Scholarlink

<p>Existing research on spacecraft Pursuit–Evasion Game (PEG) mainly focuses on short-distance scenarios, while the long-distance PEG problem in non-coplanar elliptical orbits under high-dynamics and time constraints remains largely unexplored. This problem is characterized by a vast strategy space and significant dynamical effects, which poses a formidable challenge for spacecraft in terms of long-term planning and real-time decision-making. Therefore, this paper proposes a method based on the combination of Transformer network architecture and Deep Reinforcement Learning (DRL). This framework employs DRL for real-time sequential decision-making in adversarial environments, with an integrated Transformer module addressing DRL’s inherent challenges in long-term credit assignment and low exploration efficiency. This approach involves establishing a multi-constrained game model and designing precise reward functions. The uncertain environment is modeled, and simulation analysis is conducted using the Monte Carlo (MC) method. Simulation results demonstrate that the method in this paper has significant performance advantages over the original DRL, develops a more proactive and efficient game policy, and enhances the agent’s exploration ability and training efficiency. In uncertain environment, various indicators significantly outperform the baseline algorithm, demonstrating that the proposed method has stronger robustness and environmental adaptability. This study provides a viable DRL-based solution for long-distance spacecraft PEG.</p>

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

Multi-agent spacecraft long-distance pursuit–evasion game based on deep reinforcement learning and transformer

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Peng Di,
  • Zheng Lin,
  • Ye Yao,
  • Zengshan Yin

摘要

Existing research on spacecraft Pursuit–Evasion Game (PEG) mainly focuses on short-distance scenarios, while the long-distance PEG problem in non-coplanar elliptical orbits under high-dynamics and time constraints remains largely unexplored. This problem is characterized by a vast strategy space and significant dynamical effects, which poses a formidable challenge for spacecraft in terms of long-term planning and real-time decision-making. Therefore, this paper proposes a method based on the combination of Transformer network architecture and Deep Reinforcement Learning (DRL). This framework employs DRL for real-time sequential decision-making in adversarial environments, with an integrated Transformer module addressing DRL’s inherent challenges in long-term credit assignment and low exploration efficiency. This approach involves establishing a multi-constrained game model and designing precise reward functions. The uncertain environment is modeled, and simulation analysis is conducted using the Monte Carlo (MC) method. Simulation results demonstrate that the method in this paper has significant performance advantages over the original DRL, develops a more proactive and efficient game policy, and enhances the agent’s exploration ability and training efficiency. In uncertain environment, various indicators significantly outperform the baseline algorithm, demonstrating that the proposed method has stronger robustness and environmental adaptability. This study provides a viable DRL-based solution for long-distance spacecraft PEG.