Gaining insight into crew rostering instances through ML-based sequential assignment

Crew scheduling is typically performed in two stages. First, solving the crew pairing problem generates sequences of flights called pairings. Then, the pairings are assigned to crew members to provide each person with a full schedule. A common way to do this is to solve an optimization problem called the crew rostering problem (CRP). However, before solving the CRP, the problem instance must be parameterized appropriately while taking different factors such as preassigned days off, crew training, sick leave, reserve duty, or unusual events into account. In this paper, we present a new method for the parameterization of CRP instances for pilots by scheduling planners. A machine learning-based sequential assignment procedure (seqAsg) whose arc weights are computed using a policy over state–action pairs for pilots is implemented to generate very fast solutions. We establish a relationship between the quality of the solutions generated by seqAsg and that of solutions produced by a state-of-the-art solver. Based on those results, we formulate recommendations for instance parameterization. Given that the seqAsg procedure takes only a few seconds to run, this allows scheduling workers to reparameterize crew rostering instances many times over the course of the planning process as needed.

Mots clés

crew rostering; crew scheduling; discrete optimization; evolutionary algorithm; machine learning; reinforcement learning