Optimizing Irrigation with Semi‐centralized Multi‐agent RL

This chapter develops a semi-centralized multi-agent reinforcement learning framework for daily irrigation scheduling in spatially heterogeneous fields. The approach decomposes the daily scheduling task into two levels of decision-making: a centralized coordinator agent leverages field-wide information to select a binary action that represents whether to irrigate on a given day, while decentralized local agents leverage local information to determine irrigation amounts for their assigned management zones. To handle the non-stationarity that arises from interactions between the coordinator and local agents, the chapter adopts a state-augmentation strategy in which the coordinator's action is incorporated into the observation of the local agents. The agents are trained using proximal policy optimization and the approach is evaluated using several case studies. The results illustrate that the semi-centralized framework can improve irrigation water-use efficiency relative to a decentralized alternative. The chapter also discusses how employing the semi-centralized approach alongside model predictive control can further improve irrigation performance and water-use efficiency.