Scholarlink

<p>As Multi-Agent Reinforcement Learning (MARL) systems are increasingly deployed in safety-critical applications, understanding <i>why</i> agents make decisions and <i>how</i> they collectively achieve intelligent behavior becomes paramount. However, existing explainable AI (XAI) methods fail to address the unique challenges of multi-agent settings: attributing collective outcomes to individual agents, quantifying emergent behaviors, and accounting for complex agent interactions. We present MACIE (Multi-Agent Causal Intelligence Explainer), a principled framework that unifies structural causal models, interventional counterfactuals, and Shapley values to provide comprehensive explanations of multi-agent systems. MACIE addresses three fundamental questions: (1) <i>What is each agent’s causal contribution to collective outcomes?</i> through interventional attribution scores; (2) <i>Does the system exhibit emergent intelligence?</i> via novel synergy metrics that distinguish collective effects from individual contributions; and (3) <i>How can explanations be made actionable for stakeholders?</i> through natural language generation that synthesizes causal insights into human-interpretable narratives. We evaluate MACIE across four diverse MARL scenarios spanning cooperative and mixed-motive settings (including tasks where cooperative objectives can produce competitive interference). Our results demonstrate that MACIE accurately attributes outcomes to individual agents with high consistency, successfully detects positive emergence in cooperative tasks while identifying interference in competitive scenarios, and achieves remarkable computational efficiency suitable for real-time deployment on standard CPU hardware. Compared to existing attribution methods, MACIE uniquely combines causal rigor, emergence quantification, and multi-agent support while maintaining practical feasibility for real-time deployment. Our framework represents a significant advance toward interpretable, trustworthy, and accountable multi-agent AI systems.</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

Macie a multi-agent causal intelligence explainer for collective behavior understanding

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月18日

查看全文
  • Abraham Itzhak Weinberg

摘要

As Multi-Agent Reinforcement Learning (MARL) systems are increasingly deployed in safety-critical applications, understanding why agents make decisions and how they collectively achieve intelligent behavior becomes paramount. However, existing explainable AI (XAI) methods fail to address the unique challenges of multi-agent settings: attributing collective outcomes to individual agents, quantifying emergent behaviors, and accounting for complex agent interactions. We present MACIE (Multi-Agent Causal Intelligence Explainer), a principled framework that unifies structural causal models, interventional counterfactuals, and Shapley values to provide comprehensive explanations of multi-agent systems. MACIE addresses three fundamental questions: (1) What is each agent’s causal contribution to collective outcomes? through interventional attribution scores; (2) Does the system exhibit emergent intelligence? via novel synergy metrics that distinguish collective effects from individual contributions; and (3) How can explanations be made actionable for stakeholders? through natural language generation that synthesizes causal insights into human-interpretable narratives. We evaluate MACIE across four diverse MARL scenarios spanning cooperative and mixed-motive settings (including tasks where cooperative objectives can produce competitive interference). Our results demonstrate that MACIE accurately attributes outcomes to individual agents with high consistency, successfully detects positive emergence in cooperative tasks while identifying interference in competitive scenarios, and achieves remarkable computational efficiency suitable for real-time deployment on standard CPU hardware. Compared to existing attribution methods, MACIE uniquely combines causal rigor, emergence quantification, and multi-agent support while maintaining practical feasibility for real-time deployment. Our framework represents a significant advance toward interpretable, trustworthy, and accountable multi-agent AI systems.