Calibration-First Reward-Component Auditing for Reinforcement Learning Control in Smart Greenhouses

2026-07-12

AI Summary

This study addresses the lack of interpretability and cross-scenario consistency in existing reinforcement learning approaches for intelligent greenhouse control, particularly regarding policy-driven actions such as heating, CO₂ supplementation, and ventilation. The authors propose a “calibrate-then-audit” reward decomposition framework that disentangles the scalar reward into semantically meaningful components—namely temperature, CO₂, humidity, vapor pressure deficit, screen usage, and actuator proxies—and ensures their semantic consistency across simulation training, facility adaptation, real-log replay, and rule distillation. Built upon the GreenLight-Gym platform and integrating environmental model calibration, conditional reward decomposition, and climate trajectory alignment, this work achieves, for the first time, cross-domain consistent decomposition and auditability of reward constituents in greenhouse control. Experiments on real-world data from the 2nd Autonomous Greenhouse Challenge demonstrate the framework’s interpretability, consistency, and practical utility across diverse scenarios.

Abstract

Greenhouse reinforcement learning can test climate-control ideas at a speed and scale that is difficult to achieve with crop experiments alone. For smart-greenhouse control, however, a single simulator return is not enough: a grower or control engineer also needs to know when the policy heats, enriches CO2, vents, manages humidity, deploys screens, or uses lamps.We propose a reproducible calibration-first reward audit framework that keeps named greenhouse-control reward components comparable across simulator training, facility-adapted rollouts, logged Autonomous Greenhouse Challenge records, and actuator-rule distillation. In GreenLight-Gym, the framework decomposes the scalar reward into conditional temperature, CO2, humidity and vapor-pressure-deficit, screen, and actuation-proxy terms; adapts GreenLight to the second Autonomous Greenhouse Challenge logged climate traces; and scores the same components on logged greenhouse data.

Problem

• reinforcement learning
• smart greenhouses
• reward decomposition
• climate control
• actuator auditing

Innovation

• reward decomposition
• calibration-first
• reinforcement learning
• smart greenhouse control
• interpretable policy