File size: 3,721 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 | """Post hoc summaries of the goal-set-B held-out data (not part of the fixed primary test; no new simulation). Sim2Sim.
1. Paired delta tau between the primary TR and each other condition (in particular the calibrated deterministic ROM).
2. Pair resolution relative to Chrono: for each condition, the pairs that Chrono resolves
(win-rate 95 % range excludes 0.5) split into: resolved by the condition in the same direction, resolved in the
opposite direction, unresolved by the condition.
3. Steering chatter in nominal ROM versus Chrono success (Spearman across policies).
Writes results/posthoc_confirm.json.
"""
import json
import numpy as np
from scipy.stats import spearmanr
import fixed_design as design
import rankstats as rs
from analyze_confirm import B_REPS, B_SEED, load_chrono, load_rom
from common import RESULTS
from run_rom import POOLS
rom_dir, chrono_dir = RESULTS / 'rom_B_confirm16', RESULTS / 'chrono_B_confirm16'
goals = json.loads((rom_dir / 'goals.json').read_text())
gids = [g['goal_id'] for g in goals]
gr = np.array([g['goal_range'] for g in goals])
conf = json.loads((RESULTS / 'confirm_analysis.json').read_text())
beh = json.loads((RESULTS / 'behavior_B_confirm16.json').read_text())
boot = rs.boot_indices(gr, B_REPS, B_SEED)
rom = {n: load_rom(rom_dir, n) for n in design.CONDITIONS}
all_pols = rom['nominal'][0]
Tc, Sc, _ = load_chrono(chrono_dir, all_pols, gids)
out = dict(label='post hoc (not part of the fixed primary test), derived from the goal-set-B held-out outputs, Sim2Sim', pools={})
for pname, pool in [('primary', POOLS['confirm']), ('secondary', POOLS['confirm16'])]:
idx = [all_pols.index(p) for p in pool]
Rc = rs.per_goal_ranks(Tc[idx])
R = {n: rs.per_goal_ranks(rom[n][2][idx]) for n in rom}
res, taus = rs.compare_to_truth(R, Rc, boot)
o = dict(delta_tau_primary_tr_minus={})
for n in R:
if n == design.PRIMARY_TR:
continue
d = taus[design.PRIMARY_TR] - taus[n]
o['delta_tau_primary_tr_minus'][n] = dict(value=res[design.PRIMARY_TR]['kendall_tau_b'] - res[n]['kendall_tau_b'], ci95=rs.ci(d))
P = conf['pools'][pname]
cpairs = {(p['a'], p['b']): p for p in P['chrono']['pairs']}
o['pair_resolution_vs_chrono'] = {}
for n in R:
same = opp = unres = 0
for p in P['conditions'][n]['rom']['pairs']:
c = cpairs[(p['a'], p['b'])]
if not c['resolved']:
continue
if not p['resolved']:
unres += 1
elif (p['win_a'] - 0.5) * (c['win_a'] - 0.5) > 0:
same += 1
else:
opp += 1
o['pair_resolution_vs_chrono'][n] = dict(chrono_resolved=same + opp + unres, same_direction=same, opposite_direction=opp, unresolved=unres)
chat = [beh[p]['mean_abs_steer_change'] for p in pool]
srate = Sc[idx, :, 0]
o['chatter_vs_chrono_success'] = {
rng_: dict(zip(['spearman', 'p_value'], [float(x) for x in spearmanr(chat, srate[:, gr == rng_].mean(1) if rng_ != 'all' else srate.mean(1))]))
for rng_ in ['id', 'ood', 'all']}
out['pools'][pname] = o
(RESULTS / 'posthoc_confirm.json').write_text(json.dumps(out, indent=1) + '\n')
for pname, o in out['pools'].items():
print(f'== {pname}')
for n, v in o['delta_tau_primary_tr_minus'].items():
print(f" tau(tr_rate) - tau({n}) = {v['value']:+.3f} {np.round(v['ci95'], 3)}")
for n, v in o['pair_resolution_vs_chrono'].items():
print(f" {n:16s} of {v['chrono_resolved']} Chrono-resolved pairs: same {v['same_direction']}, opposite {v['opposite_direction']}, unresolved {v['unresolved']}")
print(' chatter vs Chrono success', o['chatter_vs_chrono_success'])
|