Comparative Analysis of Reinforcement Learning and Rule-Based System Approaches for Irrigation in Horticulture

Irrigation optimization in horticulture is highly relevant due to the high sensitivity of crops to water availability. Methods such as Reinforcement Learning (RL) and Rule-Based Systems (RBS) have been explored for automated irrigation; however, direct comparisons between these approaches for horticultural crops remain limited. This study compares RL and RBS for optimizing irrigation in lettuce cultivation using the AquaCrop-OSPy model. A Q-Learning RL agent dynamically adjusts irrigation decisions based on a customized reward function, while the RBS applies predefined water balance rules. Simulations were conducted over 30 days in each of the four seasons. Results indicate that RL achieved an average dry yield of 5.84 tonne/ha with 186.25 mm of water, while RBS produced 2.35 tonne/ha with 92.01 mm. RL consistently delivered higher productivity but at the cost of increased water use, whereas RBS demonstrated greater water efficiency under conservative irrigation strategies. The findings highlight RL’s adaptability to varying conditions and RBS’s reliability for water-limited scenarios. Future research can explore hybrid systems that integrate RL’s flexibility with RBS’s efficiency. The code and datasets are available at https://github.com/GabyQueiroz/RBS_RL_Irrigation.