Abstract: 

No (external) regret is a standard performance guarantee in online learning. It ensures that, in hindsight, a learner performs nearly as well as the best fixed action. However, in strategic environments such as repeated auctions, this guarantee can be fragile: a strategic auctioneer can manipulate many widely adopted no-regret learning algorithms and extract the full welfare of a bidder using them, despite the bidder satisfying no regret. This vulnerability has motivated research on two central questions: which learning guarantees remain meaningful in strategic settings, and how can we design algorithms that are robust to strategic manipulation? In this talk, I will present two results that contribute to our understanding of these questions. First, I will describe a characterization of game-agnostic, non-manipulable algorithms in general repeated Bayesian games. Then, I will introduce a meta-algorithm that transforms any no-regret learning algorithm into a non-manipulable bidding algorithm for repeated auctions, while preserving its no-regret guarantee.