MSER Exploratory Research: Implementations, Virtual Laboratory Development, and Parameterization Analysis

It is well known that the Mean Squared Error Rule (MSER) is an efficient and effective method for mitigating initialization bias in the output analysis of steady-state, discrete-event simulation. However, the application of this method in research and practice has been delayed or misunderstood even by experienced simulation modelers. To address this issue, we develop the MSER Laboratory—a permanent website that provides user-friendly sample codes, as well as information needed to apply MSER intelligently. MSER modules for three commercial software packages, and standalone MSER codes in five popular programming languages, have been written, validated, and made publically available via the Laboratory.
In addition, we use these codes to address open issues in the selection of the parameters needed to apply MSER. These issues include the selection of the MSER truncation threshold, batch size, and batching scheme (overlapping or non-overlapping batch means), in conjunction with the determination of an initial run length for simulation replications. Experiments are conducted using three test models that pose differing challenges for the successful determination of a warm-up period. We confirm that, given adequate run lengths, MSER is both effective and robust in all cases. We also illustrate various consequences of foreshortened replications for each of the three models.