variable, defined as

1, if MD k is associated with BS m in l direction,
χlk,m = (7.4)
0, otherwise,
1 Assume each cloudlet’s computational resources are equally shared among all tasks when two or
more tasks are offloaded to the same cloudlet.
2 Interference-free wireless backhaul links are assumed, and the backhaul links will not cause inter-
ference to MD-BS links.

3 Note that if the UL and DL serving BSs are different, the BO output bits are communicated to the
k
DL serving BS via the backhaul link.
4 There is a waiting time before a cloudlet begins to execute a task, but most works assume it to be
zero [189, 200].

where l ∈ {ul, dl}. Moreover, it should be noted that each MD k ∈ K can be associated
with one BS in each link direction (i.e. ∑m∈M χlk,m = 1). Also, let λk,n be defined as

1, if subchannel n is allocated to MD k in l direction,
l
λk,n = (7.5)
0, otherwise.
Thus, the data rate of MD k ∈ K in the UL is
ul hul
!
Pk,m,n k,m,n
Rul
k =B ∑ ∑ χulk,mλulk,n log2 1+ ul
, (7.6)
m∈M n∈N Ik,m,n + σ2
where B is the RB bandwidth, and σ2 = N0 B is the variance of the AWGN, with N0

ul
being the noise spectral density. Moreover, Pk,m,n is the transmission power of MD k
when associated with BS m over subchannel n, and hul k,m,n is the corresponding channel
ul
gain in the UL. The interference Ik,m,n received at BS m on RB n for MD k is written
as
ul
Ik,m,n = ∑ λulj,n Pul ul
j,m,n h j,m,n . (7.7)
j∈K , j6=k
Similarly, the data rate at MD k in the DL is expressed as
dl hdl
!
Pk,m,n k,m,n
Rdl
k =B ∑ ∑ χdlk,mλdlk,n log2 1+ dl
, (7.8)
m∈M n∈N Ik,m,n + σ2
dl
where Pk,m,n is the transmission power of BS m to MD k over subchannel n, and hdl
m,k,n
is the corresponding channel gain. The interference received at MD k on RB n is given
by
dl
Im,k,n = ∑ λdl dl dl
k,i Pi,k,n hi,k,n . (7.9)
i∈M ,i6=m
Notably, MD k’s UL and DL data rates Rul dl

k and Rk are related to MD k’s serving BS,
transmission power, and the interference level over the assigned subchannel in each
link direction.
The latency (in seconds) necessary to complete the UL transmission for MD k is

defined as
BIk
Lkul = ul , (7.10)
Rk
while the DL transmission time is
BO
Lkdl = k
. (7.11)
Rdl
k
The backhaul delay is only related to the data size [200], and thus, the backhaul latency
is determined as
bh Bbh
Lk = kbh , (7.12)
C
where Bbh
k is MD k’s backhaul data size, defined as

BO , decoupled access,
bh k
Bk = (7.13)
0, coupled access.
The execution latency for a typical MD k is
BCk
Lkexe = , (7.14)
Fk
where Fk = F̄k /K̄, with F̄k being the computation capacity (in CPU cycles/second) of
the cloudlet attached to MD k’s UL serving BS (i.e. F̄k = F M for MBS, and F̄k = F S
for SBS), and K̄ is the number of MDs associated with that BS. Lastly, the total latency
of MD k is calculated as
BIk BCk Bbh k BOk

Lk = + + + . (7.15)
Rul
k
Fk C bh R dl
k
7.2.6 Problem Formulation

In this work, the cell association, subchannel allocation and transmission power control
are considered in the UL only, as DUDe can only improve the UL performance. As
for DL, biased RSRP cell association is along with greedy subchannel allocation, and
equal power allocation over all subchannels, and all BSs transmit with their maximum
power.
In the UL, our objective function is to minimize the sum-latency, which is defined
X , Λ , P ) , Σk∈K Lk . Specifically, X = {χul
as Z(X ul ul
k,m }, Λ = {λk,n }, and P = {Pk,m,n }.
Moreover, the constraints are as follows:
• transmission power of each MD does not exceed the maximum transmission

power Pmax .
• Each MD can associate with only one BS in each link direction.
• Each MD-BS link can be assigned one subchannel.
Therefore, the formulated problem can be expressed as
Q1: (7.16)
min X , Λ , P)
Z(X (7.16a)
X , Λ, P
s.t. ∑ χk,m = 1, ∀k ∈ K , (7.16b)
m∈M
∑ λk,n = 1, ∀k ∈ K , (7.16c)
n∈N
ul
0 ≤ Pk,m,n ≤ Pmax , ∀k ∈ K , ∀m ∈ M , ∀n ∈ N . (7.16d)
Problem Q1 is non-convex and NP-hard. To efficiently solve it, it is split into two
subproblems: (1) joint cell association and subchannel allocation (JCASA), and (2)
power control (PC). Specifically, the FPC scheme is applied to allocate the initial UL
transmission power P (0) . Then, the JCASA problem can be obtained as
Q2: (7.17)

(0)
min Z X , Λ , P (7.17a)
X,Λ
s.t. (7.16b), (7.16c), (7.17b)
which is a combinatorial problem. After obtaining the cell association and subchannel
allocation solutions X ∗ , and Λ ∗ —by solving problem Q2—the PC problem can be
formulated as
Q3: (7.18)
7.3. JCASA BASED ON SPA 163
min Z (XX ∗, Λ∗, P) (7.18a)

P
s.t. (7.16d). (7.18b)
which is a non-convex and non-linear programming problem.

In Section 7.3, the JCASA problem (i.e. problem Q2) is modeled and solved via
the student-project allocation (SPA) matching model. After that, the sum-latency min-
imizing power allocation algorithm for solving problem Q3 is devised in Section 7.4.
7.3 JCASA Based on SPA
7.3.1 Student-Project Allocation Matching Model

In the two-sided SPA matching problem, the students have preferences over the projects,
while the lecturers have preferences over the students [201]. In this chapter, we con-
sider the MDs U = {u1 , . . . , uk , . . . , uK } as the students, the subchannels C = {c1 , . . . , cN×(M+1) }
as the projects5 , and the BSs S = {s0 , s1 , . . . , sm , . . . , sM } as the lecturers. Each MD uk
ranks the subchannels by preference according to the SINR, i.e. if MD uk prefers
subchannel ci to c j , then this implies that the SINR on subchannel ci —based on the al-
located power—is higher than that on c j . The top N × Mk subchannels with the highest
SINR form an acceptable subchannel set Ck ⊆ C , where Mk is an adaptive parameter
that decides if the subchannels offered by the nearest Mk BSs are acceptable. Then,
each BS sm ranks the set of MDs that find a project offered by it acceptable, and forms
a list denoted by Um ⊆ U , which consists of the set of MDs that have the suitable task
size. Specifically, since the MBS has higher computation capability, it prefers the MDs
that demand more computing resources, which helps reduce the network computation
latency. If an MBS prefers MD ui to u j , then the number of CPU cycles needed to
process MD ui ’s task is more than that of MD u j . On the other hand, if an SBS prefers
MD ui to u j , then the number of CPU cycles needed to process MD ui ’s task is less
than that of MD u j . For each subchannel cn , let Umn denote its preference list, which is
obtained from Um by deleting those MDs who do not find cn acceptable.
Each BS sm is assumed to have a capacity constraint Ωm , indicating the maximum
number of MDs that it can serve at the same time, which should not exceed the total
5 The N subchannels are assumed to be offered by the M + 1 BSs as unique subchannels. In turn,
the N subchannels are replicated at each BS, such that a subchannel can be assigned to users associated
with different base-stations. Thus, there are N × (M + 1) subchannels in total.
number of subchannels N (i.e. 1 ≤ Ωm ≤ N, ∀sm ∈ S ). In practice, Ωm is set as

νm K/(M + 1), which represents the average number of MDs per BS. Moreover, M+1 K ≤
M+1
νm ≤ K N is set to a high value if a BS sm is rich with computation resources, and vice
versa. Also, each subchannel has a capacity constraint of 1, indicating each subchannel
can be occupied by one MD when associated with a BS.
Definition 7.3.1 (Assignment). An assignment M is a subset of U × C such that:
(1) (uk , cn ) ∈ M (i.e. MD uk finds subchannel cn acceptable.
(2) |(uk , cn ) ∈ M : cn ∈ C | ≤ 1 (i.e. MD uk is assigned at most one subchannel).
If (uk , cn ) ∈ M , where BS sm offers subchannel cn , then MD uk is assigned to sub-

channel cn , and associated with BS sm . For any MD uk , M (uk ) refers to the subchannel
it is assigned to. For any subchannel cn , M (cn ) refers the MD assigned to it. For any
BS sm , M (sm ) denotes the set of MDs associated with BS sm .
Definition 7.3.2 (Matching). A matching M is an assignment, such that |M (uk )| ≤ 1,

|M (cn )| ≤ 1 and |M (sm )| ≤ Ωm , ∀uk ∈ U , ∀cn ∈ Cm , and ∀sm ∈ S .
Consequently, under matching M , each MD uk is assigned to at most one subchan-

nel, no subchannel cn ∈ C is assigned to more than one MD, and no BS sm is assigned
to more than Ωm MDs.
Definition 7.3.3 (Subscription). A subchannel cn ∈ Cm is said to be under-subscribed,

full, or over-subscribed if |M (cn )| < 1, |M (cn )| = 1, or |M (cn )| > 1, respectively.
Similarly, a BS sm ∈ S is considered to be under-subscribed, full, or over-subscribed if
|M (sm )| < Ωm , |M (sm )| = Ωm , or |M (sm )| > Ωm , respectively.
Definition 7.3.4 (Blocking). The pair |(uk , cn ) ∈ (U × C )\M is said to block a match-
ing M if:
(a) cn ∈ Ck (i.e. uk finds cn acceptable).
(b) Either uk is unassigned in M , or uk prefers cn to M(uk ).
(c) Either
(c1) cn is under-subscribed and sm is under-subscribed, or
(c2) cn is under-subscribed, sm is full, and either uk ∈ M (sm ) or sm prefers uk to
the worst MD in M (sm ), or
(c3) cn is full and sm prefers uk to the worst MD in M (cn ), where sm is the BS who
provides cn .
Definition 7.3.5 (Stable matching). A matching M is considered stable if M contains

no blocking pairs.
Different from most previous works, which consider BS association and subchan-
nel allocation separately, the SPA algorithm performs them jointly and simultaneously.
Initially, all MDs are free, and all subchannels and BSs are unsubscribed. As long as
there is an MD, uk , that is free and with a non-empty preference list, it can apply to the
first subchannel cn on Ck . Let sm be the BS that offers cn . Immediately, uk becomes
provisionally assigned to cn (and to sm ). If cn is over-subscribed, then sm rejects the
worst MD ur assigned to cn and the pair (ur , cn ) will be deleted. Similarly, if BS sm
is over-subscribed, then sm rejects its worst assigned MD ur and the pair (ur , ct ) will
be deleted from M , where ct was the subchannel assigned to ur . On the other hand, if
cn is full and ur is the worst MD assigned to cn , then delete (ut , cn ), where ut is each
successor of ur on Umn . Similarly, if sm is full and ur is the worst MD assigned to sm ,
then delete (ut , cv ), where ut is each successor of ur on Um and cv is each subchannel
offered by sm that ut finds acceptable. The SPA algorithm is described in Algorithm
96 .
The SPA algorithm converges with polynomial-time complexity of O (|U | × |C |),
where |U | = K and |C | = N ×(M +1) are the number of MDs and subchannels, respec-
tively. The stable matching resulting from the SPA algorithm is optimal with respect
to each assigned MD [201]. This is because each user is assigned to its most preferred
subchannel available on its preference list, and no stable pair is deleted during the exe-
cution of the SPA algorithm [201]. In turn, each user is simultaneously assigned to the
best channel it can get in any stable matching.
7.3.2 Swap Matching

SPA-based JCASA is with externalities (also known as peer effects) [204]. The prefer-
ence lists of the MDs and subchannels are initially constructed without considering the
inter-cell interference over each subchannel, since no BS association or subchannel al-
location initially exists. Thus, although Algorithm 9 yields a stable matching, as soon
6 The BS association and subchannel allocation are applied after the execution of the SPA algorithm
outlined in Algorithm 1. It should be noted that the accept and/or reject processes only occur during
the execution of Algorithm 1 at a network centralized controller (e.g. a MBS), and not during the
network operation. That is, the network operation takes place after the completion of the execution of the
proposed scheme shown in Fig. 2. Therefore, it will not cause call drops or failed handovers, and there
is no data exchange between MDs and BSs during the execution of the SPA algorithm. Additionally,
once a subchannel is occupied during the network operation, it will not be re-allocated to other MDs.
Algorithm 9 SPA based JCASA

1: Input: Preference lists Um , Um n and C , for ∀s ∈ S , ∀c ∈ C and ∀u ∈ U .
k m n k
2: Initialization: All MDs are free, and all subchannels and BSs are unsubscribed.
3: while (Some MD uk is free and has a non-empty list Ck ) do
4: cn = first subchannel on Ck ;
5: sm = the BS that offers cn ;
6: provisionally assign subchannel cn and BS sm to MD uk ;
7: if (cn is over-subscribed) then
8: ur = worst MD assigned to subchannel cn ;
9: sm = the BS that offers subchannel cn ;
10: sm rejects ur and break assignment (ur , cn );
11: else
12: if (sm is over-subscribed) then
13: ur = worst MD assigned to BS sm ;
14: ct = the subchannel assigned to MD ur ;
15: sm rejects ur and break assignment (ur , ct );
16: end if
17: end if
18: end while
19: if (cn is full) then
20: ur = worst MD assigned to subchannel cn ;
21: ul = each of the successor of ur on Umn ;
22: Break assignment (ul , cn );
23: end if
24: if (sm is full) then
25: ur = worst MD assigned to BS sm ;
26: ul = each successor of ur on Um ;
27: cv = each subchannel offered by sm that ut finds acceptable;
28: Break assignment (ul , cv );
29: end if
30: Output: Stable matching (uk , cn ) ∈ U × C .
as the subchannels are assigned to the MDs, the appearance of inter-cell interference
terms may not necessarily preserve the stability of the assignment. For this reason,
swap matching is adopted among the MDs associated with the same BS.
Definition 7.3.6 (Swap-Blocking Pair). Two subchannels cn (assigned to MD uk ) and

cn0 (assigned to MD uk0 , or not occupied)—for n0 6= n and k0 6= k—offered by BS sm
form a swap-block pair if:
(a) MD uk can get higher data rate on subchannel cn0 ,
(b) subchannel cn0 is assigned to MD uk0 , but MD uk0 can get higher data rate on
subchannel cn , and the sum UL transmission latency of MDs uk and uk0 is lower,
and
(c) sum-latency of all MDs transmitting over subchannel cn and cn0 is lower after
the swap7 .
Given a matching M , when a swap-blocking pair (cn , cn0 ) exists, MDs uk and
uk0 swap their subchannels, while keeping the other MD and subchannel assignments
unchanged. In turn, the updated matching is obtained as

M = M \ (uk , M (uk )) , (ur , M (ur ))

∪ (uk , M (ur )) , (ur , M (uk )) , (7.19)
On the other hand, if cn0 ,m is not occupied, then

M = M \ (uk , M (uk )) ∪ (uk , M (ur )) . (7.20)
Definition 7.3.7 (Two-Sided Exchange-Stability). A matching M is said to be two-

sided exchange-stable if it does not contain any swap-blocking pairs [205].
The swap matching algorithm is described as follows. Each MD searches if there

is a subchannel that can provide a higher data rate. If there is such a subchannel,
check if it can form a swap-blocking pair. If a swap-blocking pair is found, then a
swap-operation is performed, and the matching is updated. The process is given in
Algorithm 10. It is obvious that Algorithm 10 converges to a matching M in a finite
7 Notethat swapping users within one BS may trigger swaps for users in other BS. This is due to the
frequency re-use.
number of iterations, since the number of subchannel pairs that can be swapped for
each MD pair is finite. Note that there are N2 = 12 N 2 − N subchannel pairs. Thus,

the worst-case complexity is of order O N 2 per user pair.

Algorithm 10 Swap Matching for MDs over the same BS

1: Input: Matching M
2: for Each MD assign to cn do
3: if there is a subchannel cn0 that can provide higher data rate then
4: if (cn , cn0 ) is a swap-blocking pair then
5: Perform a swap-operation;
6: Update matching M ;
7: end if
8: end if
9: end for
10: Output: Updated matching M .
7.4 Power Allocation

Since this work is focussed on minimizing the sum-latency, and does not optimize the
DL latency, problem Q3 can be simplified as
Q4: (7.21)
!
BIk BCk Bbh
min ∑ + + kbh (7.21a)
P k∈K Rul
k (P) F k C
s.t. 0 ≤ Pkul ≤ Pmax , ∀k ∈ K . (7.21b)
BCk Bbh
For each MD, when X,λ
λ are fixed, the term Fk + Ckbh is also fixed. Thus, Q4 can be
converted to
Q5: (7.22)
BIk
min ∑ ul (7.22a)
P k∈K Rk (P)
s.t. 0 ≤ Pkul ≤ Pmax , ∀k ∈ K . (7.22b)

7.4. POWER ALLOCATION 169
Problem Q5 is non-convex and NP-hard. Alternatively, it can be transformed as
Q6: (7.23)
min ∑ τk (7.23a)
P, τ k∈K
BIk
s.t. ≤ τk , ∀k ∈ K , (7.23b)
Rul
k (P)
0 ≤ Pkul ≤ Pmax , ∀k ∈ K . (7.23c)
Proposition 1. If (P∗ , τ ∗ ) is the solution of Problem Q6, then there exists λ ∗ = [λ1 , λ2 , ..., λk ]
such that P∗ satisfies the KKT conditions of the following problem upon setting λ = λ ∗
and τ = τ ∗ ,
Q7: (7.24)
min
P
∑ λk (BIk − τk Rulk (P)) (7.24a)
k∈K
s.t. 0 ≤ Pkul ≤ Pmax , ∀k ∈ K . (7.24b)
Proof : The Lagrangian of Problem Q6 is expressed as

I ul
L (λλ, P, τ) = ∑ τk + ∑ λk Bk − τk Rk (P) . (7.25)
k∈K k∈K
If (P∗ , τ ∗ ) is the solution of Problem Q6, then there exists λ ∗ satisfying the following
KKT conditions, ∀k ∈ K , such that
∂L (λ
λ, P, τ )
= 1 − λ∗k Rul ∗
k (P ) = 0, (7.26)
∂τk
∂L (λ
λ, P, τ ) ∂Rul (P∗ )
= − ∑ λ∗k τ∗k k = 0, (7.27)
∂Pk k∈K ∂P k
∗
λ∗k (BIk − τ∗k Rul
k (P)) = 0, (7.28)
∗
BIk − τ∗k Rul
k (P) ≤ 0, (7.29)
λ∗k ≥ 0, (7.30)
and
0 ≤ Pkul ≤ Pmax . (7.31)
According to (7.26), it can be verified that
1
λ∗k = > 0. (7.32)
Rul ∗
k (P )
Furthermore, (7.28) implies that
BIk
τ∗k = > 0. (7.33)
Rul ∗
k (P )
Lastly, since (7.26), (7.27) and (7.31) are also the KKT conditions of Problem Q7,
Proposition 1 is proved.
Based on Proposition 1, when λ = λ ∗ and τ = τ ∗ , P∗ satisfies the following con-

straints
1
λk = ul ∗ , ∀k ∈ K , (7.34)
Rk (P )
and
BIk
τk = , ∀k ∈ K . (7.35)
Rul ∗
k (P )
In this way, Problem Q6 is transformed into Q7, which can be solved in two steps.
Firstly, fix λ and τ , and obtain P by solving Problem Q7 via Algorithm 12, which
will be discussed shortly. Then, fix P, and update λ and τ via the modified Newton’s
algorithm until convergence [206, 207]. The process is given in Algorithm 11, where
∗
ρk (λk ) = λk Rul
k (P ) − 1, ∀k ∈ K , (7.36)
and
∗
κk (τk ) = τk Rul I
k (P ) − Bk , ∀k ∈ K . (7.37)
In order to solve Problem Q7, which is non-convex since the rate function Rulk (P)
is non-convex, it is transformed into a convex problem. Note that the rate function in
(7.6) can be more conveniently be written as (7.38).
 
Pkul hul
k,m,n
Rul
k (P) = B χul ul
k,m λk,n log2 1 + (7.38)
 
∑ ∑ λulj,n Pul 2
j h j,m,n + σ

m∈M n∈N ∑
j∈K , j6=k
7.4. POWER ALLOCATION 171
Algorithm 11 Optimal Power Allocation

1: Initialization: P(0) , t = 0, ζ ∈ (0, 1), ε ∈ (0, 1), calculate λk and τk by (7.34) and
(7.35), respectively.
2: repeat
3: Update P(t+1) via Algorithm 4;
4: Update λk and τk as follows

i(t+1) (t)
ζ ρk λk
(t+1) (t)
λk = λk − , ∀k ∈ K ,
Rul
k (P
(t+1) )
and
(t+1) (t)
ζi κk τk
(t+1) (t)
τk = τk − , ∀k ∈ K ,
Rul
k (P
(t+1) )
where i(t+1) is the smallest integer among i ∈ {1, 2, 3, ...} satisfying


(t)

ζi ρk λk 2
(t)
∑ ρk λk − Rul (P(t+1) ) 
k∈K k

(t)

ζi κk τk 2
(t)
+ ∑ κk τk − ul (t+1) 

k∈K Rk (P )
2
i 2
(t) 2 (t)
≤ 1 − εζ ∑ ρk λk + κk τk ;
k∈K
5: Set t = t + 1;
6: until the following conditions are satisfied:

(t) (t)
λk Rul
k P − 1 = 0, ∀k ∈ K ,
and
(t) (t)
τk Rul
k P − BIk = 0, ∀k ∈ K ;
7: Output: Optimal P∗ = P(t) .
Now, consider the lower-bound approximation [156]
log2 (1 + γ) ≥ µ1 log2 (γ) + µ2 , (7.39)
where γ ≥ 0, and the bound is tight for γ = γ̄. Moreover,
γ̄
µ1 = , (7.40)
γ̄ + 1
and
µ2 = log2 (1 + γ̄) − µ1 log2 (γ̄). (7.41)
ul
By using the variable substitution Pkul = 2Qk , the rate function can be lower-bounded
as (7.42).
ul
Rul
k (P) ≥ Rk (Q)
    
2 Qul
k hul
k,m,n
χul ul 
    
, B  + µ2,k 
∑ ∑ k,m λk,n µ1,k log2 

Qulj
λulj,n 2 h j,m,n + σ2
  
m∈M n∈N ∑
j∈K , j6=k
! !!
Qulj
=B ∑ ∑ χulk,m λulk,n µ1,k Qul ul
k + µ1,k log2 (hk,m,n ) − µ1,k log2 ∑ λulj,n 2 h j,m,n + σ2 + µ2,k
m∈M n∈N j∈K , j6=k
(7.42)
Hence, Problem Q7 can be rewritten as

Q8: (7.43)
ul
min L̄(Q) = ∑ λk (BIk − τk Rk (Q)) (7.43a)
Q k∈K
ul
s.t. 0 ≤ 2Qk ≤ Pmax , ∀k ∈ K . (7.43b)
ul
Remark 10. For the rate function Rk (Q) in (7.42), thenegative log-sum-exp
term
ul ul
is concave in Q [155]. Thus, Rk (Q) is concave and λk BIk − τk Rk (Q) is convex.
More importantly, the sum of the convex functions
is also convex
[157], and hence, the
ul
lower-bounded objective function ∑k∈K λk BIk − τk Rk (Q) is convex in Q. Also, the
constraint set of Problem Q8 is convex.
Accordingly, Problem Q8 can be solved optimally for fixed values of µ1,k and µ2,k
via any standard convex optimization package. By iteratively updating µ1,k and µ2,k
via (7.40) and (7.41), respectively, in which case P(t+1) in Algorithm 11 is obtained
via Algorithm 12. Hence, the global optimal solution of problem Q5 can be obtained.
The complexity of Algorithm 11 is mainly dependent on Step 3, as all the other
steps are based on explicit expressions. Since a convex optimization problem is solved
in each iteration in Algorithm 12, it has polynomial-time complexity [157]. As for
convergence, Algorithm 11 is guaranteed to converge in a finite number of iterations
[206].
Finally, a flow-chart of the proposed joint BS association, subchannel allocation,
and power control scheme is given in Fig. 7.2. Particularly, the proposed scheme
Algorithm 12 Solution of Problem Q8

1: Initialization: Set error tolerance ε ∈ (0, 1), iteration index t = 0, µ1,k = 1, µ2,k =
0, ∀k ∈ K , select a feasible Q(0) , and calculate L̄(Q(0) ).
2: repeat
3: Set t = t + 1;
(t) (t)
4: Update µ1,k and µ2,k by (7.40) and (7.41), ∀k ∈ K ;
5: Compute L̄(Q(t) ) by solving Problem Q8;
6: until L̄(Q(t) ) − L̄(Q(t−1) ) ≤ ε
(t)
7: Output: P? = 2Q
starts by performing BS association and subchannel allocation via the SPA algorithm
given in Algorithm 9. After that, if at least one swap-blocking pair is found, then
the corresponding MDs’ subchannels are swapped via Algorithm 10. Then, optimal
power allocation is applied using Algorithms 11 and 12. After the power allocation,
if a swap-blocking pair is found, Algorithm 10 is applied again, which is followed by
optimal power allocation, and so on. This is to ensure stability after power allocation
until no further swap-blocking pairs can be found8 .

In this section, the performance of the proposed DUDe access scheme is evaluated and
compared with benchmark schemes in terms of latency and energy-efficiency (EE).
The network energy-efficiency in the UL is determined as [78]
∑k∈K Rul
k
EE = ul
. (7.44)
∑k∈K ∑m∈M ∑n∈N Pk,m,n
In what follows, the proposed SPA algorithm along with swap matching (SM), and
optimal power allocation (OPA) are compared to the CUDA scheme, which is based
on biased RSRP for cell association, greedy (G) subchannel allocation [132]9 , and
FPC. Moreover, a decoupled access scheme, called Min-PL-F-FPC is also compared,
which is based on minimum path-loss (min-PL) cell association [2], greedy subchannel
allocation, and FPC. In the min-PL criterion, the MDs are connected in the UL to the
8 Generally speaking, it has been determined that the possibility of finding a swap-blocking pair to
perform swap matching is very low after the first loop.
9 In the greedy subchannel allocation algorithm, subchannels with high SINR are preferentially as-
signed to the MDs [132].

Figure 7.2: Flow-chart of the proposed scheme.
BS with the lowest path-loss. Particularly, a typical MD k ∈ K is associated with BS

m ∈ M in the UL if
−1
ul
Pk,mWm PLk,m −1 ≥ Pk,m
ul
0 Wm0 PLk,m0 , ∀m, m0 ∈ M , (7.45)
where W is the UL cell bias value, which is positive and refers to expanding the cov-
erage of the cells. Moreover, PLk,m is the UL path-loss between MD k and BS m.
Table 7.2 summarizes the evaluated resource allocation schemes, while the simulation
parameters are given in Table 7.3.
Fig. 7.3 depicts the relationship between the UL sum transmission latency and the
Table 7.2: Resource Allocation Schemes

Scheme Cell- Subchannel Al- Power Control
Association location
CUDA Biased RSRP Greedy FPC
Min-PL-G-FPC Min-PL Greedy FPC
SPA-FPC SPA FPC
SPA-SM-FPC SPA-SM FPC
SPA-SM-OPA SPA-SM OPA

Parameters MD MBS SBS
Maximum transmission power 23 dBm 46 dBm 30 dBm
Spatial density 250/km2 5/km 2 25/km2
Spatial distribution Uniform distribution
Lognormal shadowing µ = 0, σ = 4 dB [148]
Path-loss exponent 3 [148]
Operating frequency 2 GHz
UL bandwidth 5 MHz [84]
DL bandwidth 5 MHz [84]
Subcarrier spacing 15 kHz [84]
Number of subchannels 25
Noise spectral density -174 dBm/Hz
Target received power P0 -80 dBm
PL compensation factor w 0.7
Computation capacity N/A 36 GHz 3.6GHz [200]
Backhaul link capacity Cbh 10 Mbits/s [200]
Offloaded task size BIk (5 Mbits, 10 Mbits)
Output to input bits proportion αk 0.2 [189]
CPU cycles per bit βk 330, 960, or 1900 [198]
Mk 2 N/A N/A
νm N/A 2 0.8
4500
Min-PL-G-FPC
4000 CUDA
SPA-FPC
Sum UL Trassmission Latency (s)
3500 SPA-SM-FPC
SPA-SM-OPA
3000
2500
2000
1500
1000
500
0
3 3.5 4 4.5 5 5.5 6
Number of SBSs
Figure 7.3: UL sum transmission latency.
280
Min-PL-G-FPC
SPA-SM-FPC
260
SPA-SM-OPA
Sum UL Trassmission Latency (s)
240
220
200
180
160
140
120
3 3.5 4 4.5 5 5.5 6
Number of SBSs
Figure 7.4: UL sum transmission latency.

number of SBSs. As can be seen, the UL transmission latency of the CUDA scheme
is several times higher than all its decoupled counterparts. This is because the decou-
pled schemes shorten the MD-BS distance, and reduce the interference in the network.
Due to peer effects, the sum-latency of the SPA scheme decreases with the number
of SBSs, which demonstrates the importance of swap matching. Since the latency of
the CUDA and SPA-FPC schemes are significantly higher than other schemes, Fig.
7.4 focuses on the Min-PL-G-FPC, SPA-SM-FPC, and SPA-SM-OPA schemes. Par-
ticularly, the UL transmission latency of the SPA-SM-FPC is higher than the Min-PL
scheme, since some computation-intensive tasks are offloaded to the MBS cloudlet to
reduce the computation latency, but the communication latency increases at the same
time. Furthermore, with the aid of the proposed OPA scheme, the transmission latency
of SPA-SM-OPA decreases by as much as 20% and is lower than the latency of the
Min-PL-G-FPC scheme.
350
Min-PL-G-FPC
CUDA
SPA-FPC
300
Sum Computation Latency (s)
250
200
150
100
3 3.5 4 4.5 5 5.5 6
Number of SBSs
Figure 7.5: Sum computation latency of Min-PL-G-FPC, CUDA and SPA-FPC

schemes.
Fig. 7.5 illustrates the sum computation latency of the different schemes. Since
the SPA-FPC, SPA-SM-FPC, and SPA-SM-OPA schemes choose the same UL serving
BS for each MD, and have the same computation latency, only the SPA-FPC scheme
is considered. When the number of SBSs increases, there are more computational re-
sources, and so the computation latency decreases. It can be seen that the SPA-FPC
4500
Min-PL-G-FPC
4000 CUDA
SPA-FPC
3500 SPA-SM-FPC
SPA-SM-OPA
3000
Sum-latency (s)
2500
2000
1500
1000
500
0
3 3.5 4 4.5 5 5.5 6
Number of SBSs
Figure 7.6: Sum-latency.
scheme has the lowest computation latency, because most computation-intensive tasks
are offloaded to the MBS cloudlet that has high computation capability. Most of the
MDs are associated with the MBS by the CUDA scheme, and so the computation la-
tency for the CUDA scheme is also low. As for the Min-PL-G-FPC scheme, it offloads
most MDs to the SBSs and does not offload all the computation-intensive tasks to the
MBS cloudlet, thus its computation latency is the highest.
Fig. 7.6 illustrates the sum-latency of all schemes, which is the sum of UL and DL
transmission latency, backhaul latency, and computation latency. As the DL latency
of all the schemes is the same and the backhaul latency is very small compared to
communication and computation latency, the sum-latency mainly depends on the UL
transmission latency and computation latency. It is clear from Fig. 7.6 that the latency
of all the DUDe schemes is much lower than that of the CUDA scheme. The sum-
latency of the SPA-SM-OPA scheme is the lowest, and only 15% to 60% of the CUDA
scheme. According to Fig. 7.7, the sum-latency of SPA-SM-OPA scheme is around
15% lower than that of the Min-PL-G-FPC scheme.
Another important metric to consider in this scenario is the fairness metric. Fig.
7.8 and 7.9 utilizes the Jain’s fairness index [208] to evaluate the UL transmission
latency and computation latency fairness, respectively. As the BSs are uniformly dis-
tributed in our simulations, the Min-PL scheme that selects the UL serving BS from the
800
Min-PL-G-FPC
750 SPA-SM-FPC
SPA-SM-OPA
700
650
Sum Latency (s)
600
550
500
450
400
350
3 3.5 4 4.5 5 5.5 6
Number of SBSs
Figure 7.7: Sum-latency of Min-PL-G-FPC, SPA-SM-FPC and SPA-SM-OPA.
0.9
0.8
0.7
0.6
0.5
Min-PL-G-FPC
CUDA
0.4 SPA-FPC
SPA-SM-FPC
SPA-SM-OPA
0.3
3 3.5 4 4.5 5 5.5 6
Number of SBSs
Figure 7.8: Jain’s Fairness Index of UL Transmission Latency vs. number of SBSs.
0.65
Min-PL-G-FPC
CUDA
SPA-FPC
0.6
0.55
0.5
3 3.5 4 4.5 5 5.5 6
Number of SBSs
Figure 7.9: Jain’s Fairness Index of Computation Latency vs. number of SBSs.
perspective of the UL path-loss can guarantee the UL transmission fairness to the most
extent. The SPA scheme jointly considers the computation and transmission latency
and offloads the computationally-intensive tasks to the MBS, it may cause high trans-
mission latency to those MDs, so the fairness index is lower, but the value is still higher
than the traditional coupled UL and DL access (CUDA) scheme, especially when com-
bining SPA with swap matching, which helps reduce the inter-cell interference. As for
the computation latency fairness, since the CUDA scheme associates most MDs with
the MBS, and the limited SBS computation resources are allocated to the rest of the
MDs, the fairness index of CUDA scheme is the highest. The SPA scheme allocates
the computationally-intensive tasks to the MBSs and the others to the nearby SBSs,
which can also achieve similar fairness as the CUDA scheme10 . The Min-PL scheme
does not consider the computation capabilities of different kinds of BSs, the MDs near
the SBSs but have computationally-intensive tasks may suffer from high computation
latency, so its fairness index is the lowest.
Fig. 7.10 depicts the UL transmission EE of the different schemes, which increases
with the number of SBSs. It can be seen that the normalized EE of the Min-PL-G-FPC
scheme is the highest, since it associates each MD with their nearest BS; while the
10 SPA-FPC, SPA-SM-FPC and SPA-SM-OPA schemes have the same computation latency, as the
MD-BS pairs are the same in these schemes.
1
Min-PL-G-FPC
0.9 CUDA
SPA-FPC
Normalized UL Transmission EE SPA-SM-FPC
0.8 SPA-SM-OPA
0.7
0.6
0.5
0.4
0.3
0.2
3 3.5 4 4.5 5 5.5 6
Number of SBSs
Figure 7.10: Normalized UL transmission EE vs. number of SBSs.
EE of the CUDA scheme is the lowest, which associates most MDs with the MBS.
Furthermore, swap matching can improve the SINR of some MDs by swapping their
subchannels, and thus improves their EE. The EE of the SPA-SM-FPC scheme is twice
higher than the Min-PL-G-FPC scheme. The aim of the proposed power allocation
scheme is to reduce the network sum-latency, and so the transmission powers of some
MDs are increased, which reduces EE.
Fig. 7.11 depicts the 10th , 20th , 50th , 80th , and 90th percentile data rates per MD.
It can be seen that the CUDA scheme has the lowest data rates, especially the 10th per-
centile data rates, because it allocates too many MDs to the MBS; and for those MDs
at the MBS edge, their UL data rates are relatively low. The Min-PL scheme selects
the UL serving BS by the UL path-loss, which results in a balanced MD distribution
among BSs, and the data rate distribution is the most balanced (e.g. the 90th percentile
data rate of the Min-PL scheme is only 70% higher than it 10th percentile data rate).
The SPA-SM scheme has lower 10th , 20th , 50th percentile data rates as it offloads more
MDs to the MBS. However, its 80th and 90th percentile data rates are higher, which is
due to its better inter-cell interference control. Compared with the SPA-SPA-SM-FPC
scheme, the proposed optimal power allocation scheme (OPA) increases the 10th , 20th ,
50th percentile data rates at the expense of the 80th and 90th percentile data rates, which
reduces the overall UL transmission latency, as shown in Fig. 7.4.
106
3.5
Min-PL-G-FPC
CUDA
3 SPA-FPC
SPA-SM-FPC
UL Data Rate per MD (bps)
SPA-SM-OPA
2.5
1.5
0.5
0
Percentile
Figure 7.11: 10th , 20th , 50th , 80th , and 90th percentile MD data rates based on different
schemes.
7.6 Summary
This chapter has focused on exploring the utilization of DUDe in MEC networks. Ex-
isting research works considered BS association and subchannel allocation separately,
and studied MEC under UL/DL coupled single BS association. In contrast, this chapter
has provided a new perspective on resource allocation in MEC-enabled heterogeneous
networks. Our results have successfully demonstrated that the latency of all the DUDe
schemes are much lower than that of the CUDA scheme. Specifically, the network
sum-latency of the SPA-SM-OPA scheme is the lowest among all the DUDe schemes,
which is only 15% to 60% of the CUDA scheme, and is around 15% lower than that
of the Min-PL-G-FPC scheme. The fairness of the UL transmission latency of the
SPA-SM-OPA scheme is also higher than that of the CUDA scheme. The EE of the
SPA-SM-FPC scheme is two times higher than the CUDA scheme. The users’ data
rates are increased, especially for those cell edge MDs.
Chapter 8
Conclusions and Future Research

Directions
8.1 Conclusions
This thesis studies the application of the DUDe technique and explore the feasibility
of exploiting its potential within other cutting-edge communication schemes. Specifi-
cally, a DUDe BS association and subchannel allocation algorithm based on capacity
maximization were proposed in this thesis, and the application of DUDe on multi-
connectivity, mmWave communication, D2D communication, cellular-enabled UAV
communication, MEC, were investigated. It was demonstrated throughout that DUDe
offers consistently improved performance under the various scenarios considered.
Specifically, in Chapter 3, an adaptive decoupling and multi-BS association scheme
for the mmWave/UHF hybrid HetNets has been proposed, and its performance has
been compared its performance for a wide range of metrics with benchmark single and
dual connectivity alternatives. The results provide an insight into the performance of
DUDe combined with mmWave and multi-connectivity techniques. It was shown that
capacity-based cell association schemes (CDA, SMBA, CDA&SMBA) can achieve
higher data rates than path-loss based cell association. Dual connectivity does not
necessarily improve the sum-rate, especially when the density of UHF BSs is low, and
the advantage of decoupled association is less effective in dual connectivity, because
the UEs distribution among the BSs is more balanced in the dual connectivity scenario
than in the single connectivity scenario. In mmWave/UHF hybrid HetNets, since the
available mmWave bandwidth is much wider than UHF bandwidth, the network sum-
rate mostly depends on the density of mmWave BSs, but UHF SBSs and MBSs are
183
184 CHAPTER 8. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS
still important to provide umbrella coverage to guarantee a consistent service.

While, in Chapter 4, a joint DL/UL decoupled cell-association, subchannel allo-
cation and power control scheme for D2D-underlay HetNets has been proposed, and
its performance has been compared with state-of-the-art alternatives. In the proposed
scheme, the Min-PL based DUDe has been used for cell-association in the UL. A
greedy coloring and a modified Munkres algorithm have been employed to solve the
subchannel allocation problem. The D.C. programming is utilized to maximize the
network sum-rate while meeting UE maximum transmission power and minimum data
rate constraints. Numerical results show that the DUDe association schemes achieve
higher data rates than their coupled counterparts. The proposed subchannel allocation
scheme has been shown to outperform the Greedy and Random subchannel allocation
schemes, and power allocation using D.C. programming dramatically improves the
network sum-rate. It was also found that power optimization based on D.C. program-
ming tends to allocate more power to UEs with higher channel quality at the expense
of the UEs with lower channel quality, and it may decrease the EE if the power opti-
mization stage continues to allocate power to the UEs when the data rate improvement
is marginal.
On the other hand, in Chapter 5, a new access scheme in heterogeneous cellular
networks with UAVs has been proposed. Particularly, to meet the different require-
ments of UAV data links and CNPC links, as well as GUE uplinks and downlinks, we
separate the UAV CNPC links from the high-capacity data communication links, and
separate the GUE uplinks from the downlinks. All the CNPC signals are connected to
the GCSs operating in the L/C band, while the data signals are communicated to cellu-
lar BSs via the mmWave band. To further reduce interference between the UAVs and
GUEs, the GUE downlinks and uplinks are also decoupled in terms of operating fre-
quency and serving BS. Then, an optimal power allocation algorithm has been devised
to optimize the EE of the DUDe access scheme. Numerical results demonstrated that
the DUDe access scheme can achieve several times higher sum-rates in both UL and
DL than its coupled counterparts. The proposed power allocation scheme is shown to
achieve a 15% higher GUE sum-rate and twice higher GUE EE than the conventional
fractional power control scheme. Furthermore, it should be noted that for noise-limited
UAV communications, there is a trade-off between sum-rate and EE as simultaneously
maximizing both is unattainable.
In Chapter 6, two power allocation schemes based on QL and DQL have been
proposed to optimize the EE of the DUDe network, and their performances have been
8.2. FUTURE WORK 185
compared with the fractional power control scheme (which is applied in 4G and 5G
networks) and the optimal power allocation scheme we proposed in Chapter 5. The QL
and DQL algorithms have shown promising results in complex wireless environments.
It was shown that although reinforcement learning methods can achieve optimal results
in theory, practically for some applications with predominant dynamic environment
and have a limited time to execute the optimization process, QL and DQL with limited
number of iterations may not achieve the optimal results. However it may achieve a
near-optimal solution for EE as demonstrated in the results at Fig. 7.7 and Fig. 7.8.
Nonetheless, the machine learning methods have achieved better EE performance than
the baseline FPC method by around 100% and 70% for UHF band, and by around
160% and 130% for mmWave band respectively in compare to conventional fractional
power control.
Finally, in Chapter 7, a joint BS association and subchannel allocation algorithm
based on a student-project allocation (SPA) matching has been proposed, which takes
the communication and computational disparity of SBS and MBS cloudlets into con-
sideration and based on DUDe access. Then, an optimal power allocation algorithm is
proposed to minimize the sum-latency of the network MDs. The formulated prob-
lem is a sum-of-ratios problem, which is non-convex and NP-hard. To efficiently
and optimally tackle it, the proposed algorithm tightly approximates the problem as
a convex optimization problem, and successively solves it until convergence to the
global optimal power allocation solution. Numerical results have successfully demon-
strated that the latencies of the DUDe schemes are much lower than that of the coupled
scheme. Specifically, the network sum-latency of the SPA-SM-OPA scheme is the low-
est among all the DUDe schemes, which is only 15% to 60% of the CUDA scheme,
and is around 15% lower than that of the Min-PL-G-FPC scheme. The EE of the
SPA-SM-FPC scheme is two times higher than the CUDA scheme.
8.2 Future Work

The results provide an insight into the performance of DUDe combined with mmWave,
dual connectivity, and D2D, but considerably more work still needs to be done. Some
possible areas of future research are list as follows.
• We applied independent Q-learning in this paper to solve the multi-agent prob-

lem, where each UE independently learns its own policy, treating other agents
as part of the environment. It makes the environment become non-stationary
186 CHAPTER 8. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS
from the point of view of each UE, as it contains other UEs who are them-
selves learning, the convergence can not be guaranteed. Fortunately, substantial
empirical evidence has shown that independent Q-learning often works well in
practice [184]. As for DQL, we assume the states and actions of all UEs as
an input. Although these methods can solve the EE optimization problem, it is
necessary to investigate other multi-agent reinforcement learning methods. In-
terested reader may refer to [209] for a comprehensive overview in this field.
• In Chapter 7, we complete the BS association and subchannel allocation in one

step by SPA algorithm. Looking still further ahead, we may consider if we
can jointly solve the BS association, subchannel allocation and power control
problems, for example, using 3-dimensional matching algorithms. Moreover,
we assume each cloudlet’s computational resources are equally shared among all
tasks when two or more tasks are offloaded to the same cloudlet. More complex
computational resource allocation schemes can be investigated in the future.
• We have taken cell load, computing power and communication capability into
consideration when studying cell association. However, we have not taken ac-
count of the backhaul capacity of different BSs, which should be another con-
straint in cell association.
• Most of the proposed schemes in this thesis are centralized in nature hence re-
quire a control node. If there is no such node in the network, then distributed
alternatives must be sought.
• DUDe is still far from the engineering implementation. In this report, we as-
sumed DUDe can be achieved perfectly . However, in real networks, to achieve
DUDe, BSs are supposed to communicate with each other. This will bring addi-
tional signaling overhead to the network and need new protocols to support this
kind of communication.
• DUDe improves the overall network EE and fairness at the expense of some UEs
with very high EE by improving the EE of the majority of UEs with lower EE.
For a few special scenarios, the application of DUDe will affect the experience
of those customers who require extremely high data rate or low latency, thus a
compromise and correction mechanism should be introduced.
• As DUDe increases the access and handover frequency, it will offset the energy
saving to some degree. Further strategies to reduce energy consumption, such
8.2. FUTURE WORK 187
as the BS sleeping strategy, should be considered. Moreover, user mobility is

a big challenge in DUDe, which further increases the switching frequency, thus
should be investigated in the future.
• With the significant development of ultra-dense low Earth orbit (LEO) satellite
constellations, satellite access networks have shown their potential as an expan-
sion of terrestrial 5G and beyond cellular networks. UEs can transmit their data
directly to LEO satellites via C-band with low data rates, or using a terrestrial-
satellite terminal as an access point and transmit data to the satellite via Ka-
band with high data rates. In both cases, the LEO satellite forwards the received
data to the BS or the earth gateway station, which is connected to the core net-
work [210]. Unlike traditional SCells (TSC) with small coverage and limited
backhaul capacity connected to the core network via multihop backhaul links,
the LEO-based SCelles (LSC) have vast coverage and large backhaul capacity
supported by the Ka-band transmission. Explore the feasibility of DUDe for the
HetNet with LSC could be interesting.
Bibliography
[1] F. Boccardi, J. Andrews, H. Elshaer, M. Dohler, S. Parkvall, P. Popovski, and

S. Singh. Why to decouple the uplink and downlink in cellular networks and
how to do it. IEEE Communications Magazine, 54(3):110–117, March 2016.
[2] H. Elshaer, F. Boccardi, M. Dohler, and R. Irmer. Downlink and uplink decou-
pling: A disruptive architectural design for 5G networks. In 2014 IEEE Global
Communications Conference, pages 1798–1803, Dec 2014.
[3] Jeffrey G Andrews. Seven ways that HetNets are a cellular paradigm shift. IEEE
Communications Magazine, 51(3):136–144, 2013.
[4] Sandra Ondrusova and Dongwook Kim. Huawei’s “5G implementation guide-
lines: NSA option 3. https://www.gsma.com/futurenetworks/wiki/
5g-implementation-guidelines/. Accessed Nov. 27, 2020.
[5] Tim Hatt and Emanuel Kolta. 5G energy efficiencies: green is the new
black. https://data.gsmaintelligence.com/research/research/
research-2020/5g-energy-efficiencies-green-is-the-new-black.
Accessed Nov. 27, 2020.
[6] Mohammed W Baidas and Emad Alsusa. Power allocation, relay selection
and energy cooperation strategies in energy harvesting cooperative wireless net-
works. Wireless Communications and Mobile Computing, 16(14):2065–2082,
2016.
[7] Ritesh Madan, Jaber Borran, Ashwin Sampath, Naga Bhushan, Aamod Khan-
dekar, and Tingfang Ji. Cell association and interference coordination in hetero-
geneous LTE-A cellular networks. IEEE Journal on Selected Areas in Commu-
nications, 28(9):1479–1489, 2010.
188
BIBLIOGRAPHY 189
[8] Arman Shojaeifard, Khairi Ashour Hamdi, Emad Alsusa, Daniel KC So, Jie
Tang, and Kai-Kit Wong. Design, modeling, and performance analysis of multi-
antenna heterogeneous cellular networks. IEEE Transactions on Communica-
tions, 64(7):3104–3118, 2016.
[9] D Colombi, B Thors, and C Törnevik. Implications of EMF exposure limits on

output power levels for 5G devices above 6 GHz. IEEE Antennas and Wireless
Propagation Letters, 14:1247–1249, 2015.
[10] Zhilin Chen, Xueying Hou, and Chenyang Yang. Training resource allocation
for user-centric base station cooperation networks. IEEE Transactions on Ve-
hicular Technology, 65(4):2729–2735, 2015.
[11] Subramanya Chandrashekar, Andreas Maeder, Cinzia Sartori, Thomas Höhne,

Benny Vejlgaard, and Devaki Chandramouli. 5g multi-rat multi-connectivity ar-
chitecture. In 2016 IEEE International Conference on Communications Work-
shops (ICC), pages 180–186. IEEE, 2016.
[12] Yao Shi, Emad Alsusa, Aysha Ebrahim, and Mohammed Wael Baidas. Uplink
performance enhancement through adaptive multi-association and decoupling
in uhf-mmwave hybrid networks. IEEE Transactions on Vehicular Technology,
68(10):9735–9746, 2019.
[13] Azade Fotouhi, Haoran Qiang, Ming Ding, Mahbub Hassan, Lorenzo Galati
Giordano, Adrian Garcia-Rodriguez, and Jinhong Yuan. Survey on UAV cel-
lular communications: Practical aspects, standardization advancements, regu-
lation, and security challenges. IEEE Communications Surveys & Tutorials,
2019.
[14] Y. Shi, E. Alsusa, and M. W. Baidas. An energy-efficient decoupled access

scheme for cellular-enabled UAV communication systems. IEEE Systems Jour-
nal, 34(11):2851–2865, Nov. 2020.
[15] Imran Ashraf, Federico Boccardi, and Lester Ho. Sleep mode techniques for
small cell deployments. IEEE Communications Magazine, 49(8):72–79, 2011.
[16] 3GPP. Evolved Universal Terrestrial Radio Access (E-UTRA) and Evolved
Universal Terrestrial Radio Access Network (E-UTRAN); Overall description;
Stage 2 . Technical Report (TR) 36.300, 3rd Generation Partnership Project
(3GPP), 11 2011. Version 10.5.0.
190 BIBLIOGRAPHY
[17] Mahmoud Kamel, Walaa Hamouda, and Amr Youssef. Ultra-dense networks:
A survey. IEEE Communications Surveys & Tutorials, 18(4):2522–2545, 2016.
[18] Joseph Boccuzzi Michael. Femtocells: design & application. 2018.
[19] David Lopez-Perez, Ismail Guvenc, Guillaume De la Roche, Marios Koun-

touris, Tony QS Quek, and Jie Zhang. Enhanced inter-cell interference coordi-
nation challenges in heterogeneous networks. arXiv preprint arXiv:1112.1597,
2011.
[20] Assen Golaup, Mona Mustapha, and Leo Boonchin Patanapongpibul. Femtocell
access control strategy in UMTS and LTE. IEEE Communications Magazine,
47(9):117–123, 2009.
[21] Aruna Prem Bianzino, Claude Chaudet, Dario Rossi, and Jean-Louis Rougier.
A survey of green networking research. IEEE Communications Surveys & Tu-
torials, 14(1):3–20, 2010.
[22] David Lopez-Perez, Ismail Guvenc, and Xiaoli Chu. Mobility management
challenges in 3GPP heterogeneous networks. IEEE Communications Magazine,
50(12):70–78, 2012.
[23] Holger Claussen, Florian Pivit, and Lester TW Ho. Self-optimization of femto-
cell coverage to minimize the increase in core network mobility signalling. Bell
Labs Technical Journal, 14(2):155–183, 2009.
[24] Haijun Zhang, Xiangming Wen, Bo Wang, Wei Zheng, and Yong Sun. A novel
handover mechanism between femtocell and macrocell for LTE based networks.
In 2010 second international conference on communication software and net-
works, pages 228–231. IEEE, 2010.
[25] J. Sangiamwong, Y. Saito, N. Miki, T. Abe, S. Nagata, and Y. Okumura. Inves-

tigation on cell selection methods associated with inter-cell interference coordi-
nation in heterogeneous networks for LTE-advanced downlink. In Proc. of 17th
European Wireless 2011 - Sustainable Wireless Technologies, pages 1–6, April
2011.
[26] M. Sternad, T. Ottosson, A. Ahlen, and A. Svensson. Attaining both cov-

erage and high spectral efficiency with adaptive OFDM downlinks. In
BIBLIOGRAPHY 191
2003 IEEE 58th Vehicular Technology Conference. VTC 2003-Fall (IEEE Cat.
No.03CH37484), volume 4, pages 2486–2490 Vol.4, Oct 2003.
[27] Ekram Hossain, Mehdi Rasti, Hina Tabassum, and Amr Abdelnasser. Evolution
towards 5G multi-tier cellular wireless networks: An interference management
perspective. arXiv preprint arXiv:1401.5530, 2014.
[28] Sundeep Rangan, Theodore S Rappaport, and Elza Erkip. Millimeter

wave cellular wireless networks: Potentials and challenges. arXiv preprint
arXiv:1401.2560, 2014.
[29] I. A. Hemadeh, K. Satyanarayana, M. El-Hajjar, and L. Hanzo. Millimeter-

wave communications: Physical channel models, design considerations, an-
tenna constructions, and link-budget. IEEE Communications Surveys Tutorials,
20(2):870–913, 2018.
[30] Yaniv Azar, George N Wong, Kevin Wang, Rimma Mayzus, Jocelyn K Schulz,
Hang Zhao, Felix Gutierrez, DuckDong Hwang, and Theodore S Rappaport.
28 ghz propagation measurements for outdoor cellular communications using
steerable beam antennas in new york city. In 2013 IEEE international confer-
ence on communications (ICC), pages 5143–5147. IEEE, 2013.
[31] Theodore S Rappaport, James N Murdock, and Felix Gutierrez. State of the
art in 60-ghz integrated circuits and systems for wireless communications. Pro-
ceedings of the IEEE, 99(8):1390–1436, 2011.
[32] Stanley Chia, Max Gasparroni, and Patrick Brick. The next challenge for cellu-
lar networks: Backhaul. IEEE Microwave Magazine, 10(5):54–66, 2009.
[33] Eldad Perahia, Carlos Cordeiro, Minyoung Park, and L Lily Yang. IEEE 802.11
ad: Defining the next generation multi-Gbps wi-fi. In 2010 7th IEEE consumer
communications and networking conference, pages 1–5. IEEE, 2010.
[34] Steven J Vaughan-Nichols. Gigabit wi-fi is on its way. Computer, 43(11):11–14,

2010.
[35] Robert C Daniels, James N Murdock, Theodore S Rappaport, and Robert W

Heath. 60 ghz wireless: Up close and personal. IEEE Microwave magazine,
11(7):44–50, 2010.
192 BIBLIOGRAPHY
[36] Tuncer Baykas, Chin-Sean Sum, Zhou Lan, Junyi Wang, M Azizur Rahman,
Hiroshi Harada, and Shuzo Kato. Ieee 802.15. 3c: the first ieee wireless standard
for data rates over 1 gb/s. IEEE Communications Magazine, 49(7):114–121,
2011.
[37] Theodore S Rappaport, James N Murdock, and Felix Gutierrez. State of the art
in 60-GHz integrated circuits and systems for wireless communications. Pro-
ceedings of the IEEE, 99(8):1390–1436, 2011.
[38] James N Murdock and Theodore S Rappaport. Consumption factor and power-
efficiency factor: A theory for evaluating the energy efficiency of cascaded
communication systems. IEEE Journal on Selected Areas in Communications,
32(2):221–236, 2013.
[39] Zhao Qingling and Jin Li. Rain attenuation in millimeter wave ranges. In 2006
7th International Symposium on Antennas, Propagation & EM Theory, pages
1–4. IEEE, 2006.
[40] Theodore S Rappaport, Shu Sun, Rimma Mayzus, Hang Zhao, Yaniv Azar,
Kevin Wang, George N Wong, Jocelyn K Schulz, Mathew Samimi, and Fe-
lix Gutierrez. Millimeter wave mobile communications for 5G cellular: It will
work! IEEE access, 1:335–349, 2013.
[41] Zhouyue Pi and Farooq Khan. An introduction to millimeter-wave mobile

broadband systems. IEEE Communications Magazine, 49(6):101–107, 2011.
[42] Hua Wang, Claudio Rosa, and Klaus I Pedersen. Dual connectivity for LTE-
advanced heterogeneous networks. Wireless Networks, 22(4):1315–1328, 2016.
[43] Anna Zakrzewska, David López-Pérez, Stepan Kucera, and Holger Claussen.
Dual connectivity in LTE HetNets with split control-and user-plane. In 2013
IEEE Globecom Workshops (GC Wkshps), pages 391–396. IEEE, 2013.
[44] Hisham Elshaer, Federico Boccardi, Mischa Dohler, and Ralf Irmer. Load &
backhaul aware decoupled downlink/uplink access in 5G systems. In 2015 IEEE
International Conference on Communications (ICC), pages 5380–5385. IEEE,
2015.
[45] Roha Masroor, Mohammed Burhan Abrar, and Xin Gui. Device to device
(D2D) communication : Interference management perspective. 2017.
BIBLIOGRAPHY 193
[46] D. Feng, L. Lu, Y. Yuan-Wu, G. Y. Li, S. Li, and G. Feng. Device-to-device com-
munications in cellular networks. IEEE Communications Magazine, 52(4):49–
55, April 2014.
[47] A. Asadi, Q. Wang, and V. Mancuso. A survey on device-to-device communica-

tion in cellular networks. IEEE Communications Surveys Tutorials, 16(4):1801–
1819, Fourthquarter 2014.
[48] Yong Zeng, Jiangbin Lyu, and Rui Zhang. Cellular-connected UAV: Poten-
tial, challenges, and promising technologies. IEEE Wireless Communications,
26(1):120–127, 2018.
[49] S. Zhang, Y. Zeng, and R. Zhang. Cellular-enabled UAV communication: A

connectivity-constrained trajectory optimization perspective. IEEE Transac-
tions on Communications, 67(3):2580–2604, March 2019.
[50] G. Zhang, Q. Wu, M. Cui, and R. Zhang. Securing UAV communications via
joint trajectory and power control. IEEE Transactions on Wireless Communica-
tions, 18(2):1376–1389, Feb 2019.
[51] Helin Yang and Xianzhong Xie. Energy-efficient joint scheduling and resource
management for UAV-enabled multicell networks. IEEE Systems Journal,
14(1):363–374, 2019.
[52] Dingkun Hu, Qi Zhang, Quanzhong Li, and Jiayin Qin. Joint position, decod-
ing order, and power allocation optimization in UAV-based NOMA downlink
communications. IEEE Systems Journal, 14(2):2949–2960, Jun. 2019.
[53] Xiaobo Zhou, Qingqing Wu, Shihao Yan, Feng Shu, and Jun Li. UAV-enabled
secure communications: Joint trajectory and transmit power optimization. IEEE
Trans. Veh. Technol., 68(4):4069–4073, 2019.
[54] Shuowen Zhang, Yong Zeng, and Rui Zhang. Cellular-enabled UAV commu-
nication: A connectivity-constrained trajectory optimization perspective. IEEE
Transactions on Communications, 67(3):2580–2604, 2018.
[55] 3GPP. Enhanced LTE support for aerial vehicles. http://www.3gpp.org/

ftp//Specs/archive/36_series/36.777/. Accessed Aug. 15, 2019.
194 BIBLIOGRAPHY
[56] B. V. Der Bergh, A. Chiumento, and S. Pollin. LTE in the sky: Trading off prop-
agation benefits with interference costs for aerial nodes. IEEE Communications
Magazine, 54(5):44–50, May 2016.
[57] 3GPP. 5G; NR; base station (BS) radio transmission and reception. http:
//www.3gpp.org/. Accessed Feb. 28, 2019.
[58] C. Zhang, W. Zhang, W. Wang, L. Yang, and W. Zhang. Research challenges

and opportunities of UAV millimeter-wave communications. IEEE Wireless
Communications, 26(1):58–62, Feb. 2019.
[59] L. Zhang, H. Zhao, S. Hou, Z. Zhao, H. Xu, X. Wu, Q. Wu, and R. Zhang.
A survey on 5G millimeter wave communications for UAV-assisted wireless
networks. IEEE Access, 7:117460–117504, 2019.
[60] Ofir Nachum, Mohammad Norouzi, Kelvin Xu, and Dale Schuurmans. Bridging
the gap between value and policy based reinforcement learning. In Advances in
Neural Information Processing Systems, pages 2775–2785, 2017.
[61] China Unicom Network Technology Research Insti-

tute. China unicom edge computing technology white pa-
per. https://builders.intel.com/docs/networkbuilders/
china-unicom-edge-computing-technology-white-paper.pdf. Ac-
cessed Nov. 7, 2020.
[62] Pavel Mach and Zdenek Becvar. Mobile edge computing: A survey on architec-
ture and computation offloading. IEEE Communications Surveys & Tutorials,
19(3):1628–1656, 2017.
[63] Yuyi Mao, Changsheng You, Jun Zhang, Kaibin Huang, and Khaled B Letaief.
A survey on mobile edge computing: The communication perspective. IEEE
Communications Surveys & Tutorials, 19(4):2322–2358, 2017.
[64] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimiza-
tion. Cambridge university press, 2004.
[65] Aylwin. 3g & 4g frequency list. https://support.alwaysonlinewireless.

com/. Accessed Feb. 28, 2019.
BIBLIOGRAPHY 195
[66] D Colombi, B Thors, and C Törnevik. Implications of emf exposure limits on

output power levels for 5g devices above 6 ghz. IEEE Antennas and Wireless
Propagation Letters, 14:1247–1249, 2015.
[67] Sarabjot Singh, Xinchen Zhang, and Jeffrey G Andrews. Joint rate and SINR
coverage analysis for decoupled uplink-downlink biased cell associations in
HetNets. IEEE Transactions on Wireless Communications, 14(10):5360–5373,
2015.
[68] Katerina Smiljkovikj, Petar Popovski, and Liljana Gavrilovska. Analysis of the
decoupled access for downlink and uplink in wireless heterogeneous networks.
IEEE Wireless Communications Letters, 4(2):173–176, 2015.
[69] Lan Zhang, Weili Nie, Gang Feng, Fu-Chun Zheng, and Shuang Qin. Uplink
performance improvement by decoupling uplink/downlink access in HetNets.
IEEE Transactions on Vehicular Technology, 66(8):6862–6876, 2017.
[70] H. Elshaer, M. N. Kulkarni, F. Boccardi, J. G. Andrews, and M. Dohler. Down-

link and uplink cell association with traditional macrocells and millimeter wave
small cells. IEEE Transactions on Wireless Communications, 15(9):6244–6258,
Sep. 2016.
[71] Osama Waqar Bhatti, Haris Suhail, Uzair Akbar, Syed Ali Hassan, Haris Per-
vaiz, Leila Musavian, and Qiang Ni. Performance analysis of decoupled cell
association in multi-tier hybrid networks using real blockage environments. In
2017 13th International wireless communications and mobile computing con-
ference (IWCMC), pages 62–67. IEEE, 2017.
[72] M. A. Lema, E. Pardo, O. Galinina, S. Andreev, and M. Dohler. Flexible dual-

connectivity spectrum aggregation for decoupled uplink and downlink access in
5G heterogeneous systems. IEEE Journal on Selected Areas in Communica-
tions, 34(11):2851–2865, Nov 2016.
[73] Jeffrey G Andrews, François Baccelli, and Radha Krishna Ganti. A tractable
approach to coverage and rate in cellular networks. IEEE Transactions on com-
munications, 59(11):3122–3134, 2011.
[74] M. K. Samimi, G. R. MacCartney, S. Sun, and T. S. Rappaport. 28 GHz

millimeter-wave ultrawideband small-scale fading models in wireless channels.
196 BIBLIOGRAPHY
In 2016 IEEE 83rd Vehicular Technology Conference (VTC Spring), pages 1–6,
May 2016.
[75] Mathworks. Multipath fading channel. https://www.mathworks.com. Ac-

cessed Feb. 28, 2019.
[76] N. Kolehmainen, J. Puttonen, P. Kela, T. Ristaniemi, T. Henttonen, and

M. Moisio. Channel quality indication reporting schemes for UTRAN long
term evolution downlink. In Proc. of IEEE Vehicular Technology Conference
(VTC) - Spring, pages 2522–2526, May 2008.
[77] M. Shi, K. Yang, C. Xing, and R. Fan. Decoupled heterogeneous networks with
millimeter wave small cells. IEEE Transactions on Wireless Communications,
17(9):5871–5884, Sep. 2018.
[78] O. Arnold, F. Richter, G. Fettweis, and O. Blume. Power consumption model-

ing of different base station types in heterogeneous cellular networks. In Proc.
Future Network Mobile Summit, pages 1–8, June 2010.
[79] 3GPP. Physical layer procedures. https://www.etsi.org/. Accessed Feb.

28, 2019.
[80] S. Berger, B. Almeroth, V. Suryaprakash, P. Zanier, I. Viering, and G. Fet-

tweis. Dynamic range-aware uplink transmit power control in LTE networks:
Establishing an operational range for LTE’s open-loop transmit power control
parameters(α, p0 ). IEEE Wireless Communications Letters, 3(5):521–524, Oct
2014.
[81] A. Ebrahim, E. Alsusa, and W. Pramudito. Joint interference aware resource par-
titioning and cellular association for femtocell networks. In 2015 IEEE Global
Communications Conference (GLOBECOM), pages 1–5, Dec 2015.
[82] M. Fallgren. On the complexity of maximizing the minimum shannon capacity

in wireless networks by joint channel assignment and power allocation. In 2010
IEEE 18th International Workshop on Quality of Service (IWQoS), pages 1–7,
June 2010.
[83] Martin Schlueter. Midaco software performance on interplanetary trajectory

benchmarks. Advances in Space Research, 54:744–754, 08 2014.
BIBLIOGRAPHY 197
[84] Campos J. Understanding the 5G NR physical layer. https:

//www.keysight.com/upload/cmc_upload/All/Understanding_the_
5G_NR_Physical_Layer.pdf. Accessed Feb. 28, 2019.
[85] Ericsson. Designing for the future: the 5G NR physical layer. https:
//www.ericsson.com/en/ericsson-technology-review/archive/2017/
designing-for-the-future-the-5g-nr-physical-layer. Accessed Feb.
28, 2019.
[86] Wikipedia. Percentile. https://en.wikipedia.org/wiki/Percentile. Ac-

cessed Oct. 31, 2018.
[87] Z. Kuang, G. Liu, G. Li, and X. Deng. Energy efficient resource allocation algo-
rithm in energy harvesting-based D2D heterogeneous networks. IEEE Internet
of Things Journal, 6(1):557–567, Feb. 2019.
[88] H. Tang and Z. Ding. Mixed mode transmission and resource allocation
for D2D communication. IEEE Transactions on Wireless Communications,
15(1):162–175, Jan. 2016.
[89] H. Chour, E. A. Jorswieck, F. Bader, Y. Nasser, and O. Bazzi. Global opti-

mal resource allocation for efficient FD-D2D enabled cellular network. IEEE
Access, 7:59690–59707, 2019.
[90] Klaus Doppler, Chia-Hao Yu, Cassio B Ribeiro, and Pekka Janis. Mode se-
lection for device-to-device communication underlaying an LTE-advanced net-
work. In 2010 IEEE Wireless Communication and Networking Conference,
pages 1–6. IEEE, 2010.
[91] Min-Hong Han, Byung-Gook Kim, and Jang-Won Lee. Subchannel and trans-
mission mode scheduling for D2D communication in OFDMA networks. In
2012 IEEE Vehicular Technology Conference (VTC Fall), pages 1–5. IEEE,
2012.
[92] Shangwen Xiang, Tao Peng, Ziyang Liu, and Wenbo Wang. A distance-
dependent mode selection algorithm in heterogeneous D2D and IMT-advanced
network. In 2012 IEEE Globecom Workshops, pages 416–420. IEEE, 2012.
[93] Rui Yin, Guanding Yu, Caijun Zhong, and Zhaoyang Zhang. Distributed re-
source allocation for D2D communication underlaying cellular networks. In
198 BIBLIOGRAPHY
2013 IEEE International Conference on Communications Workshops (ICC),

pages 138–143. IEEE, 2013.
[94] Chen Xu, Lingyang Song, Zhu Han, Qun Zhao, Xiaoli Wang, and Bingli Jiao.
Interference-aware resource allocation for device-to-device communications as
an underlay using sequential second price auction. In 2012 IEEE international
conference on communications (ICC), pages 445–449. IEEE, 2012.
[95] Chen Xu, Lingyang Song, Zhu Han, Dou Li, and Bingli Jiao. Resource allo-
cation using a reverse iterative combinatorial auction for device-to-device un-
derlay cellular networks. In 2012 IEEE Global Communications Conference
(GLOBECOM), pages 4542–4547. IEEE, 2012.
[96] Lingyang Song, Dusit Niyato, Zhu Han, and Ekram Hossain. Game-theoretic
resource allocation methods for device-to-device communication. IEEE Wire-
less Communications, 21(3):136–144, 2014.
[97] Mohammad Zulhasnine, Changcheng Huang, and Anand Srinivasan. Efficient

resource allocation for device-to-device communication underlaying LTE net-
work. In 2010 IEEE 6th International conference on wireless and mobile com-
puting, networking and communications, pages 368–375. IEEE, 2010.
[98] Pekka Janis, Visa Koivunen, Cassio Ribeiro, Juha Korhonen, Klaus Doppler,
and Klaus Hugl. Interference-aware resource allocation for device-to-device
radio underlaying cellular networks. In VTC Spring 2009-IEEE 69th Vehicular
Technology Conference, pages 1–5. IEEE, 2009.
[99] Bin Wang, Li Chen, Xiaohang Chen, Xin Zhang, and Dacheng Yang. Resource
allocation optimization for device-to-device communication underlaying cellu-
lar networks. In 2011 IEEE 73rd vehicular technology conference (VTC Spring),
pages 1–6. IEEE, 2011.
[100] Chia-Hao Yu, Klaus Doppler, Cassio B Ribeiro, and Olav Tirkkonen. Resource
sharing optimization for device-to-device communication underlaying cellular
networks. IEEE Transactions on Wireless communications, 10(8):2752–2763,
2011.
[101] Anders Gjendemsjø, David Gesbert, Geir E Øien, and Saad G Kiani. Binary
power control for sum rate maximization over multiple interfering links. IEEE
Transactions on Wireless Communications, 7(8):3164–3173, 2008.
BIBLIOGRAPHY 199
[102] C-H Yu, Olav Tirkkonen, Klaus Doppler, and Cássio Ribeiro. Power optimiza-
tion of device-to-device communication underlaying cellular communication.
In 2009 IEEE international conference on communications, pages 1–5. IEEE,
2009.
[103] H Wang and X Chu. Distance-constrained resource-sharing criteria for device-

to-device communications underlaying cellular networks. Electronics letters,
48(9):528–530, 2012.
[104] Sami Hakola, Tao Chen, Janne Lehtomaki, and Timo Koskela. Device-to-device
(D2D) communication in cellular network-performance analysis of optimum
and practical communication mode selection. In 2010 IEEE wireless communi-
cation and networking conference, pages 1–6. IEEE, 2010.
[105] Hongnian Xing and Sami Hakola. The investigation of power control schemes
for a device-to-device communication integrated into OFDMA cellular system.
In 21st Annual IEEE International Symposium on Personal, Indoor and Mobile
Radio Communications, pages 1775–1780. IEEE, 2010.
[106] Bo Peng, Chunjing Hu, Tao Peng, and Wenbo Wang. Optimal resource alloca-
tion for multi-D2D links underlying OFDMA-based communications. In 2012
8th International conference on wireless communications, networking and mo-
bile computing, pages 1–4. IEEE, 2012.
[107] Namyoon Lee, Xingqin Lin, Jeffrey G Andrews, and Robert W Heath. Power
control for D2D underlaid cellular networks: Modeling, algorithms, and analy-
sis. IEEE Journal on selected areas in communications, 33(1):1–13, 2014.
[108] Hafiz A Mustafa, Muhammad Z Shakir, Yusuf A Sambo, Khalid A Qaraqe,

Muhammad A Imran, and Erchin Serpedin. Spectral efficiency improvements in
HetNets by exploiting device-to-device communications. In 2014 IEEE Globe-
com Workshops (GC Wkshps), pages 857–862. IEEE, 2014.
[109] Amal Algedir and Hazem H Refai. Adaptive D2D resources allocation un-
derlaying (2-tier) heterogeneous cellular networks. In 2017 IEEE 28th Annual
International Symposium on Personal, Indoor, and Mobile Radio Communica-
tions (PIMRC), pages 1–6. IEEE, 2017.
200 BIBLIOGRAPHY
[110] Hui Gao, Min Wang, and Tiejun Lv. Energy efficiency and spectrum efficiency
tradeoff in the D2D-enabled HetNet. IEEE Transactions on Vehicular Technol-
ogy, 66(11):10583–10587, 2017.
[111] H. Dai, Y. Huang, C. Li, K. Song, and L. Yang. Resource allocation for device-
to-device and small cell uplink communication networks. In 2016 IEEE Wire-
less Communications and Networking Conference, pages 1–6, April 2016.
[112] Francesco Malandrino, Zana Limani, Claudio Casetti, and Carla-Fabiana Chi-
asserini. Interference-aware downlink and uplink resource allocation in het-
nets with D2D support. IEEE Transactions on Wireless Communications,
14(5):2729–2741, 2015.
[113] Kai Wen, Yongli Chen, and Yunhan Hu. A resource allocation method for D2D
and small cellular users in HetNet. In 2017 3rd IEEE International Conference
on Computer and Communications (ICCC), pages 628–632. IEEE, 2017.
[114] Yifei Huang, Ali A Nasir, Salman Durrani, and Xiangyun Zhou. Mode selec-
tion, resource allocation, and power control for D2D-enabled two-tier cellular
network. IEEE Transactions on Communications, 64(8):3534–3547, 2016.
[115] Mukesh Kumar Giluka, M Sibgath Ali Khan, Vanlin Sathya, and A Antony
Franklin. Leveraging decoupling in enabling energy aware D2D communi-
cations. In 2016 IEEE International Conference on Advanced Networks and
Telecommunications Systems (ANTS), pages 1–6. IEEE, 2016.
[116] A. Celik, R. M. Radaydeh, F. S. Al-Qahtani, and M. Alouini. Joint interference

management and resource allocation for device-to-device (D2D) communica-
tions underlying downlink/uplink decoupled (DUDe) heterogeneous networks.
In Proc. of IEEE International Conference on Communications (ICC), pages
1–6, 2017.
[117] François Bourgeois and Jean-Claude Lassalle. An extension of the munkres

algorithm for the assignment problem to rectangular matrices. Communications
of the ACM, 14(12):802–804, 1971.
[118] Derrick Wing Kwan Ng, Ernest S Lo, and Robert Schober. Energy-efficient re-
source allocation in multi-cell OFDMA systems with limited backhaul capacity.
IEEE Transactions on Wireless Communications, 11(10):3618–3631, 2012.
BIBLIOGRAPHY 201
[119] Laurence A. Wolsey and George L. Nemhauser. Integer and combinatorial op-
timization. John Wiley & Sons, 1999.
[120] Y.-F. Liu and Y.-H. Dai. On the complexity of joint subcarrier and power alloca-
tion for multi-user OFDMA systems. IEEE Transactions on Signal Processing,
62(3):583–596, 2014.
[121] M. Coupechoux and J.-M. Kelif. How to set the fractional power control com-
pensation factor in LTE ? Proc. of 34th IEEE Sarnoff Symposium, pages 1–5,
May 2011.
[122] M. Kubale. Graph Colorings. Contemporary mathematics (American Mathe-

matical Society) v. 352. American Mathematical Society, 2004.
[123] James Munkres. Algorithms for the assignment and transportation problems.
Journal of the society for industrial and applied mathematics, 5(1):32–38, 1957.
[124] Yi. Cao. Munkres’ assignment algorithm, modified for rectangular ma-
trices. http://csclab.murraystate.edu/˜bob.pilgrim/445/munkres.
html. Accessed Feb. 28, 2019.
[125] Jin Kue Wong. A new implementation of an algorithm for the optimal assign-
ment problem: An improved version of munkres’ algorithm. BIT Numerical
Mathematics, 19(3):418–424, 1979.
[126] Rainer E Burkard, Mauro Dell’Amico, and Silvano Martello. Assignment prob-
lems. Springer, 2009.
[127] Hoang Tuy, Tuy Hoang, Tuy Hoang, Viêt-nam Mathématicien, Tuy Hoang, and
Vietnam Mathematician. Convex analysis and global optimization. Springer,
1998.
[128] Ha Hoang Kha, Hoang Duong Tuan, and Ha H Nguyen. Fast global optimal
power allocation in wireless networks by local DC programming. IEEE Trans-
actions on Wireless Communications, 11(2):510–515, 2011.
[129] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University

Press, 2003.
[130] M. Frank and P. Wolfe. An algorithm for quadratic programming. Naval Re-
search Logistics Quaterly, 3(1):95–110, Mar. 1956.
202 BIBLIOGRAPHY
[131] D. P. Palomar and Y. C.Eldar. Convex Optimization in Signal Processing and

Communications. Cambridge University Press, New York, USA, 2010.
[132] Y. Tao, J. Sun, and S. Shao. Radio resource allocation based on greedy algorithm
and successive interference cancellation in device-to-device (D2D) communi-
cation. In IET International Conference on Information and Communications
Technologies (IETICT), pages 452–458, Apr. 2013.
[133] MIDACO-Solver. Mixed integer distributed ant colony optimization. Online:

http://www.midaco-solver.com.
[134] W. Mei and R. Zhang. Uplink cooperative NOMA for cellular-connected UAV.
IEEE Journal of Selected Topics in Signal Processing, 13(3):644–656, June
2019.
[135] W. Wang, J. Tang, N. Zhao, X. Liu, X. Y. Zhang, Y. Chen, and Y. Qian. Joint
precoding optimization for secure SWIPT in UAV-aided NOMA networks.
IEEE Transactions on Communications, 68(8):5028–5040, 2020.
[136] N. Zhao, Y. Li, S. Zhang, Y. Chen, W. Lu, J. Wang, and X. Wang. Security en-
hancement for NOMA-UAV networks. IEEE Transactions on Vehicular Tech-
nology, 69(4):3994–4005, 2020.
[137] X. Liu, Y. Liu, Y. Chen, and L. Hanzo. Trajectory design and power control
for multi-uav assisted wireless networks: A machine learning approach. IEEE
Transactions on Vehicular Technology, 68(8):7957–7969, 2019.
[138] Y. Zeng and R. Zhang. Energy-efficient UAV communication with trajectory

optimization. IEEE Transactions on Wireless Communications, 16(6):3747–
3760, Jun. 2017.
[139] D. Yang, Q. Wu, Y. Zeng, and R. Zhang. Energy tradeoff in ground-to-UAV

communication via trajectory design. IEEE Transactions on Vehicular Technol-
ogy, 67(7):6721–6726, Jul. 2018.
[140] S. Zhang, Y. Zeng, and R. Zhang. Cellular-enabled UAV communication: A

connectivity-constrained trajectory optimization perspective. IEEE Transac-
tions on Communications, 67(3):2580–2604, Mar. 2019.
BIBLIOGRAPHY 203
[141] S. Yin, S. Zhao, Y. Zhao, and F. R. Yu. Intelligent trajectory design in UAV-
aided communications with reinforcement learning. IEEE Transactions on Ve-
hicular Technology, 68(8):8227–8231, Aug. 2019.
[142] K. Venugopal, M. C. Valenti, and R. W. Heath. Device-to-device millimeter

wave communications: Interference, coverage, rate, and finite topologies. IEEE
Transactions on Wireless Communications, 15(9):6175–6188, Sep. 2016.
[143] Ruiqian Ma, Weiwei Yang, Yu Zhang, and Songqing Wang. Secure on-off
transmission in UAV relay-assisted mmwave networks. Applied Sciences,
9(19):4138, 2019.
[144] Y. Zhu, G. Zheng, and M. Fitch. Secrecy rate analysis of UAV-enabled

mmWave networks using Matérn hardcore point processes. IEEE Journal on
Selected Areas in Communications, 36(7):1397–1409, July 2018.
[145] S. Singh, M. N. Kulkarni, A. Ghosh, and J. G. Andrews. Tractable model for

rate in self-backhauled millimeter wave cellular networks. IEEE Journal on
Selected Areas in Communications, 33(10):2196–2211, 2015.
[146] A. Al-Hourani, S. Kandeepan, and S. Lardner. Optimal LAP altitude for max-
imum coverage. IEEE Wireless Communications Letters, 3(6):569–572, Dec
2014.
[147] T. Bai and R. W. Heath. Coverage and rate analysis for millimeter-wave cellular
networks. IEEE Transactions on Wireless Communications, 14(2):1100–1114,
Feb. 2015.
[148] J. Chakareski, S. Naqvi, N. Mastronarde, J. Xu, F. Afghah, and A. Razi. An en-

ergy efficient framework for UAV-assisted millimeter wave 5G heterogeneous
cellular networks. IEEE Transactions on Green Communications and Network-
ing, 3(1):37–44, March 2019.
[149] David W Matolak and Ruoyu Sun. Unmanned aircraft systems: Air-ground
channel characterization for future applications. IEEE Vehicular Technology
Magazine, 10(2):79–85, 2015.
[150] D. W. Yun and W. C. Lee. LTE-TDD interference analysis in spatial, time and
frequency domain. In Proc. of Ninth International Conference on Ubiquitous
and Future Networks (ICUFN), pages 785–787, July 2017.
204 BIBLIOGRAPHY
[151] Hossein Shokri-Ghadikolaei, Carlo Fischione, Gabor Fodor, Petar Popovski,

and Michele Zorzi. Millimeter wave cellular networks: A MAC layer perspec-
tive. IEEE Transactions on Communications, 63(10):3437–3458, 2015.
[152] Jeffrey G Andrews, Tianyang Bai, Mandar N Kulkarni, Ahmed Alkhateeb, Ab-
hishek K Gupta, and Robert W Heath. Modeling and analyzing millimeter wave
cellular systems. IEEE Transactions on Communications, 65(1):403–430, 2016.
[153] 3GPP. User equipment radio transmission and reception. https:

//www.etsi.org/deliver/etsi_ts/138100_138199/13810101/15.
02.00_60/ts_13810101v150200p.pdf. Accessed Aug. 15, 2019.
[154] E Tejaswi and B Suresh. Survey of power control schemes for LTE uplink.
International Journal of Computer Science and Information Technologies (IJC-
SIT), 4(2):369–373, 2013.
[155] Alessio Zappone, Eduard Jorswieck, et al. Energy efficiency in wireless net-
works via fractional programming theory. Foundations and Trends in Commu-
nications and Information Theory, 11(3-4):185–396, 2015.
[156] J. Papandriopoulos and J. S. Evans. SCALE: a low-complexity distributed pro-

tocol for spectrum balancing in multiuser DSL networks. IEEE Transactions
on Information Theory, 55(8):3711–3724, Aug. 2009.
[157] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge

university press, 2004.
[158] W. Dinkelbach. On nonlinear fractional programming. Management Science,

13(7):492–498, Mar. 1967.
[159] Jean-Pierre Crouzeix, Jacques A. Ferland, and Van Hien Nyugen. Revisting
Dinkelbach-type algorithms for generalized fractional programs. OPSEARCH,
45(2):97–110, Jun. 2008.
[160] A. Zappone, E. Bjornson, L. Sanguinetti, and E. Jorswieck. Globally optimal

energy-efficient power control and receiver design in wireless networks. IEEE
Transactions on Signal Processing, 65(11):2844–2859, Jun. 2017.
[161] Fanzi Zeng, Zhenzhen Hu, Zhu Xiao, Hongbo Jiang, Siwang Zhou, Wenping
Liu, and Daibo Liu. Resource allocation and trajectory optimization for QoE
BIBLIOGRAPHY 205
provisioning in energy-efficient UAV-enabled wireless networks. IEEE Trans-

actions on Vehicular Technology, 69(7):7634–7647, 2020.
[162] Lijun Deng, Gang Wu, Jingwei Fu, Yizhong Zhang, and Yifu Yang. Joint re-
source allocation and trajectory control for UAV-enabled vehicular communi-
cations. IEEE Access, 7:132806–132815, 2019.
[163] Zhendong Li, Ying Wang, Man Liu, Ruijin Sun, Yuanbin Chen, Jun Yuan, and
Jiuchao Li. Energy efficient resource allocation for UAV-assisted space-air-
ground Internet of Remote Things networks. IEEE Access, 7:145348–145362,
2019.
[164] Y. Azar, G. N. Wong, K. Wang, R. Mayzus, J. K. Schulz, H. Zhao, F. Gutierrez,

D. Hwang, and T. S. Rappaport. 28 GHz propagation measurements for out-
door cellular communications using steerable beam antennas in New York city.
In Proc. of IEEE International Conference on Communications (ICC), pages
5143–5147, June 2013.
[165] Bin Li, Zesong Fei, and Yan Zhang. Uav communications for 5g and beyond:
Recent advances and future trends. IEEE Internet of Things Journal, 6(2):2241–
2263, 2018.
[166] Jingjing Cui, Yuanwei Liu, and Arumugam Nallanathan. Multi-agent reinforce-
ment learning-based resource allocation for UAV networks. IEEE Transactions
on Wireless Communications, 19(2):729–743, Feb. 2020.
[167] Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduc-

tion. MIT press, 2018.
[168] C. J. C. H. Watkins and P. Dayan. Q-learning. Machine Learning, 8(3):279–292,

May 1992.
[169] V. Mnih and et al. Human-level control through reinforcement learning. Nature,
518(7540):529–533, 2015.
[170] C. Zhang, P. Patras, and H. Haddadi. Deep learning in mobile and wireless
networking: A survey. IEEE Communications Surveys & Tutorials, 21(3):2224–
2287, Mar. 2019.
206 BIBLIOGRAPHY
[171] N. C. Luong, D. T. Hoang, S. Gong, D. Niyato, P. Wang, Y.-C. Liang, and

D. I. Kim. Applications of deep reinforcement learning in communications and
networking: A survey. IEEE Communications Surveys & Tutorials, 21(4):3133–
3174, May 2019.
[172] G. L. Santos, P. T. Endo, D. Sadok, and J. Klener. When 5G meets deep learn-
ing: A systematic review. Algorithms, 13(208):1–34, Aug. 2020.
[173] 3GPP. Enhancement for Unmanned Aerial Vehicles; Stage 1. Technical Report
(TR) 22.829, 3rd Generation Partnership Project (3GPP), 09 2019. Version
17.1.0.
[174] U. Challita, W. Saad, and C. Bettstetter. Interference management for cellular-

connected UAVs: A deep reinforcement learning approach. IEEE Transactions
on Wireless Communications, 18(4):2125–2140, Apr. 2019.
[175] H. Qi, Z. Hu, H. Huang, X. Wen, and Z. Lu. Energy efficient 3-D UAV con-
trol for persistent communication service and fairness: A deep reinforcement
learning approach. IEEE Access, 8:53172–53184, Mar. 2020.
[176] C. H. Liu, Z. Chen, J. Tang, J. Xu, and X. Piao. Energy-efficient UAV control
for effective and fair communication coverage: A deep reinforcement learning
approach. IEEE Journal on Selected Areas in Communications, 36(9):2059–
2070, Sept. 2018.
[177] Hasini Viranga Abeywickrama, Ying He, Eryk Dutkiewicz, Beeshanga Abewar-
dana Jayawickrama, and Markus Mueck. A reinforcement learning approach
for fair user coverage using uav mounted base stations under energy constraints.
IEEE Open Journal of Vehicular Technology, 1:67–81, 2020.
[178] Jin Qiu, Jiangbin Lyu, and Liqun Fu. Placement optimization of aerial base sta-
tions with deep reinforcement learning. In ICC 2020-2020 IEEE International
Conference on Communications (ICC), pages 1–6. IEEE, 2020.
[179] Y. S. Nasir and D. Guo. Multi-agent deep reinforcement learning for dynamic
power allocation in wireless networks. IEEE Journal on Selected Areas in Com-
munications, 37(10):2239–2250, Oct. 2019.
[180] R. E. Bellman. Dynamic programming. Princeton University Press, 1957.

BIBLIOGRAPHY 207
[181] M. Kearns and S. Singh. Near-optimal reinforcement learning in polynomial

time. Machine Learning, 49:209–232, Nov. 2002.
[182] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative

agents. In Proceedings of the tenth international conference on machine learn-
ing, pages 330–337, 1993.
[183] Singh S, T. Jaakkola, M. L. Littman, and C. Szepesvari. Convergence results for

single-step on-policy reinforcement-learning algorithms. Machine Learning,
38:287–308, Mar. 2000.
[184] Laetitia Matignon, Guillaume J Laurent, and Nadine Le Fort-Piat. Indepen-

dent reinforcement learners in cooperative markov games: a survey regarding
coordination problems. Knowledge Engineering Review, 27(1):1–31, 2012.
[185] F Richard Yu and Ying He. Deep Reinforcement Learning for Wireless Net-
works. Springer, 2019.
[186] Fa-Long Luo. Machine Learning for Future Wireless Communications. John
Wiley & Sons, 2020.
[187] Chi Harold Liu, Zheyu Chen, Jian Tang, Jie Xu, and Chengzhe Piao. Energy-
efficient uav control for effective and fair communication coverage: A deep
reinforcement learning approach. IEEE Journal on Selected Areas in Commu-
nications, 36(9):2059–2070, 2018.
[188] Y. LeCun, Y. Bengio, and G. Hinton. Nature. Deep Learning, 521:436––444,

May 2015.
[189] Liang Huang, Xu Feng, Luxin Zhang, Liping Qian, and Yuan Wu. Multi-server
multi-user multi-task computation offloading for mobile edge computing net-
works. Sensors, 19(6):1446, 2019.
[190] Quoc-Viet Pham, Tuan Leanh, Nguyen H Tran, Bang Ju Park, and Choong Seon
Hong. Decentralized computation offloading and resource allocation for mobile-
edge computing: A matching game approach. IEEE Access, 6:75868–75885,
2018.
208 BIBLIOGRAPHY
[191] W. Almughalles, R. Chai, J. Lin, and A. Zubair. Task execution latency

minimization-based joint computation offloading and cell selection for mec-
enabled HetNets. In 2019 28th Wireless and Optical Communications Con-
ference (WOCC), pages 1–5, 2019.
[192] Fagui Liu, Zhenxi Huang, and Liangming Wang. Energy-efficient collaborative
task computation offloading in cloud-assisted edge computing for IoT sensors.
Sensors, 19(5):1105, 2019.
[193] Liang Huang, Xu Feng, Anqi Feng, Yupin Huang, and Li Ping Qian. Distributed
deep learning-based offloading for mobile edge computing networks. Mobile
Networks and Applications, pages 1–8, 2018.
[194] L. Huang, S. Bi, and Y. J. Zhang. Deep reinforcement learning for online
computation offloading in wireless powered mobile-edge computing networks.
IEEE Transactions on Mobile Computing, pages 1–1, 2019.
[195] O. Muñoz, A. Pascual-Iserte, and J. Vidal. Optimization of radio and computa-

tional resources for energy efficiency in latency-constrained application offload-
ing. IEEE Transactions on Vehicular Technology, 64(10):4738–4755, 2015.
[196] T. Q. Dinh, J. Tang, Q. D. La, and T. Q. S. Quek. Offloading in mobile edge

computing: Task allocation and computational frequency scaling. IEEE Trans-
actions on Communications, 65(8):3571–3584, 2017.
[197] Y. Wang, M. Sheng, X. Wang, L. Wang, and J. Li. Mobile-edge computing: Par-
tial computation offloading using dynamic voltage scaling. IEEE Transactions
on Communications, 64(10):4268–4282, 2016.
[198] Antti P Miettinen and Jukka K Nurminen. Energy efficiency of mobile clients
in cloud computing. HotCloud, 10(4):19, 2010.
[199] Surong Xiao, Chubo Liu, Kenli Li, and Keqin Li. System delay optimization
for mobile edge computing. Future Generation Computer Systems, 2020.
[200] Ali Al-Shuwaili and Ahmed Lawey. Achieving low-latency mobile edge com-
puting by uplink and downlink decoupled access in HetNets. arXiv preprint
arXiv:1809.04717, 2018.
BIBLIOGRAPHY 209
[201] David J Abraham, Robert W Irving, and David F Manlove. Two algorithms for
the student-project allocation problem. Journal of Discrete Algorithms, 5(1):73–
90, 2007.
[202] 3GPP. Evolved Universal Terrestrial Radio Access (E-UTRA); Physical layer
procedures. Technical Report (TR) 36.213, 3rd Generation Partnership Project
(3GPP), 04 2017. Version 14.2.0.
[203] 3GPP. 5G; NR; Physical layer procedures for control. Technical Report (TR)
38.213, 3rd Generation Partnership Project (3GPP), 05 2019. Version 15.5.0.
[204] Elizabeth Bodine-Baron, Christina Lee, Anthony Chong, Babak Hassibi, and
Adam Wierman. Peer effects and stability in matching markets. In International
Symposium on Algorithmic Game Theory, pages 117–129. Springer, 2011.
[205] Abdulkadir Celik, Ming-Cheng Tsai, Redha M Radaydeh, Fawaz S Al-Qahtani,

and Mohamed-Slim Alouini. Distributed user clustering and resource alloca-
tion for imperfect NOMA in heterogeneous networks. IEEE Transactions on
Communications, 67(10):7211–7227, 2019.
[206] Y Jong. An efficient global optimization algorithm for nonlinear sum-of-ratios

problem. Optimization Online, 2012.
[207] S. He, Y. Huang, L. Yang, and B. Ottersten. Coordinated multicell multiuser

precoding for maximizing weighted sum energy efficiency. IEEE Transactions
on Signal Processing, 62(3):741–751, 2014.
[208] Rajendra K Jain, Dah-Ming W Chiu, William R Hawe, et al. A quantitative

measure of fairness and discrimination. Eastern Research Laboratory, Digital
Equipment Corporation, Hudson, MA, 1984.
[209] Yaodong Yang and Jun Wang. An overview of multi-agent reinforcement learn-
ing from game theoretical perspective. arXiv preprint arXiv:2011.00583, 2020.
[210] Boya Di, Lingyang Song, Yonghui Li, and H Vincent Poor. Ultra-dense leo: In-
tegration of satellite access networks into 5g and beyond. IEEE Wireless Com-
munications, 26(2):62–69, 2019.
[211] L. Venturino, A. Zappone, C. Risi, and S. Buzzi. Energy-efficient scheduling

and power allocation in downlink OFDMA networks with base station coordi-
nation. IEEE Transactions on Wireless Communications, 14(1):1–14, Jan. 2015.
Appendix A
Proof in Chapter 5
A.1 Proof of Lemma 1

Proof: The proof proceeds by noting that for fixed values of αi,n and βi,n , ∀i ∈ I G , and
G G
∀n ∈ NUHF , functions R (Q) and P (Q) are concave
and convex, respectively. This
G
implies that −P (Q) is concave, and thus F λ(`) is concave in Q for each fixed
G
value of λ(`) . Moreover, Algorithm 5 successively solves a concave maximization

problem with a convex constraints set, and hence is guaranteed to converge to the
global optimal solution of problem T-GUEs-EE-MAX.
A.2 Proof of Lemma 2

Proof: The proof of Lemma 2 is two-fold: (1) showing that Algorithm 6 converges in
a finite number of iterations, and (2) illustrating that P? is the global optimal solution
that satisfies the Karush-Kuhn-Tucker (KKT) conditions for both problems, GUEs-
EE-MAX and T-GUEs-EE-MAX.
(`)
Firstly, note that in the `th iteration, the obtained solution P̂(`) = 2Q̂ maximizes
the objective function of problem T-GUEs-EE-MAX (via Algorithm 5), which is
based on the convexified constraints of problem GUEs-EE-MAX, and for fixed values
of αi,n and βi,n . Additionally, recall that Q̂` is the global optimal solution of problem
T-GUEs-EE-MAX, as per Lemma 1. Furthermore, and according to (5.17), the eval-
G

uated objective function EE P̂
c (`) G
= EE P̂ (`) , where the equality is due to the
updated values of αi,n and βi,n , ∀n ∈ NUHF and ∀i ∈ I G . Similarly, in the (`+1)th itera-
(`+1)
tion, P̂(`+1) = 2Q̂ maximizes the objective function of problem T-GUEs-EE-MAX,
210
A.2. PROOF OF LEMMA 2 211
c G P̂(`+1) ≥ EE c G P̂(`) . Moreover, and since EE c G P̂(`+1) is a

and hence EE
c G P̂(`+1) .

lower-bounded value of problem GUEs-EE-MAX, then EEG P̂(`+1) ≥ EE
By induction, one can verify that the following inequality holds [211]
c G P̂(`)
c G P̂(`+1) ≥ EE

· · · = EEG P̂(`+1) ≥EE
(A.1)
= EEG P̂(`) ≥ · · · ,
G

and hence EE P̂
c (`) monotonically increases in each iteration until convergence (i.e.
G G

when EE P
c (`) − EE P
c (`−1) ≤ ε).
Secondly, the KKT conditions of the reformulated problem T-GUEs-EE-MAX can
be obtained via the Lagrange function, which is defined as
G
L ({Qi,n } , {πi } , {ξi } , {ζi,n }) = EE (Q)
!

G
+ ∑ πi Ri (Q) − RG
min − ∑Gξi ∑ 2Qi,n − Pmax
G
(A.2)
i∈I G i∈I n∈NUHF
+ ∑ ∑G ζi,n2Qi,n ,
n∈NUHF i∈I
with {πi }, {ξi }, and {ζi,n } being the corresponding Lagrange multipliers. In turn, the
KKT conditions can be obtained by differentiating L ({Qi,n } , {πi } , {ξi } , {ζi,n }) with
respect to Pi,n , and setting the expression to zero. Particularly, the KKT conditions can
be shown to be

G
πi Ri (Q) − RG min = 0, ∀i ∈ I ,
G
!
ξi ∑ 2Qi,n − Pmax
G
= 0, ∀i ∈ I G ,
n∈NUHF
ζi,n 2Qi,n = 0, ∀i ∈ I G , ∀n ∈ NUHF ,

G
Ri (Q) ≥ RG G
min , ∀i ∈ I , (A.3)
∑ 2Qi,n ≤ Pmax
G
, ∀i ∈ I G ,
n∈NUHF
Pi,n ≥ 0, ∀i ∈ I G , ∀n ∈ NUHF ,
πi , ξi ≥ 0, ∀i ∈ I G ,
ζi,n ≥ 0, ∀i ∈ I G , ∀n ∈ NUHF ,
212 APPENDIX A. PROOF IN CHAPTER 5
while
G G G
!
∂EE (Q) 1 ∂R (Q) G ∂P (Q)
= G − EE (Q) . (A.4)
∂Qi,n P (Q) ∂Qi,n ∂Qi,n
G G
∂P (Q) ∂R (Q)
Moreover, ∂Qi,n = ln(2)2Qi,n , and ∂Qi,n is obtained as given as
G
∂R (Q) 1
= B αi,n − 2Qi,n |hi,m,n |2 ∑ α j,n Q (A.5)
∂Qi,n j∈I G , j6=i ∑ 2 k,n |hk,m,n |2 + σ2
k∈I G ,k6= j
On the other hand, the first order optimality conditions of problem GUEs-EE-
MAX must be verified in the Q-space by setting P = 2Q [211]. It can be easily verified
that the KKT conditions of problem GUEs-EE-MAX are identical to that of problem
T-GUEs-EE-MAX but with the substitution of P = 2Q . Particularly, it can be shown
that
∂EEG (Q) ∂RG (Q) ∂PG (Q)

1 G
= G − EE (Q) , (A.6)
∂Qi,n P (Q) ∂Qi,n ∂Qi,n
∂PG (Q) ∂RG (Q)
where ∂Qi,n = ln(2)2Qi,n , and ∂Qi,n is determined as given
∂RG (Q)

γ̄i,n γ̄ j,n 1
=B − 2Qi,n |hi,m,n |2 ∑ Q
∂Qi,n 1 + γ̄i,n j∈I G , j6=i
1 + γ̄ j,n ∑ 2 |hk,m,n |2 + σ2
k,n
k∈I G ,k6= j
(A.7)
where
2Q j,n |h j,m,n |2
γ̄ j,n = . (A.8)
∑ 2Qk,n |hk,m,n |2 + σ2
k∈I G ,k6= j
By comparing (A.5) and (A.7), it is evident that
γ̄i,n
αi,n = , (A.9)
1 + γ̄i,n
which is in agreement with (5.12). Moreover, βi,n can also be obtained via (5.12), i ∈
G G
I G , and ∀n ∈ NUHF . Additionally, ∂P∂Q(Q)
i,n
= ∂P∂Q(Q)
i,n
, and with the update of the αi,n and
G G
βi,n values, R (Q) → RG (Q), and hence, EE (Q) → EEG (Q). Lastly, since the same
A.2. PROOF OF LEMMA 2 213
KKT conditions apply to both GUEs-EE-MAX and T-GUEs-EE-MAX problems,

?
then the obtained solution P? = 2Q is the global optimal solution to problem GUEs-
EE-MAX.

5G AND BEYOND WIRELESS NETWORKS OPTIMIZATION Thesis

Uploaded by

Copyright:

Available Formats

5G AND BEYOND WIRELESS NETWORKS OPTIMIZATION Thesis

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

5G AND BEYOND WIRELESS NETWORKS OPTIMIZATION Thesis

Uploaded by

Copyright:

Available Formats

5G AND BEYOND WIRELESS

A THESIS SUBMITTED TO THE U NIVERSITY OF M ANCHESTER

2 Background and State of the Art 28

3 Capacity Enhancement Decoupled Access 58

4 Interference Management for D2D based on DUDe 86

5 DUDe Cellular-Enabled UAV Communications 114

6 Q-Learning and DQL based Power Control for UAVs 136

7 DUDe Access in Mobile Edge Computing 154

8 Conclusions and Future Research Directions 183

A Proof in Chapter 5 210

Word Count: 40480

2.1 The Parameters of Various Types of Nodes in the LTE-A Network . . 31

3.1 Cell Association Schemes . . . . . . . . . . . . . . . . . . . . . . . . 71

4.1 Resource Allocation Schemes . . . . . . . . . . . . . . . . . . . . . 105

5.1 Antenna parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

6.1 QL Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

7.1 Summary of MEC Scenarios and Objectives . . . . . . . . . . . . . . 155

1.1 Evolution of the mobile communication system. . . . . . . . . . . . . 20

2.1 A multi-tier network model. . . . . . . . . . . . . . . . . . . . . . . 29

3.1 Illustration of SMBA association scheme. . . . . . . . . . . . . . . . 67

4.1 Interference between MUEs, SUEs and DUEs . . . . . . . . . . . . . 93

6.1 Proposed power control scheme using DQL . . . . . . . . . . . . . . 144

7.1 A DUDe MEC model. . . . . . . . . . . . . . . . . . . . . . . . . . 156

AMC Adaptive Modulation and Coding

I would also like to thank my collaborators, Mutasem Q Hamdan. He is enthusiastic

My acknowledgment would not be complete without thanking my parents for their

Figure 1.1: Evolution of the mobile communication system.

DUDe has numerous advantages. BS decoupling shortens the distance between

Mobile edge computing: Low latency is also an important requirement for 5G

1.2 Main Contributions

• In Chapter 3, we investigate the feasibility of decoupling the UL and DL serv-

• In Chapter 4, we investigate the performance of DUDe in D2D-underlay Het-

which happens to be non-convex and NP-hard, and thus is computationally-

• In Chapter 5, we propose a novel DUDe based access scheme for cellular-

• In Chapter 7, we analyze the application of DUDe on MEC, the communication

1.3 Thesis Organization

of mmWave communication and multi-connectivity in DUDe. Chapter 4 proposes

1.4 Publication List

2. Shi Y, Alsusa E, and Baidas M W, ”Downlink-Uplink Decoupled Access

Background and State of the Art

2.1 Heterogeneous Networks

Figure 2.1: A multi-tier network model.

2.2 Small Cell

Figure 2.2: LTE-A heterogeneous network topology.

2.2.1 The Advantages of SCells

• Better support high-frequency resources: In order to achieve large-scale cover-

2.2.2 The Challenges of SCell

• Interference management: After deploying SCells to the MCell network, the

• Mobility management: Mobility management in MCell-Scell hybrid networks

• Energy efficiency management: Energy efficiency has always been an important

• Backhaul: SCell has a variety of backhaul methods, such as fiber, microwave,

2.3 Cell Association Criteria

2.3.1 Signal to Interference Plus Noise Power Ratio (SINR) based

iSINR = arg max γi , (2.1)

2.3.2 Reference Signal Received Power (RSRP) based Cell Associ-

iRSRP = arg max si , (2.2)

2.3.3 Reference Signal Received Quality (RSRQ) based Cell Asso-

iRSRQ = arg max ρi , (2.4)

2.3.4 Cell Range Extension

Figure 2.3: The phenomenon of the Uplink and downlink imbalance.