Learning to enumerate shifts for large-scale flexible personnel scheduling problems

Personnel scheduling consists of determining employee work schedules (sequences of work shifts and days off) to cover the demand for multiple jobs over a planning horizon. We consider finding a near-optimal set of personnel schedules via the solution of a generalized set-covering model with side constraints in a flexible context where a large number of potential shifts can be considered as in the retail industry. Commercial solvers applied to this model often require very long computational times for practical problem sizes and as such rely on enumeration heuristics for filtering non-promising shifts/schedules and, thus, reducing the problem size. We propose deep learning-based heuristics to drive the enumeration of promising potential shifts based on the information collected from previously solved instances. Our models predict a subset of time points at which promising shifts are more likely to either start or end, thus filtering out those that do not start nor end at those time points. Our computational results on real-life instances show that personnel scheduling problems can be solved considerably faster with an acceptable optimality gap if shifts are enumerated according to the time points predicted by our models.