Active Learning for Deep Gaussian Process Surrogates

Deep Gaussian processes (DGPs) are increasingly popular as predictive models in machine learning for their nonstationary flexibility and ability to cope with abrupt regime changes training data. Here, we explore DGPs surrogates computer simulation experiments whose response surfaces exhibit similar characteristics. In particular, transport a DGP’s automatic warping of the input space full uncertainty quantification, via novel elliptical slice sampling Bayesian posterior inferential scheme, through active strategies that distribute runs nonuniformly space—something an ordinary (stationary) GP could not do. Building up design sequentially this way allows smaller sets, limiting both expensive evaluation simulator code mitigating cubic costs DGP inference. When data sizes kept small careful acquisition, parsimonious layout latent layers, framework can be effective computationally tractable. Our methods illustrated on two real varying dimensionality. We provide open source implementation deepgp package CRAN.