University of Glasgow
Training-efficient deep reinforcement learning for safe autonomous driving
Abstract
dc:description.abstractAutonomous driving holds the potential to transform the transportation industry, offering significant improvements in safety, efficiency, and convenience. However, traditional model-based planning approaches struggle to address the complexities and uncertainties of real-world driving environments. This thesis employs deep reinforcement learning (DRL) to achieve safe and efficient autonomous driving using realistic simulation settings and evaluation based on rational criteria. The proposed framework integrates five key factors—driving safety, driving efficiency, training efficiency, unselfishness, and interpretability (DDTUI) to ensure reliable and optimal decision-making across various driving scenarios. The research addresses two primary applications: highway driving and autonomous racing. In highway driving, the DRL-based framework demonstrates superior performance compared to popular baseline algorithms, improving safety and efficiency. In autonomous racing, an extreme case of autonomous driving, the framework is adapted to manage high velocities and safe control, achieving fewer collisions, faster lap times, and reduced training time in comparison to benchmark algorithms. This thesis contributes to the field by advancing RL-based planning techniques and establishing a design methodology for integrating key factors in autonomous driving. The results of this study provide evidences of the development of safer, more efficient, and interpretable autonomous driving systems. Finally, key achievements are summarized, limitations are discussed, and future research directions are proposed.
Degree
thesis:*- Level dc:type.qualificationlevel
- PhD
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zhen, Tian
Subjects
dc:subject × 2Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.85027