Currently, I am working in two areas: multiclass classification and model-based reinforcement learning, but my research interests include supervised learning, online learning and reinforcement learning. I have been approaching machine learning research in a theoretical way, and I am eager to explore problems where practical challenges and empirical observations guide the development of theory, and theory guides the design of better algorithms.

Multiclass classification. My interests in this problem stem out from calibration and how to design/evaluate convex losses for multiclass classification, but touch on other issues such as sample complexity. I am also curious to see how developments in calibration and sample complexity work in structured prediction problems.

Model-based reinforcement learning. We wish to design/describe methods for planning in MDPs. What should the model of an MDP be like? What are the properties of a good model? How flexible can we make our models, without losing these properties?