The slides for this talk were not shared by the speaker.

Abstract: I will especially focus on the use of multi-armed bandit theory in the context of linearly correlated distributions, optimization of Lipschitz functions, and the problem of sequential ranking. The talk will be based on recent results obtained by Stefan Magureanu and by Audrey Durand during their PhD.

You can download the slides of this talk.

Abstract: We will review how techniques from reinforcement learning (bandits, Markov decision processes) naturally occur in a wide variety of tasks related to intensional data management, i.e., management of data whose access is associated to a cost. This covers a range of applications, from Web crawling to crowdsourcing, from query optimization to management of virtual machines. In contrast with traditional reinforcement learning applications, data management scenarios often involve a very large but heavily structured state or action space, requiring adaptations of traditional techniques.