File size: 5,949 Bytes
0c3dd1d | 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 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 | """Collect the Case 2 numbers reported in the paper from the analysis outputs (no simulation). Sim2Sim.
Reads results/exploration_tau_table.json, results/confirm_analysis.json, results/posthoc_confirm.json,
results/behavior_B_confirm16.json and the Chrono records. Writes results/paper_numbers.json and prints a summary.
Usage: python paper_numbers.py
"""
import json
import fixed_design as design
import variants as V
from common import RESULTS
from run_rom import POOLS
R = lambda name: json.loads((RESULTS / name).read_text())
expl, conf, ph, beh = (R('exploration_tau_table.json'), R('confirm_analysis.json'), R('posthoc_confirm.json'),
R('behavior_B_confirm16.json'))
r2 = lambda x: round(float(x), 2)
out = {}
# Exploration, goal set A (16 policies: 12 design policies + 4 released checkpoints)
e = expl['all16']
out['exploration'] = dict(
test_conditions_excluding_nominal=len([n for n in V.EXPLORE if n != 'nominal']),
chrono_episodes=sum(1 for _ in open(RESULTS / 'chrono_A_explore/episodes.jsonl')),
tau_b={n: r2(e[n]['all']) for n in ['nominal', 'paper_tr', 'rate_rand_paper', 'rate_rand', 'delay_rand03']})
# Held-out evaluation, goal set B
P, S = conf['pools']['primary'], conf['pools']['secondary']
held = P['policies']
nom, tr = P['conditions']['nominal']['rom'], P['conditions'][design.PRIMARY_TR]['rom']
csr = P['chrono']['success_rate']
out['heldout_success'] = dict(
nominal_rom_min_success_id_ood=[min(nom['success_rate'][g][p] for p in held) for g in ('id', 'ood')],
chrono_success_id_range=[r2(min(csr['id'].values())), r2(max(csr['id'].values()))],
chrono_success_ood_range=[r2(min(csr['ood'].values())), r2(max(csr['ood'].values()))],
chrono_success_all_min=r2(min(csr['all'].values())))
pr = ph['pools']['primary']['pair_resolution_vs_chrono']
table = {}
for n in design.CONDITIONS:
v = P['conditions'][n]['vs_chrono']
table[n] = dict(tau_b=r2(v['all']['kendall_tau_b']), ci95=[r2(x) for x in v['all']['kendall_ci95']],
tau_id=r2(v['id']['kendall_tau_b']), tau_ood=r2(v['ood']['kendall_tau_b']),
same=pr[n]['same_direction'], reversed=pr[n]['opposite_direction'], unresolved=pr[n]['unresolved'])
out['table_case2_confirm'] = table
pt = conf['primary_test']
out['primary_test'] = dict(delta_tau=r2(pt['delta_tau']), ci95=[r2(x) for x in pt['ci95']], met=pt['supported'],
chrono_resolved_pairs=f"{P['chrono']['resolved_pairs']}/{len(P['chrono']['pairs'])}")
# Steering behavior of the held-out policies (nominal ROM) and its relation to Chrono success
chat = {p: beh[p]['mean_abs_steer_change'] for p in held}
bang = sorted(p for p in held if chat[p] > 0.1)
smooth = sorted(p for p in held if chat[p] <= 0.1)
two = sorted(held, key=chat.get)[:2]
rank = lambda order, p: order.index(p) + 1
out['steering'] = dict(
mean_abs_steer_change={p: round(chat[p], 3) for p in sorted(held, key=chat.get)},
bang_bang_count=len(bang), bang_bang_range=[r2(min(chat[p] for p in bang)), r2(max(chat[p] for p in bang))],
smooth_range=[r2(min(chat[p] for p in smooth)), r2(max(chat[p] for p in smooth))],
spearman_chatter_vs_chrono_ood_success=dict((k, round(v, 3)) for k, v in
ph['pools']['primary']['chatter_vs_chrono_success']['ood'].items()),
two_smoothest={p: dict(chrono_rank=rank(P['chrono']['order'], p), nominal_rank=rank(nom['order'], p),
tr_rank=rank(tr['order'], p)) for p in two})
d = ph['pools']['primary']['delta_tau_primary_tr_minus']['det_calibrated']
out['calibrated_rom'] = dict(tau_b=table['det_calibrated']['tau_b'], tr_minus_calibrated=r2(d['value']),
ci95=[r2(x) for x in d['ci95']])
rat_p = P['conditions']['nominal']['chrono_over_rom_time_ratio']
rat_s = S['conditions']['nominal']['chrono_over_rom_time_ratio']
out['time_ratio_chrono_over_nominal_rom'] = dict(heldout_range=[r2(min(rat_p.values())), r2(max(rat_p.values()))],
checkpoints={p: r2(rat_s[p]) for p in POOLS['released']})
# Four released checkpoints (secondary pool, 16 policies)
rel = POOLS['released']
ssr = S['chrono']['success_rate']
sub_order = lambda order: [p for p in order if p in rel]
pairs = {}
for src, pl in [('chrono', S['chrono']['pairs']), ('nominal', S['conditions']['nominal']['rom']['pairs']),
(design.PRIMARY_TR, S['conditions'][design.PRIMARY_TR]['rom']['pairs'])]:
pairs[src] = {f"{p['a']}-{p['b']}": (0 if not p['resolved'] else (1 if p['win_a'] > 0.5 else -1)) for p in pl
if p['a'] in rel and p['b'] in rel}
cmp = lambda m: dict(same=[k for k, s in pairs[m].items() if s and s == pairs['chrono'][k]],
reversed=[k for k, s in pairs[m].items() if s and s == -pairs['chrono'][k]],
unresolved=[k for k, s in pairs[m].items() if not s])
out['checkpoints'] = dict(
chrono_success_id_pct={p: round(100 * ssr['id'][p]) for p in rel},
chrono_success_ood_pct={p: round(100 * ssr['ood'][p]) for p in rel},
order_chrono=sub_order(S['chrono']['order']),
order_nominal=sub_order(S['conditions']['nominal']['rom']['order']),
order_tr=sub_order(S['conditions'][design.PRIMARY_TR]['rom']['order']),
chrono_resolved_pairs=sum(1 for s in pairs['chrono'].values() if s),
nominal_vs_chrono=cmp('nominal'), tr_vs_chrono=cmp(design.PRIMARY_TR))
sv = {n: S['conditions'][n]['vs_chrono']['all'] for n in ['nominal', design.PRIMARY_TR, 'tr_rate_half', 'tr_rate_double']}
out['pool16'] = dict(tau_b={n: r2(v['kendall_tau_b']) for n, v in sv.items()},
delta_tr_minus_nominal=r2(sv[design.PRIMARY_TR]['delta_tau_vs_nominal']),
ci95=[r2(x) for x in sv[design.PRIMARY_TR]['delta_tau_ci95']])
(RESULTS / 'paper_numbers.json').write_text(json.dumps(out, indent=1) + '\n')
print(json.dumps(out, indent=1))
|