Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Optimizer-boundary decision experiment

Recorded campaign: 20 independent root seeds.

Corrected Nelder–Mead experiment

Scores below are medians over independent optimizer seeds and use held-out validation for both ReBop variants. Lower is better. de-head is the exact DE prefix supplied to the hybrid; it separates improvement by the tail from differences in DE random streams.

ProtocolProblemDEDE headDE→NMNM serialNM multistart
parallel-r60-w16cfd-ventilation1.524981.527271.527271.668761.53134
parallel-r60-w16optical-lens86.2012132.923131.3051.751e+05239.045
parallel-r60-w16rebop-crn3.040613.18783.18784.475263.21081
parallel-r60-w16rebop-resampled3.089973.063483.008044.608393.32624
serial-r160-w1cfd-ventilation1.552871.570371.569931.63327
serial-r160-w1optical-lens455.142557.987540.134792.397
serial-r160-w1rebop-crn3.21633.323633.401354.10756
serial-r160-w1rebop-resampled3.453143.557643.320324.13449

Paired DE→NM evidence

W/L/T compares held-out DE→NM against full-budget DE for the same root seed. The effect is the median log score ratio; negative favors DE→NM. The interval is a deterministic paired bootstrap over seeds.

ProtocolProblemW/L/Tsign pmedian log ratio95% bootstrap CItail vs DE head W/L/T
parallel-r60-w16cfd-ventilation2/18/00.00046.986e-05[2.038e-05, 3.609e-04]7/0/13
parallel-r60-w16optical-lens0/20/00.00000.425033[0.233049, 0.497756]18/0/2
parallel-r60-w16rebop-crn3/10/70.09230.00612956[0, 0.0642285]2/0/18
parallel-r60-w16rebop-resampled10/10/01.00000.017258[-0.0472225, 0.0963837]10/9/1
serial-r160-w1cfd-ventilation8/11/10.64761.023e-04[-5.485e-04, 0.0025482]18/0/2
serial-r160-w1optical-lens9/11/00.82380.017737[-0.0266009, 0.0951969]20/0/0
serial-r160-w1rebop-crn6/7/71.00000[-0.0081631, 0.0044426]5/4/11
serial-r160-w1rebop-resampled11/8/10.6476-0.00418485[-0.0388108, 0.0494711]12/6/2

Resource accounting is exact. With one worker, a 16-member DE generation costs 16 rounds. With 16 workers it costs one round. A serial simplex costs one round per objective call; multistart runs one simplex per worker. Every NM tail receives at least 2 × (dimension + 1) calls, so every reported simplex is initialized and has a genuine descent budget.

Equal-wall-time Bayesian experiment

The objective landscape is evaluated normally, while each best-so-far trace records optimizer overhead with simulator time subtracted. For an assumed per-evaluation latency c, call i completes at overhead(i) + i·c. DE and EGO are then compared at the same deadline nominal calls × c; BO cannot spend its modelling time twice.

Problemnominal callslatencyDE score (calls)BO score (calls)BO vs DE W/L/Tsign p
cfd-ventilation251 ms1.65432 (24)1.65411 (2)9/11/00.8238
cfd-ventilation2510 ms1.65432 (24)1.63271 (19)17/3/00.0026
cfd-ventilation25100 ms1.65432 (24)1.63103 (23)17/3/00.0026
cfd-ventilation25500 ms1.65432 (24)1.63103 (24)18/2/00.0004
cfd-ventilation251000 ms1.65432 (24)1.63103 (24)18/2/00.0004
cfd-ventilation601 ms1.63566 (59)1.63428 (17)13/7/00.2632
cfd-ventilation6010 ms1.63566 (59)1.6195 (36)15/5/00.0414
cfd-ventilation60100 ms1.63566 (59)1.60179 (55)17/3/00.0026
cfd-ventilation60500 ms1.63566 (59)1.60179 (58)17/3/00.0026
cfd-ventilation601000 ms1.63566 (59)1.60179 (59)17/3/00.0026
cfd-ventilation1501 ms1.5562 (149)1.62481 (26)6/14/00.1153
cfd-ventilation15010 ms1.5562 (149)1.60179 (74)7/13/00.2632
cfd-ventilation150100 ms1.5562 (149)1.59458 (130)7/13/00.2632
cfd-ventilation150500 ms1.5562 (149)1.59458 (145)7/13/00.2632
cfd-ventilation1501000 ms1.5562 (149)1.59458 (147)7/13/00.2632
optical-lens251 ms2.568e+05 (24)n/a (0)0/0/201.0000
optical-lens2510 ms2.568e+05 (24)1.290e+05 (18)15/5/00.0414
optical-lens25100 ms2.568e+05 (24)1.186e+05 (23)15/5/00.0414
optical-lens25500 ms2.568e+05 (24)1.126e+05 (24)15/5/00.0414
optical-lens251000 ms2.568e+05 (24)1.126e+05 (24)15/5/00.0414
optical-lens601 ms21511.8 (59)1.368e+05 (17)6/14/00.1153
optical-lens6010 ms21511.8 (59)67300.3 (34)7/13/00.2632
optical-lens60100 ms21511.8 (59)55604.3 (54)7/13/00.2632
optical-lens60500 ms21511.8 (59)55170.4 (58)7/13/00.2632
optical-lens601000 ms21511.8 (59)55170.4 (59)7/13/00.2632
optical-lens1501 ms455.142 (149)1.126e+05 (24)0/20/00.0000
optical-lens15010 ms455.142 (149)51802.8 (69)0/20/00.0000
optical-lens150100 ms455.142 (149)24423.2 (126)2/18/00.0004
optical-lens150500 ms455.142 (149)24423.2 (143)2/18/00.0004
optical-lens1501000 ms455.142 (149)24423.2 (146)2/18/00.0004

Measured optimizer overhead at the end of the trace

ProblemDEEGO
cfd-ventilation0.0001 s2.9605 s
optical-lens0.0001 s4.0208 s

Raw per-seed and per-evaluation artifacts are refiner-raw.tsv and bo-trace.tsv. They, rather than this rendered table, are authoritative.