MARLens: Understanding Multi-Agent Reinforcement Learning for Traffic Signal Control via Visual Analytics

Area of Science:

• Artificial Intelligence
• Urban Planning
• Data Visualization

Background:

• Traffic congestion hinders urban development, with intelligent traffic signal control (TSC) offering a solution.
• Reinforcement learning (RL) shows promise for TSC, but current evaluation metrics lack depth.
• Existing visual analysis tools are insufficient for multi-agent reinforcement learning (MARL) in complex traffic systems.

Purpose of the Study:

• To address the interpretability challenge in MARL for TSC.
• To introduce MARLens, a visual analytics system designed for MARL-based TSC.
• To provide researchers with a tool for exploring MARL decision-making and agent interactions in traffic management.

Main Methods:

• Developed MARLens, a visual analytics system with multiple visualization views.
• Integrated a traffic simulation module for replaying training scenarios.
• Conducted case studies, expert interviews, and a user study for validation.

Main Results:

• MARLens provides a versatile platform for exploring MARL features from multiple perspectives.
• The system reveals decision-making processes and inter-agent interactions in TSC.
• Validation through case studies and user feedback confirms the system's utility.

Conclusions:

• MARLens enhances the understanding of MARL-based TSC systems.
• The system facilitates more informed and efficient traffic management strategies.
• MARLens supports both RL and TSC researchers in practical implementation and further development.