In this talk, we study the problem of learning unknown input distributions when there is some prior knowledge about the inputs, and a cost on obtaining each training example that can be used to improve this knowledge during the exploration phase. Problems of this nature are ubiquitous in clinical trial design, optimal classifier selection, and in improving system performance by judiciously learning about system parameters.

This talk presents an overview of active learning frameworks, and surveys policy design for important special cases. We then present some new results on designing approximately optimal policies.