Analysis of variance reduction techniques in various systems

Abstract

In this thesis, we consider four different Variance Reduction Techniques (VRTs): Antithetic Variates (AV), Latin Hypercube Sampling (LHS), Control Variates (CV), and Poststratified Sampling (PS). These methods individually or in combination are applied to the steady state simulation of three well-studied systems. These systems are M/M/1 Queuing System, a Serial Line Production System, and an (s,S) Inventory Policy. Our results indicate that there is no guarantee of a reduction in variance or an improvement in precision in estimates. The performance of VRTs totally depends on the system characteristics. Nevertheless, CV performs better than PS, AV and LHS on the average. Therefore, instead of altering the input part of the simulation, extracting more information by CV should be more effective. However, if any extra information about the system is not available, AV or LHS can be favored since they do not require additional knowledge about the system. Furthermore, since the analysis of output data through CV or PS requires a negligible time compared to the simulation run time, applying CV and PS at all possible cases and then selecting the best one can be the best strategy in the variance reduction. The use of the combination of methods provides more improvement on the average.