TITLE:  Sequential learning in computer and other experiments, with a flexible additive mode

SPEAKER: Hugh Chipman

ABSTRACT:

Sequential design, or "active learning" can be an effective way to plan a experiment, so as to gain maximal information about a response model. The data-generating mechanism as well as the scientific objective can have important influence on the way in which the design is generated, and the estimated response model. For example, if the objective is to maximize response, we may only be interested in accurate estimates near the maximum. In computer experiments, Gaussian process models are a common approach, and have been used for sequential design and optimization. Instead we use an adaptive nonparametric regression model ("Bayesian Additive Regression Trees", or BART) to deal with nonstationarities and other complex relationships. By providing both point estimates and uncertainty bounds for prediction, BART provides a basis for sequential design criteria to ?nd optima with few function evaluations. Other applications, including sequential design in high-throughput screening for drug discover will also be discussed.