Multiobjective Optimisation and Machine Learning based Resource Allocation for 5G Massive MIMO Systems
- 주제(키워드) Machine Learning , 5G Communication , Massive MIMO , Optimization , Artificial Intelligence
- 주제(DDC) 621.381
- 발행기관 아주대학교
- 지도교수 Wonsik Yoon
- 발행년도 2022
- 학위수여년월 2022. 8
- 학위명 박사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/ajou/000000032093
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Most of the existing works studies on fifth generation (5G) wireless communication are mainly focused on the quality of service (QoS) requirements, e.g., throughput, energy efficiency (EE), etc., but the situation becomes more complex in a multiobjective environment where several conflicting variables are present. This thesis aims to investigate and optimize the EE of massive multiple-input multiple-output (MIMO) communication network systems using a multiobjective optimization approach based on machine learning (ML)/artificial intelligence (AI) tools. Motivated by this fact, the main focus of this thesis is on solving multiobjective optimization problems (MOOP) in a 5G network. Various resource allocation strategies were used in this work to optimize the system performance of massive MIMO network. The multiobjective trade-offs between EE and spectral efficiency (SE) were investigated by Tchebycheff scalarization, joint selection of antenna and user scheduling (JASUS), switch on-off strategy and various multiobjective reinforcement learning (MORL) applied to remote radio heads (RRHs) based networks. The performance of promising 5G communication technologies cloud-radio access network (C-RAN), large scale-distributed antenna system (L-DAS), distributed and co-located MIMO systems operating under certain constraints were also verified in each case. Such analyses have hands-on significance for 5G technology that requires low latency, small power consumption, and/or the ability to simultaneously support a huge number of users. Recently, C-RAN has been considered a promising technology for 5G communication networks. The radio frequency signals from located antennas are collected by radio remote heads (RRHs) and transferred to the baseband units (BBUs) pool through front haul links. Such a centralized architecture facilitates global optimization of joint baseband processing unit (BPU) and resource allocation for every RRHs and users. For this purpose, this thesis utilizes novel schemes for energy efficient optimization of resources of 5G technologies in full-duplex (FD)-enabled massive MIMO networks. This thesis firstly extends the concept of FD communication network, where existing works have focused on optimization. In addition, this thesis also explores the use of machine learning tools such as multi-agent reinforcement learning (RL) and deep reinforcement learning (DRL) for solving MOOP in a multi-agent scenario. This thesis verified that by simultaneous transmitting and receiving, FD-enabled MIMO has the potential to significantly improve the SE of the half-duplex (HD) communication networks in MIMO networks. The proposed MOOP algorithms to study the EE-SE trade-off of the networks allow both uplink and downlink power savings, leading to an overall EE improvement in the wireless link. The findings of this thesis include enhancement in EE, capacity, QoS, and throughput of the given network using the ML and MOOP algorithms applied to the 5G MIMO systems.
more목차
CHAPTER 1. Introduction 1
1.1 Background 1
1.2 Research Contributions 3
1.3 Outline of Dissertation 5
CHAPTER 2. 5G Communication Systems 7
2.1 Background 7
2.2 MIMO Communication System 8
2.3 Large-scale Distributed Antenna System 9
2.4 Cloud-Radio Access Networks 11
2.4.1 C-RAN Architecture 11
2.4.2 Benefits of C-RAN 12
2.5 Energy Efficiency 15
2.6 Resource Allocation 16
2.6.1 Resource Allocation Techniques 18
2.7 Solution Methodology 21
2.7.1 Classical Optimization 22
2.7.2 Multiobjective Optimization 21
2.7.3 Reinforcement Learning (RL) 24
2.7.4 Multiobjective Reinforcement Learning (MORL) 25
2.7.5 Deep Reinforcement Learning (DRL) 26
2.7.6 Deep Q Network (DQN) 27
CHAPTER 3. Multiobjective Optimization Method for Resource Allocation in FD Large Distributed MIMO Systems 28
3.1 Background 28
3.2 System Model 31
3.2.1 Uplink Transmission 32
3.2.2 Downlink Transmission 34
3.3 Problem Formulation 37
3.4 Proposed Solution 38
3.4.1 JASUS Algorithm 39
3.4.2 Proposed MOOP Algorithm 43
3.4.3 Dual-Lagrangian Analysis and KKT Conditions 45
3.5 Simulation Results 47
3.6 Summary 57
CHAPTER 4. Reinforcement Learning Based Energy Saving and Resource Allocation in C-RAN enabled Massive MIMO 59
4.1 Background 59
4.2 Related Work 60
4.3 MORL Agent Environment 61
4.4 System Model 64
4.5 MORL Formulation 67
4.6 Proposed Method 70
4.7 Numerical Analysis 77
4.8 Summary 88
CHAPTER 5. Energy Efficient Power Allocation in Massive MIMO Based on Deep Reinforcement Learning 89
5.1 Background 89
5.2 System Model 92
5.2.1 Power Consumption 95
5.3 Problem Formulation 95
5.4 Multi-agent DRL Optimization Scheme 97
5.4.1 Fundamental of RL Method 98
5.4.2 Multi-agent DQN Frameworks 102
5.5 Simulation Result 105
5.6 Summary 110
CHAPTER 6. Conclusions and Future Work 111
6.1 Conclusions 111
6.2 Future Work 115
REFERENCES 116
