Minimum mean-squared-error estimators for simulation experiments

• Published in: Communications of the ACM 1981-04, Vol.24 (4), p.253-259
• Main Authors: Donnelly, James H., Shannon, Robert E.
• Format: Magazinearticle
• Language: English
• ISSN: 0001-0782; 1557-7317
• DOI: 10.1145/358598.358637

The use of the mean-squared-error of the sample mean as a criterion for simulation experiments has been suggested in the literature. Interest in this criterion has led to the development of a proposed systematic methodology for obtaining a near optimal mean-squared-error estimator for the mean of a discrete, random event, autocorrelated process which has experienced a transient start-up condition. Application of the methodology to a pilot simulation process provides an experimenter with specific guidance in terms of the number of warm-up observations to delete and the number of simulation replications to conduct. Investigations with simple autoregressive and queueing processes have indicated that the methodology may be applicable to other simulation processes. The methodology may off a possible alternative to the widely used rules-of-thumb employed by simulation practitioners for speedifying simulation model operating conditions.