Identifying Copeland Winners in Dueling Bandits with Indifferences
26-09-2023 11:30
Escuela Superior de Ingeniería, Avenida de España, Albacete, España
Organizado por José Antonio Gámez Martín
Dueling bandits, a variant of multi-armed bandits, involves finding the best choice alternative through noisy preference feedback. This approach finds utility in applications such as information retrieval and the analysis of voting behavior, where only preference comparisons can be made. This talk introduces an intriguing dimension: dealing with feedback that reflects indifference between choices. This phenomenon is commonly encountered in human-provided feedback. The talk delves into strategies for swiftly identifying the optimal selection by leveraging the Copeland score while minimizing feedback requests. It also includes an exploration of the lower bound on sample complexity and introduces the POCOWISTA algorithm, showcasing its robust empirical performance across both standard and indifference-based dueling bandits scenarios.
Ponente: Viktor Bengs (LMU Munich)
Día: 26 septiembre 11´30h
Lugar: Salón de actos del I3A