Research
Research Interests
- Intelligent Systems
- Reinforcement Learning
- Generative AI
Traffic Signal Cycle Control Under Varying Intervention Frequencies
This study introduces an innovative joint phase traffic signal cycle control method that operates effectively across varying control intervals. Our approach features an 'adjust all phases' action design, enabling simultaneous phase changes within the signal cycle, promoting immediate stability and sustained effectiveness, even at lower frequencies. Additionally, the method utilizes decentralized actors to manage the complexity of the action space, with a centralized critic ensuring coordinated phase adjustments. This approach addresses the limitations of existing reinforcement learning-based TSC systems, offering a practical solution for real-world applications. [Paper] | [Code]
The Framework of Our Method (CCDA) with Intervention Frequency.
A Universal RL-based Framework for Traffic Signal Control
Traffic congestion in urban areas demands effective traffic signal control (TSC) systems. Current Reinforcement Learning (RL) methods, while promising, struggle to generalize across different intersection structures. This research introduces a universal RL-based TSC framework, featuring a novel agent design that uses a junction matrix to represent intersection states. This design enhances the model's applicability to diverse intersections. Additionally, tailored traffic state augmentation methods are developed to improve the model's ability to manage various intersection configurations. The framework allows for efficient training of a robust base model, which can be quickly fine-tuned for specific intersections, addressing the generalization challenges of existing RL-based TSC methods. [Paper] | [Code]
The Overall Framework of Our Universal RL-Based TSC Method.