Frans A. Oliehoek | Exploiting structure in learning and planning

Reinforcement learning (RL) and more generally sequential decision making deal with problems where the decision maker ('agent') needs to take actions over time. While impressive results have been achieved on challenging domains like Atari, Go, and Starcraft, most of this work relies on neural networks to form their own internal abstractions. However, in many applications, we may be able to exploit some knowledge about the structure of the environment to guide this process. In this talk I will give an overview of some of my work, with a focus on approaches that try to exploit structure to define effective methods for planning and reinforcement learning.