Opportunities of robust regression for variance reduction in discrete event simulation

Abstract

Variance reduction increases the efficiency of discrete event simulation. The method of using control variates searches for variables with known mean which are correlated with the response variable. The more correlated they are, the more the variance of the response variable can be reduced to obtain a confidence interval. Mostly only a few replications are made. This enables outliers to give a false view on the reduction possibilities. To cope with this statistical problem, we propose the use of a robust regression method in determining optimal weights for the control variables.