My research focus is on sequential design of experiments and modeling. I am currently working on designing a method that will balance searching for features of the response and exploring the entire space. Output from earlier points are used to keep an updated model of the surface which helps guide the search. We incorporate ideas from designs that have desirable properties, minimum energy, and Gaussian process modeling.
An example of what I’m working on can be seen on this Shiny app.
Publications
Comparison of Gaussian process modeling software
|
C. B. Erickson, B. E. Ankenman, S. M. Sanchez, Comparison of Gaussian process modeling software, In European Journal of Operational Research, 2017, ISSN 0377-2217, https://doi.org/10.1016/j.ejor.2017.10.002.
[pdf] [arXiv]
Abstract
Gaussian process fitting, or kriging, is often used to create a model from a set of data. Many available software packages do this, but we show that very different results can be obtained from different packages even when using the same data and model. We describe the parameterization, features, and optimization used by eight different fitting packages that run on four different platforms. We then compare these eight packages using various data functions and data sets, revealing that there are stark differences between the packages. In addition to comparing the prediction accuracy, the predictive variance—which is important for evaluating precision of predictions and is often used in stopping criteria—is also evaluated.
|
Gradient Based Criteria for Sequential Design
|
C. B. Erickson, B. E. Ankenman, M. Plumlee, S. M. Sanchez, Gradient Based Criteria for Sequential Design, Proceedings of the 2018 Winter Simulation Conference,
Abstract
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. While
a user may want to understand the entire response surface, they may also want to focus on interesting
regions of the design space, such as where the gradient is large. In this paper we present an algorithm that
adaptively runs a simulation experiment that focuses on finding areas of the response surface with a large
gradient while also gathering an understanding of the entire surface. We consider the scenario where small
batches of points can be run simultaneously, such as with multi-core processors.
|