Thompson Sampling Procedures for Ranking and Selection with Pairwise Comparisons and Binary Outcomes

We consider ranking and selection (R&S) problems where the performance of competing designs or alternatives can only be assessed by evaluating two alternatives simultaneously with a binary outcome; the true performance of an alternative is quantified by its average probability of outperforming other alternatives. We discuss the challenges associated with applying conventional R&S techniques to this particular setting and propose heuristic algorithms based on Thompson sampling to overcome those difficulties. Through a series of simple numerical experiments, we assess the effectiveness of these heuristics and highlight open research questions.