Comparison of Traffic Control with Model Predictive Control and Deep Reinforcement Learning

Traffic congestion is among the worst causes of pollution, and the time spent in traffic can cost the world tens of billions of dollars every year. Solutions to mitigate this problem are at hand thanks to the advent of advanced control techniques and artificial intelligence (AI). Traditional traffic light control strategies based on fixed timing of the green, yellow and red phases are simple to implement, but at the same time very inefficient, in particular for busy intersections. This paper discusses both a model predictive control (MPC) approach and a model-free deep reinforcement learning (DRL) algorithm for controlling the traffic lights at a single intersection, with the aim of improving the traffic flow. Firstly, a detailed linear mathematical model of an intersection is formulated and successively tested in a MPC framework; secondly, a DRL algorithm is proposed and verified by comparing it with the currently implemented baseline controller. Finally, the results for the three approaches, MPC, DRL and the baseline controller, are validated through the SUMO (Simulation of Urban Mobility) microscopic traffic simulator.