MARLIN: Multi-Agent Reinforcement Learning with Murmuration Intelligence and LLM Guidance for Reservoir Management

Intensifying climate change and cascading uncertainties across interconnected reservoir networks pose escalating threats to global water security, demanding management systems that are both adaptive and scalable. Traditional centralized optimization becomes computationally intractable and brittle under real-world uncertainty, while existing reinforcement learning (RL) approaches are not designed for complex, multi-node hydrological systems. To address these challenges, we introduce MARLIN, a decentralized reservoir management framework that explicitly handles dual-layer uncertainty: (i) stochastic variability in physical water transfer and (ii) dynamic, human-environmental perturbations. MARLIN embeds bio-inspired alignment, separation, and cohesion rules into a multi-agent RL (MARL) architecture to stabilize coordination under physical uncertainty. Additionally, external conditions such as weather forecasts, regulatory updates, and stakeholder preferences introduce unstructured textual information that traditional models cannot process directly. To bridge this gap, we integrate a Large Language Model (LLM) that interprets such contextual information and dynamically adjusts the coordination parameters of the three murmuration rules, enabling rapid adaptation to evolving environmental and human requirements. Experiments on USGS data show that MARLIN improves uncertainty handling by 23%, reduces computational cost by 35%, and accelerates flood response by 68%. The framework demonstrates excellent scalability, with emergent coordination patterns increasing super-linearly as the network expands while maintaining linear computational complexity. These results highlight MARLIN’s potential as a scalable and intelligent solution for adaptive water resource management and disaster prevention.