Spring 2013 Seminar Series

DATE: March 21st
LOCATION: E51-315
TIME: 4:15pm
Reception immediately following

SPEAKER:
John Birge

TITLE
Markov Chain Monte Carlo Methods for Stochastic Optimization

ABSTRACT
Dynamic stochastic optimization models invariably suffer from the curse of dimensionality, leading to a variety of approximation techniques and various uses of Monte Carlo methods to decrease the computational burden of increased dimension. This talk will discuss an approach using Markov chain Monte Carlo (MCMC) methods which have a long history of use in deterministic optimization but less obvious interpretations for dynamic stochastic optimization. The procedure uses a risk-sensitive objective to form a proper prior, the addition of auxiliary variables to form Gaussian mixture models to represent different objective forms, and slice and split-sampling to reduce variance. Convergence results and relationships to other approaches such as robust optimization, approximate dynamic programming, and direct Monte Carlo will be presented along with computational results on test cases.