GNN-based Multi-Agent Control of Traffic Shockwaves in Sparse Vehicular Ad-hoc Networks

A decentralized MARL framework with Graph Neural Networks reduces stop-and-go wave propagation by up to 80% at only 10% CAV penetration

Prachi Nandi; Madhuri Malakar; Sonakshi Satpathy; Pabitra Mohan Khilar· 2026· DOI 10.48550/arXiv.2607.23792

The core problem

Traffic shockwaves are stop-and-go waves that propagate upstream through streams of vehicles and are a major cause of congestion, fuel inefficiency, and increased accident rates in modern transportation systems. Connected and Autonomous Vehicles (CAVs) offer a promising opportunity to mitigate such shockwaves, but most existing control strategies rely on global traffic state information. This reliance makes them impractical for early-stage deployment of Vehicular Ad-hoc Networks (VANETs), where connectivity is sparse and global information is unavailable. The authors address this gap by proposing a decentralized Multi-Agent Reinforcement Learning (MARL) framework that integrates a Graph Neural Network (GNN) to enhance the control architecture of connected and autonomous vehicles. The approach enables vehicles to learn cooperative control policies using locally available information and interaction with neighboring vehicles, making it suitable for realistic, sparse VANET conditions.

Simulation results show that the proposed GNN-based MARL framework can reduce the propagation of traffic shockwaves by up to 80%, even when only 10% of the vehicles are connected. This significant reduction is achieved without relying on global traffic state information, demonstrating the robustness of the decentralized approach in sparse VANET scenarios. The experiments were conducted under realistic highway traffic conditions, and the scalable simulation environment allowed for testing with varying penetration rates of connected vehicles. The key quantitative finding is the 80% reduction in shockwave propagation at a 10% CAV penetration rate, which highlights the potential of the method for early-stage deployment where connectivity is limited.

Traffic shockwaves are stop-and-go waves that propagate upstream through streams of vehicles and are a major cause of congestion, fuel inefficiency, and increased accident rates in modern transportation systems. Connected and Autonomous Vehicles (CAVs) offer a promising opportunity to mitigate such shockwaves, but most existing control strategies rely on global traffic state information. This reliance makes them impractical for early-stage deployment of Vehicular Ad-hoc Networks (VANETs), where connectivity is sparse and global information is unavailable. The authors address this gap by proposing a decentralized Multi-Agent Reinforcement Learning (MARL) framework that integrates a Graph Neural Network (GNN) to enhance the control architecture of connected and autonomous vehicles. The approach enables vehicles to learn cooperative control policies using locally available information and interaction with neighboring vehicles, making it suitable for realistic, sparse VANET conditions.

The proposed framework combines MARL with a GNN to enable decentralized control. Each vehicle acts as an agent that observes local state information and communicates with neighboring vehicles within its communication range. The GNN aggregates information from neighboring vehicles to produce a graph embedding that captures the local traffic topology. This embedding is then used by the MARL policy to select control actions (e.g., acceleration or deceleration) that dampen shockwave propagation. The architecture can be represented as a graph where nodes are vehicles and edges represent communication links. The GNN performs message passing:

The results indicate that integrating GNNs with MARL enables effective cooperative control using only local information, overcoming the limitations of centralized strategies that require global traffic state knowledge. The graph-based representation allows vehicles to implicitly learn the traffic topology and coordinate with neighbors, which is crucial in sparse VANETs where communication links are intermittent. The 80% reduction at 10% penetration suggests that even a small fraction of connected vehicles can significantly mitigate shockwaves if they are controlled intelligently. However, the study relies on simulation; real-world deployment may face challenges such as communication delays, packet loss, and heterogeneous vehicle dynamics. Future work could explore robustness to these factors and extend the framework to mixed traffic with human-driven vehicles. Overall, the paper presents a promising decentralized solution for traffic shockwave control in early-stage VANETs.

CS practitioners and researchers

Opening member content…