Geometric Generalisation of Surrogate Model-Based Optimisation to Combinatorial and Program Spaces

Table 3

Correlation between predicted and real fitness on a test set of randomly sampled solutions after the SMs have been trained on 50 randomly sampled data-points. The columns contain counts of significantly positive (larger than 0.15) and significantly negative correlations (less than −0.15), together with average correlation coefficients.

Instance	Hamming model		Swap model
	Pos. (Ave.)	Neg.	Pos. (Ave)	Neg.
kra30a	50 (0.26)	0	31 (0.20)	0
kra32	46 (0.22)	0	34 (0.20)	0
lipa30a	0 (N/A)	0	0 (N/A)	0
nug30	32 (0.18)	0	21 (0.19)	0
ste36a	50 (0.23)	0	27 (0.18)	0
tho30	12 (0.18)	0	7 (0.16)	0
unih30	50 (0.77)	0	50 (0.43)	0
unis30	47 (0.24)	0	39 (0.19)	0

∗Each test was repeated 50 times.