A Multi-Phase Curriculum Learning Framework for Resilient Multi-Agent Reinforcement Learning in Water Resource Management
No Thumbnail Available
Authors
Amarnath, Vidhyalakshmi
Issue Date
2025-10-09
Type
Article
Language
en
Keywords
multi-agent reinforcement learning , curriculum learning , environmental modelling , water resource management , dynamic optimization , drought resilience
Alternative Title
Abstract
We propose a Multi-Phase Curriculum Learning framework for Multi-Agent Proximal Policy Optimization (MAPPO) to enhance resilience in dynamic environmental systems. The approach integrates inter-agent communication with staged curriculum progression, enabling decentralized controllers to learn cooperative water allocation policies that remain stable under drought stress. Using a simulated urban water distribution network with real inflow and demand data, our method reduced water shortages by approximately 45% and operational violations by 97% compared with heuristic and baseline MARL strategies. While water management provides the case study, the framework contributes a generalizable training methodology for resilient cooperative reinforcement learning in safety-critical environmental domains.
Description
Preprint article submitted to "Engineering Applications of Artificial Intelligence (EAAI)"
Citation
Publisher
License
Attribution-NonCommercial-NoDerivs 3.0 United States
openAccess
openAccess