Robust Linear Optimization and Coherent Risk Measures
David Brown
Department of Electrical Engineering and Computer Science
Massachusetts Institute of Technology
Friday, October 21, 2005
1:30 - 2:45 PM
Terman Engineering Center, Room 453
Abstract:
The field of robust optimization is relatively silent on the issue of
how to construct uncertainty sets; in this work, we propose an
axiomatic methodology for constructing uncertainty sets within the
framework of robust linear optimization. We take as primitive for such
problems a coherent risk measure, and utilize a well-known duality
theorem for these measures to construct corresponding uncertainty sets
in the robust optimization framework. Our approach focuses on the
practical scenario in which we only have observations of the uncertain
data but no detailed distributional information. We study classes of
risk measures which give rise to polyhedral uncertainty sets. A
subclass of these measures corresponds to centrally symmetric,
polyhedral uncertainty sets and naturally induces a norm space. We
show that the class of risk measures corresponding to tail moments
leads to robust optimization problems with norm-bounded uncertainty
sets. We also find that the converse construction holds; that is,
given a convex uncertainty set, there exists a corresponding coherent
risk measure which is equivalent. Finally, we show some computational
evidence illustrating the value of this approach.
Operations Research Colloquia: http://or.stanford.edu/oras_seminars.html