Urban traffic congestion incurs high economic and environmental costs, including fuel wastage, increased greenhouse gas emissions, and driver frustration. Current genetic algorithm-based traffic routing systems do not adapt when real-time conditions deviate from their offline optimization assumptions, exacerbating congestion during incidents and demand surges. These systems lack three critical capabilities: online adaptation without retraining, coordination to prevent route oscillation, and transfer across cities with different topologies. To address these gaps, we formulate adaptive traffic routing as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) and propose MA-GRL, a multi-agent graph reinforcement learning framework. MA-GRL combines Graph Attention Networks (GAT) with Multi-Agent PPO under the Centralized Training with Decentralized Execution (CTDE) paradigm. Each vehicle executes a decentralized policy that processes local traffic observations through a 3-hop GAT encoder. A novel coordination reward implicitly penalizes simultaneous route switches by nearby agents, enabling stable cooperation without explicit communication. Experiments on the Monaco MoST scenario with 100 vehicles over 3600-second episodes show that MA-GRL reduces average travel time by 11. 1% compared to genetic algorithm routing (p < 0. 001, Cohen’s d = 0. 80), recovers from 10% road closures within 45 steps, and achieves 87% zero-shot transfer retention across Monaco, Luxembourg, and Bologna city scenarios. Our framework demonstrates the potential for scalable, adaptive, and transferable traffic management systems that can alleviate congestion and its associated societal burdens.
Mahmoud et al. (Thu,) studied this question.