Scholarlink

<p>Multi-Agent Reinforcement Learning (MARL) for water distribution networks often exhibits policy brittleness when exposed to out-of-distribution stressors. This study introduces a Multi-Phase Curriculum Learning (MPCL) framework designed to enhance resilience under stochastic drought conditions. Leveraging real-world hydrological data from Seattle, WA, agents are transitioned through staged training phases that progressively expose them to increasing environmental stress. Statistical validation across five independent seeds (<i>n</i> = 5) demonstrates that MPCL achieves a 95% reduction in violation variance (σ: 39.3 → 1.8) compared to standard MAPPO baselines, substantially improving cross-seed operational stability. Furthermore, the framework reduces system oversupply by 52.3% during severe drought while suppressing oscillatory depletion dynamics, representing a significant gain in hydraulic efficiency. Behavioral analysis identifies an emergent coordination mechanism, termed “Predictive Throttling,” wherein agents utilize latent communication to coordinate resource preservation as hydraulic stress intensifies. Run-length analysis further reveals suppression of oscillatory collapse-recovery behavior and convergence toward a temporally stable low-storage operating regime. These findings demonstrate that curriculum-staged training induces stable degradation dynamics and suppresses oscillatory collapse–recovery behavior, supporting the deployment of robust, coordinated AI in critical municipal infrastructure 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

Robust Control of Water Distribution Networks Under Drought Stress: A Multi-Agent Curriculum Learning Approach

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Vidhyalakshmi Amarnath

摘要

Multi-Agent Reinforcement Learning (MARL) for water distribution networks often exhibits policy brittleness when exposed to out-of-distribution stressors. This study introduces a Multi-Phase Curriculum Learning (MPCL) framework designed to enhance resilience under stochastic drought conditions. Leveraging real-world hydrological data from Seattle, WA, agents are transitioned through staged training phases that progressively expose them to increasing environmental stress. Statistical validation across five independent seeds (n = 5) demonstrates that MPCL achieves a 95% reduction in violation variance (σ: 39.3 → 1.8) compared to standard MAPPO baselines, substantially improving cross-seed operational stability. Furthermore, the framework reduces system oversupply by 52.3% during severe drought while suppressing oscillatory depletion dynamics, representing a significant gain in hydraulic efficiency. Behavioral analysis identifies an emergent coordination mechanism, termed “Predictive Throttling,” wherein agents utilize latent communication to coordinate resource preservation as hydraulic stress intensifies. Run-length analysis further reveals suppression of oscillatory collapse-recovery behavior and convergence toward a temporally stable low-storage operating regime. These findings demonstrate that curriculum-staged training induces stable degradation dynamics and suppresses oscillatory collapse–recovery behavior, supporting the deployment of robust, coordinated AI in critical municipal infrastructure systems.