Optimizing Markov decision process state design for deep reinforcement learning manufacturing scheduling using Bayesian optimization

Abstract This study investigates the application of Bayesian optimization for feature selection in Markov decision processes when applied to production scheduling problems. Traditional supervised learning feature selection methods are unsuitable due to the absence of explicit target values and the dynamic nature of scheduling environments. To address this, a bi-level optimization framework is proposed, with Bayesian optimization at the upper level for feature selection and reinforcement learning at the lower level for evaluation. Experimental results conducted in dynamic flexible job shop and thin-film transistor liquid-crystal display production scheduling environments demonstrate that the framework enhances efficiency by focusing on impactful features, reducing computational complexity, and improving decision-making. The findings highlight the significance of aligning state representations with scheduling dynamics and provide a foundation for future research on systematic feature selection in complex environments.