Attention-based Spatio-Temporal Graph Convolutional Networks-enhanced deep reinforcement learning for adaptive...

Area of Science:

• Intelligent Transportation Systems
• Artificial Intelligence
• Urban Planning

Background:

• Conventional traffic signal control systems lack adaptability to dynamic urban traffic flows, leading to congestion.
• Existing deep reinforcement learning (DRL) methods often overlook complex spatio-temporal dependencies in road networks.
• Graph-based models are typically used for traffic prediction, not signal control optimization.

Purpose of the Study:

• To develop a novel adaptive traffic signal control framework.
• To address limitations in current DRL and graph-based approaches for traffic signal control.
• To improve urban traffic flow efficiency and reduce congestion.

Main Methods:

• Integration of Attention-based Spatio-Temporal Graph Convolutional Networks (ASTGCN) for feature extraction.
• Application of Multi-Agent Deep Deterministic Policy Gradient (MADDPG) for cooperative control.
• Utilizing the SUMO simulation platform for experiments with real-world and stochastic traffic data.

Main Results:

• The proposed framework reduced traffic congestion by 20%-25% compared to baseline methods.
• Demonstrated superior performance in convergence speed, generalization, and resilience to extreme traffic loads.
• Ablation studies confirmed the effectiveness of the ASTGCN component for spatio-temporal feature extraction.

Conclusions:

• The ASTGCN-MADDPG framework offers a scalable, data-driven solution for intelligent traffic signal control.
• This approach enhances urban mobility by adapting to dynamic traffic conditions.
• The study advances the development of smart city infrastructure and traffic management systems.