Simulation analytics treats stochastic simulation as data analytics for systems that do not yet exist, and extends traditional performance estimation and system optimization to uncovering underlying patterns and the key drivers and dynamics of system behavior by retaining the sample paths generated throughout simulation runs. My dissertation addresses two research areas in the field of simulation analytics: input uncertainty and virtual statistics, with a focus on the latter. For the input uncertainty study, we propose the first single-experiment method for quantifying the input-uncertainty variance -- both overall variance and the contribution to it of each input model, by using the sample paths generated from the nominal experiment. For the virtual statistics study, we describe a $k$-nearest-neighbor approach for estimating the properties including mean, variance, and derivative of virtual performance that are conditional on the occurrence of an event for a (possibly) nonstationary stochastic process post-simulation from the retained sample paths. We study both the asymptotic and small-sample behaviors for all of the virtual statistics and propose data-driven approaches for parameter tuning. The performance of the virtual statistics are evaluated via controlled studies.

Last modified

• 04/11/2018

Creator

• Yujing Lin

DOI

• https://doi.org/10.21985/N2Q11R

Subject

• Industrial Engineering and Management Sciences

Keyword

• Industrial engineering

Date created

• 2017-01-01

Resource type

• Dissertation

Rights statement

• In Copyright