Nurse Rostering Constrained by Physiological-Psychological State Evolution
提出一种新的护士排班方法，用长短期记忆网络估计护士的生理和心理状态，并设计Q学习辅助的进化算法最大化状态奖励，在基准和真实案例上优于传统方法。
Nurse rostering is a typical personnel scheduling problem that involves a variety of constraints. Existing rostering methods that aim to minimize the penalty of soft constraint violations suffer from several limitations including the use of crisp parameters, inflexibility to changes, and a lack of consideration for individual nurse states. This is because violations of the widely-used soft constraints cannot truly reflect the actual working states of nurses that determine the quality of schedules. In this paper, we reformulate the nurse rostering problem by using a new objective function that maximizes the overall reward for achieving good physiological and psychological states of nurses, which are estimated based on their personal characteristics and work assignments using long short-term memory. To solve the problem, we propose an evolutionary algorithm assisted by Q-learning to dynamically adjust the fitness function based on constraint violations to balance positive rewards to good states and negative rewards to bad states during the evolution process. Our results on benchmark and real-world instances demonstrate that our method obtains significantly better nurse states compared to selected popular rostering methods.