Trois variantes du problème de rotations pour une approche semi-intégrée de la planification d'horaires de personnel aérien

Aircrew scheduling is usually performed according to a two-step sequential procedure: crew pairing and crew rostering. While the crew pairing problem (CPP) finds a set of pairings that covers the legs of a given period at minimum cost, the crew rostering problem (CRP) uses those pairings in order to create a personalized schedule for each crew member. A pairing is a sequence of legs, deadheads, connections and rests spanning over one or multiple days, and that can be assigned to a crew member. A pairing must also begin and end at the same crew base (airport where crew members are stationed), and comply with many rules imposed by airline authorities as well as collective agreements. The crew schedules must cover all pairings, and are also subject to many regulations. The main drawback of this sequential approach is that the set of pairings produced by the CPP is often ill-suited for the CRP. For instance, the CPP solution might assign too much work to a given base, resulting in an imbalance in the work distribution among the bases. Ideally, both steps would be integrated into a single optimization problem. Even though many such approaches have been proposed in the literature, computing times required to solve those integrated problems are prohibitive, even for small-sized instances. Semi-integrated approaches are designed to overcome some limitations of the sequential approaches, without unduly increasing computing times. The main idea is to solve a variant of the CPP that includes some elements that traditionally belong in the CRP. This enables the CPP to create pairings that are better-suited for the CRP. In this thesis, we study three such CPP variants. Each variant is formulated as a setpartitionning problem with additional constraints, in which the main decision variables are associated with feasible pairings. These problems are solved by a column generation algorithm that uses a restricted master problem to select the pairings and multiple subproblems to generate the pairings to add to the restricted master problem. In the first subject of this thesis, we study the CPP with base constraints (CPPBC). Base constraints penalize excess work performed at each crew base in order to evenly distribute the workload among them. Although most commercial softwares include base constraints in the CPP, no academic research has studied their impact on the existing solution methods. Preliminary tests show that when base constraints are very restrictive, the heuristic branching algorithms traditionally used struggle to find a good-quality integer solution: they take a larger number of risky branching decisions, which negatively impact the quality of the solutions. We develop a new heuristic branching scheme, called retrospective branching, that identifies risky branching decisions in the branch-and-bound tree, and removes poor branching decisions when the gap between the current and the best fractional solution becomes too large, without backtracking. The proposed method is tested on seven weekly instances. We show that the retrospective branching algorithm produces solutions of better quality than with the other commonly used branching methods, in reasonable computing times. The retrospective branching is currently implemented in a commercial crew scheduling software, and has been used to obtain good-quality solutions to monthly instances containing tens of thousands of legs. In the second subject of this thesis, we propose a variant of the CPP, called the CPP with complex features (CPPCF) which takes into account legs and vacations preferences of crew members, with the aim of increasing the number of preferences awarded in the CRP, and thus, crew member satisfaction towards their schedule. We identify six pairing features related to those preferences, which could be beneficial to the CRP. Pairings containing one or more of those features are granted a bonus in order to promote their presence in the solutions. The solution method for the CPP is adapted to the CPPCF. We modify the dominance rules of the labeling algorithm used to solve the subproblems, based on the values of new state resources. The proposed method is tested on seven monthly instances. We show that using the CPPCF allows for a significantly higher number of awarded preferences in the CRP. The third subject of this thesis deals with language constraints — constraints on the language qualifications of the crew operating some legs. Satisfying these constraints is essential for airlines, which would otherwise have to pay high penalties, or even cancel some legs. Current methods are inadequate to deal with problems containing a large number of language constraints and few crew members with language qualifications. This is because the CPP does not account for language constraints, resulting in a spreading of the legs with language constraints among many pairings. We study this problem in the context of cabin crew scheduling. We formulate a CPP variant, called CPP with language constraints (CPPLC), which favors the grouping of legs with similar language constraints within the same pairing. The main challenge in solving the CPPLC is the combinatorial explosion in the number of subproblems. We put forward a subproblem selection strategy in which only a fraction of these subproblems are solved at each column generation iteration. We show that taking into account the language constraints in the CPP allows for a significant reduction of the number of language constraint violations in the CRP solutions. Although this study was conducted only for language constraints, the proposed method can be applied to many types of qualification constraints for cabin crews as well as pilots and copilots.