Estimating Gaussian Process Models for Large Data Sets in JMP® Pro 13
Aug 4, 2016 12:17 PM
Ryan Parker, PhD, JMP Research Statistician Developer, SAS
Gaussian process models are a popular way to emulate the output from deterministic computer models that, given inputs to some scientific process, construct the output associated with these inputs. Computer experiments are designed to collect samples from these models over the range of the inputs. We demonstrate how to build a Gaussian process emulator for large data sets collected from these computer experiments, such as those with more than 2,500 observations, by using an approximate likelihood technique to block the observations that is available in the Gaussian Process platform in JMP Pro 13. This technique allows for the use of parallel processing to dramatically reduce the computing time needed to estimate these models. Also, we demonstrate another new feature of the Gaussian process platform: the ability to estimate models with categorical inputs.