Attention-based multi-agent reinforcement learning for traffic flow stability in mountainous tunnel entrances

Area of Science:

• Intelligent Transportation Systems
• Artificial Intelligence
• Traffic Engineering

Background:

• Mountain tunnel entrances face complex road conditions and nonlinear coupling effects, leading to unstable traffic flow, reduced efficiency, and safety concerns.
• Traditional rule-based traffic regulation methods struggle to address the dynamic and complex nature of traffic flow in these environments.

Purpose of the Study:

• To propose a novel Multi-agent Fusion Double-Dueling-Deep Q-Network Traffic Flow (MF3DQN-TF) framework for optimizing traffic flow in mountain tunnel entrances for Connected and Autonomous Vehicles (CAVs).
• To enhance traffic efficiency, stability, and safety by leveraging multi-agent deep reinforcement learning and attention mechanisms.

Main Methods:

• Developed a Multi-agent Fusion Double-Dueling-Deep Q-Network Traffic Flow (MF3DQN-TF) framework integrating multi-agent deep reinforcement learning and attention mechanisms.
• Conducted comparative experiments and simulations to evaluate the framework's performance against traditional rule-based methods in various traffic scenarios.

Main Results:

• The MF3DQN-TF framework demonstrated significant improvements, boosting traffic flow stability by over 15% and vehicle efficiency by approximately 20%.
• Congestion time was reduced by 18%, average vehicle speed increased by 25%, and the traffic congestion index decreased by 22% compared to conventional methods.
• The attention mechanism enhanced intelligent agents' decision-making efficiency, enabling real-time vehicle interaction and coordination optimization.

Conclusions:

• The proposed MF3DQN-TF framework effectively stabilizes traffic flow and alleviates common traffic issues at mountain tunnel entrances.
• The framework enhances the overall adaptability of intelligent transportation systems, supporting their development and deployment.