Top-k selection with pairwise comparisons

In this work we consider active, pairwise top- selection, the problem of identifying the highest quality subset of given size from a set of alternatives, based on the information collected from noisy, sequentially chosen pairwise comparisons. We adapt two well known Bayesian sequential sampling techniques, the Knowledge Gradient policy and the Optimal Computing Budget Allocation framework for the pairwise setting and compare their performance on a range of empirical tests. We demonstrate that these methods are able to match or outperform the current state of the art racing algorithm approach