There are many opportunities to automate real-world control problems, based on the increase of data capture and storage, coupled with recent advances in machine learning. For example, in control of wind turbines, an autonomous agent could use wind, acceleration and strain sensors to learn and continually adapt the wind turbine controller, improving efficiency and reducing structure load. In this talk, I will discuss two key aspects of developing autonomous, adaptive learning agents: off-policy reinforcement learning and convex representation learning. I will discuss how reinforcement learning is a simple yet general formalism for learning controllers from massive streams of data. However, the success of machine learning techniques, including reinforcement learning techniques, often hinge on the quality of the representation of the data. I will discuss my contributions to representation learning particularly for noisy, high-dimensional data. Tackling these two important problems concurrently is a significant step towards the deployment of autonomous agents in data-rich, real-world applications.

Martha White is a Ph.D. candidate at the University of Alberta, in the Department of Computing Science, pursuing research in machine learning. Her M.Sc. research work was on reducing variance in measuring agent skill in game theory; since then, her focus has been convex representation learning and reinforcement learning. She is passionate about developing autonomous learning agents, particularly for computational sustainability.