Corelab Seminar

Chara Podimata
Learning in the presence of strategic agents

Machine Learning currently dominates almost every research activity from the most experimental works, to the most theoretical ones. Learning in the context of strategic presence can be significantly different than learning under random or adversarial noise. On the other hand, real life scenarios (like auctions) impose extra informational structure for the learning agents and should help drive new developments in the already existent Machine Learning literature. In an effort to cover both the aforementioned aspects, in this talk I am going to first outline families of Linear Regressors robust to strategic noise and secondly, a bidding strategy for learning bidders that do not know their valuation for the item(s) being auctioned a priori that achieves exponentially better convergence guarantees compared to generic multi-armed bandit algorithms. This talk is based on joint works with Yiling Chen & Nisarg Shah and Zhe Feng & Vasilis Syrgkanis.