Real World Traffic Optimization by Reinforcement Learning: A Concept

Due to the growing urban population, the existing infrastructure and traffic control are successively reaching their limits, making an optimization of the traffic flow by intelligent control of Traffic Lights increasingly important. Previous research has already shown the basic suitability of Deep Reinforcement Learning (DRL) methods for TL control, for both, the optimization of single intersections and the optimization of traffic networks using Multi Agent Reinforcement Learning (MARL). A major gap in research concerning this area is the training and usage in real-life systems due to several challenges: (1) Training in real systems is difficult since agents cannot perform unrestricted arbitrary actions. (2) It cannot always be guaranteed that the learned policies are sufficiently robust. (3) DRL controllers must ensure that existing safety and operational constraints are enforced at all times.