haidang2405's picture
Mirror audited SenseShift-RL v1.1 research release
1a81987 verified
|
Raw
History Blame Contribute Delete
6.9 kB

SenseShift-RL v1.1 Confirmatory Results

Experiment audit: PASS

Experiment-base release SHA-256: 79fcf1d7700af2c647dc2cfb7920a9349985a0931c1132484d3c478c867150cb
Method-lock SHA-256: 7eebff5c140d33e516d4499f6d1678d7a108662d9306570d88310e25d797bc37

Locked claim decision

The pre-locked nested paired intervals are positive on Future time and Future time + unseen geography; the two test tracks are not pooled. The corresponding Holm-adjusted exact sign-flip tests do not cross 0.05. In the post-hoc crossed-seed sensitivity, a wholly positive interval remains only on Future time + unseen geography; evidence robust to both resampling schemes is limited to that track.

The selected learned family was chosen using calibration only. Test policy parameters and normalization were frozen, and scalar evaluation rewards were not returned as policy feedback.

Calibration-only selection

Policy Locked configuration Equal-track calibration return/day
DQN dqn_b -0.1374
PPO ppo_b -0.1261
Recurrent PPO rppo_b -0.1299

Calibration selected PPO.

Default-cost test results

World-macro mean return/day [10,000-replicate pre-locked nested hierarchical bootstrap 95% interval]. Higher is better. Crossed-seed sensitivity is reported for paired claims below, not for these absolute policy intervals.

Policy Future time Future time + unseen geography
No sensing -0.2510 [-0.2750, -0.2271] -0.2784 [-0.3250, -0.2370]
Random budget -0.1495 [-0.1601, -0.1403] -0.1522 [-0.1698, -0.1361]
Static spatial -0.1416 [-0.1519, -0.1326] -0.1614 [-0.1821, -0.1412]
Gaussian information -0.1316 [-0.1459, -0.1175] -0.1276 [-0.1465, -0.1087]
Tabular Q -0.1529 [-0.1735, -0.1335] -0.1572 [-0.1887, -0.1265]
DQN -0.1362 [-0.1497, -0.1233] -0.1436 [-0.1653, -0.1219]
PPO -0.1288 [-0.1448, -0.1145] -0.1408 [-0.1638, -0.1180]
Recurrent PPO -0.1310 [-0.1473, -0.1159] -0.1406 [-0.1645, -0.1174]
Clairvoyant oracle -0.0674 [-0.0723, -0.0633] -0.0731 [-0.0840, -0.0628]

Robust aggregates and lower-tail worlds

Mean, median, and IQM are across world-seed units; world P10 is the 10th percentile after averaging seeds within each world.

Track Policy Mean Median IQM World P10
Future time No sensing -0.2510 -0.2585 -0.2498 -0.3241
Future time Random budget -0.1495 -0.1392 -0.1415 -0.1832
Future time Static spatial -0.1416 -0.1328 -0.1353 -0.1716
Future time Gaussian information -0.1316 -0.1292 -0.1321 -0.1717
Future time Tabular Q -0.1529 -0.1333 -0.1442 -0.2199
Future time DQN -0.1362 -0.1249 -0.1297 -0.1802
Future time PPO -0.1288 -0.1088 -0.1147 -0.1639
Future time Recurrent PPO -0.1310 -0.1130 -0.1177 -0.1814
Future time Clairvoyant oracle -0.0674 -0.0630 -0.0635 -0.0861
Future time + unseen geography No sensing -0.2784 -0.2508 -0.2678 -0.3543
Future time + unseen geography Random budget -0.1522 -0.1464 -0.1480 -0.1793
Future time + unseen geography Static spatial -0.1614 -0.1587 -0.1605 -0.1931
Future time + unseen geography Gaussian information -0.1276 -0.1330 -0.1262 -0.1533
Future time + unseen geography Tabular Q -0.1572 -0.1601 -0.1530 -0.2002
Future time + unseen geography DQN -0.1436 -0.1345 -0.1396 -0.1754
Future time + unseen geography PPO -0.1408 -0.1338 -0.1338 -0.1685
Future time + unseen geography Recurrent PPO -0.1406 -0.1329 -0.1318 -0.1702
Future time + unseen geography Clairvoyant oracle -0.0731 -0.0721 -0.0724 -0.0890

The clairvoyant oracle reads hidden current truth and is an unattainable upper bound, not a deployable method.

Locked paired comparisons

Delta is left minus right return/day; positive favors the left. Holm correction is within each track.

Track Left Right Delta [locked CI] Crossed-seed CI P(improvement) Holm p
Future time Recurrent PPO PPO -0.0021 [-0.0113, 0.0079] [-0.0121, 0.0088] 0.309 0.5941
Future time Recurrent PPO DQN 0.0052 [-0.0018, 0.0114] [-0.0039, 0.0122] 0.931 0.1459
Future time PPO Static spatial 0.0128 [0.0008, 0.0234] [-0.0003, 0.0238] 0.980 0.1175
Future time Gaussian information Static spatial 0.0100 [0.0013, 0.0192] [0.0010, 0.0190] 0.988 0.1459
Future time + unseen geography Recurrent PPO PPO 0.0002 [-0.0139, 0.0116] [-0.0205, 0.0174] 0.558 1
Future time + unseen geography Recurrent PPO DQN 0.0029 [-0.0193, 0.0189] [-0.0259, 0.0213] 0.659 1
Future time + unseen geography PPO Static spatial 0.0206 [0.0090, 0.0355] [0.0061, 0.0414] 1.000 0.125
Future time + unseen geography Gaussian information Static spatial 0.0338 [0.0279, 0.0408] [0.0279, 0.0408] 1.000 0.125

Value/age-history ablation

  • Future time: recurrent PPO minus values/ages-masked recurrent PPO = -0.0095 [-0.0183, -0.0029].
  • Future time + unseen geography: recurrent PPO minus values/ages-masked recurrent PPO = 0.0039 [-0.0117, 0.0206].

Scientifically important outcomes

  • Gaussian information had the strongest observed deployable mean on future time + unseen geography (-0.1276 versus PPO -0.1408); deep RL was not the strongest observed deployable approach. This contrast is descriptive and was not in the locked comparison family.
  • Recurrent PPO did not demonstrate an advantage over PPO; both locked paired intervals include zero.
  • Masking values/ages improved recurrent-PPO performance on future time, while the direction on future time + unseen geography was inconclusive.
  • PPO's descriptive future-unseen minus future-known world gap was -0.0120 return/day. It is not a paired or pure-geography effect.

Interpretation boundaries

  • The sensing-cost sweep re-scores frozen neural and fixed-spatial decisions. Gaussian information instead reoptimizes its analytic subset at each candidate cost.
  • The unseen-world track also occurs in 2023-2024; it measures joint future-time plus geographic shift, not a pure geographic effect.
  • Six held-out geographic worlds limit inferential resolution; per-world effects are released.
  • The inverse-distance reward is an authored proxy, not operational monitoring utility.
  • Historical replay cannot model action-dependent sensor failures or pollution interventions.
  • The measurements are real EPA records with natural missingness; the benchmark is retrospective and EPA records can be revised.

See SENSESHIFT_RL_WORKSHOP_PAPER.pdf, analysis.json, paired_world_effects.parquet, and experiment_audit.json for the full method and evidence.