Name: Boesel
Firstname: Justin
Title: Sr. Simulation and Modeling Engineer
Institution: The MITRE Corporation
Street: 7515 Colshire Dr
City: McLean, VA
Zip-Code: 22102-7508
Country: USA
Phone: 703 883 6993
Fax: 703 883 1911
Email: boesel@mitre.org
Authors: Justin Boesel
Title: Combinations of subset selection and indifference-zone procedures to select the best of a large number of systems.
Abstract: This talk presents statistically valid extensions for combining subset-selection and indifference-zone (IZ) procedures to find the best system among a large number of stochastic systems. These extensions take greater advantage of mean and variance information gained during the course of experimentation. A subset-selection procedure returns a random-sized subset that contains the best of the k systems. An IZ procedure guarantees to select the best system by doing sampling in two or more stages. In a typical IZ procedure, the total sample size required of each system increases with the total number of systems being compared and decreases with the initial sample size. Subset selection and IZ procedures can be combined by "restarting" the experiment by performing a complete two-stage IZ procedure on only those systems returned by the subset procedure. One can also perform a "rolling screen," where second-stage samples of a few systems are taken before subset selecti! on; these systems, with their reduced sample variance due to second-stage sampling, are better able to screen out inferior systems. The first extension described in this talk improves the restart procedure by using variance information from the first-stage data-information that is usually discarded-to optimize the initial sample size for the restarted experiments. The second extension improves the rolling screen procedure by sorting the systems by their first-stage (search) sample means before performing the rolling screen.