Multi Agent Reinforcement Learning System for Vehicular and Pedestrian Traffic Control with Visible Light Communication

With the modernization of vehicles, new opportunities have emerged when it comes to automotive innovation. Due to the technological evolution of today’s new vehicles, it is possible for vehicles to be connected (CV) to each other, exchanging traffic and safety information.New communication technologies between vehicles and/or infrastructures can increase safety, and comfort while traveling. With a connected traffic environment, it is expected to obtain important information about the vehicular and pedestrian traffic flow in order to optimise it. Visible light communication (VLC) is then proposed as an integrated approach to optimize traffic flow at urban intersections.With this amount of information collected via VLC it is necessary to have a system that intelligently controls in real time all traffic and responds to all incoming vehicle’s manoeuvring requests. A two connected four-arm intersection scenario is considered. The simulation process involves using the SUMO urban mobility simulator. A multi agent reinforcement learning system, with a Deep Q-Learning algorithm is used to control both intersections, optimising traffic flows and minimise congestion during peak hours. A comparative analysis of the simulation results highlights the benefits of the proposed approach in terms of throughput, reducing delays and minimising vehicle stops, revealing improved patterns for optimising signals and trajectories.