|
Abstract
|
We consider knowledge-gradient methods, a new class of fully sequential Bayesian information collection methods. These methods may be used whenever we must choose which information to collect, and how to balance the cost of collecting information with the benefits it may provide.
Knowledge-gradient methods have several attractive qualities: they are myopically optimal in general; they are asymptotically optimal and convergent in a broad class of problems; and they are flexible and perform well numerically. We demonstrate how these methods may be used for global optimization when function evaluation is expensive, and describe their use in two applied contexts: the development of a new cancer drug in collaboration with a medical group at Georgetown University Hospital, and the calibration of an optimization-based stochastic logistics model used by the truckload carrier Schneider National.
|