Scholarlink

In multi-agent reinforcement learning (MARL), the action-state space grows exponentially as the number of agents increases, resulting in the curse of dimensionality. This leads to a significant decline in sample efficiency and algorithm performance. In this paper, we propose a novel network model, GAIN, which integrates graph theory with action interactions. Using a self-attention mechanism, GAIN generates a weighted graph-based action interaction network that captures the complex interaction dynamics between agents, thereby mitigating the complexity of the action-state space. Experimental results demonstrate that, in six scenarios of StarCraft II, GAIN achieves a winning rate close to 100%, outperforming the state-of-the-art methods. Furthermore, ablation experiments in the Multi-Agent Particle Environment (MPE) from OpenAI show that GAIN substantially enhances system performance.

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

Improving the Efficiency and Scalability of Multi-agent Reinforcement Learning via Graph-Action Interaction Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuehai Wang,
  • Yajun Han,
  • Na Xing,
  • Xinyi Zhao

摘要

In multi-agent reinforcement learning (MARL), the action-state space grows exponentially as the number of agents increases, resulting in the curse of dimensionality. This leads to a significant decline in sample efficiency and algorithm performance. In this paper, we propose a novel network model, GAIN, which integrates graph theory with action interactions. Using a self-attention mechanism, GAIN generates a weighted graph-based action interaction network that captures the complex interaction dynamics between agents, thereby mitigating the complexity of the action-state space. Experimental results demonstrate that, in six scenarios of StarCraft II, GAIN achieves a winning rate close to 100%, outperforming the state-of-the-art methods. Furthermore, ablation experiments in the Multi-Agent Particle Environment (MPE) from OpenAI show that GAIN substantially enhances system performance.