Fall 2006 Seminar Series

DATE: Thursday, October 26, 2006
LOCATION: E40-298
TIME: 4:15pm
Reception immediately following in the Philip M. Morse Reading Room, E40-106

SPEAKER:
Daniel Bienstock

TITLE
Experiments in Robust Optimization

ABSTRACT
We present ongoing experimental work involving Robust Optimization in various applications, primarily (in this talk) Portfolio Optimization.

 

Robust Optimization is a framework for managing optimization in the presence of data uncertainty, and is typically offered as a counterpoint to Stochastic Programming, which requires known and precise information on stochastic distributions of data. In principle, the basic machinery of Robust Optimization can be extended to any form of data uncertainty, no matter what its magnitude or structure might be. A criticism that has been leveled against Robust Optimization  is that it can prove too conservative, or not conservative enough, and, in particular, it has been criticized for assigning the same "weight" or importance, to all possible realizations of data in the uncertainty set. Part of this difficulty stems from the fact that the typical uncertainty sets that have been examined in the literature are "well behaved" whereas in several practical settings data can misbehave in "malformed" or "lumpy" ways.

 

In this talk we will focus on the use of Benders' decomposition (or, more generally, cutting plane algorithms) to handle realistic and complex uncertainty models.  Our implementations prove successful on large, difficult portfolio optimization problems.