Multi-Agent Reinforcement Learning for Optimal Resource Allocation in Space-Air-Ground Integrated Networks

• Published in: IEEE internet of things journal p. 1
• Main Authors: Huynh, Dang Van, Khosravirad, Saeed R., Cotton, Simon L., Shin, Hyundong, Duong, Trung Q.
• Format: Journal Article
• Language: English
• Published: IEEE 2025
• Subjects: Dynamic scheduling; Edge computing; multi-agent reinforcement learning; non-terrestrial networks; Optimization; Quality of service; Reliability; Resource management; Satellites; Servers; Space-air-ground integrated networks; Training
• ISSN: 2327-4662; 2327-4662
• DOI: 10.1109/JIOT.2025.3598597

This paper addresses the problem of reliable task offloading in space-air-ground integrated network (SAGIN)-assisted edge computing systems, with the goal of maximising the ratio of tasks successfully offloaded and executed within quality-of-service (QoS) constraints. In the considered system, ground users offload computation tasks to a satellite-mounted edge server via unmanned aerial vehicles (UAVs) acting as relays. The formulated optimisation problem jointly considers task offloading portions and bandwidth allocations across ground-to-air and air-to-space links, subject to constraints on transmission rates, total bandwidth, energy budgets, and the satellite's computational capacity. The resulting problem is non-linear, non-convex, and mixed-integer, making it challenging to solve with traditional optimisation techniques. To this end, we propose a deep reinforcement learning (DRL)-based solution to learn optimal offloading and resource allocation policies in dynamic environments. Furthermore, to enhance scalability and decentralised coordination, we develop a multi-agent DRL framework that enables cooperative decision-making across UAVs. Simulation results demonstrate that both the single-agent and multi-agent approaches achieve stable training performance, and the proposed method improves the reliable task offloading ratio by up to two times compared to benchmark schemes, while also achieving more efficient resource utilisation in complex SAGIN scenarios.