МЕТОДИ ОПТИМІЗАЦІЇ РОБОТИ СВІТЛОФОРІВ НА РЕГУЛЬОВАНИХ ПЕРЕХРЕСТЯХ ЗА ДОПОМОГОЮ ГЛИБИННОГО НАВЧАННЯ З ПІДКРІПЛЕННЯМ

Context. Intersections are the most critical areas of a road network, where the largest number of collisions and the longest waiting times are observed. The development of optimal methods for traffic light control at signalized intersections is necessary for improving the flow of traffic at existing urban intersections, reducing the chance of traffic collisions, the time it takes to cross the intersection, and increasing the safety for drivers and pedestrians. Developing such an algorithm requires simulating and comparing the work of different approaches in a simulated environment.Objective. The aim of this study is to develop an effective deep reinforcement-learning model aimed at optimizing traffic lightcontrol at intersections.Method. A custom simulation environment is designed, which is compatible with the OpenAI Gym framework, and two types of algorithms are compared: Deep Q-Networks and Proximal Policy Optimization. The algorithms are tested on a range of scenarios, involving ones with continuous and discrete action spaces, where the set of actions the agent may take are represented either by different states of the traffic lights, or by the length of traffic light signal phases. During training, various hyperparameters were also tuned, and different reward metrics were considered for the models: average wait time and average queue length. The developed environment rewards the agent during training according to one of the metrics chosen, while also penalizing it for any traffic rule violations.Results. A detailed analysis of the test results of deep Q network and Proximal Policy Optimization algorithms was provided. In general, the Proximal Policy Optimization algorithms show more consistent improvement during training, while deep Q network algorithms suffer more from the problem of catastrophic forgetting. Changing the reward function allows the algorithms to minimize different metrics during training. The developed simulation environment can be used in the future for testing other types of algorithms on the same task, and it is much less computationally expensive compared to existing solutions. The results underline the need to study other methods of traffic light control that may be integrated with real-life traffic light systems for a more optimal and safer traffic flow.Conclusions. The study has provided a valuable comparison of different methods of traffic light control in a signalized urban intersection, tested different ways of rewarding models during training and reviewed the effects this has on the traffic flow. The developed environment was sufficiently simple for the purposes of the research, which is valuable due to the large computational requirements of the models themselves, but can be improved in the future by expanding it with more complex simulation features, such as various types of intersections that aren’t urban, creating a road network of intersections that would all be connected to each other, adding pedestrian crossings, etc. Future work may be done to refine the simulation environment, expand the range of considered algorithms, consider the use of models for vehicle control in addition to traffic light control.