Towards efficient and robust reinforcement learning via synthetic environments and offline data

Thesis

Over the past decade, Deep Reinforcement Learning (RL) has driven many advances in sequential decision-making, including remarkable applications in superhuman Go-playing, robotic control, and automated algorithm discovery. However, despite these successes, deep RL is also notoriously sample-inefficient, usually generalizes poorly to settings beyond the original environment, and can be unstable during training. Moreover, the conventional RL setting still relies on exploring and learning tab...

Expand abstract

Authors: Lu, C

• DOI: 10.5287/ora-6g1v05xvz
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: world models; generative modeling; deep reinforcement learning; offline reinforcement learning
• Subjects: Machine Learning; Reinforcement Learning