Scholarlink

With the rapid development of cooperative multi-agent reinforcement learning (MARL), its application in real-world scenarios has garnered increasing attention in recent years. Current multi-agent algorithms largely rely on the success of deep neural networks; however, the vulnerability of deep neural networks to adversarial attacks significantly undermines the robustness of multi-agent systems. Traditional approaches often employ adversarial training to enhance the network's robustness against external attacks, but these methods overlook the interdependence and mutual influence among agents within the system, making it difficult to achieve optimal robustness. To leverage this characteristic, we proposes a defense module that significantly enhances system robustness by reconstructing the observations of attacked agents using the shared observational features among similar agents within the system. Building on this module, we further develop the ROMAT adversarial training framework, which generates adversarial attacks against agent observations progressively, aiding agents in learning robust cooperative strategies. Experimental results in the StarCraft and Predator-Prey environments demonstrate the effectiveness of ROMAT, showing a 4.3% performance improvement over traditional robust training methods.

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

Robust Multi-agent Reinforcement Learning Methods Based on Adversarial Training

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Fangfang Xu,
  • Fulong Jing,
  • Lei Yuan,
  • Zhong Liu,
  • Haiyin Piao,
  • Huiming Yang,
  • Riyao Dong,
  • Lin Li,
  • Simiao Fei

摘要

With the rapid development of cooperative multi-agent reinforcement learning (MARL), its application in real-world scenarios has garnered increasing attention in recent years. Current multi-agent algorithms largely rely on the success of deep neural networks; however, the vulnerability of deep neural networks to adversarial attacks significantly undermines the robustness of multi-agent systems. Traditional approaches often employ adversarial training to enhance the network's robustness against external attacks, but these methods overlook the interdependence and mutual influence among agents within the system, making it difficult to achieve optimal robustness. To leverage this characteristic, we proposes a defense module that significantly enhances system robustness by reconstructing the observations of attacked agents using the shared observational features among similar agents within the system. Building on this module, we further develop the ROMAT adversarial training framework, which generates adversarial attacks against agent observations progressively, aiding agents in learning robust cooperative strategies. Experimental results in the StarCraft and Predator-Prey environments demonstrate the effectiveness of ROMAT, showing a 4.3% performance improvement over traditional robust training methods.