Akademik Veri Yönetim Sistemi
Science, Engineering Management and Information Technology First International Conference, SEMIT 2022, Virtual, Online, 2 - 03 Şubat 2022, cilt.1809 CCIS, ss.107-120, (Tam Metin Bildiri)
In vehicular networks, one of the traffic light signal parameters is the current inefficient traffic light control and causes problems such as long delay and energy waste. To improve traffic efficiency, dynamically adjusting the traffic light duration, taking into account real-time traffic information, is a logical and reasonable method. In this study; a deep reinforcement learning model is proposed to control the traffic light. In order to reduce the waiting time of the intersection users, the model and timing of the changes were optimized using deep reinforcement learning for the signals. In addition to the existing studies, Shibuya Crossing is chosen as an exemplary intersection application, focusing on encrypted intersections as the application target of traffic control with deep reinforcement learning. A traffic simulation SUMO is used to create the perimeter of Shibuya Crossing. Traffic signals are optimized using DQN, A2C and PPO algorithms. As a result, by using reinforcement learning, the waiting time has been reduced by about four times compared to the signal patterns currently used. In the study, the behavior of the optimized signal is also analyzed, explaining how the accuracy of the learning process changes when the method or condition observation is changed.