AI Cuts Traffic Jams by 16%

Traffic congestion in cities leads to longer travel times, increased emissions, and driver frustration, especially during peak hours. While building new infrastructure is costly, emerging technologies like GPS and internet connectivity offer economical solutions through algorithmic optimizations. Current navigation services, such as Google Maps and Waze, rely on Shortest Path First (SPF) algorithms, which are optimal for individual vehicles in static conditions but perform poorly in dynamic, multi-vehicle settings, often worsening congestion by routing many vehicles along identical paths. A new study introduces a multi-agent reinforcement learning (MARL) system that enables coordinated, network-aware navigation, reducing average travel times by up to 15.9% in heavy traffic scenarios.

The researchers developed two key models: the Adaptive Navigation (AN) model and its scalable extension, the Hierarchical Hub-based Adaptive Navigation (HHAN). AN assigns a reinforcement learning agent to each intersection in a road network, providing routing guidance based on local traffic states and neighborhood information modeled with Graph Attention Networks (GAT). This allows vehicles to make decisions that consider congestion and coordinate implicitly. For larger networks, HHAN places agents only at strategically selected hubs—critical intersections—decomposing journeys into hub-to-hub segments. This hierarchical approach uses the Attentive Q-Mixing (A-QMIX) framework under a centralized training and decentralized execution paradigm, aggregating asynchronous decisions via attention mechanisms to handle coordination efficiently.

Methodology involved simulating traffic using the Simulation of Urban Mobility (SUMO) platform on synthetic grids and real-world maps from Toronto and Manhattan. The AN model processed routing queries at intersections, with agents using Q-learning and GAT layers to share traffic information. HHAN employed hub selection based on K-Medoids clustering and incorporated flow-aware features, such as vicinity speed and congestion ratios, to enable proactive routing. Training used Adam optimizer with a discount factor of 0.99 and ε-greedy policies, with experiments run over hundreds of episodes to ensure stable convergence.

Results showed that AN reduced average travel time by up to 25.7% compared to SPF baselines on smaller networks, achieving a 100% success rate in vehicles reaching their destinations. In the 5x6 grid test, AN (with one-hop GAT) achieved an average travel time of 96.8 seconds, a 28% improvement over SPF with rerouting (134.8 seconds). HHAN scaled to networks with hundreds of intersections, like Manhattan's 320-intersection map, cutting average travel time by 15.9% in heavy traffic conditions—from 360.8 seconds with SPF rerouting to 303.5 seconds. The models maintained high success rates, avoiding gridlock that plagued uncoordinated methods like Q-Routing, which failed entirely in some scenarios.

This research matters because it offers a practical path to reducing urban congestion without expensive infrastructure changes, potentially lowering emissions and improving daily commutes. By leveraging vehicle-to-infrastructure communication, the system could integrate with existing technologies in conventional and autonomous vehicles. However, limitations include the use of uniform traffic patterns in simulations, which may not capture real-world heterogeneity, and the need for further testing in larger metropolitan areas. The study's open-source code supports reproducibility, but real-world deployment complexities remain unexplored.