Performance comparison of MSER-5 and N-Skart on the simulation start-up problem

Abstract

We summarize some results from an extensive performance comparison of the procedures MSER-5 and N-Skart for handling the simulation start-up problem. We assume a fixed-length simulation-generated time series from which point and confidence-interval (CI) estimators of the steady-state mean are sought. MSER-5 uses the data-truncation point that minimizes the half-length of the usual batch-means CI computed from the truncated data set. N-Skart uses a randomness test to determine the data-truncation point beyond which spaced batch means are approximately independent of each other and the simulation's initial condition; then using truncated nonspaced batch means, N-Skart exploits separate adjustments to the CI half-length that account for the effects on the distribution of the underlying Student's t-statistic arising from skewness and autocorrelation of the batch means. In most of the test problems, N-Skart's point estimator had smaller bias than that of MSER-5; moreover in all cases, N-Skart's CI estimator outperformed that of MSER-5.