Combined correlation induction strategies for designed simulation experiments
This dissertation deals with variance reduction techniques (VRTs) for improving the reliability of the estimators of interest through a controlled laboratory-like simulation experiment. This research concentrates on correlation methods of VRTs which include common random numbers, antithetic variates...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | en |
| Published: |
Virginia Tech
2014
|
| Subjects: | |
| Online Access: | http://hdl.handle.net/10919/39056 http://scholar.lib.vt.edu/theses/available/etd-08062007-094405/ |
| Summary: | This dissertation deals with variance reduction techniques (VRTs) for improving the reliability of the estimators of interest through a controlled laboratory-like simulation experiment. This research concentrates on correlation methods of VRTs which include common random numbers, antithetic variates and control variates. The basic idea of these methods is to utilize the linear correlation either between the responses or between the response and control variates in order to reduce the variance of estimators of certain system parameters. Combining these methods, we develop procedures for estimating a system parameter of interest.
First, we develop three combined methods utilizing antithetic variates and control variates for improving the estimation of the mean response in a single population model. We explore how these methods may reduce the variance of the estimator of interest. A combined method (Combined Method 1) using antithetic variates for the non-control variate stochastic components and independent streams for the control variates yields better results than by applying methods of either antithetic variates or control variates individually for several selected models.
Second, we develop variance reduction techniques for improving the estimation of the model parameters in a multipopulation simulation model. We extend Combined Method 1 showing good performance in estimating the mean response of a single population model to the multipopulation context with independent simulation runs across design points. We also develop another extension of Combined Method 1 that incorporates the Schruben-Margolin method to estimate the parameters of a multipopulation model. Under certain conditions, this method is superior to the Schruben-Margolin method. Finally, we propose a new approach (Extended Schruben-Margolin Method) utilizing the control variates under the Schruben-Margolin strategy for improving the estimation in a first-order linear model. Extended Schruben-Margolin Method yields better results than the Schruben-Margolin method in estimating the model parameters of interest. === Ph. D. |
|---|