Statistics Seminars: Some thoughts on the design of computer experiments
19 February 2008 14:15 in CM221
Experiments carried out in a computer are often expensive to run, sometimes taking months of computer time. If we are to perform a statistical analysis of such experiments we need to design them carefully. The standard analysis is to build an emulator for the computer code and use this to perform statistical inference. In the design we have two competing priorities: to span the parameter space of the computer code and to estimate the smoothness of the emulator. So far most designs have concentrated on the former problem looking at ways of covering a multidimensional space efficiently. I will discuss whether we can add a small number of points to such a design to produce better estimates of the smoothing (roughness) parameters.
Contact email@example.com for more information