Gaussian Process Regression for Bayesian Fusion of Multi-Fidelity Information Sources

2018, American Institute of Aeronautics and Astronautics Inc, AIAA. All rights reserved. Predictions and design decisions for complex systems often can be made or informed by a variety of information sources. These information sources describe the quantity of interest with different levels of fidelity. We propose a Bayesian approach in which prior beliefs about the information sources are represented in terms of Gaussian processes and we utilize these sources to generate a fused model with superior predictive capability than any of the constituent models. For this, we implement a multifidelity co-Kriging model aimed at constructing an accurate estimate of the quantity of interest by exploiting data from all models. The key feature of our proposed approach is the relaxation of the assumption of hierarchical relationships among information sources. Instead, we consider an autore-gressive model where each information source is related to the highest fidelity information source. The approach is demonstrated on a one-dimensional example test problem and an aerodynamic design problem.

Gaussian Process Regression for Bayesian Fusion of Multi-Fidelity Information Sources Conference Paper