@inproceedings {pub3601,
	title = {Multi-fidelity surrogate model approach to optimization},
	author = {Sander van Rijn AND Sebastian Schmitt AND Thomas B{\"a}ck AND Markus Olhofer AND Matthijs  van Leeuwen},
	year = {2018},
	abstract = {Gaussian Process regression, also known as Kriging, has established itself as a valuable surrogate model for computationally expensive optimization problems. An extension of this is \textit{co-Kriging} whereby the response surface between evaluation methods of varying fidelity is modelled, as multiple fidelities are usually available when the evaluation is based on numerical simulations. Only recently has the use of Radial Basis Functions (RBF) been introduced as an optional alternative to Kriging in this context. However, training the Kriging or RBF model can become prohibitively expensive when working with high dimensionalities or many training points. In this paper, we compare the performance of Random Forest-based \textit{co-surrogates} to the previously introduced co-Kriging and co-RBF.

Performance in a CMA-ES-based optimization setting is determined empirically using a set of bi-fidelity benchmark problems in 2, 4 and 8 dimensions. Our results show that there is a minimal overall difference between the different co-surrogate models with regards to final performace, although the training of Random Forests takes much less time compared to the Kriging and RBF methods.},
	publisher = {ACM, New York, USA},
	booktitle = {GECCO 2018}
}