Scholarlink

Coordinating agents in dynamic environments is a salient challenge in many fields, including robotics, economics and planning. Although many Multi-Agent Reinforcement Learning (MARL) algorithms achieve high levels of optimality in coordination tasks, there are significant practical issues with deploying them. Most MARL algorithms are limited in their scalability and when deployed, they require reliable communications for sharing state information. In real-world scenarios, communication is often unreliable or unavailable. MARL with Epistemic Priors (MARL-EP) reduces or eliminates the reliance on communication. We propose an extension to MARL-EP called Hierarchical MARL-EP (H-MARL-EP). H-MARL-EP overcomes scaling limitations while remaining communication-agnostic. The main contribution is to allow MARL to scale without reliance on communication by employing a multilevel hierarchical shared mental model. Our results show that H-MARL-EP can solve navigation tasks that are unsolvable with existing baseline methods, and scale to hundreds of agents.

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

Hierarchical Multi-agent Reinforcement Learning with Epistemic Priors for Scalable Communicationless Coordination of Teamable Agents

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Thayne T. Walker,
  • Jaime S. Ide

摘要

Coordinating agents in dynamic environments is a salient challenge in many fields, including robotics, economics and planning. Although many Multi-Agent Reinforcement Learning (MARL) algorithms achieve high levels of optimality in coordination tasks, there are significant practical issues with deploying them. Most MARL algorithms are limited in their scalability and when deployed, they require reliable communications for sharing state information. In real-world scenarios, communication is often unreliable or unavailable. MARL with Epistemic Priors (MARL-EP) reduces or eliminates the reliance on communication. We propose an extension to MARL-EP called Hierarchical MARL-EP (H-MARL-EP). H-MARL-EP overcomes scaling limitations while remaining communication-agnostic. The main contribution is to allow MARL to scale without reliance on communication by employing a multilevel hierarchical shared mental model. Our results show that H-MARL-EP can solve navigation tasks that are unsolvable with existing baseline methods, and scale to hundreds of agents.