Scholarlink

<p>Effective organization of kill chains in modern air–sea combat necessitates decision-making that is not only intelligent and adaptive but also strictly trustworthy. While deep reinforcement learning (DRL) offers potential for dynamic coordination, its inherent opacity and sparse rewards often yield tactically irrational behaviors and unstable convergence. To bridge this gap, we propose the pattern-driven reinforcement learning (PDRL) framework, which synergizes a novel sliding window closed frequent stream mining algorithm with DRL. Instead of relying on static rules, PDRL dynamically extracts intrinsic causal temporal dependencies from combat streams and embeds them as behavioral constraints via a hybrid reward mechanism. This effectively aligns the opaque agent with tactically traceable logic. Simulations in high-fidelity anti-missile scenarios demonstrate that PDRL doubles the interception success rate (stabilizing at 0.40–0.45 vs. 0.20 baseline) and significantly maximizes the Integrity Score of protected assets. By bridging data-driven adaptability with knowledge-guided operational controllability, this framework provides a robust paradigm for reliable autonomous decision-making in complex kill webs.</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

Pattern-driven reinforcement learning for tactically traceable kill chain decision making in air combat

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Tongyao Mu,
  • Hanlin Li,
  • Yuqian Wu,
  • Miao Wang,
  • Guoqing Wang

摘要

Effective organization of kill chains in modern air–sea combat necessitates decision-making that is not only intelligent and adaptive but also strictly trustworthy. While deep reinforcement learning (DRL) offers potential for dynamic coordination, its inherent opacity and sparse rewards often yield tactically irrational behaviors and unstable convergence. To bridge this gap, we propose the pattern-driven reinforcement learning (PDRL) framework, which synergizes a novel sliding window closed frequent stream mining algorithm with DRL. Instead of relying on static rules, PDRL dynamically extracts intrinsic causal temporal dependencies from combat streams and embeds them as behavioral constraints via a hybrid reward mechanism. This effectively aligns the opaque agent with tactically traceable logic. Simulations in high-fidelity anti-missile scenarios demonstrate that PDRL doubles the interception success rate (stabilizing at 0.40–0.45 vs. 0.20 baseline) and significantly maximizes the Integrity Score of protected assets. By bridging data-driven adaptability with knowledge-guided operational controllability, this framework provides a robust paradigm for reliable autonomous decision-making in complex kill webs.