The paper, posted on arXiv, examines expected improvement methods built on stochastic kriging surrogates, which are widely used for continuous simulation optimization. A central complication is that noise is heteroscedastic and unknown, making the standard formulation less straightforward.
The authors argue that three basic design choices become especially consequential under these conditions, beginning with how the incumbent is defined. The abstract does not detail the remaining choices, but it signals that a principled framework is needed to guide them.
Because the source is a single paper, there are no conflicting findings to compare. The main takeaway is that careful, deliberate design of expected improvement can matter as much as the choice of surrogate model itself.