CS 789 THEORY SEMINAR [home]
Speaker: Eric
Friedman
Affiliation: Computer Science, Cornell University
Date: Monday, February 4, 2002
Title: Selfishness, Learning, and Mechanism Design on the Internet
Abstract:
In this paper we consider the foundations of
game-theoretic analysis for many problems arising on the Internet. We find that,
in general, selfish learning algorithms do not converge to the Nash equilibrium.
However, they do converge to the serially unoverwhelmed set, which we define. We
then consider the mechanism design problem for the Internet, in which one
designs network protocols to achieve good system performance under selfish
learning. While such protocols are easy to design if one assumes
convergence to the Nash equilibrium, they are much more difficult to design for
the selfish learning case, in which play does not converge to the Nash
equilibrium. In fact, we prove that the set of learnable social choice functions
(achievable performance levels) is quite restrictive under selfish learning.
joint work with Scott Shenker