This project provides parameterized implementations of various multi-fidelity Gaussian Process Regression algorithms:
- Nonlinear autoregressive multi-fidelity Gaussian Processes [1]
- Gaussian Processes with Data Fusion and Delays [2]
Underfitted models can be efficiently improved using an entropy reduction method called Adaptation. This repo also provides a Polynomial Chaos Expansion implementation, which can be performed on the mean prediction functions of MFGPs. Linking PCE and multi-fidelity models leads to equal precisions as direct PCE but needs much less high-fidelity model evaluations. This saves a significant amount of computation effort.
This library also supports applications:
- Multifidelity forward uncertainity quantification
- Multifidelity No-Turn Sampling
- Multifildelity Bayesian Optimization
[1] Perdikaris, Paris, et al. "Nonlinear information fusion algorithms for data-efficient multi-fidelity modelling." Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 473.2198 (2017): 20160751.
[2] S. Lee, F. Dietrich, G. E. Karniadakis, and I. G. Kevrekidis. “Linking Gaussianprocess regression with data-driven manifold embeddings for nonlinear datafusion.” In:Interface focus9.3 (2019), p. 20180083.