Work

Knowledge Gradient Procedure to Select the Best System Under Pairwise Comparisons

Year: 2025

Type: article

Abstract: ABSTRACT This article considers fixed‐budget ranking and selection (R&S) problems where the performance of alternative designs can only be assessed through pairwise comparisons, a setting encountered in many applications, including player ranking in games, sports tournaments, recommender systems, image‐based search, public choice models such as voting schemes or decision rules in committees, and market research. Assuming Gaussian sampling noise, we successfully extend the knowledge gradient (KG) procedure to adaptively allocate the sampling budget for identifying the best design or the top‐ designs. The proposed algorithms inherit several attractive features of KG such as enabling exact computation and achieving asymptotic mean‐variance trade‐offs. Additionally, we investigate the asymptotically optimal sampling budget allocation ratio for the pairwise R&S problem within a large deviations framework. The derived balance conditions for the optimal ratio resemble those of traditional R&S but include specific features pertinent to pairwise comparisons. Lastly, we perform extensive numerical experiments to demonstrate that the proposed algorithms deliver competitive and robust finite‐budget performance compared with several other state‐of‐the‐art procedures.

Source: Naval Research Logistics (NRL)

DOI: 10.1002/nav.22268

Authors: Dongyang Li, Ek Peng Chew, Haobin Li, Enver Yücesan, Chun‐Hung Chen

Institutions: National University of Singapore, INSEAD, George Mason University

Language: English

Topic: Advanced Bandit Algorithms Research

Subfield: Management Science and Operations Research

Field: Decision Sciences

Domain: Social Sciences

Open Access status: hybrid

Estimated APC paid: $3,700

Funders: National Science Foundation, Singapore Maritime Institute