• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 1
  • Tagged with
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
1

Multi-armed bandits with unconventional feedback / Bandits multi-armés avec rétroaction partielle

Gajane, Pratik 14 November 2017 (has links)
Dans cette thèse, nous étudions des problèmes de prise de décisions séquentielles dans lesquels, pour chacune de ses décisions, l'apprenant reçoit une information qu'il utilise pour guider ses décisions futures. Pour aller au-delà du retour d’information conventionnel tel qu'il a été bien étudié pour des problèmes de prise de décision séquentielle tels que les bandits multi-bras, nous considérons des formes de retour d’information partielle motivées par des applications pratiques.En premier, nous considérons le problème des bandits duellistes, dans lequel l'apprenant sélectionne deux actions à chaque pas de temps et reçoit en retour une information relative (i.e. de préférence) entre les valeurs instantanées de ces deux actions.En particulier, nous proposons un algorithme optimal qui permet à l'apprenant d'obtenir un regret cumulatif quasi-optimal (le regret est la différence entre la récompense cumulative optimale et la récompense cumulative constatée de l’apprenant). Dans un second temps, nous considérons le problème des bandits corrompus, dans lequel un processus de corruption stochastique perturbe le retour d’information. Pour ce problème aussi, nous concevons des algorithmes pour obtenir un regret cumulatif asymptotiquement optimal. En outre, nous examinons la relation entre ces deux problèmes dans le cadre du monitoring partiel qui est un paradigme générique pour la prise de décision séquentielle avec retour d'information partielle. / The multi-armed bandit (MAB) problem is a mathematical formulation of the exploration-exploitation trade-off inherent to reinforcement learning, in which the learner chooses an action (symbolized by an arm) from a set of available actions in a sequence of trials in order to maximize their reward. In the classical MAB problem, the learner receives absolute bandit feedback i.e. it receives as feedback the reward of the arm it selects. In many practical situations however, different kind of feedback is more readily available. In this thesis, we study two of such kinds of feedbacks, namely, relative feedback and corrupt feedback.The main practical motivation behind relative feedback arises from the task of online ranker evaluation. This task involves choosing the optimal ranker from a finite set of rankers using only pairwise comparisons, while minimizing the comparisons between sub-optimal rankers. This is formalized by the MAB problem with relative feedback, in which the learner selects two arms instead of one and receives the preference feedback. We consider the adversarial formulation of this problem which circumvents the stationarity assumption over the mean rewards for the arms. We provide a lower bound on the performance measure for any algorithm for this problem. We also provide an algorithm called "Relative Exponential-weight algorithm for Exploration and Exploitation" with performance guarantees. We present a thorough empirical study on several information retrieval datasets that confirm the validity of these theoretical results.The motivating theme behind corrupt feedback is that the feedback the learner receives is a corrupted form of the corresponding reward of the selected arm. Practically such a feedback is available in the tasks of online advertising, recommender systems etc. We consider two goals for the MAB problem with corrupt feedback: best arm identification and exploration-exploitation. For both the goals, we provide lower bounds on the performance measures for any algorithm. We also provide various algorithms for these settings. The main contribution of this module is the algorithms "KLUCB-CF" and "Thompson Sampling-CF" which asymptotically attain the best possible performance. We present experimental results to demonstrate the performance of these algorithms. We also show how this problem setting can be used for the practical application of enforcing differential privacy.

Page generated in 0.1263 seconds