자료유형 | 학위논문 |
---|---|
서명/저자사항 | Adaptive Computer Experiments for Metamodeling. |
개인저자 | Erickson, Collin B. |
단체저자명 | Northwestern University. Industrial Engineering and Management Sciences. |
발행사항 | [S.l.]: Northwestern University., 2019. |
발행사항 | Ann Arbor: ProQuest Dissertations & Theses, 2019. |
형태사항 | 173 p. |
기본자료 저록 | Dissertations Abstracts International 81-03B. Dissertation Abstract International |
ISBN | 9781085643016 |
학위논문주기 | Thesis (Ph.D.)--Northwestern University, 2019. |
일반주기 |
Source: Dissertations Abstracts International, Volume: 81-03, Section: B.
Advisor: Ankenman, Bruce E. |
이용제한사항 | This item must not be sold to any third party vendors. |
요약 | Computer simulation experiments are commonly used as an inexpensive alternative to real-world experiments to form a metamodel that approximates the input-output relationship of the real-world experiment. The metamodel can be useful for decision making and making predictions for inputs that have not been evaluated yet since it can be evaluated much faster than the actual simulation. The two main components of computer experiments are choosing which input points to evaluate and building a statistical model, called a metamodel, using the data that can be used to approximate the simulation output. In this dissertation, we study three problems in computer experiments.First, we investigate Gaussian process models, one of the most commonly used types of metamodel. We find that, despite implementing nearly the same model, different software implementations fit to the same data can provide very different predictions. The difference in time it takes to fit each model can also vary by orders of magnitude across the software implementations.Second, we propose a new algorithm for running sequential computer experiments when the user wants to have better prediction accuracy in regions where the simulation output varies the most. In sequential experiments, the data is gathered in batches, and data from previous batches can help inform the choice of which points to select in following batches. We assert that practitioners often have a goal of fitting the entire surface reasonably well, but want to have better prediction accuracy in regions that are more interesting to them. This goal can be achieved by changing the criterion that is used each iteration to choose which points to evaluate next.Third, we devise a new algorithm for adaptive computer experiments that allows for the construction of a metamodel using large amounts of data. Gaussian process models are infeasible for more than a couple thousand points because of computational demands. Using the sparse grid designs of Plumlee [2014], Gaussian process inference can be done on over 100,000 points. We build upon this work to allow for data to be added adaptively in order to focus simulation effort in the input dimensions that are harder to predict. |
일반주제명 | Industrial engineering. Statistics. |
언어 | 영어 |
바로가기 | ![]() |