Combining Correlation Induction and Control Variates in Screening and Selection Procedures

• Main Authors: Chen-HaoKuo, 郭宸豪
• Other Authors: Shing-Chih Tsai
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/76595930276664800928

碩士 === 國立成功大學 === 工業與資訊管理學系碩博士班 === 98 === The propose of using Ranking and Selection Procedures is to find superior systems form all of candidate systems. However, if the variance of system output is large, we need to sampling more sample to find best system to guarantee confidence level. Therefore, we apply Variance Reduction Technique to procedures, replacing the origin estimator, sample mean, to accomplish the propose of reducing variance; and further, decrease the number of samples we need. In our research, we establish four model of combining Correlation Induction and Control Variates in Screening Procedure, Multistage Selection Procedure, Two-stage Selection Procedure, and Fully Sequential Selection Procedure, and by inference or proving that each procedure will guarantee confidence level, and we also analyze in what condition our combine procedure will be better than CV procedure. Through empirical results and a realistic illustration, we find that the probability of correct selection of all procedures will conform to confidence level guarantee, and when the problem is more complicated, then we can get more benefit form our combine procedures. In the end of our research we conclude that, for Screening Procedure, we recommend using Model 3 when the number of samples and the number of controls is close; using Model 4 when the problem is complex with setting large samples to do; in the other condition we suggest using Model 1. For Multistage Selection Procedure and Fully Sequential Selection Procedure, we recommend using Model 1 when facing a complicated problem, but otherwise using the CV procedure. Finally, for Two-stage Procedure we recommend using Model 1 in all situations.