CCGN: Centralized collaborative graphical transformer multi-agent reinforcement learning for multi-intersection...

Area of Science:

• Artificial Intelligence
• Traffic Engineering
• Network Science

Background:

• Current multi-agent reinforcement learning models for traffic signal control suffer from local optimization, leading to inefficient traffic flow and increased travel times.
• These models lack inter-intersection communication and adaptability to dynamic traffic conditions, necessitating a more integrated approach.

Purpose of the Study:

• To develop a centralized collaborative graph network (CCGN) for traffic signal control.
• To achieve a signal-free corridor by enabling seamless traffic flow between intersections.

Main Methods:

• The proposed CCGN model integrates local policy networks (LPN) and global policy networks (GPN).
• LPN utilizes Transformer and Graph Convolutional Network (GCN) for intersection-level action prediction.
• GPN employs GCN and Q-network to manage intersections and facilitate signal-free corridors, with a specific implementation called Deep Graph Convolution Q-Network (DGCQ).

Main Results:

• The CCGN model, particularly DGCQ, demonstrated superior performance in creating signal-free traffic corridors.
• Evaluations on real-world traffic networks showed that the proposed model outperforms existing state-of-the-art methods.
• The system effectively leverages GCN for intersection collaboration and DQN for information aggregation.

Conclusions:

• The centralized collaborative graph network (CCGN) offers a significant advancement in traffic signal control.
• This approach effectively reduces traffic congestion and enhances travel efficiency by creating signal-free corridors.
• The model's ability to adapt and collaborate across intersections provides a robust solution for complex traffic scenarios.