In the railway industry, as well as in other branches of transportation, the construction of yearly schedules for drivers is a key step of the planning process. One of the main criteria to assess the quality of such schedules is the unproductive time of drivers (rest periods, idle times and relocation trips). The objective function of the problem of building yearly schedules for freight train drivers at the SNCF (the main French railway company) is the total number of working days, which captures directly the above criterion. Moreover, the yearly schedules of train drivers at the SNCF must take the form of cyclic rosters, each shared by every driver of a team and operated on a weekly basis.
Due to the high complexity of the problem, the SNCF applies a classical two-step approach. In a first step---the Crew Scheduling---, trains (assumed to be the same every week) are combined together to form daily duties. In a second step---the Cyclic Crew Rostering---, these duties are organized so as to form cyclic rosters, one per team of drivers, with the objective of minimizing the total number of working days.

An important drawback of sequential approaches is the inherent inability to find optimal solutions. This is a well-identified issue in the operations research community, and complete works have actually been devoted to this question. Here, the sequential approach makes it moreover difficult for the SNCF to account for uncertainties: the trains and duties cancellation, as an outcome of train delays' propagation on a roster, can only be measured by considering simultaneously the duties and their organization within the roster. Yet, train delays occur frequently and significantly affect the final cost incurred by the company.

In this work, we propose to modify the two-step approach currently used at the SNCF by handling the two steps simultaneously and by including stochastic constraints, so as to tackle both suboptimalities and train delays. This ``integrated'' approach relies on column generation and models exactly the impact of uncertainty on rosters via the introduction of a quantitative metric that we call fragility (assuming known the distribution of the delays).
In the deterministic case (with no uncertainties), the integrated approach decreases systematically by a significant margin the number of working days on all instances of the SNCF with respect to the current method. When there are uncertainties, preliminary results show that the integrated approach is able to decrease substantially the fragility of rosters, without deteriorating the value of the objective function from the current approach.