Conditioning Sparse Variational Gaussian Processes for Online\n Decision-making

摘要

With a principled representation of uncertainty and closed form posterior\nupdates, Gaussian processes (GPs) are a natural choice for online decision\nmaking. However, Gaussian processes typically require at least\n$\\mathcal{O}(n^2)$ computations for $n$ training points, limiting their general\napplicability. Stochastic variational Gaussian processes (SVGPs) can provide\nscalable inference for a dataset of fixed size, but are difficult to\nefficiently condition on new data. We propose online variational conditioning\n(OVC), a procedure for efficiently conditioning SVGPs in an online setting that\ndoes not require re-training through the evidence lower bound with the addition\nof new data. OVC enables the pairing of SVGPs with advanced look-ahead\nacquisition functions for black-box optimization, even with non-Gaussian\nlikelihoods. We show OVC provides compelling performance in a range of\napplications including active learning of malaria incidence, and reinforcement\nlearning on MuJoCo simulated robotic control tasks.\n