Reinforcement Learning for Rice Harvester Dispatching Under Typhoon Disruptions

Harvesting operations are time-sensitive and vulnerable to weather-related disruptions, making effective field dispatching important for minimizing crop loss. This study extends a previously developed multi-agent discrete-event simulation of Japanese rice harvesting by integrating a forecast-aware reinforcement learning (RL) policy for dynamic harvester dispatching. The simulation combines real-world field configurations, historical weather, process-based crop development, machinery and logistics processes, and synthetic planting schedules and typhoon disruptions. A Double Deep Q-Network was used to train a shared policy that selects fields using spatial, crop, machinery, system, and forecast-related information. The policy was evaluated on 100 held-out scenarios and compared with Closest-Field and first-in-first-out (FIFO) dispatching policies under normal and typhoon-disrupted conditions. Under normal conditions, the RL policy achieved crop preservation comparable to the Closest-Field policy and better than FIFO. Under typhoon disruption, the median crop-loss rate decreased from 8.59% with Closest-Field to 7.52% with RL, corresponding to a 12.5% relative reduction. The paired difference was statistically significant and remained significant after accounting for typhoon duration, total scenario area, and scenario-level variability. These results demonstrate the potential of context-dependent RL dispatching to reduce crop loss under capacity-constrained and weather-disrupted harvesting conditions.