MULTIVARIATE ANALYSIS OF EXTRAPOLATING TIME-INVARIANT DATA WITH UNCERTAINTY

摘要

Data analysis deciphers phenomena and system behaviors within a large number of experimental realizations. Transforming these massive quantities of raw data into knowledge about the data is made possible thanks to continuously improved computing techniques. In science and engineering,there is particular interest concerning surrogate models for system behavior prediction and data extrapolation. These models tend toward underfitting or overfitting when confronted with a complex dataset or a dataset embedded with uncertainty. In this paper,we suggest an approach to treat experimental data under uncertainty prior to creation of any surrogate model. We especially focus on extrapolation as an attempt to estimate the true underlying phenomena. Our approach quantifies the uncertainty quantity through eigenvalues, copies the behavior of the data through its covariance matrix, and reproduces an almost identical dataset whose particularity perfectly correlates inputs and output. This new dataset is then used as the basis for the creation of a surrogate model. Our approach can be used to show consistency in patterns of a dataset where there are data produced under uncertainty. An approach to perform an extrapolation of data with uncertainty prior to construction of a surrogate model allows for improved predictions in that it reveals behavior of the dataset overall, while preserving a method to consider the behavior of each data point.