A Deep Reinforcement Learning Approach for Autonomous Highway Driving. Issue 5 (2020)

• Record Type: Journal Article
• Authors: Zhao, Junwu; Qu, Ting; Xu, Fang
• Is Part Of: IFAC-PapersOnLine. Volume 53:Issue 5(2020)
• Page Start: 542
• Page End: 546
• Publication Date: 2020
• Subjects: Autonomous Driving -- Deep Reinforcement Learning -- Vehicle Control -- Traffic Modeling -- Traffic Simulator; Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2021.04.142
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: Autonomous driving has been the trend. In this paper, a Deep Reinforcement Learning (DRL) method is exploited to model the decision making and interaction between vehicles on highway driving. To avoid the overestimate action values induced by Q-learning, we use the Double Deep Q-Network (DDQN) for the training of the host vehicle. The agent learns from trial and interactions with the environment. A simulation platform based on the Simulation of Urban Mobility (SUMO) is also established, it helps facilitate the variation of control algorithms. The results show that the proposed framework can simulate highway driving, and the trained agent can accomplish the driving task with ease after training and can approximate the highest safe driving speed as defined without collision.