This thesis is a review of bandit algorithms in information retrieval. In information retrieval a result list should include the most relevant documents and the results should also be non-redundant and diverse. To achieve this, some form of feedback is required. This document describes implicit feedback collected from user interactions by using interleaving methods that allow alternative rankings of documents to be presented in result lists. Bandit algorithms can then be used to learn from user interactions in a principled way. The reviewed algorithms include dueling bandits, contextual bandits, and contextual dueling bandits. Additionally coactive learning and preference learning are described. Finally algorithms are summarized by using regret as a performance measure.
