物理感知与课程学习辅助的低轨卫星通信混合波束成形
Physics Aware and Curriculum Learning Aided Hybrid Beamforming for LEO Satellite Communications
王心怡 1蒋挺1
作者信息
- 1. 北京邮电大学信息与通信工程学院,北京 100876
- 折叠
摘要
低地球轨道卫星网络凭借全球覆盖和低延迟优势,已成为下一代无线系统的核心。然而,密集星座的高动态移动性及同频干扰为波束成形设计带来了巨大挑战。本文提出一种基于课程学习的物理感知多智能体软行为者-评论者（CP-MASAC）混合波束成形算法。首先,将多星协作波束成形建模为去中心化部分可观测马尔可夫决策过程,并通过设计包含不确定性统计量的增强状态空间,提升了系统在非完美信道状态信息下的鲁棒性。其次,设计了连续语义动作空间,并引入正交匹配追踪算法作为物理感知投影层,实现了强化学习控制信号向底层硬件参数的有效映射。此外,通过几何预热、QoS增强及干扰抑制三个演进阶段设计课程学习策略,动态重塑优化目标以引导智能体收敛。仿真结果表明,CP-MASAC的平均信干噪比较传统多智能体深度强化学习算法提升了5.1至7.1倍；与WMMSE算法相比,其推理时间从90.48毫秒降至49.29毫秒,有效解决了传统迭代算法的实时性瓶颈。
Abstract
Low Earth Orbit (LEO) satellite communication networks have emerged as a cornerstone of next-generation wireless systems, owing to their advantages of ubiquitous coverage and low transmission latency. However, the high dynamic mobility characteristics of LEO satellites and the issue of co-channel interference in large-scale dense constellations pose significant challenges to beamforming design. To address these issues, this paper proposes a Curriculum Learning-based Physics-aware Multi-Agent Soft Actor-Critic (CP-MASAC) hybrid beamforming algorithm. First, this paper formulates the multi-satellite cooperative beamforming problem as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) and improves system robustness under imperfect CSI by designing an augmented state space incorporating uncertainty statistics. Second, a continuous semantic action space is designed, and the Orthogonal Matching Pursuit (OMP) algorithm is introduced as a physics-aware projection layer to achieve an effective mapping from the continuous control signals generated by reinforcement learning to the underlying hardware parameters. Furthermore, a Curriculum Learning (CL) strategy is designed. By utilizing three evolutionary stages—geometric warm-up, QoS enhancement, and interference suppression—this strategy dynamically reshapes the optimization landscape to guide the agents toward convergence. Finally, extensive simulations were conducted to validate the effectiveness of the proposed framework. CP-MASAC achieves a 5.1 to 7.1 times gain in average SINR over MADRL baselines. Compared to WMMSE, it reduces inference time from 90.48 ms to 49.29 ms, effectively resolving the real-time bottleneck of iterative approaches.关键词
卫星通信/多智能体强化学习/课程学习Key words
Satellite Communications/ Multi-agent reinforcement learning/ Curriculum learning引用本文复制引用
王心怡,蒋挺.物理感知与课程学习辅助的低轨卫星通信混合波束成形[EB/OL].(2026-02-12)[2026-02-14].http://www.paper.edu.cn/releasepaper/content/202602-71.学科分类
航空航天技术
评论