Chergui Oussama2025-05-192025-05-192024-10-05https://repository.univ-msila.dz/handle/123456789/46183As urban landscapes evolve rapidly, marked by exponential growth in both population and vehicles, this thesis addresses the pressing issue of traffic congestion through the implementation of an innovative multi agent adaptive control algorithm for traffic lights In a smart city. Leveraging VANET technology for real-time communication and data-sharing across multiple signalized intersections and vehicles, our algorithm, named Self-Attention Multi-Agents Proximal Policy Optimization (SAMAPPO), aims to alleviate congestion in intersections with diverse traffic flows and traffic network map sizes. Utilizing the simulation tool SUMO for realistic traffic simulation, our algorithm demonstrates its efficacy in varying traffic flows as well as small and large network maps. An additional strength of our algorithm lies in its scalability, showcasing superior performance in larger networks without requiring retraining the whole model, thanks to the incorporation of transfer learning, which reduces the computation costs associated with training. Our implementation proves to be a practical solution for congestion in smart cities, offering scalability to accommodate higher traffic flows and larger network maps. The success of our algorithm suggests its potential to address the traffic congestion challenge posed by evolving urban traffic scenarios.ensmart cityvehicular ad hoc network (VANET)Adaptive traffic lights controlreinforcement learningmulti-agentsSelf-attentionProximal Policy OptimizationThesis