The main methodological contributions of the MAGnUM project are related to three interrelated topics.
Network traffic modeling: Earlier developments for large-scale network models partition a city in multiple regions and track flow exchanges to represent traffic dynamics. We proposed a new methodological framework that keeps the parsimony of regional traffic dynamics but individualizes trips inside the multimodal transportation system. This trip-based approach is proven more accurate in free-flow and saturated traffic. To adequately handle congestion spillbacks in oversaturated regimes, we mix trip-based with the original accumulation-based approach at the region perimeters. Our framework can handle multiple trip lengths in each region. We also developed new methods for calibrating large-scale models considering innovative datasets like mobile phones or drone footages.
Network equilibrium and traffic assignment: We introduced a new regional framework to handle route choices across regional paths. The main challenge was scaling-up the road network characteristics into relevant aggregate figures. In particular, we design methods to identify the prevailing regional paths and estimate their lengths considering both the network topology and the local routing discipline due to congestion. We also revisited the network equilibrium problem by introducing metaheuristics (simulating annealing and genetic algorithms) from the optimization field. Such a computational framework permits parallelizing some calculations and leads to significant time savings while improving accuracy. We then propose a new framework to identify trips that reduce the most network performance. Based on machine learning technics, we characterize these trips only based on their features and network characteristics. It paves the way for their direct identification in real-time. Lastly, a significant methodological contribution is the introduction of Mean Field Game (MFG) theory to solve the optimal departure time problem. Classical methods suffer from scaling issues that prevent them from addressing large-scale samples. They also resort to restrictive assumptions that reduce the solution space exploration. Combining MFG with the trip-based modeling framework blew up those restrictions and efficiently solved realistic test cases with more than 40000 trips.
User behaviors, data collection, and analysis: The simulation game we developed provides priceless behavioral observations. It cannot replace on-field experiments, but it simultaneously collects information for many users considering multiple network configurations. Furthermore, as the simulation is fully controlled and finely tuned, it permits to have a perfect vision of the local and global network scales. Based on such unprecedented datasets, we performed thorough data analysis and implemented advanced clustering technics to define the most representative OD pairs and investigate bounded rationality at the individual level.
A cutting-edge contribution in data analysis is the concept of 3D congestion maps. We benchmarked different clustering algorithms to reduce the link speeds variation into a small bunch of homogeneous space-time domains. We were also able to cluster days with homogeneous patterns and apply consensual learning to determine each cluster's typical congestion map. Our test case, i.e. Amsterdam's city, shows that only four consensual maps can describe the city traffic dynamics over 35 days. We also studied two other datasets in Zurich and Luzerne over an entire year. Here, we use a new similarity measure for MFD based on dynamic time warping to characterize and compare traffic states' daily time evolution. Again, a few consensual patterns – 7 - suffice to describe all days. We related these patterns to different bottleneck activation at the city level. It results in an early identification system, which identifies the more likely pattern for the current day.
