Apprentissage par renforcement multi-agents pour la régulation de la circulation routière dans les carrefours : application à la mobilité urbaine dans la ville de Kinshasa

Urban mobility in the city of Kinshasa is becoming a serious problem given the congestion of the roads by vehicles and pedestrians, leaving its users with various concerns. How can road traffic be regulated in such a way as to make automobile traffic smooth and safe? Several techniques or methods have been developed to solve this problem, such as light signals, variable message signs, presence of traffic police, etc. In this article we plan to propose the technique of light lights integrated into intelligent transport systems. This regulation technique works in a multi-agent environment, the main agent of which remains the traffic controller. This agent is in fact equipped with two very important elements including lights (effectors) and sensors (detectors). This union produces an available, safe, fast, comfortable and economical regulation service to its environment (vehicles and pedestrians). To make this system intelligent, we used one of the machine learning techniques called reinforcement learning. This technique helped us train our agents through the trial-and-error process using our table of values called (Q-values). The following methods were used: the Multi-agent System Engineering (MaSE) method to set up the multi-agent system, the yolov7 model for real-time object detection and the Q-Learning algorithm for learning. We have thus built a model which regulates road traffic according to traffic and major priorities but not according to the timer system or non-adaptive regulation as is the case for other models. We chose a game of 1000 episodes to simulate our model and we noted that after training our model, the agent gradually acquires performance from negative to positive score, which means that our model is good and the agent actually learns.