matCercola18's picture
Add files using upload-large-folder tool
929e74d verified
Raw History Blame Contribute Delete
1.19 kB
multi-solution pendulum dataset: 2600 train plants (the original 2600), 32 starts x 160 steps x 6 label ICs per plant
S_scale: original 219.63, new 238.23
plants with >=1 certified solution in the +5% pool: 2600/2600; rows: tier(<=+2%,<=8) 2611, pool(<=+5%,<=16) 3113
stable starts per plant (of 32): mean 25.1, median 26, min 10
distinct certified solutions within +2% (<=8, rel dist>0.10): mean 1.00, median 1, p10/p90 1/1; share with >=2: 4.3%, >=4: 0.0%, ==8: 0.0%
distinct certified within +5% (<=16, rel dist>0.10): mean 1.20, median 1, p10/p90 1/2; share with >=2: 18.7%, >=4: 0.0%, ==8: 0.0%
distinct in +5% pool at rel-dist>0.05: mean 1.17, share>=2 16% (300 plants subsample)
distinct in +5% pool at rel-dist>0.1: mean 1.17, share>=2 16% (300 plants subsample)
distinct in +5% pool at rel-dist>0.25: mean 1.01, share>=2 1% (300 plants subsample)
distinct in +5% pool at rel-dist>0.5: mean 1.00, share>=2 0% (300 plants subsample)
old single label vs new best on the 6 label ICs: stable 100.0%, excess median 1.96%, p90 3.91%, share old within 2% of new best 54.2%
per-plant compute: see build logs (about 9-10 s/plant single core, 32 starts x 160 steps x 6 ICs, batched 4 plants)