Date and Location:
Monday, December 19, 12:00, A 0.23
Presenter:
Vasilis Gkatzelis
Title:
Decentralized
Welfare-Optimizing Mechanisms for Scheduling Games
Author:
Vasilis Gkatzelis
(NYU)
Abstract:
Game Theory and its algorithmic
counterpart, Mechanism Design, are by now standard tools for studying massive
decentralized engineering systems such as the Internet.In this context, one of
the most prominent issues concerns the design of mechanisms inducing socially
efficient outcomes, such as that of Vickrey. Unfortunately these socially
optimal mechanisms more often than not require full information and
prohibitively large computational resources, something opelessly impractical in
large distributed environments.
In this
work we study simple mechanisms that require only local information to operate
properly. The specific setting in which we describe our work is a classic
computationally hard scheduling problem in computer science (namely that of
unrelated machine scheduling to minimize the total weighted completion time;
here, selfish jobs choose a machine on which to be processed in order to
minimize the time at which they are released), though we expect that our results
extend to other settings with similar characteristics. Our findings indicate
that one can design simple and local mechanisms inducing outcomes whose social
cost is bounded by a small constant times that of the socially efficient
solution. Furthermore, we exhibit a somewhat counter-intuitive phenomenon by
noting that mechanisms yielding Pareto dominated outcomes may in fact enhance
the overall performance of the system. The latter is explained by observing that
the additional costs are actually well aligned with negative externalities that
players impose on others. We also find that the use of randomization in the
design of local mechanisms can lead to yet more improvement in the total
welfare.
Most of
the policies we consider have attractive properties--notably, they induce exact
potential games. This fact, along with the other game-theoretic insights gained,
allow us to give a new combinatorial approximation algorithm for the underlying
optimization problem. This constitutes an example where ideas from economics are
key to the development of algorithms for computationally difficult combinatorial
problems.