Traffic signal control(TSC) has become an important issue for urban traffic management. Recent studies utilize reinforcement learning(RL) in traffic signal control since it has the advantages of no requirement for prior knowledge and real-time control. However, these studies mainly focused on control performance of training scenarios, but ignore the adaptability to different intersection topologies and flow distributions. Furthermore, most studies employed an impractical phase-selection scheme with an unfixed phase order that may confuse human drivers. To address these issues, we propose PhaseLight, a method combining lane-based representation, sophisticated network structure and advanced reinforcement learning algorithm. It is capable of adapting to various intersection topologies and flow distributions without additional training. Meanwhile, it employs a phase-switching scheme to improve practicality with little performance loss. Comprehensive experiment are conducted using the Simulation of Urban MObility(SUMO) simulator. The results in both training and testing scenarios demonstrate the effectiveness of PhaseLight, indicate its potential in real-world applications.
PhaseLight: An Universal and Practical Traffic Signal Control Algorithms Based on Reinforcement Learning
2023-09-24
2490324 byte
Conference paper
Electronic Resource
English
Reinforcement learning-based traffic signal control
IEEE | 2024
|Deep Reinforcement Learning-based Traffic Signal Control
IEEE | 2023
|Traffic signal control method based on reinforcement learning
European Patent Office | 2023
|