Graph Reinforcement Learning-Based Decision-Making Technology for Connected and Autonomous Vehicles: Framework,...

Area of Science:

• Intelligent Transportation Systems
• Artificial Intelligence
• Robotics

Background:

• Connected and autonomous vehicles (CAVs) are vital for future transport, but mixed autonomy traffic (CAVs and human-driven vehicles) presents challenges.
• Effective decision-making for CAVs is crucial for safety and efficiency during the transition to full autonomy.
• Deep reinforcement learning (DRL) has shown promise, but graph reinforcement learning (GRL) offers superior capabilities for modeling complex vehicle interactions.

Purpose of the Study:

• To provide a comprehensive review of GRL-based methods for CAV decision-making.
• To establish a generic GRL framework for understanding autonomous driving decision technologies.
• To identify challenges and future research directions in GRL for autonomous driving.

Main Methods:

• Reviewing GRL methods from the perspective of mixed autonomy traffic construction.
• Examining graph representation techniques for dynamic driving environments.
• Summarizing related works on graph neural networks (GNN) and DRL in autonomous driving.
• Compiling validation methods for assessing decision-making performance.

Main Results:

• GRL methods demonstrate significant potential for improving CAV decision-making by accurately representing inter-vehicle effects and dynamic traffic.
• The review categorizes GRL approaches based on traffic composition, environmental representation, and underlying network architectures.
• Key validation strategies for GRL-based autonomous driving systems are identified.

Conclusions:

• GRL is a promising approach for advancing CAV decision-making in mixed traffic environments.
• Further research is needed to address the challenges and fully leverage the potential of GRL for autonomous driving.
• This review serves as a foundational resource for researchers developing GRL-based solutions for autonomous vehicles.