Decentralized graph attention multi-agent reinforcement learning for adaptive urban traffic routing.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Computer Science

Background:

• Urban traffic congestion leads to significant economic and environmental costs.
• Existing genetic algorithm routing systems fail to adapt to real-time traffic changes, worsening congestion during incidents.
• Current systems lack online adaptation, inter-agent coordination, and cross-city transferability.

Purpose of the Study:

• To develop an adaptive traffic routing system that addresses the limitations of current methods.
• To formulate adaptive traffic routing as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP).
• To propose MA-GRL, a multi-agent graph reinforcement learning framework for intelligent traffic management.

Main Methods:

• MA-GRL combines Graph Attention Networks (GAT) with Multi-Agent PPO using the CTDE paradigm.
• Vehicles use decentralized policies with GAT encoders for local traffic observations.
• A novel coordination reward encourages stable cooperation without explicit communication.

Main Results:

• MA-GRL reduced average travel time by 11.1% compared to genetic algorithms.
• The system recovered from 10% road closures within 45 steps.
• Achieved 87% zero-shot transfer retention across different city scenarios.

Conclusions:

• MA-GRL offers a scalable, adaptive, and transferable solution for traffic management.
• The framework effectively alleviates urban congestion and its societal impacts.
• This approach shows promise for future intelligent transportation systems.