Bayesian Approaches to Optimal Information Gathering

The problem of gathering information in an ecient manner is critical in a wide range of elds. Animals and robots must keep track of rapidly changing environments with limited number of sensors in real-time. Therefore, choosing the right place to look at or to sense is important. Similarly in statistical modeling, selecting data to label or experiments to test hypotheses in an ecient manner is essential, especially when the labels and experiments are expensive. In this paper, we survey Bayesian approaches on optimal information gathering developed in various elds including experiment design, sensor placement, active sensing, and active learning. In particular, we cover the classical solutions for linear problems as well as new approximate methods emerging from the reinforcement learning and optimal control literature.