Development of machine learning based graph algorithm for smart intersections

In our world where the urban population is rapidly increasing, controlling resource consumption in traffic is one of the most important problems to be solved in terms of the sustainability of cities. The rapidly increasing city population brings with it the number of vehicles and the waiting time at intersections. Since changes in physical traffic topology take a long time, current studies are moving towards the development of adaptive traffic signaling methods. The main purpose of these studies is to optimize the number of vehicles passing per unit time by reducing the waiting times of vehicles in traffic. An efficient traffic signaling method makes a positive contribution to many areas such as fuel consumption, carbon monoxide gas emission, contribution to the individual economy, contribution to the national economy, individual time and efficiency. Intersection signaling optimization is a real-world problem affected by a large amount of real-time variable data. Maintaining maximum flow at a single intersection may cause congestion at the next intersection. For this reason, within the scope of the thesis, a Deep Reinforcement Learning (DRL) based graph method that can optimize phase and duration in order to ensure optimal flow throughout a line or area has been proposed. This method was developed on Simulation of Urban Mobility (SUMO) by integrating vehicle data directly into scaled city intersections. The method uses information on two-day vehicle traffic at two consecutive intersections. In this approach, the phase sequence is calculated by the DRL method. Phase duration is calculated using the maximum flow finding method of the Ford-Fulkerson algorithm. Signaling is carried out by combining phase sequence and duration. The proposed method was tested in the SUMO simulator by running parallel models as separate models on consecutive intersections. It has been observed that this approach reduces the queue length at intersections individually by 31%to 73%by using real maps and real data, and produces efficient results in solving traffic congestion by reducing the queue length by 61%at the general traffic level. The development and application of the method based on test results obtained from real vehicle data as well as synthetically generated vehicle data is promising.