Scholarlink

To address the challenges of complex decision-making, algorithmic convergence difficulties, and poor stability in multi-objective coordinated combat operations with multiple unmanned aerial vehicles (UAVs), this paper proposes a hierarchical reinforcement learning-based decision optimization method for dynamic game scenarios. First, an XGBoost classification model is used to categorize four distinct tactical strategies. Subsequently, based on the classification outcomes, the corresponding multi-agent deep deterministic policy gradient (MADDPG) model is selected for strategy generation. Furthermore, a strategy distillation technique is introduced to significantly reduce computational overhead and enhance decision-making speed. Experimental results demonstrate that, compared to a single MADDPG model, the proposed hierarchical reinforcement learning approach effectively mitigates convergence issues in complex tasks and provides a novel solution framework for adversarial decision-making problems.

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

Hierarchical Reinforcement Learning-Based Multi-UAV Dynamic Game Cooperative Decision-Making Optimization in Game-Theoretic Frameworks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yicheng Shen,
  • Rongrong Hu,
  • Ya Zhang

摘要

To address the challenges of complex decision-making, algorithmic convergence difficulties, and poor stability in multi-objective coordinated combat operations with multiple unmanned aerial vehicles (UAVs), this paper proposes a hierarchical reinforcement learning-based decision optimization method for dynamic game scenarios. First, an XGBoost classification model is used to categorize four distinct tactical strategies. Subsequently, based on the classification outcomes, the corresponding multi-agent deep deterministic policy gradient (MADDPG) model is selected for strategy generation. Furthermore, a strategy distillation technique is introduced to significantly reduce computational overhead and enhance decision-making speed. Experimental results demonstrate that, compared to a single MADDPG model, the proposed hierarchical reinforcement learning approach effectively mitigates convergence issues in complex tasks and provides a novel solution framework for adversarial decision-making problems.