Development of Autonomous Lane Changing Method using Deep Reinforcement Learning in Complex Random Driving Road

Abstract

There have been many studies focusing on autonomous driving; particularly a lane changing system has become an important technology. This thesis proposes lane changing methods using deep reinforcement learning (DRL) in complex random driving road. Among DRL algorithms, Deep-Q Network (DQN) and Proximal Policy Optimization (PPO) are proposed and applied in the ego vehicle to make decisions in slope and curved road with random driving environments. In order to apply curve driving, the steering control for straight driving is modified by coordinate transformation and incremental change of the radius of the curved road. For complex driving, a reward combining position and velocity of vehicles is designed and an action composed of a four-step acceleration and a three-lane change is developed. To test the proposed methods, a simulator is developed as Google Colab based animation. Vehicles and lanes are displayed in graphic images and the lane changing process is demonstrated in animation platform. The existing driving models, Intelligent Driving Model (IDM) and Minimizing Overall Braking Induced by Lane changes (MOBIL) are compared with the proposed methods under the same driving conditions. Their performances are evaluated by the number of successful steps and the received rewards. This research contributes to design autonomous driving methods for safely coping with complex random driving road and to develop a Google Colab based simulator for the vehicle driving test.