Fairness constraints and reward manipulation in stochastic multi-armed bandits

Κονταλέξη, Μαρίνα; Kontalexi, Marina

dc.contributor.author	Κονταλέξη, Μαρίνα	el
dc.contributor.author	Kontalexi, Marina	en
dc.date.accessioned	2025-01-13T12:13:45Z
dc.date.available	2025-01-13T12:13:45Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/60718
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.28414
dc.rights	Default License
dc.subject	Multi-armed bandits	en
dc.subject	Algorithmic fairness	en
dc.subject	Regret	en
dc.subject	Adversarial corruption	en
dc.title	Fairness constraints and reward manipulation in stochastic multi-armed bandits	en
heal.type	bachelorThesis
heal.classification	Αλγόριθμοι	el
heal.classification	Online learning	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2024-06-12
heal.abstract	This thesis studies the stochastic multi-armed problem, where a learner plays a sequential game with an environment for T rounds. In each round the learner chooses one of the K available arms to pull and receives a stochastically generated reward. The goal of the learner is to perform as the best policy in hindsight. Optimal algorithms can guarantee that the learner’s regret is bounded by O (\sqrt{KT}), which matches the lower bound obtained by information theory. Joseph et al. [1] imposed a fairness constraint on the learner’s actions, that restricts her from favoring an arm (i.e., pull it with higher probability) unless the arm is of greater merit. Our work proposes an ε-relaxation of their fairness definition and a fair algorithm that achieves O(\sqrt{1/ε}\sqrt{KT}) regret. Applications where fairness is sought after (like recommendation systems) are vulnerable to adversarial attacks (e.g., fake reviews) thus we present the behaviour of known algorithms in the mixed model and aspire to connect fair algorithms with robustness to adversarial corruption.	en
heal.advisorName	Φωτάκης, Δημήτρης	el
heal.committeeMemberName	Ποδηματά, Χαρά	el
heal.committeeMemberName	Παγουρτζής, Αριστείδης	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών. Τομέας Τεχνολογίας Πληροφορικής και Υπολογιστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	69 σ.	el
heal.fullTextAvailability	false