Modern machine learning systems are driving progress across multiple application domains, with the most striking successes in domains where vast amounts of data are readily available. Data are often expensive yet typically not gathered all at once. By being strategic about the data we aim to accelerate learning. In fact, we design the learning system itself to interactively conduct controlled experiments to collect most useful training data for the problem at hand.
We study questions of sample complexity and algorithm design for interactive learning in a variety of tasks including clinical trials, A/B testing, and policy identification in models for sequential strategic interaction. The mathematical analysis combines statistics, game theory, information theory and online optimization.
People
