CS-ECON Seminar Series

Mechanism Design with Correlated Distributions: Limited Correlation and Uncertain Distributions

Speaker:Michael Albert
Date: Tuesday, May 31, 2016
Time: 12:00pm - 1:00pm
Location: Gross Hall 304B, Duke
Lunch will be served.

Abstract

The standard assumption in the mechanism design literature is that the agents valuations over the outcome are independently distributed. However, in many situations, it is natural to assume that agents valuations are correlated. When valuations are correlated, many of the standard negative results in mechanism design do not apply. Notably, in a monopolistic auction with n bidders, the seller can extract the full social surplus as revenue, given certain assumptions about the correlation structure. However, existing results have two limitations: 1) They provide sufficient, but not necessary conditions for full surplus extraction, and 2) they require perfect knowledge of the distribution over agents' valuations. In this talk, I will present research where we provide necessary and sufficient conditions for full surplus extraction under correlated valuations for the case of a monopolistic seller selling one good. Further, we extend the concept of incentive compatibility and individual rationality to encompass uncertainty over the distribution of agents' valuations and present a polynomial time algorithm in the agent types for designing these mechanisms. Finally, we demonstrate experimentally that these "robust" mechanisms can significantly outperform other mechanism design paradigms when the true distribution is estimated.

Biography

Michael Albert received his PhD in finance from Duke University in 2013. After receiving his PhD, he was a visiting assistant professor at The Ohio State University finance department in the Fisher School of Business from 2013 through 2015. Since then, he has been a post-doctoral researcher at the Learning Agents Research Group in the Department of Computer Science at the University of Texas at Austin working with Peter Stone.

Hosted by:
Vince Conitzer