Design and analysis of experiments in large scale computer simulations

拟议的研究涵盖两个领域。一是计算机实验的设计与分析。计算机模型被广泛应用于几乎所有的科学和工程领域来模拟物理过程。由于运行代码的计算负担，通常需要构建代码的统计近似值。本提案中的项目涉及审查与建立一个良好的代码近似值有关的各种问题。例如，生态系统健康和土地利用模型涉及大量投入；在这种情况下，有必要了解这些投入如何影响代码的输出，特别是如何校准模型以与研究中的特定地理区域相匹配。该模型可用于研究气候变化对生态系统的影响，也可用于土地利用和作物决策。当统计仿真器不能很好地近似计算机模型的输出时，还考虑对更多计算机模型运行的顺序收集。此外，还考虑了为物理过程开发有效的设计。特别是，我们考虑使用顺序技术来增加实验的运行次数，以改进对潜在活性因素的估计。主要关注用于工业产品改进的析因实验设计的有效策略。

Two areas are covered in the proposed research. The first is the design and analysis of computer experiments. Computer models are widely used in almost every area of science and engineering to model physical processes. Due to the computational burden of running the code, it is often necessary to build a statistical approximation of the code. The projects in this proposal involve examining various issues relating to building a good approximation of the code. For example, models for ecosystem health and land use involve a large number of inputs; in such cases it is necessary to understand how these inputs impact the output of the code and in particular how one can calibrate the model to match the particular geographic area under study. The calibrated model can be used to investigate the e?ects of climate change on ecosystems and can also be used to make decisions regarding land use and crops. When the output from the computer model is not well approximated by the statistical emulator, sequential collection of more computer model runs is also considered. In addition, considerations are given to developing efficient designs for physical processes. In particular we consider sequential techniques to add more runs to an experiment to improve estimates of potentially active factors. The main focus is on efficient strategies for the design of factorial experiments used for industrial product improvement.