|
Abstract:
|
For the past twenty years the Bayesian calibration of computer models has been an active area of research and application. In this scenario a computer code has two types of inputs: controllable (x) and calibration (w) variables. The goal of calibration is to use experimental/field data at specified x-vectors in order to obtain a posterior distribution of the true value of w and subsequently predict at new values of x. In the absence of controllable inputs x, prediction can no longer be a goal— only estimation. Furthermore, the discrepancy function cannot be a Gaussian process on the x-space. In this talk I discuss calibration of these “parameter-only” computer models which have non-functional multivariate output. Motivated by Nonlinear Regression, the issues of identifiability, non-informative priors, and the importance of understanding the input-output map will be covered. The ideas will be illustrated using a nuclear physics model.
|
