CHIDYN: a clustering-based hierarchical approach for dynamic traffic assignment

Dynamic Traffic Assignment (DTA) plays a crucial role in the efficient management of traffic, as it is adaptable to changing conditions, such as weather, accidents, and variations in demand. However, conventional DTA algorithms tend to be sluggish, which hinders their real-time effectiveness. Therefore, the development of faster algorithms is imperative to enable timely and precise adjustments in traffic assignment. This study outlines the initial phases of creating a distributed multi-agent system designed to handle dynamic scenarios. In this research, we utilized the Sioux Falls network and its two different scales as representative networks for conducting User Equilibrium (UE) computations. Specifically, we applied the Method of Successive Averages (MSA) algorithm in two distinct configurations: one without clustering and the other with clustering. In the case without clustering, we computed the average travel time (ATT) and total traffic flow, whereas with clustering, we further analyzed these values within each cluster. Additionally, we measured the total computation time for both MSA with clustering and MSA without clustering. The results revealed substantial improvements in overall runtime when employing clustering, with a remarkable 14-fold enhancement for the Sioux Falls network and a 35% improvement for the mixed network. However, it's important to note that the clustering process led to some loss of information. To quantify this lost information, we employed the KL-divergence metric, which characterizes the information loss from the perspective of the distribution of vehicle flows per link. According to the KL-divergence analysis, the lost information amounted to 31% for a sub-network of Sioux Falls, 28% for the Sioux Falls network as a whole, and merely 1% for the mixed network. These findings underscore the effectiveness of our approach, as it allows for UE values to be computed an order of magnitude faster while preserving a relatively minimal amount of information loss during the clustering process.