User-Centric Green Light Optimized Speed Advisory with Reinforcement Learning

We address Green Light Optimized Speed Advisory (GLOSA), an application in the field of Intelligent Transportation Systems (ITS) for improving traffic flow and reducing emissions in urban areas. The aim of this study is to improve GLOSA, both by including traffic condition information, more specifically queue length, into the calculation of an optimal speed as well as by applying Reinforcement Learning (RL). We incorporate rule-based classic GLOSA and RL-based GLOSA in a common comparable simulation environment. In doing so, performance is also examined considering action frequency in order to create a user-centric GLOSA system for settings of non-automated driving. Results show that incorporating queue information positively influences the performance of both, RL-agents and classic GLOSA systems. Both algorithms achieve the best results at the lowest investigated action frequency of an update every second. As the frequency decreases, the improvement compared to the baseline without any GLOSA diminishes. However, the decline is more pronounced for the RL-agent, so the classic GLOSA algorithm delivers better results on average when the action frequency reaches five seconds. We make the source code of this work available under: github.com/urbanAIthi/GLOSA-RL.