Coordination of Urban Traffic Lights based on the Analysis of Mean Travel Speeds by Artificial Neural Networks and Deep Reinforcement Learning: A Case Study in the City of Yaoundé

This study focuses on the prototype of a traffic light control system to enhance signalized intersection coordination in urban environments using artificial intelligence techniques to analyze mean travel speeds. The system looks at integrate an artificial neural network for mean travel speeds prediction and a reinforcement learning model for speed-based traffic signal control. Field data were collected from Google Maps API for a road section in Yaoundé, Cameroon, namely the road axis joining the Messassi junction to the Nlongkak roundabout. The latter was used to simulate traffic conditions. It comes out that the speed regression neural network shows a high predictive accuracy with a mean square error of 1.095 and a R² of 0.908 for real speeds. The reinforcement learning traffic light control model, implemented in a simulation in the SUMO software and guided by rewards based solely on mean travel speeds, generated adaptive signal policies that outperformed fixed-cycle traffic light control systems. Simulation results revealed a reduction in the value of the average coefficient of variation of speeds—47.75% in one direction and 18.85% in the other, namely due to left-turns distribution along the axis and between directions—indicating improved coordination. These findings demonstrate the potential of AI-driven approaches combined with mean travel speeds as an alternative to the use of physical sensor-based data, to optimize urban traffic flow and reduce congestions.