Optimal decision making in black-box online auctions markets

A commonly studied problem in the context of Algorithmic Game Theory is the problem of behavior optimization in infinitely repeated games. While the computation of equilibria in the case of non-repeated games is generally a well-defined problem, the computation of equilibria in infinitely repeated games can be quite difficult. Multiple methods have however been developed in order to tackle this problem, notably in the context of repeated online ad auctions, but most of them require strong assumptions about the behavior of the other players. In this thesis, we propose a method for behavior optimization in repeated ad auctions basing ourselves on the assumption that the highest adverse bids for an ad slot are independent realizations of unknown random variables. This assumption is used to show that our problem of behavior optimization can be seen as a problem of censored regression of these unknown random variables. We use our estimation of the different distributions to formulate a probabilistic optimization problem. Finally, we validate the performance of our algorithm using simulated data and conclude this thesis by providing numerical results proving the convergence of our algorithm to the optimal bid.