Abstract
Sub-optimal control policies in intersection traffic signal controllers (TSC)contribute to congestion and lead to negative effects on human health and theenvironment. Reinforcement learning (RL) for traffic signal control is apromising approach to design better control policies and has attractedconsiderable research interest in recent years. However, most work done in thisarea used simplified simulation environments of traffic scenarios to trainRL-based TSC. To deploy RL in real-world traffic systems, the gap betweensimplified simulation environments and real-world applications has to beclosed. Therefore, we propose LemgoRL, a benchmark tool to train RL agents asTSC in a realistic simulation environment of Lemgo, a medium-sized town inGermany. In addition to the realistic simulation model, LemgoRL encompasses atraffic signal logic unit that ensures compliance with all regulatory andsafety requirements. LemgoRL offers the same interface as the wellknown OpenAIgym toolkit to enable easy deployment in existing research work. To demonstratethe functionality and applicability of LemgoRL, we train a state-of-the-artDeep RL algorithm on a CPU cluster utilizing a framework for distributed andparallel RL and compare its performance with other methods. Our benchmark tooldrives the development of RL algorithms towards real-world applications.