Reinforcement Learning-Based Feeder Link Switchover for Gateway Load Balancing in NTN
- 주제(키워드) Reinforcement learning , LEO , Feeder link switchover
- 주제(DDC) 004.6
- 발행기관 아주대학교 일반대학원
- 지도교수 Jae-Hyun Kim
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 AI융합네트워크학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035644
- 본문언어 한국어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
This thesis proposes a multi-agent deep reinforcement learn- ing (MARL) framework for feeder link switchover (FLSO) to address gateway (GW) load balancing in low Earth orbit (LEO) non-terrestrial networks (NTN). Conventional heuris- tic FLSO strategies, which rely on simple distance or signal strength, often lead to severe load imbalance across ground GWs and inefficient resource utilization. To address this lim- itation, this thesis formulates the FLSO decision problem as a centralized multi-agent Markov decision process (MMDP) where ground GWs act as intelligent agents to jointly opti- mize link stability, signal quality, and traffic load distribu- tion. This thesis introduces a novel algorithm based on a multi-agent dueling double deep Q-Network architecture, en- hanced with prioritized experience replay (PER) and Polyak averaging. The agents learn to make cooperative switchover decisions using locally observable metrics, including satellite ephemeris, L3-filtered reference signal received power (RSRP), and real-time traffic load. The reward function is designed to maximize the network-wide utility by balancing conflicting objectives. Using extensive simulations based on a realistic Starlink Gen2 constellation and live GW locations, this thesis demonstrates the superiority of the proposed approach. Com- pared to 3rd generation partnership project (3GPP) heuris- tics and standard RL baselines, the proposed method achieves the most robust performance trade-off. It effectively suppresses FLSO comparable to conservative heuristics while achiev- ing high link quality and load fairness superior to aggressive signal-based methods. Crucially, it unlocks the network’s la- tent capacity, achieving the highest traffic throughput and connection utilization without causing congestion, thereby validating its potential for efficient autonomous mobility man- agement in next-generation NTN.
more목차
1 Introduction 1
1.1 Background and Motivation 1
1.2 Contributions 2
1.3 Overview 3
2 Related Works 4
2.1 LEO Satellite FLSO Schemes 4
2.2 Load Balancing in LEO Satellite Networks 9
2.3 RL for LEO Mobility Management 10
3 Proposed MARL-based FLSO Algorithm 12
3.1 System Model 12
3.2 Problem Formulation 16
3.3 Proposed Method 19
3.3.1 MMDP Modeling 19
3.3.2 Proposed Multi-Agent Dueling DDQN Algorithm 21
4 Performance Evaluation 27
4.1 Simulation Setup 27
4.2 Key Performance Indicators 31
4.3 Results and Discussion 33
5 Conclusion 48
References 49
