Reducing Urban Traffic Congestion Using Deep Learning and Model Predictive Control.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Control Systems

Background:

• Traffic congestion is a significant urban problem.
• Existing traffic signal control methods struggle with dynamic traffic conditions.
• Advanced control algorithms are needed for efficient traffic management.

Purpose of the Study:

• To propose a novel deep learning (DL)-based control algorithm for traffic congestion reduction.
• To integrate DL for system identification with velocity-based model predictive control (VMPC) for traffic signal control.
• To evaluate the effectiveness of the proposed DL-VMPC algorithm under slowly time-varying traffic signal controls.

Main Methods:

• Developed a DL-VMPC algorithm combining DL-based system identification and VMPC for traffic signal control.
• Utilized a modeling error entropy loss criterion for DL training, inspired by stochastic distribution control (SDC).
• Conducted simulations to assess the algorithm's performance in reducing traffic congestion.

Main Results:

• The proposed DL-VMPC algorithm effectively reduced traffic congestion with slowly time-varying traffic signal inputs.
• Ablation studies confirmed the algorithm's superiority over other model-based controllers.
• Demonstrated advantages in prediction error, signal varying speed, and overall control effectiveness.

Conclusions:

• The DL-VMPC algorithm presents a promising approach for intelligent traffic signal control.
• This method offers improved traffic management solutions for dynamic urban environments.
• The integration of DL and VMPC enhances control performance and reduces congestion.