File size: 2,767 Bytes
0a4f337 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | """Paired uncertainty summaries over the eight independently seeded systems."""
from pathlib import Path
import csv
import json
import numpy as np
def main():
root=Path(__file__).resolve().parents[1]
rows=list(csv.DictReader((root/'results/v2_program_search.csv').open()))
seeds=sorted({int(r['seed']) for r in rows})
def values(method,generation,key):
return np.array([float(next(r[key] for r in rows if int(r['seed'])==seed and
r['method']==method and int(r['generation'])==generation)) for seed in seeds])
rng=np.random.default_rng(811827)
indices=rng.integers(0,len(seeds),size=(100000,len(seeds)))
results={}
comparisons=[('proposal_vs_frozen','mean_proposed_heldout_loss_ratio','frozen_proposer',5),
('proposal_vs_own_initial','mean_proposed_heldout_loss_ratio','corrected',1),
('proposal_vs_no_feedback','mean_proposed_heldout_loss_ratio','no_feedback',5),
('proposal_vs_gateblind','mean_proposed_heldout_loss_ratio','gate_blind',5),
('proposal_vs_direct_tournament','mean_proposed_heldout_loss_ratio','direct_tournament',5),
('best_found_vs_frozen','heldout_loss_ratio','frozen_proposer',5)]
for name,key,comparison,generation in comparisons:
ours=values('corrected',5,key);base=values(comparison,generation,key)
difference=base-ours
boot=difference[indices].mean(axis=1)
relative=(base[indices].mean(axis=1)-ours[indices].mean(axis=1))/base[indices].mean(axis=1)
results[name]={'positive_favors_corrected':True,'system_ids':seeds,
'paired_differences':difference.tolist(),'mean_difference':float(difference.mean()),
'bootstrap_95_percentile_interval':np.quantile(boot,[.025,.975]).tolist(),
'relative_reduction_of_mean':float((base.mean()-ours.mean())/base.mean()),
'bootstrap_relative_95_percentile_interval':np.quantile(relative,[.025,.975]).tolist(),
'systems_favoring_corrected':int(np.count_nonzero(difference>0)),
'systems_favoring_comparator':int(np.count_nonzero(difference<0))}
out={'independent_units':'eight seeded quadratic systems, not individual states or candidate programs',
'bootstrap_resamples':100000,'bootstrap_seed':811827,
'warning':'Descriptive paired percentile intervals for this small constructed task distribution; not evidence of general-purpose RSI or broad benchmark superiority. The pilot and repair were not a controlled single-factor comparison.',
'comparisons':results}
(root/'results/v2_paired_analysis.json').write_text(json.dumps(out,indent=2)+'\n')
print(json.dumps(results,indent=2))
if __name__=='__main__':main()
|