A principal component approach to analyzing simulation output

Abstract

A new procedure, called the principal component method, is developed to handle the problem of data correlation in simulation output analysis. The method is derived from matrix diagonalization theorems, which allow for an orthogonal transformation of data with an estimated covariance structure into a version of the data with uncorrelated structure. Matrix manipulation of this uncorrelated version of the data yields a derivation of an unbiased estimate of the underlying process mean and an estimate of the standard error of the mean. Using the Central Limit Theorem, the confidence interval is constructed. The performance of this confidence interval methodology is empirically tested over several independent replications of M/M/1 queueing models set at various utilization rates and of time series models with known correlation structures. Compared to the batched mean procedure, the principal component method provides good coverage, acceptable half-width information, and excellent bias information.