An imitation learning approach for personnel scheduling with flexibility for on-call shifts

Predictive models are playing an increasingly pivotal role in optimizing decision-making. This paper introduces an innovative approach based on DAgger, an imitation learning algorithm to train a policy iteratively with data collected from experts for stochastic sequential decision problems. Such problems can be hard, and a good expert may be expensive or unavailable. We focus on designing a good expert within the DAgger framework, which will based on deterministic solutions obtained from a sequence of contextual scenarios generated at each decision epoch. Subsequently, a predictive model is constructed that imitates the expert to facilitate real-time decision-making. The applicability of this methodology is demonstrated through a novel dynamic employee call-timing problem for the scheduling of casual personnel for on-call work shifts. The critical decision in the problem involves determining the optimal time to wait before placing a call to the next employee in the order of seniority, allowing them to select a preferred work shift. The uncertainty is associated with the time taken by an employee to respond and choose a work shift. The aim is to strike a balance between minimizing schedule changes induced by sending too many notifications or calls early in the planning period and having unassigned shifts due to sending calls/notifications too late. In contrast to conventional predict-and-optimize approaches, our methodology leverages optimization to train learning models that establish connections between the current state of the system and the wait time from the expert.