The economic and social development of modern cities relies on the efficiency, mobility and resilience of their transportation systems. The latter has thus be- come a major research challenge involving multiple disciplines, related to urban activities. Old infrastructures and their limited capacity make cities more and more vulnerable to unpredictable events and increasing demand. Congestions are more frequent, as a consequence of the growth of the urban population, vehicle emissions and air pollution create high stress on the infrastructures and increase time waste for travelers.Solutions to improve traffic conditions, in terms of health, security and traffic management are more and more precise, embracing the gen- eralized use of Artificial Intelligence, jointly with Big Data technologies for data collection, storage and computing. Moreover, traffic simulations are now based on various data sources and on more accurate information to better reproduce traffic dynamics and travelers’ behaviors. However, analyzing such complex data in a large scale context is still a significant research challenge that requires solutions based on agent-based modelling, distribution and parallelization. Moreover, the characterization and modelling of transport vulnerabilities for improving human mobility is still at early stages of research.

To prevent congestion and identify vulnerable locations, i.e. areas or sections where failures would have high cost consequences, two types of vulnerability anal- yses are most common in the domain of transport: dynamic system-based, and static topological based. They are both studied in this thesis. The first approach is the dynamic system-based representation that simulates travelers, their trips and the infrastructure over a given period of time, supported by the large volume of data now collected. The second approach is a topological analysis based on graph theory, and static topological considerations. To reduce vulnerability in- side road networks, we propose in this thesis a control strategy that dynamically protect identified areas and recommends new routes to drivers to avoid creation of congestion in such zones. Our strategy relies on a hierarchical cooperative multi- agent algorithm. Road infrastructures and vehicles are modeled as agents that dynamically react to traffic conditions. This control strategy enables congestion avoidance and a reduction of the congestion duration. We take into consideration drivers behaviours to find a balance between system performance improvement (system optimum) and individual travel choices (individual optimum), as well as privacy constraints that are now necessary for realistic applications. We prove the robustness of our approach by testing it on different demand scenarios and show that identifying and protecting critical spots of the network improves our strategy. To identify such vulnerable spots, our solution integrates the compu- tation of Betweenness Centrality (BC), a metric usually studied with topological approaches. It is indeed quite unusual to include BC in dynamic congestion avoid- ance approaches whereas the BC is a popular metric in many domains for critical spot identification in the context of static graph analysis. This is due to the high computation time and the difficulty of computing it on large graphs in a context of real-time applications. This second problem of computation of BC for static vulnerability analysis is addressed in this thesis with a distributed algorithm for the exact and fast computation of BC developed for large graphs. We provide mathematical proofs of our algorithm exactness and show the high scalability of our approach, developed in an optimized framework for parallel computation. Through distributed approaches, we can design a robust solution, based on a combination of control and topological study, to dynamically reduce vulnerability inside cities in a real-time context. The proposed solution for computation of BC on large-scale graphs can be extended for real-time computation of this met- ric on time-varying weighted graphs and further enhance our control solution for congestion avoidance based on dynamic vulnerability detection of road networks.