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

Short summary

This paper proposes a calibration-first reward audit framework for smart-greenhouse RL that decomposes scalar rewards into named components—temperature, CO2, humidity, screen, and actuation-proxy terms—keeping them comparable across simulator training, facility-adapted rollouts, and logged greenhouse records. The framework is implemented in GreenLight-Gym and adapts the GreenLight simulator to Autonomous Greenhouse Challenge climate traces. It enables growers and engineers to understand not just overall return but when and how a policy heats, vents, or deploys screens.

• •Calibration-first reward audit framework decomposes scalar RL reward into interpretable greenhouse-control components
• •Components stay comparable across simulator training, facility rollouts, and logged greenhouse challenge data
• •Implemented in GreenLight-Gym, adapted to Autonomous Greenhouse Challenge climate traces

Generated with AI, which can make mistakes.