[from the authors]

• new selection of points to be evaluated by the original fitness using an uncertainty criterion: selecting the union of the $k$ nearest neighbors of every point for which the fitness should be predicted ($k$ is maximal such that the total number of selected points does not exceed $N_{max}$)
• bugfix in the surrogate selection algorithm (compared to the 2017 version)
• sligthly different parameter settings (maximum radius for selecting points to the training set = 4 times the 0.99-quantile of the length of vectors from $N(\mathbf{0},\mathbf{I})$, maximum number of points in the training set = $20 D$, probability of improvement as the surrogate selection criterion)