Fig. 4

Comparison of objective values for our multi-GP optimization (top/middle) and SNOBFIT (bottom). Both methods use the same training set of 100 randomly sampled inputs, and both return 21 points for evaluation in a subsequent round. In the second round (R2 region) of search, our method recovers 3 low RMSD points, i.e. the blue square, pink asterisk and purple diamond. These three points, displayed in the middle figures with 95% confidence intervals for each dimension, minimize acquisition functions for distinct GPs and exploration/exploitation trade-off parameters. Displayed for SNOBFIT are 7 rounds of search in which it fails to recover equally low RMSD points. With default settings, SNOBFIT returned points of three types (distinguished by color), two of which result from local searches, and the remaining from a global search