Show simple item record

dc.contributor.advisorBarman, Siddharth
dc.contributor.authorSawarni, Ayush
dc.date.accessioned2024-04-29T04:57:35Z
dc.date.available2024-04-29T04:57:35Z
dc.date.submitted2024
dc.identifier.urihttps://etd.iisc.ac.in/handle/2005/6499
dc.description.abstractWe study regret in online learning from a welfarist perspective and explore an application of bandit algorithms in causal inference. We introduce Nash regret, which measures the difference between the optimal action choices and the algorithm’s performance in terms of the Nash social welfare function. By providing bounds on Nash regret, we establish principled fairness guarantees for online learning algorithms. We investigate different online learning settings and derive tight bounds on Nash regret. Furthermore, we study the problem of finding optimal interventions in causal graphs by providing theoretical guarantees within the context of the causal bandit problem. In the first part, we focus on the classic multi-armed bandit (MAB) framework and develop an algorithm that achieves a tight Nash regret bound. Specifically, given a horizon of play T , our algorithm achieves a Nash regret of O ✓q k log T T ◆ , where k represents the number of arms in the MAB instance. The lower bound on average regret applies to Nash regret as well, making our guarantee essentially tight. Additionally, we propose an anytime algorithm with a Nash regret guarantee of O ✓q k log T T log T ◆ . In the second part, we study the stochastic linear bandits problem with non-negative, ⌫-sub Poisson rewards. We present an algorithm that achieves a Nash regret bound of O ⇣q d⌫ T log(T |X|) ⌘ , where X denotes the set of arms in ambient dimension d and T represents the number of rounds. Furthermore, for linear bandit instances with potentially infinite arm sets, we derive a Nash re- gret upper bound of O ⇣ d 5/4 ⌫ 1/2 pT log(T ) ⌘ . Our algorithm builds upon the successive elimination method and incorporates novel techniques such as tailored concentration bounds and sampling via the John ellipsoid in conjunction with the Kiefer-Wolfowitz optimal design. In the third part, we investigate Nash regret in the context of online concave optimization and the Experts problem, assuming adversarially chosen reward functions. Our algorithm achieves Nash regret of O log N T for the Experts problem where N is the number of experts. We provide a lower bound for this setting that is essentially tight with respect to the upper bound. Additionally, for online concave optimization, we provide a Nash regret guarantee of O d log T T , where d denotes the ambient dimension. In the final part of this thesis, we focus on the causal bandit problem, which involves iden- tifying near-optimal interventions in a causal graph. Previous works have provided a bound of eO(N/pT ) for simple regret for causal graphs with N vertices, constant in-degree, and Bernoulli random variables. In this work, we present a new approach for exploration using covering in- terventions. This allows us to achieve a significant improvement and provide a tighter simple regret guarantee of eO(pN/T ). Furthermore, we extend our algorithm to handle the most general case of causal graphs with unobservable variablesen_US
dc.language.isoen_USen_US
dc.relation.ispartofseries;ET00508
dc.rightsI grant Indian Institute of Science the right to archive and to make available my thesis or dissertation in whole or in part in all forms of media, now hereafter known. I retain all proprietary rights, such as patent rights. I also retain the right to use in future works (such as articles or books) all or part of this thesis or dissertationen_US
dc.subjectNash regreten_US
dc.subjectmulti-armed banditen_US
dc.subjectKiefer-Wolfowitz optimal designen_US
dc.subjectalgorithmsen_US
dc.subject.classificationResearch Subject Categories::TECHNOLOGY::Information technology::Computer science::Computer scienceen_US
dc.titleBandit Algorithms: Fairness, Welfare, and Applications in Causal Inferenceen_US
dc.typeThesisen_US
dc.degree.nameMTech (Res)en_US
dc.degree.levelMastersen_US
dc.degree.grantorIndian Institute of Scienceen_US
dc.degree.disciplineEngineeringen_US


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record