Optimal Control of Traffic Flow Based on Reinforcement Learning

Abstract

We study approaches to use (real-time) data, communicated between cars and infrastructure, to improve and to optimize traffic flow in the future and, thereby, to support holistic, efficient and sustainable mobility solutions. To set up virtual traffic environments ranging from artificial scenarios up to complex real world road networks, we use microscopic traffic models and traffic simulation software SUMO. In particular, we apply a reinforcement learning approach, in order to teach controllers (agents) to guide certain vehicles or to control infrastructural guidance systems, such as traffic lights. With real-time information obtained from other vehicles, the agent iteratively learns to improve the traffic flow by repetitive observation and algorithmic optimization. For the RL approach, we consider different control policies including widely used neural nets but also Linear Models and Radial Basis Function Networks. Finally, we compare our RL controller with other control approaches and analyse the robustness of the RL traffic light controller, especially under extreme scenarios.