Estimation of percentiles in discrete event simulation

Abstract

Percentiles are convenient measures of the entire range of values of simulation outputs. Analysts find percentiles partic ularly useful in establishing reasonable capacities for facil ities, comparing the overall performance of alternative de signs, or establishing minimum standards. Unfortunately, per centiles are not easy to estimate in most discrete simulations, since the observations are not independent or normally distrib uted.

Several methods are available for estimating percentiles in re generative simulations. Moore's method, which uses a com plex assignment of observations to subsamples, has the best statistical properties but requires a relatively large amount of computer time and storage. Methods that use grids or divide the observations into batches do not perform as well statisti cally but are more efficient computationally. For nonregener ative simulations, an approach is recommended which dis cards the transient period and divides the remaining observa tions into batches.