Verkehrseffiziente und energieoptimierte Längsführung eines autonomen Fahrzeugs mittels Deep Reinforcement Learning

Autoren

T. Brinkmann, M.Sc., M. Wegener, M.Sc., L. Koch, M.Sc., Univ.-Prof Dr.-Ing. J. Andert, Research and Teaching Area for Mechatronics in Mobile Propulsion, RWTH Aachen University

Zusammenfassung

Potentially, the future emergence of connected and automated vehicles in real-world traffic environments can not only improve road safety, but also achieve significant energy savings for the entire transportation system. This contribution investigates the effect of the connectivity and automation of a single vehicle on the energy consumption in urban traffic. This vehicle is equipped with a longitudinal controller whose control strategy is generated based on a model free Reinforcement Learning (RL) agent. The goal of this agent is to maximize a predefined reward function to reduce the vehicle's energy demand by using sensor data and vehicle-to-infrastructure information. Unlike other state of the art algorithms in the context of optimized longitudinal control, the implemented RL approach does not require an explicit model to predict the future traffic situation. The urban traffic environment is modelled using the microscopic traffic simulator SUMO with a probabilistic driver model for non-automated vehicles. Using a microscopic simulator enables the quantification of both, the impact on the ego vehicle's and the system’s energy consumption. The results of this investigation indicate energy savings of up to 30 % with a 3 % increase in travel time for the ego-vehicle compared to a human driver. In contrast to human driving behavior, the agent’s strategy showed an independence of the energy consumption from favorable traffic light timings. Beyond these results, a slightly positive impact on the energy demand for the surrounding vehicles was shown.

Mitglieder-Login