چكيده به لاتين
Intersections are an essential part of the urban streets network. According to research conducted by the Tehran Municipality, 24% of the users’ travel is spent waiting at intersections. Therefore, optimizing intersections can lead to reducing travel time and thus help improve citizens’ daily commute. In this research, by using reinforcement learning, an attempt has been made to achieve this goal. Adaptation to the existing conditions is the essential feature of using these algorithms. Accordingly, in this research, a learning environment and a traffic light controller algorithm have been developed. In the first step, the learning environment, including a part of Tehran with 20 square kilometers area, was built that could simulate seven consecutive days with the demand, which was obtained by analyzing the data collected from the induction loop detectors. In the next step, an algorithm based on reinforcement learning was developed and linked to SUMo. Compared to the base network, the proposed algorithm reduced travel time by 62% and improve intersection throughput by 3%. It was also shown that the algorithm used in this study has successfully increased network throughput over time and improved the performance of intersections in terms of reducing queue length.
