Linde Institute/SISL Seminar: Haifeng Xu, USC

Persuasion, defined as the act of exploiting an informational advantage in order to influence the decisions of others, is ubiquitous. Indeed, persuasive communication has been estimated to account for a considerable share of all economic activity in the US. In this talk, I will examine persuasion through a computational lens and look to design algorithms for optimal or approximately optimal persuasion. 

I will start with one of the most fundamental models in this space, namely the Bayesian Persuasion model of Kamenica and Gentzkow. Here there are two players, a sender and a receiver. The receiver must take one of a number of actions with a-priori unknown payoff. The sender, on the other hand, has access to additional information regarding the payoffs of various actions for both players, and looks to persuade the receiver to take an action that is more preferable by the sender. We examine the sender's optimization problem in three of the most natural input models and pin down the computational complexity for each. Next, I will move to a natural generalization of the above model where one sender wishes to persuade multiple receivers in order to optimize a global objective that depends on all the receivers' actions. Here, we focus on a basic setting where each receiver takes a binary action and the sender's objective function is monotone submodular. We propose efficient approximate algorithms that match the complexity-theoretic upper bound. Finally, I will also briefly mention some applications of the Bayesian persuasion model in real world.

Based on joint work with Shaddin Dughmi, Zinovi Rabinovich, Milind Tambe.

The talk will be mainly based on the following two papers: Algorithmic Bayesian Persuasion and Private Bayesian Persuasion with Monotone Submodular Objectives.