Scholarlink

Low earth orbit (LEO) satellite networks offer significant advantages for global communications, including low latency, extensive coverage, and high bandwidth capabilities. However, the highly dynamic nature of these networks, characterized by frequent topology changes and complex link dynamics, coupled with uneven traffic distribution across satellite coverage areas, presents unprecedented challenges for routing algorithms. This paper proposes a multimodal deep q-network (MMDQN) routing algorithm for LEO satellite networks that integrates graph neural networks (GNNs) and convolutional neural networks (CNNs) to enhance the model’s representational capabilities. By processing topological structures and traffic information through complementary neural architectures, our approach effectively captures the heterogeneous characteristics of satellite network data, enabling a more comprehensive understanding of network states. Extensive simulation results demonstrate that our algorithm significantly outperforms existing benchmark methods, achieving reduced end-to-end delay and lower packet rejection rates, particularly under highly imbalanced load conditions.

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

MMDQN: A Multimodal Deep Reinforcement Learning Approach for Load-Balanced Routing in LEO Satellite Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Qinghe Cai,
  • Chuxiong Sun,
  • Shuaijun Liu,
  • Lixiang Liu

摘要

Low earth orbit (LEO) satellite networks offer significant advantages for global communications, including low latency, extensive coverage, and high bandwidth capabilities. However, the highly dynamic nature of these networks, characterized by frequent topology changes and complex link dynamics, coupled with uneven traffic distribution across satellite coverage areas, presents unprecedented challenges for routing algorithms. This paper proposes a multimodal deep q-network (MMDQN) routing algorithm for LEO satellite networks that integrates graph neural networks (GNNs) and convolutional neural networks (CNNs) to enhance the model’s representational capabilities. By processing topological structures and traffic information through complementary neural architectures, our approach effectively captures the heterogeneous characteristics of satellite network data, enabling a more comprehensive understanding of network states. Extensive simulation results demonstrate that our algorithm significantly outperforms existing benchmark methods, achieving reduced end-to-end delay and lower packet rejection rates, particularly under highly imbalanced load conditions.