GloMPO (Globally Managed Parallel Optimization): a tool for expensive, black-box optimizations, application to ReaxFF reparameterizations

Table 4 Timing tests on the Disulfide and 20D Schwefel functions showing the number of function evaluations possible within a fixed time-limit

Function	Ver.^a	Push Freq.^b	Function evaluations achieved^c	GloMPO cost (ms/eval)
Disulfide\(^{\mathrm{d}}\)			57 ± 0
Disulfide	v2.1	1	56 ± 0	93.98
Disulfide	v3.1	1	57 ± 0	18.53
Schwefel\(^{\mathrm{d}}\)			9415963 ± 032073
Schwefel	v2.1	1	60227 ± 154	4.95
		10	516510 ± 1614	0.55
Schwefel	v3.1	1	744309 ± 30812	0.37
		10	1646672 ± 5801	0.15

Timing tests are not optimizations, but an infinite loop of evaluations of the same parameter vector. Functions were tested managed and unmanaged for 300 s on a single 2.60 GHz Intel Xeon E5-2650 v2 CPU. Every configuration test repeated 5 times. GloMPO cost estimates are constructed by assuming the mean unmanaged evaluation rate to be representative of the intrinsic function evaluation rate, and further assuming that the balance of the 300s evaluation time can be attributed solely GloMPO management costs
\(^{\mathrm{a}}\) GloMPO version number
\(^{\mathrm{b}}\) Optimizer configured to send every nth evaluation to the manager
\(^{\mathrm{c}}\) Average and standard deviation over the 5 repeats
\(^{\mathrm{d}}\) Unmanaged run

ISSN: 1758-2946