Deep reinforcement learning based multi-autonomous vehicle control

Abstract

The method of determining the algorithm to control the unmanned autonomous vehicle was gradually achieved through a learning - based approach in the existing rule - based system. As the road traffic ratio of autonomous vehicles increases, there is an advantage in reducing the accident rate, improving fuel economy, and reducing congestion. In the weaving section, the interactions between the vehicles due to the lane change of the vehicle frequently occur, so that the process of making decisions based on interests is very complicated. In order to learn the autonomous driving model, labeled big data is essential. In this study, the vehicle driving simulator is produced to continuously generate data necessary for learning. We study the deep reinforcement learning algorithm that controls the vehicle through the generated data, and evaluate the traffic engineering performance in the weaving section.