Increase of Traffic Efficiency by Mutual Concessions of Autonomous Driving Cars Using Deep Q-Network

In recent years, autonomous operation technology has been actively developed in various research institutions and companies. Many experiments have been conducted on public roads to confirm whether an autonomous driving car can drive safely. However, there is a lack of research on autonomous driving operation for improving traffic efficiency with inter-vehicle communication. In our research, we implement mutual concessions of autonomous driving cars with Deep Q-Network (DQN), which is a deep neural network structure used for estimating the Q-value of the Q-learning method. Mutual concessions are a collective behavior in which a vehicle sometimes gives way to other vehicles and sometimes is given way by other vehicles. To verify the influence of mutual concessions, an experiment environment has been developed with radio control (RC) cars. Our experiment environment consists of up to 16 RC cars equipped with infrared LED markers and RaspberryPi3, an infrared camera for location estimation of the RC cars, a laptop controlling the RC cars through Wi-Fi, and a course of 6 m in length and 6 m in width. In this paper, mutual concessions of autonomous cars are implemented at the confluence at a roundabout. DQN is applied for the decision-making mechanism to decide speed at the roundabout based on the status of other cars. As a result of the experiment in our experiment environment, it is confirmed that mutual concessions at the roundabout were acquired with DQN, and that mutual concessions can increase traffic efficiency.