Essays in Simulation Modeling and Optimization

仿真建模及優化的研究

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date11 Jun 2018

Abstract

Simulation modeling and optimization techniques are widely used in the field of operations research and management science for performance evaluation (prediction) and decision making. This thesis investigates two issues in simulation modeling and optimization, one mainly aims to improve the metamodel-based performance prediction, and the other facilitates the personalized decision making.

Metamodels are constructed to approximate computationally expensive simulation models to support fast performance evaluation. As a popular metamodeling technique, stochastic kriging typically treats the simulation model as a black box in practice and often fails to capture the highly nonlinear response surfaces that frequently arise from queueing simulations. In the first essay of this thesis, we propose a simple but effective approach to improve the performance of stochastic kriging for queueing simulation by incorporating stylized queueing models, which contain useful information about the shape of the response surfaces. We provide simple methods for constructing highly stylized models for a large class of queueing networks and several statistical tools to measure the usefulness and effectiveness of a stylized model. It is shown that even a relatively crude stylized model can improve the prediction accuracy of stochastic kriging substantially.

As a special type of simulation optimization, ranking and selection (R&S) is concerned with selecting the best with the largest or smallest mean performance among finite alternatives via running simulation experiments. With exponential growth of customer-specific data nowadays, leveraging personalized information as covariates enables the decisions to be tailored at individual level. However, the conventional R\&S that considers the mean performances of all alternatives as constants is inadequate for handling covariates. In the second essay of this thesis, we introduce the ranking and selection with covariates (R&S-C) setting where the mean performances of all alternatives are functions of the covariates and,therefore, the identity of the best alternative is a function of the covariates as well. Under the linearity assumption, we develop two-stage selection procedures to produce selection policies, which are proved to be statistically valid in terms of unconditional probability of correct selection. The advantage and potential usefulness of R&S-C are demonstrated. Other issues such as least favorable configuration and experimental design are also investigated.