Tag
This paper evaluates whether an LLM can serve as a calibration agent for expensive engineering simulators, finding that physics-labeled prompts cut early calibration error 22-fold but lose their advantage after about ten runs, so a hybrid warm-start plus numerical optimizer workflow reaches target accuracy in 9 runs versus 17 for Nelder-Mead.
Proposes a calibration-first reward-component auditing framework for reinforcement learning control in smart greenhouses, enabling decomposition of scalar rewards into interpretable components and validating them across simulated and real-world data.