Simulation-based Stochastic Optimization on Discrete Domains: Integrating Optimal Computing and Response Surfaces

Simulation can be a very powerful tool to help decision making in many applications but exploring multiple courses of actions can be time consuming. Numerous ranking & selection (R&S) procedures have been developed to enhance the simulation efficiency of finding the best design. This dissertation explores the potential of further enhancing R&S efficiency by incorporating simulation information from across the domain into a regression metamodel. Under some common conditions in most regression-based approaches, our new method provides approximately optimal rules that determine the design locations to conduct simulation runs and the number of samples allocated to each design location for problems with only one partition. In addition to utilizing concepts from the design of experiments (DOE) literature, it introduces the probability of correct selection (PCS) optimality criterion that underpins our new R&S method to the DOE literature. This dissertation then extends the method by incorporating simulation information from across a partitioned domain into a regression based metamodel. Our new method provides approximately optimal rules for between and within partitions that determine the number of samples allocated to each design location. Numerical experiments demonstrate that our new approaches for one partition domains and for multiple partition domains can dramatically enhance efficiency over existing efficient R&S methods.

Author

• Brantley, Mark W

• Stochastic Optimization
• Optimal Computing
• Response Surfaces
• Design of Experiments
• Discrete Event Stimulation
• Optimal Simulation Design

Identifier

hdl:1920/6347

OAI identifier oai:identifier

oai:MARS:1920/6347