File size: 5,515 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 | """Write results/exploration_report.md and results/exploration_tau_table.json: every ROM condition tried on
goal set A versus Chrono (exploration, Sim2Sim)."""
import json
import numpy as np
import rankstats as rs
import variants as V
from analyze import load_chrono, load_rom
from common import RESULTS
B = 2000
goals = json.loads((RESULTS / 'rom_A_explore/goals.json').read_text())
gids = [g['goal_id'] for g in goals]
gr = np.array([g['goal_range'] for g in goals])
pols, _, _, _ = load_rom('A', 'explore', 'nominal')
Tc, Sc = load_chrono('A', 'explore', pols, gids)
boot = rs.boot_indices(gr, B, 20261104)
sub_boot = {s: rs.boot_indices(gr[gr == s], B, 20261104 + 7 + i) for i, s in enumerate(['id', 'ood'])}
pools = {'all16': list(range(16)), 'new12': [i for i, p in enumerate(pols) if p not in ('rel1', 'rel2', 'rel3', 'rel4')]}
names = list(V.EXPLORE)
data = {n: load_rom('A', 'explore', n) for n in names}
def spec_text(n):
K, spec = V.EXPLORE[n]
if not spec:
return 'none (training ROM)'
parts = []
for k, d in spec.items():
if d[0] == 'const':
parts.append(f'{k}={d[1]:g}')
elif d[0] == 'normal':
parts.append(f'{k}~N({d[1]:g},{d[2]:g})')
elif d[0] == 'choice':
parts.append(f'{k} in {{{",".join(f"{x:g}" for x in d[1])}}}')
else:
parts.append(f'{k}~{"LogU" if d[0] == "loguniform" else "U"}({d[1]:.3g},{d[2]:.3g})')
return '; '.join(parts)
res = {}
for pname, idx in pools.items():
Rc = rs.per_goal_ranks(Tc[idx])
R = {n: rs.per_goal_ranks(data[n][2][idx]) for n in names}
out = {}
for sub, gi in {'all': np.arange(len(gids)), 'id': np.flatnonzero(gr == 'id'), 'ood': np.flatnonzero(gr == 'ood')}.items():
b = boot if sub == 'all' else sub_boot[sub]
r, taus = rs.compare_to_truth({n: R[n][:, gi] for n in names}, Rc[:, gi], b)
for n in names:
o = out.setdefault(n, {})
o[sub] = r[n]['kendall_tau_b']
if sub == 'all':
o['ci'] = r[n]['kendall_ci95']
o['d'] = r[n]['kendall_tau_b'] - r['nominal']['kendall_tau_b']
o['dci'] = rs.ci(taus[n] - taus['nominal'])
o['sr'] = float(data[n][3][idx].mean())
res[pname] = out
L = ['# Case 2 exploration on goal set A (Sim2Sim)', '',
'Generated by `report_explore.py` from `results/rom_A_explore/*.npz` and `results/chrono_A_explore/episodes.jsonl`.',
'Goal set A: 50 ID (10-20 m) + 50 OOD (20-40 m) goals. Pools: all 16 (released 1-4 + 12 new) and the 12 new policies.',
'Kendall tau-b of the ROM mean-rank order with the Chrono mean-rank order; 95 % goal-bootstrap ranges (2000 replicates,',
'stratified by goal range, paired). Delta = tau(condition) - tau(nominal ROM). These numbers were used to choose',
'the TR design that was then fixed for the held-out evaluation on goal set B. They are optimistic for the chosen',
'condition (selection among all rows); the held-out evaluation tests the choice on new policies and goals.',
"Condition 'paper_tr' is the noise-only TR (action and powertrain noise).", '',
'Parameter units: sig_* = sd of additive noise on normalized throttle/steering per 0.1 s step; steer_rate = max change',
'of normalized steering per 0.1 s step; delay and hold in 0.1 s steps; wheelbase in m; vmax in m/s; tau0 in N m;',
'omega0 in rad/s; yaw0, g_rot in rad; obs_pos_sig in m; gx_scale = goal-radius factor. K = trials per goal.', '',
'| Condition | Perturbation (per trial) | K | ROM success, 16 | tau, 16 (95 % range) | Delta, 16 (95 % range) | tau ID, 16 | tau OOD, 16 | tau, new 12 (95 % range) | Delta, new 12 (95 % range) |',
'|---|---|---:|---:|---|---|---:|---:|---|---|']
f = lambda x: f'{x:.2f}'
for n in sorted(names, key=lambda n: -res['all16'][n]['all']):
a, b = res['all16'][n], res['new12'][n]
L.append(f"| {n} | {spec_text(n)} | {V.EXPLORE[n][0]} | {a['sr']:.2f} | {f(a['all'])} ({f(a['ci'][0])} to {f(a['ci'][1])}) | "
f"{a['d']:+.2f} ({a['dci'][0]:+.2f} to {a['dci'][1]:+.2f}) | {f(a['id'])} | {f(a['ood'])} | "
f"{f(b['all'])} ({f(b['ci'][0])} to {f(b['ci'][1])}) | {b['d']:+.2f} ({b['dci'][0]:+.2f} to {b['dci'][1]:+.2f}) |")
# per-policy table
an = json.loads((RESULTS / 'analysis_A_explore.json').read_text())
ch = an['chrono']
L += ['', '## Per policy, goal set A (exploration)', '',
'Mean rank among 16 policies (1 = best). Success = share of goals (ROM TR: share of goal-trials).', '',
'| Policy | Chrono success ID | Chrono success OOD | Chrono mean rank | Nominal ROM success | Nominal ROM mean rank | Rate-limit TR (rate_rand_paper) success | Rate-limit TR mean rank | Noise-only TR (paper_tr) mean rank |',
'|---|---:|---:|---:|---:|---:|---:|---:|---:|']
for p in sorted(pols, key=lambda p: ch['mean_rank'][p]):
c = an['conditions']
L.append(f"| {p} | {ch['success_rate']['id'][p]:.2f} | {ch['success_rate']['ood'][p]:.2f} | {ch['mean_rank'][p]:.2f} | "
f"{c['nominal']['rom']['success_rate']['all'][p]:.2f} | {c['nominal']['rom']['mean_rank'][p]:.2f} | "
f"{c['rate_rand_paper']['rom']['success_rate']['all'][p]:.2f} | {c['rate_rand_paper']['rom']['mean_rank'][p]:.2f} | "
f"{c['paper_tr']['rom']['mean_rank'][p]:.2f} |")
(RESULTS / 'exploration_report.md').write_text('\n'.join(L) + '\n')
json.dump(res, open(RESULTS / 'exploration_tau_table.json', 'w'), indent=1)
print('\n'.join(L[:60]))
|