Laboratoire de Génie Informatique et d’Automatique de l’Artois

This PhD thesis focuses on the optimisation of Multi-Attribute Vehicle Routing Problems (MAVRP), an important issue in modern logistics, particularly in the seafood sector. The complexity and rapid evolution of contemporary logistics systems, coupled with the stringent requirements of handling perishable products, make these problems particularly challenging but also fascinating for the application of advances in machine learning. This PhD focuses on using Graph Neural Networks (GNNs), a cutting-edge machine learning technique, for their unique ability to model complex relationships and interpret the multiple attributes that characterise MAVRPs. Unlike traditional approaches, which are often limited by their rigidity and inability to manage various constraints simultaneously or adapt to dynamic situations, GNNs offer a promising approach to flexible, adaptable and much more efficient solutions. Building on recent studies that highlight the superior effectiveness of MAVRP-based solutions for a wide range of logistics problems, this work proposes to go a step further by integrating transfer learning. This approach would allow GNN models to be adapted to different variants of MAVRPs without the need for extensive new data collection, offering unrivalled flexibility and the ability to handle different applications.

Seafood logistics provides the application framework for this thesis. Seafood products require optimised transport routes not only to minimise costs and lead times, but also to ensure that product quality is preserved, involving factors such as distance, transit time and storage conditions. This thesis therefore aims to explore the ability of GNNs to navigate this complexity, developing models capable of proposing optimal solutions that take into account all the dynamic attributes and constraints of seafood logistics.