A reinforcement-learning-based column generation algorithm for integrated operating room planning and scheduling
提出混合整数规划模型和列生成算法，结合强化学习与遗传算法生成初始列，解决手术室规划与排程问题，在合成和真实数据上平均最优性缺口分别为1.23%和1.49%，优于现有方法。
Operating room planning and scheduling are vital components of hospital management, contributing to improved efficiency, patient satisfaction, staff well-being, and overall quality of care delivery. In this paper, we propose a novel mixed integer programming model to formulate integrated operating room planning and scheduling problems, where several mandatory and elective surgeries are to be assigned and scheduled in operating rooms on different days. Aside from the standard working hours in each operating room, we also take into account the potential for performing surgeries in overtime periods. In addition, our approach also takes into account the availability of surgeons by considering their allowed surgical time on each day. We propose a column generation (CG) algorithm to solve large-scale instances. In order to enhance the CG, we integrate the Reinforcement Learning Algorithm and the Genetic Algorithm and develop a hybrid algorithm to generate initial columns for the CG algorithm. For our analysis, we employed two sets of test instances: one consisting of synthetic data and the other based on real-world cases from a local hospital in Naples, Italy. Computational experiments demonstrate that our proposed model and methodology yields an average optimality gap of 1.23% for synthetic instances and 1.49% on real-world scenarios, significantly outperforming previous solution methodologies in the literature. Additionally, we demonstrate that the developed CG algorithm provides a high-quality solution for large-scale instances where other models and methods fail to obtain even a feasible solution. To further evaluate robustness under uncertainty, we examined scenarios with ±20% variability in surgery durations. The results indicate that incorporating a 120-minute buffer time minimises the overall cost. Moreover, we investigated the impact of emergency surgeries by either introducing additional cases or escalating surgical priorities. For synthetic instances, the inclusion of emergency surgeries increased the total rescheduling cost by 4.13%, whereas in the real-world Naples cases, priority escalation led to only a 0.11% increase, highlighting the resilience of our proposed model in practical hospital settings.