현실 제약 조건을 반영한 다중 교차로에서의 협력적 강화학습 기반 교통 신호 제어

As extreme weather events such as heatwaves and heavy rainfall caused by climate change increase, the importance of reducing greenhouse gas emissions has grown. This study proposes a cooperative reinforcement learning-based traffic signal control model at multi intersection to alleviate traffic congestion, one of the causes of greenhouse gas emissions in the transportation sector. The proposed model configures the state of the model and the Q-Network so that the agent in reinforcement learning considers the state and actions of its neighbors. Additionally, real-world constraints such as signal sequence and minimum green time are included to enhance real-world applicability. For validation, the proposed model was compared with a general reinforcement learning model and a model without constraints. As a result, the model with added multi-agent reinforcement learning was shown to be more efficient in signal control at multi intersection, as it showed less vehicle waiting time and CO2 emission compared to the general reinforcement learning model. Although the model without constraints showed relatively less vehicle waiting time and CO2 emission than the proposed model, but showed a large number of vehicle stops.