Relu Patrascu, Craig Boutilier
Department of Computer Science
University of Toronto
Toronto, ON M5S 3H5
email: relu,cebly@cs.toronto.edu
Rajarshi Das, Jeffrey O. Kephart, Gerald Tesauro,
William E. Walsh
IBM T. J. Watson Research Center
19 Skyline Dr.
Hawthorne, NY 10532, USA
email: rajarshi,kephart,gtesauro,wwalsh1@us.ibm.com
Abstract
Autonomic (self-managing) computing systems
face the critical problem of resource allocation to
different computing elements. Adopting a recent
model, we view the problem of provisioning resources
as involving utility elicitation and optimization
to allocate resources given imprecise utility
information. In this paper, we propose a new
algorithm for regret-based optimization that performs
significantly faster than that proposed in earlier
work. We also explore new regret-based elicitation
heuristics that are able to find near-optimal
allocations while requiring a very small amount of
utility information from the distributed computing
elements. Since regret-computation is intensive,
we compare these to the more tractable Nelder-Mead optimization
technique w.r.t. amount of utility
information required.
To appear, AAAI-05
Return to List of Papers