Adaptive urban traffic signal control based on enhanced deep reinforcement learning

Area of Science:

• Intelligent Transportation Systems
• Artificial Intelligence
• Traffic Engineering

Background:

• Current intelligent transportation systems focus on traffic signal control (TSC) for urban efficiency.
• Deep Reinforcement Learning (DRL) is widely used but faces challenges with slow convergence due to inefficient sample selection and requires enhanced robustness for varied traffic conditions.

Purpose of the Study:

• To propose an enhanced method for traffic signal control (TSC) based on Deep Reinforcement Learning (DRL).
• To improve training efficiency, model robustness, and overall performance in urban traffic management.

Main Methods:

• Utilized dueling network and double Q-learning to mitigate DRL's overestimation issues.
• Implemented a priority sampling mechanism for efficient memory utilization.
• Integrated noise parameters into the neural network for enhanced robustness.
• Represented traffic data as matrices and employed a phase-cycled action space.
• Designed a reward function reflecting real-world traffic scenarios.

Main Results:

• Demonstrated faster convergence compared to standard DRL methods.
• Achieved optimal performance in reducing queue length and waiting times.
• Experimental results confirmed the method's robustness across diverse traffic flow scenarios.

Conclusions:

• The proposed DRL-based TSC method offers significant improvements in training efficiency and performance.
• The enhancements lead to more robust and effective traffic signal control, particularly in complex urban environments.