Deep Reinforcement Learning-Based Traffic Signal Control Using High-Resolution Event-Based Data

Area of Science:

• Intelligent Transportation Systems
• Traffic Engineering
• Artificial Intelligence

Background:

• Traffic congestion is a major urban problem.
• Current traffic signal control methods have limitations in data granularity and efficiency.
• Reinforcement learning (RL) shows promise for adaptive traffic signal control.

Purpose of the Study:

• To propose a cost-effective and efficient adaptive traffic signal control method using deep reinforcement learning (DRL).
• To leverage high-resolution event-based data for improved state definition in RL-based traffic signal control.
• To overcome the limitations of coarse or difficult-to-measure data in existing RL traffic control literature.

Main Methods:

• Developed a deep reinforcement learning (DRL) model for traffic signal control.
• Utilized high-resolution event-based data from vehicle-actuated detectors.
• Employed deep learning techniques to extract relevant traffic features from event data.
• Benchmarked the proposed DRL method against fixed-time and actuated control strategies.

Main Results:

• The proposed DRL method significantly outperforms fixed-time and actuated control strategies.
• High-resolution event-based data provides informative features for traffic signal control.
• Deep learning effectively extracts useful features for adaptive control.
• The method demonstrates cost-effectiveness and efficiency in alleviating traffic congestion.

Conclusions:

• Deep reinforcement learning with high-resolution event-based data offers a superior approach to traffic signal control.
• The proposed method provides a more efficient and cost-effective solution for adaptive traffic signal control.
• This approach has the potential to significantly alleviate urban traffic congestion.