Abstract
Urban traffic congestion presents significant economic and public health challenges, with costs in major metropolitan areas often reaching billions of pounds annually. While decentralized routing applications like Google Maps offer individual benefits, they frequently fail to optimize the road network from a systemic perspective. This paper introduces a centralized traffic routing (CTR) framework that integrates automated planning with the Simulation of Urban MObility (SUMO) environment. By identifying congestion hotspots and utilizing PDDL-based planning to coordinate vehicle flows, the proposed system achieves a 22.5% reduction in average travel times compared to historical baselines in a real-world Dublin scenario.
Introduction
Traffic congestion is an escalating crisis. In London alone, the economic impact of congestion exceeded £5 billion in 2020. Current navigation solutions rely on "selfish" routing, where individual agents seek the shortest path for themselves, often leading to sub-optimal outcomes for the city as a whole. Centralized traffic routing (CTR) aims to address this by adopting a "global" perspective, optimizing routes for all vehicles to maximize network efficiency. However, the computational complexity of calculating optimal routes for thousands of vehicles across vast urban networks has historically limited the practicality of this approach. This research addresses these challenges by focusing on problematic regions and simplifying the search space for automated planning algorithms.
Methodology
The proposed framework is designed as a modular wrapper around the SUMO (Simulation of Urban MObility) simulator, consisting of three primary stages:
- Region Identification: To manage computational load, the system identifies congestion hotspots rather than managing the entire city. It calculates a "Congestion Index" for road segments and employs Gravitational Clustering to group these segments into distinct problematic regions.
- Network Simplification and Smart Routing: The framework pre-computes a set of promising routes for vehicles entering identified regions. It utilizes a variant of the A* algorithm to discover diverse paths and applies DBSCAN clustering (using the Jaccard Index for path similarity) to select the most distinct and efficient options.
- PDDL-Based Planning: The routing problem is translated into PDDL (Planning Domain Definition Language). A specialized solver, the Mercury planner, determines the optimal distribution of vehicles across the pre-computed routes to minimize overall traffic intensity.
Experimental Evaluation
The framework was evaluated using a real-world dataset from the Dublin metropolitan area. The study focused on the morning rush hour (7:00 AM to 9:00 AM), involving a total of 45,000 vehicles. Approximately 6,100 of these vehicles were routed via the centralized system to test the impact of coordinated planning on known congestion points.
Results and Discussion
The simulation results demonstrate that the CTR approach significantly outperforms historical baselines across all key performance indicators. As shown in the data below, the CTR approach reduced average travel time to 581.55 seconds, representing a 22.5% improvement over the original historical data (749.68 seconds). Furthermore, average waiting time was reduced by approximately 33%, dropping from 482.92s to 324.42s.
While a decentralized "Dynamic User Optimal" (DUO) approach—representing a scenario where all vehicles use advanced GPS—performed better for the city as a whole, the CTR approach was 10% more effective than DUO when specifically analyzing performance within the targeted problematic regions. This suggests that centralized coordination is particularly valuable in high-density, high-friction areas where individual optimization leads to gridlock.
Conclusion and Future Work
This research demonstrates that strategic simplification—filtering for vehicles in problematic regions and limiting their options to "smart" routes—makes complex automated planning feasible in near-real-time. However, challenges remain regarding computational scaling for even larger urban densities. Future work will focus on integrating these routing decisions with intelligent traffic light control and exploring the dynamics of mixed autonomy, where the system must account for human drivers who may not comply with centralized routing suggestions.
Summary updated