Salma Bouhlla soutiendra sa thèse le jeudi 10 juin 2026 à 10h en salle B014 du LISIC à Calais.
La thèse est co-dirigée par Rym Guibadj et Aziz Moukrim.
Titre :
Optimization and Learning Approaches for Vehicle Routing Problem in Retail Distribution Systems
Résumé :
This thesis was carried out under a CIFRE partnership involving Mobivia Supply Solutions (M2S), the Laboratory of Computer Science, Signal and Image of the Opal Coast (LISIC) at the University of the Littoral Opal Coast (ULCO), and the Heuristics and Diagnostics for Complex Systems laboratory (Heudiasyc) at the University of Technology of Compiègne (UTC). It addresses optimization problems arising in the transport and distribution activities of a large-scale retail logistics network dedicated to the automotive aftermarket. As the entity responsible for coordinating the logistics operations of the Mobivia group, M2S manages the distribution of products to nearly 2,000 points of sale. This distribution relies on a network characterized by heterogeneous flows, strict delivery requirements, limited transport capacities, and the use of external logistics providers. In this context, routing decisions, transport-flow assignment, and workload distribution among carriers are strongly interdependent, leading to combinatorial optimization problems that extend beyond classical vehicle routing settings. The first part of this thesis focuses on urgent distribution flows between retail stores. We model this problem as a Multi-Pickup and Delivery Problem with Time Windows (MPDPTW), a generalization of the Pickup and Delivery Problem with Time Windows (PDPTW) in which a request may include multiple pickup locations associated with a single delivery location. This feature gives rise to more complex precedence constraints. We then propose heuristic and metaheuristic approaches to efficiently tackle the MPDPTW. In particular, we develop a GRASP-based framework enriched with memory mechanisms, local search procedures, and a post-optimization phase in order to produce high-quality routing plans adapted to real insdustrial constraints industrial requirements. The second part addresses the Capacitated Multi-Pickup and Delivery Problem with Time Windows (CMPDPTW). For this setting, we propose a hybrid framework that combines combinatorial optimization techniques with learning-based decision mechanisms. More specifically, we introduce a reinforcement learning-based variable search space approach in which an agent dynamically guides the selection of neighborhood operators throughout the search process. The results show that incorporating learning into the metaheuristic allows the search behaviour to adapt through a more informed operator selection mechanism. This hybrid framework improves both search efficiency and solution quality. The third part addresses a tactical planning problem arising in the M2S distribution network, namely the assignment of store orders to the most appropriate transport mode. To support this decision, we develop a decision-support framework to assist logistics planners in assigning each order placed by an M2S store to the most appropriate transport mode : a conventional traction-distribution scheme organized as a two-echelon system, and a direct-delivery scheme from the central warehouse. This decision involves jointly considering routing feasibility, consolidation opportunities, order characteristics, and carrier operating constraints. We therefore develop dedicated model and algorithmic approach to support this decision-making process under realistic operational constraints. Overall, this thesis provides methodological contributions at the intersection of operations research, combinatorial optimization, and machine learning for real-world vehicle routing problems arising in retail distribution systems. It demonstrates the industrial relevance of combining combinatorial optimization and machine learning, and highlights their potential for developing effective decision-support tools for real operational problems.
Jury :
Mme Caroline Prodhon, Professeure des Universités, LOSI, Université de Technologie de Troyes, France — Rapporteure
M. Ammar Oulamara, Professeur des Universités, LORIA, Université de Lorraine, France — Rapporteur
M. Matthieu Basseur, Professeur des Universités, LISIC, Université du Littoral Côte d’Opale, France — Examinateur
M. Nikolay Tchernev, Professeur des Universités, LIMOS, Université Clermont Auvergne, France — Examinateur
Mme Rym Guibadj, Maître de conférences HDR, LISIC, Université du Littoral Côte d’Opale, France — Co-directrice de thèse
M. Aziz Moukrim, Professeur des Universités, Heudiasyc, Université de Technologie de Compiègne, France — Directeur de thèse