Most physical models need to be designed with high accuracy with the absence of certain features. It can be a challenge to obtain high accuracy in exchange for high computational and resource cost when dealing with complex systems. In this paper we explore methods related to Multi-Fidelity Gaussian Models to tackle these problems. We will first design a multi-fidelity Gaussian process regression model which augments high-fidelity data with low-fidelity data for an approximation of the high-fidelity model. We also make use of composite kernel functions as prior covariance for their ability to improve computation cost and for more complex learning and better approximations. We also implement our work on a Deep Gaussian Process architecture in order to take advantage of the stacked learning nature of a Deep Neural Network. For this paper, we make use of two levels of fidelity. We observed that we can approximate the high-fidelity function with more degree of accuracy using multi-fidelity method as compared to single fidelity. Moreover, we also show that this can be used to provide efficient approximation of uncertainty for the forward uncertainty quantification problem.     «

Most physical models need to be designed with high accuracy with the absence of certain features. It can be a challenge to obtain high accuracy in exchange for high computational and resource cost when dealing with complex systems. In this paper we explore methods related to Multi-Fidelity Gaussian Models to tackle these problems. We will first design a multi-fidelity Gaussian process regression model which augments high-fidelity data with low-fidelity data for an approximation of the high-fidel...     »