Recursive Gaussian Processes for Discrepancy Modeling
In the pioneering work of Kennedy and O'Hagan, a discrepancy function modeled by a Gaussian process (GP) was introduced to account for the model error that is inherent to every mathematical model. However, the representation of the discrepancy function by a GP entails difficulties when prior information is missing. As has recently been demonstrated in the field of multifidelity modeling, recursive Gaussian Processes (rGP) have the potential to overcome such difficulties as they proved to adequately describe complex non-linear and space-dependent interrelations between different model hierarchies. In analogy, experimental measurements can be considered at the top of a model hierarchy with descending fidelity levels, where the actual simulation model is on the next lower level and a rGP plays the role of a discrepancy function. The goal of this contribution is to investigate the applicability of rGP to the modeling of the discrepancy function on the example of a suspension-type structural system and compare it to a conventional GP representation.