ERL: Edge Based Reinforcement Learning for Optimized Urban Traffic Light Control

Traffic congestion is worsening in every major city and brings increasing costs to governments and drivers. Vehicular networks provide the ability to collect more data from vehicles and roadside units, and sense traffic in real time. They represent a promising solution to alleviate traffic jams in urban environments. However, while the collected information is valuable, an efficient solution for better and faster utilization to alleviate congestion has yet to be developed. Current solutions are either based on mathematical models, which do not account for complex traffic scenarios or small-scale machine learning algorithms. In this paper, we propose ERL, a solution based on Edge Computing nodes to collect traffic data. ERL alleviates congestion by providing intelligent optimized traffic light control in real time. Edge servers run fast reinforcement learning algorithms to tune the metrics of the traffic signal control algorithm ran for each intersection. ERL operates within the coverage area of the edge server, and uses aggregated data from neighboring edge servers to provide city-scale congestion control. The evaluation based on real map data shows that our system decreases 48.71% average waiting time and 32.77% trip duration in normally congested areas, with very fast training in ordinary servers.