Scholarlink

<p>Robust multi-agent reinforcement learning (MARL) in noisy, dynamic environments with prevalent observation errors, execution perturbations, and non-stationarity presents a formidable challenge. This is exacerbated in structurally heterogeneous systems where diverse agent attributes demand sophisticated coordination. Existing MARL methods often address robustness and heterogeneity in isolation, limiting their practical resilience and generalization. We propose RhoMARL, a novel framework engineered to jointly confront these challenges via a dual-network design and synergistic training. RhoMARL’s Robust Dynamic Individual Value Network (RDIVN) bolsters local policy stability and information utilization efficiency through permutation-relevant modeling and dual regularization against observation and action disturbances. Concurrently, its Heterogeneous Fine-grained Mixing Value Network (HFMVN) enables scalable, expressive coordination by decomposing value functions across structurally distinct agent classes. These components are trained via a unified optimization scheme balancing task performance and resilience. Comprehensive experimental evaluations demonstrate that RhoMARL yields notable improvements over state-of-the-art baselines in robustness, convergence speed, and coordination efficacy, offering a promising unified framework for robust heterogeneous MARL.</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

RhoMARL: Robust Learning for Heterogeneous Multi-agent Systems in Dynamic Environments

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Zaipeng Xie,
  • Wenhao Fang,
  • Chentai Qiao,
  • Yiming Zhao,
  • WenZhan Song

摘要

Robust multi-agent reinforcement learning (MARL) in noisy, dynamic environments with prevalent observation errors, execution perturbations, and non-stationarity presents a formidable challenge. This is exacerbated in structurally heterogeneous systems where diverse agent attributes demand sophisticated coordination. Existing MARL methods often address robustness and heterogeneity in isolation, limiting their practical resilience and generalization. We propose RhoMARL, a novel framework engineered to jointly confront these challenges via a dual-network design and synergistic training. RhoMARL’s Robust Dynamic Individual Value Network (RDIVN) bolsters local policy stability and information utilization efficiency through permutation-relevant modeling and dual regularization against observation and action disturbances. Concurrently, its Heterogeneous Fine-grained Mixing Value Network (HFMVN) enables scalable, expressive coordination by decomposing value functions across structurally distinct agent classes. These components are trained via a unified optimization scheme balancing task performance and resilience. Comprehensive experimental evaluations demonstrate that RhoMARL yields notable improvements over state-of-the-art baselines in robustness, convergence speed, and coordination efficacy, offering a promising unified framework for robust heterogeneous MARL.