Scholarlink

<p>Multi-agent reinforcement learning (MARL) for complex task environments often relies on parameter sharing mechanisms or fully connected layers for dynamic task planning. Existing parameter-sharing methods suffer from agent behavior homogenization and limited strategic flexibility, while fully connected approaches fail to model the logical relationships between subtasks and agent groups, thereby lacking the ability to capture hierarchical dependencies among subtasks. To address these issues, we propose a three-layer decoupled Coordination Graph-based Dynamic Task Allocation (CG-DTA) framework, whose core innovation lies in explicitly decoupling subtask representation from task allocation. Specifically, we introduce the Graph Attention Synergistic Coordination Network (GASCN) to construct coordination-aware dynamic subtask representations, thereby alleviating the behavior homogenization problem caused by parameter-sharing schemes and improving strategy diversity. To improve algorithmic flexibility in complex environments, we propose an agent-subtask dynamic matching strategy that enhances coordination among agents within the same group by considering group restructuring and dynamic changes in subtasks. Additionally, we propose an end-to-end optimization strategy that achieves effective policy refinement from local observations while maintaining the global Q-function’s optimality. Experiments conducted in SMAC and predator-prey scenarios demonstrate that our algorithm increases the win rate gap from 24 to 36% compared to SOTA algorithms on progressively complex maps, highlighting its superiority in handling intricate scenarios. Simultaneously, our ablation studies within SMAC validate the effectiveness of the Group-Aware Cooperative Subtask Representation Module, Group Dynamic Task Allocation Module, and Group Policy Training Module, further confirming the effectiveness of the proposed decoupled architecture and the contribution of GASCN to robust dynamic task allocation in complex multi-agent environments.</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

Efficient dynamic task allocation in multi-agent systems via coordinated graph representations

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Yang Gao,
  • Yue Wang,
  • Lingyun Tian,
  • Jialun Li,
  • Dongguang Li

摘要

Multi-agent reinforcement learning (MARL) for complex task environments often relies on parameter sharing mechanisms or fully connected layers for dynamic task planning. Existing parameter-sharing methods suffer from agent behavior homogenization and limited strategic flexibility, while fully connected approaches fail to model the logical relationships between subtasks and agent groups, thereby lacking the ability to capture hierarchical dependencies among subtasks. To address these issues, we propose a three-layer decoupled Coordination Graph-based Dynamic Task Allocation (CG-DTA) framework, whose core innovation lies in explicitly decoupling subtask representation from task allocation. Specifically, we introduce the Graph Attention Synergistic Coordination Network (GASCN) to construct coordination-aware dynamic subtask representations, thereby alleviating the behavior homogenization problem caused by parameter-sharing schemes and improving strategy diversity. To improve algorithmic flexibility in complex environments, we propose an agent-subtask dynamic matching strategy that enhances coordination among agents within the same group by considering group restructuring and dynamic changes in subtasks. Additionally, we propose an end-to-end optimization strategy that achieves effective policy refinement from local observations while maintaining the global Q-function’s optimality. Experiments conducted in SMAC and predator-prey scenarios demonstrate that our algorithm increases the win rate gap from 24 to 36% compared to SOTA algorithms on progressively complex maps, highlighting its superiority in handling intricate scenarios. Simultaneously, our ablation studies within SMAC validate the effectiveness of the Group-Aware Cooperative Subtask Representation Module, Group Dynamic Task Allocation Module, and Group Policy Training Module, further confirming the effectiveness of the proposed decoupled architecture and the contribution of GASCN to robust dynamic task allocation in complex multi-agent environments.