Asynchronous Multi-Agent Reinforcement Learning Aided Packet-Routing for Space-Air-Ground Integrated Networks
作者：刘栋 时间：2026年10月01日 13:43 点击数：
语言：English
发表刊物：IEEE Transactions on Mobile Computing
发表日期：2026年8月26日
摘要：
Routing in space-air-ground integrated networks (SAGIN) is highly challenging due to rapidly varying topology, cross-layer coupling, and inherently asynchronous operations that jointly cause severe fluctuations in link quality and end-to-end (E2E) performance. To address these challenges, this paper proposes Asynchronous Multi-agent Graph-based Routing (AMG-routing), an asynchronous multi-agent reinforcement learning (AMARL) framework for routing optimization in SAGIN. Unlike conventional synchronous MARL approaches that rely on globally aligned decision steps, AMG-routing formulates routing as a macro-action decentralized partially observable Markov decision process (MacDec-POMDP), enabling fully decentralized and event-driven coordination among heterogeneous nodes. A permutation-equivariant (PE) graph neural network (GNN)-based Q-network is further developed to ensure order-consistent neighbor-action mappings and improve policy generalization under dynamic topology variations. Moreover, cross-layer features reflecting queueing, MAC contention, and channel occupancy are incorporated into the observation space to capture dominant delay contributors. Simulation results demonstrate that AMG-routing consistently outperforms traditional and state-of-the-art MARL/GNN-based routing schemes in terms of E2E delay and throughput.
原文链接
