A Comparative Analysis of Deep Reinforcement Learning-Enabled Freeway Decision-Making for Automated Vehicles

Abstract

In application, advanced autonomous driving technologies still face numerous challenges. Deep Reinforcement Learning (DRL) has emerged as a widespread and effective approach to address artificial intelligence challenges, due to its substantial potential for autonomous learning and self-improvement. In this study, four DRL algorithms—Deep Q-Learning (DQN), along with its enhanced algorithm, Double DQL, Dueling DQL, and Priority Replay DQL(PR-DQN), are employed to address decision-making challenges for autonomous vehicles on highways, with a comprehensive comparative analysis conducted. The decision-making model is constructed as a Markov Decision Process, guided by specially designed reward functions, enabling the target vehicle to learn safe and efficient decision-making strategies through multiple environmental explorations. Through the analysis and discussion of a series of experimental results, the feasibility of DRL-based decision strategies is demonstrated. Finally, through comparing the experimental outcomes of different algorithms, the connection between autonomous driving results and the inherent learning features of these DRL technologies is analyzed.

Keywords

• Decision-making
• deep reinforcement learning
• autonomous vehicles
• DQL
• double DQL
• dueling DQL