Reinforcement learning has received significant attention due to its success in training agents that play popular games such as Go, Starcraft II, Dota 2, and others. Inefficient exploration, one of the earliest problems recognized in the field, still limits the success of reinforcement learning approaches that do not require domain knowledge. Although techniques like posterior sampling convincingly solve hard exploration problems in simple domains (https://searchworks.stanford.edu/view/11891201), scalable exploration techniques remain elusive.

In this project, you will develop principled and scalable exploration techniques based on reducing model uncertainty (https://arxiv.org/abs/1609.04436). Besides benefiting from games as excellent testbeds, this project has the potential to radically improve automated playtesting.