|
Abstract:
|
We introduce a method to construct an emulator/surrogate model for a computationally intensive spatiotemporal simulator. At its core, the method uses the higher-order singular value decomposition (HOSVD) as an extension of singular value decomposition (SVD) approaches to emulation. An advantage of the method is that it naturally allows for the use of a combination of different machine learning models, such as neural networks, random forests, and Gaussian process regression. Another advantage is that it does not require computer simulator output to be on a grid of spatiotemporal coordinates, in contrast to many other surrogate modeling approaches. Examples from glaciology and collective animal movement are used to demonstrate the efficacy of the method; the second example illustrates the method’s ability to emulate the positions of multiple agents through space and time. We additionally demonstrate the ability to perform statistical inference of important parameters governing collective movement by embedding the tensor-based emulator into a Bayesian model. The corresponding R package allows users to implement the tensor-based emulator in the model or application of their choice.
|
