March 2026 | PLoS One | Volume 21
Introduction: Greenhouse production requires precise control of temperature, irrigation, CO₂, and other environmental conditions, but conventional fixed-control strategies may struggle to balance crop productivity with efficient resource use as conditions change. Hindi and colleagues developed an AI-based autonomous greenhouse control framework combining deep learning prediction models with reinforcement learning. Using data from the 2nd International Autonomous Greenhouse Challenge, the system predicts greenhouse climate, cherry tomato growth, and resource consumption, allowing an AI agent to dynamically optimize greenhouse control settings.
Key findings: Among the reinforcement learning algorithms tested, the TD3 (Twin Delayed Deep Deterministic Policy Gradient) model achieved the highest average reward and the most stable overall performance, outperforming PPO (Proximal Policy Optimization), DDPG (Deep Deterministic Policy Gradient), and SAC (Soft Actor-Critic). Compared with the average resource use of other greenhouse strategies, TD3 reduced irrigation by 24.05% and CO₂ consumption by 1.86%, while achieving stronger crop growth performance, particularly in cumulative trusses. Although electricity and heating use increased slightly, the results show that reinforcement learning can dynamically balance crop growth and resource efficiency. The system was evaluated in a virtual greenhouse environment, highlighting the need for further validation under real-world commercial conditions..png)
