Approximate Gaussian Process Regression with Pairwise Comparison Data
Efe Sertkaya, Ilya O. Ryzhov · 2023
We use approximate Bayesian inference, together with Gaussian process regression, to create a new estimator for an unknown function in a situation where we can only observe pairwise comparisons of function values at different inputs. Preliminary experimental results suggest that, although information is heavily censored in this setting, it may still be possible to learn the local and global minima of the underlying function. We discuss possible sampling criteria, and explore the performance of the "probability of improvement" strategy numerically.