Biased Pressure: Cyclic Reinforcement Learning Model for Intelligent Traffic Signal Control

Area of Science:

• Intelligent Transportation Systems
• Artificial Intelligence in Urban Planning
• Traffic Engineering

Background:

• Inefficient traffic signal systems contribute to significant urban congestion.
• Current deep reinforcement learning (RL) traffic control models often lack real-world applicability due to complex definitions and ignored constraints.

Purpose of the Study:

• To develop a scalable and practical RL-based traffic light control model.
• To address limitations of existing RL methods by incorporating real-world traffic signal constraints.
• To introduce a novel Biased Pressure (BP) method for improved traffic signal management.

Main Methods:

• A novel RL-based multi-intersection traffic light control model is proposed.
• The model utilizes a simple yet effective combination of state, reward, and action definitions.
• A Biased Pressure (BP) method and an advantage actor-critic learning mechanism are employed.
• Decentralized definitions ensure model scalability.

Main Results:

• The proposed model demonstrates superior performance compared to existing methods on both synthetic and real-world datasets.
• Significant improvements in throughput and average travel time were observed.
• Ablation studies confirmed the effectiveness of the Biased Pressure (BP) method over traditional pressure methods.

Conclusions:

• The developed RL model offers a practical and efficient solution for real-world traffic signal control.
• The Biased Pressure (BP) method represents a significant advancement in traffic signal management strategies.
• The model's scalability and performance improvements contribute to mitigating urban traffic congestion.