Data Fusion With Latent Map Gaussian Processes

Abstract Multi-fidelity modeling and calibration are data fusion tasks that ubiquitously arise in engineering design. However, there is currently a lack of general techniques can jointly fuse multiple sets with varying fidelity levels while also estimating parameters. To address this gap, we introduce novel approach that, using latent-map Gaussian processes (LMGPs), converts into latent space learning problem where the relations among different sources automatically learned. This conversion endows our some attractive advantages such as increased accuracy reduced overall costs compared to existing need take combinatorial datasets. Additionally, have flexibility any number ability visualize correlations between sources. visualization allows an analyst detect model form errors or determine optimum strategy for high-fidelity emulation by fitting LMGP only sufficiently correlated We develop new kernel enables LMGPs not build probabilistic multi-fidelity surrogate but estimate parameters quite high consistency. The implementation use considerably simpler less prone numerical issues alternate methods. Through analytical examples, demonstrate benefits interpretable fusing (in particular more than two) data.