Scholarlink

<p>Transformer architectures have achieved remarkable success in Multi-Agent Reinforcement Learning (MARL), particularly in graphics-driven simulation environments. However, their quadratic complexity limits scalability in scenarios with a large number of agents. In contrast, Mamba models offer an efficient alternative for sequence modeling but may underperform in tasks requiring agent coordination. In this paper, we propose the Hybrid Mamba-Transformer Multi-Agent Reinforcement Learning Algorithm (HMMA), which combines the representational strengths of Transformer-based attention with the computational efficiency of the Mamba sequence model. We evaluate HMMA in challenging StarCraft II scenarios, demonstrating faster convergence, higher mean episode rewards, and improved win rates compared to mainstream baselines. HMMA achieves a 9.5% improvement in normalized mean episode reward and a 19.9% increase in win rate over the Multi-Agent Transformer, highlighting its potential for scalable MARL in complex environments. The related code of our method is available at <a href="https://github.com/origin-orange/HMMA">https://github.com/origin-orange/HMMA</a>.</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

Hybrid Mamba-Transformer Multi-Agent Reinforcement Learning for scalable coordination in complex environments

  • 期刊论文

  • 发表日期：

    2025年6月30日

查看全文
  • Kai Chen,
  • Zhihua Chen,
  • Lei Dai,
  • Zhe Wang,
  • Xin Chen

摘要

Transformer architectures have achieved remarkable success in Multi-Agent Reinforcement Learning (MARL), particularly in graphics-driven simulation environments. However, their quadratic complexity limits scalability in scenarios with a large number of agents. In contrast, Mamba models offer an efficient alternative for sequence modeling but may underperform in tasks requiring agent coordination. In this paper, we propose the Hybrid Mamba-Transformer Multi-Agent Reinforcement Learning Algorithm (HMMA), which combines the representational strengths of Transformer-based attention with the computational efficiency of the Mamba sequence model. We evaluate HMMA in challenging StarCraft II scenarios, demonstrating faster convergence, higher mean episode rewards, and improved win rates compared to mainstream baselines. HMMA achieves a 9.5% improvement in normalized mean episode reward and a 19.9% increase in win rate over the Multi-Agent Transformer, highlighting its potential for scalable MARL in complex environments. The related code of our method is available at https://github.com/origin-orange/HMMA.