Batch Size Effects in the Analysis of Simulation Output.
Abstract
This paper considers the effects of using fewer batches than are necessary to satisfy normality and independence assumptions. Using too few batches results in (1) correct probability of covering the mean, (2) an increase in expected half length, (3) an increase in the standard deviation of the half length, and (4) an increase in the probability of covering incorrect values of the mean (analogous to Type II error in hypothesis testing). These effects, quantified here, are shown to be small when at least eight to ten batches are used, with least effect on confidence intervals having low confidence values. With the effects of using too few batches quantified, a simulation practitioner can make the trade-off between the ease of using very few batches with known independence and normality versus using a batching algorithm to squeeze some remaining information from the data. For researchers developing batching algorithms, the results are useful in selecting initial batch sizes. The results may also be useful in the context of using independent replications to establish confidence intervals on the mean. Finally, some criteria and a procedure are suggested for Monte Carlo comparison of confidence interval procedures. These suggestions are not restricted to batch mean algorithms.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 1980
- Accession Number
- ADA087045
Entities
People
- Bruce Schmeiser
Organizations
- Purdue University