Scholarlink

As communication technology evolves, 6G requires seamless global connectivity, with low Earth orbit (LEO) satellite constellations playing a pivotal role. However, the dynamic and unpredictable nature of satellite movement and fluctuating traffic presents significant challenges for routing. Traditional routing algorithms struggle with poor adaptability and network congestion when applied to satellite systems. Reinforcement learning (RL)-based algorithms offer a promising solution, but there is a lack of an open-source simulator tailored to satellite networks that can thoroughly evaluate and optimize various constellation designs and communication setups. To address this, we developed and extended an open-source MA-DRL routing simulator based on event-driven simulation, supporting both traditional shortest-path algorithms and RL-based approaches like Q-routing and MA-DRL. Through extensive simulations across different constellation topologies (Kepler, Iridium Next, OneWeb, Starlink), we demonstrated that RL-based strategies, particularly MA-DRL, significantly improve end-to-end latency compared to traditional methods. These algorithms quickly adapt to changing network conditions caused by satellite movement, enhancing routing efficiency. This work provides a robust platform for satellite network routing research, facilitates the application of RL-based solutions and lays the foundation for future advancements in satellite communication systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on Multi-agent Deep Reinforcement Learning-Based Satellite Network Routing

  • 章节

  • 发表日期：

    2026年

查看全文
  • Weiwei Jiang,
  • Yang Zhang,
  • Zimeng Ni,
  • Ao Liu,
  • Haoyu Han,
  • Jianbin Mu,
  • Sai Huang

摘要

As communication technology evolves, 6G requires seamless global connectivity, with low Earth orbit (LEO) satellite constellations playing a pivotal role. However, the dynamic and unpredictable nature of satellite movement and fluctuating traffic presents significant challenges for routing. Traditional routing algorithms struggle with poor adaptability and network congestion when applied to satellite systems. Reinforcement learning (RL)-based algorithms offer a promising solution, but there is a lack of an open-source simulator tailored to satellite networks that can thoroughly evaluate and optimize various constellation designs and communication setups. To address this, we developed and extended an open-source MA-DRL routing simulator based on event-driven simulation, supporting both traditional shortest-path algorithms and RL-based approaches like Q-routing and MA-DRL. Through extensive simulations across different constellation topologies (Kepler, Iridium Next, OneWeb, Starlink), we demonstrated that RL-based strategies, particularly MA-DRL, significantly improve end-to-end latency compared to traditional methods. These algorithms quickly adapt to changing network conditions caused by satellite movement, enhancing routing efficiency. This work provides a robust platform for satellite network routing research, facilitates the application of RL-based solutions and lays the foundation for future advancements in satellite communication systems.