Gaussian Processes and Model Emulation

Sampling-based estimation of the posterior distribution is computationally demanding. We have already mentioned the continuing search for efficient MCMC algorithms. MCMC is especially slow when the model of interest is a process-based model (PBM) with a long run-time. In such cases, it may be good to replace the PBM with a faster surrogate model. The surrogate model will take the same inputs as the original model but calculate the output more quickly. However, its output cannot be exactly the same as that of the original model, so it just provides an approximation. If the surrogate model is a statistical model that produces not just the approximative prediction of what the original model would have produced, but a whole probability distribution, then it is called a statistical emulator, or just emulator for short.