Multi-Agent Deep Reinforcement Learning for Automated Highway Driving

Abstract

Recent advances in Deep Reinforcement Learning have sparked new interest in many different research topics, including Automated Highway Driving where agents model autonomous vehicles. The main advantage of Deep Reinforcement Learning is that the training algorithm is adaptable to its environment. In highway driving, researchers often simplify the framework of an agent by using lower level controllers and observers. However, agent observations do not yet include lane change intentions of surrounding vehicles. Resulting agents were able to drive on a maximum of three lanes and unfit to drive in lane changing environments. We aim to simplify the current state-of-the-art agent frameworks even further to improve performance. We also believe that observing other vehicles lane-change intent, or blinker status, is essential for collision avoidance in a highway driving environment. In this paper, we try to implement multi-agent Deep Reinforcement Learning on a six-lane highway, including lane changes. After training, agents are able to avoid collisions while reaching destination lanes. Moreover, a lane-selection strategy according to desired speed evolved from open freeway training.