Abstract: Computer models with high -imensional inputs require many samples to get accurate emulators. After many (say 10,000) samples from a computer model, classical computer model emulation techniques either break down or become completely infeasible. The common solution is to start approximating the emulation technique, which can completely eliminate the benefit of having many samples.
This talk will discuss how one can deploy adaptive but structured computer experiments that can easily scale to the 10,000 or 100,000 samples régime with minimal computational overhead and good numerical stability. The inference is exact; meaning there is no degradation your emulator caused by sloppy approximations. The R package CGGP package is now on CRAN, which implements this idea.