Deep Reinforcement Learning for Traffic Signal Control Model and Adaptation Study.

Area of Science:

• Intelligent Transportation Systems
• Artificial Intelligence in Traffic Management
• Traffic Signal Optimization

Background:

• Traditional traffic detection methods lack accuracy for complex signal optimization.
• Advanced sensor technology enables highly accurate traffic state data collection.
• Deep reinforcement learning (DRL) offers a promising approach for intersection signal control.

Purpose of the Study:

• To explore state space variations under different traffic scenarios using advanced sensor data.
• To analyze the relationship between traffic demand and actual traffic states.
• To develop a DRL model capable of generalizing to new traffic scenarios without retraining.

Main Methods:

• Utilized advanced sensor technology for real-time, accurate traffic data acquisition.
• Employed deep reinforcement learning, specifically a Deep Q-Network (DQN) model.
• Trained the model on a comprehensive state space derived from diverse traffic scenarios.
• Evaluated model performance against the popular Soft Actor-Critic (SAC) signal control model.

Main Results:

• The DQN model demonstrated effective learning from a comprehensive traffic state space.
• The model successfully generalized to new traffic scenarios without additional training.
• The proposed DQN model achieved an average delay of 5.13 seconds, outperforming the SAC model's 6.52 seconds.

Conclusions:

• Advanced sensor data enhances DRL for traffic signal optimization.
• The developed DQN model exhibits superior traffic control performance and adaptability.
• This approach offers a robust solution for complex urban traffic management challenges.