sync(report): policy_selection_upper_bound_20260719
Browse files- reports/0.54.2.0/policy_selection_upper_bound_20260719/README.md +35 -0
- reports/0.54.2.0/policy_selection_upper_bound_20260719/bank_selection.tsv +0 -0
- reports/0.54.2.0/policy_selection_upper_bound_20260719/preference_selection.tsv +0 -0
- reports/0.54.2.0/policy_selection_upper_bound_20260719/selection_upper_bound.png +3 -0
- reports/0.54.2.0/policy_selection_upper_bound_20260719/series_summary.tsv +11 -0
reports/0.54.2.0/policy_selection_upper_bound_20260719/README.md
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Fixed-index versus set-level selection upper bound
|
| 2 |
+
|
| 3 |
+
This report compares the policy stored at index `i` with the policy in
|
| 4 |
+
the same bank that maximizes normalized oracle utility for training
|
| 5 |
+
preference `w_i`. The latter is a retrospective oracle upper bound; it is not a deployable GPI selector.
|
| 6 |
+
|
| 7 |
+
## Contract
|
| 8 |
+
|
| 9 |
+
- Run table: `reports/brax_hv_common_vs_bestfit_all_envs_20260714/tables/combined_seed_runs.tsv`
|
| 10 |
+
- Frozen exploratory bounds: `reports/brax_hv_common_vs_bestfit_all_envs_20260714/tables/multi_metric_bounds.tsv`
|
| 11 |
+
- Normalization is shared within each environment; no bank is normalized independently.
|
| 12 |
+
- Bootstrap resamples complete seed-bank rows, preserving within-bank paired gains.
|
| 13 |
+
- Coverage: 1080 seed banks and 8856 preference-policy pairs.
|
| 14 |
+
|
| 15 |
+
## Series summary
|
| 16 |
+
|
| 17 |
+
| Environment | Series | Fixed utility | Set utility | Gain | 95% paired bootstrap CI | Own-policy win |
|
| 18 |
+
|---|---|---:|---:|---:|---:|---:|
|
| 19 |
+
| HalfCheetah | A0 common=best-fit | 0.5850 | 0.7100 | 0.1250 | [0.1150, 0.1356] | 16.8% |
|
| 20 |
+
| HalfCheetah | A1 best-fit | 0.5410 | 0.6607 | 0.1197 | [0.1100, 0.1304] | 12.3% |
|
| 21 |
+
| HalfCheetah | A1 common | 0.5277 | 0.6491 | 0.1214 | [0.1100, 0.1331] | 12.0% |
|
| 22 |
+
| Hopper | A0 best-fit | 0.2830 | 0.4693 | 0.1863 | [0.1595, 0.2130] | 10.9% |
|
| 23 |
+
| Hopper | A0 common | 0.2673 | 0.6652 | 0.3979 | [0.3826, 0.4108] | 9.3% |
|
| 24 |
+
| Hopper | A1 best-fit | 0.2712 | 0.4203 | 0.1491 | [0.1240, 0.1745] | 11.3% |
|
| 25 |
+
| Hopper | A1 common | 0.4766 | 0.7188 | 0.2422 | [0.2291, 0.2556] | 11.3% |
|
| 26 |
+
| Humanoid | A0 best-fit | 0.3986 | 0.5750 | 0.1765 | [0.1615, 0.1915] | 14.4% |
|
| 27 |
+
| Humanoid | A0 common | 0.3998 | 0.5771 | 0.1773 | [0.1631, 0.1920] | 13.9% |
|
| 28 |
+
| Humanoid | A1 common=best-fit | 0.7089 | 0.8575 | 0.1486 | [0.1389, 0.1584] | 13.4% |
|
| 29 |
+
|
| 30 |
+
## Files
|
| 31 |
+
|
| 32 |
+
- `bank_selection.tsv`: one paired fixed/set comparison per seed bank.
|
| 33 |
+
- `preference_selection.tsv`: selected policy index and utility per preference.
|
| 34 |
+
- `series_summary.tsv`: aggregate means, SDs and paired bootstrap intervals.
|
| 35 |
+
- `selection_upper_bound.png`: compact visual comparison.
|
reports/0.54.2.0/policy_selection_upper_bound_20260719/bank_selection.tsv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
reports/0.54.2.0/policy_selection_upper_bound_20260719/preference_selection.tsv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
reports/0.54.2.0/policy_selection_upper_bound_20260719/selection_upper_bound.png
ADDED
|
Git LFS Details
|
reports/0.54.2.0/policy_selection_upper_bound_20260719/series_summary.tsv
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
env_id environment series series_label architecture regularizer_role n_seed_banks policy_count fixed_utility_mean fixed_utility_std set_utility_mean set_utility_std set_gain_mean set_gain_std set_gain_bootstrap_ci_low set_gain_bootstrap_ci_high relative_gain_pct own_policy_win_rate_mean off_index_selection_rate_mean chance_own_policy_rate effective_selected_policies_mean selector_entropy_mean
|
| 2 |
+
mo-halfcheetah-v5 HalfCheetah a0_shared_rm_common_best_reg A0 shared-RM | common = best-fit regularizer A0 common=best-fit 108 7 0.5850413467160757 0.03893853754483206 0.7100498081475811 0.05927360380879802 0.12500846143150543 0.05461179022884299 0.11496530142960314 0.13564345762453112 21.367457553760357 0.16798941798941797 0.8320105820105821 0.14285714285714285 1.8459718078526948 0.5865418169938437
|
| 3 |
+
mo-halfcheetah-v5 HalfCheetah a1_independent_rm_bestfit_reg A1 independent-RM | best-fit regularizer A1 best-fit 108 7 0.5410262368097072 0.03463392417904543 0.6607270157582763 0.07127364925741056 0.11970077894856913 0.05421407339395568 0.1099840314504015 0.130353717718279 22.12476416197741 0.12301587301587301 0.876984126984127 0.14285714285714285 1.8166436784044555 0.5612080432658929
|
| 4 |
+
mo-halfcheetah-v5 HalfCheetah a1_independent_rm_common_reg A1 independent-RM | common regularizer A1 common 108 7 0.5276628585827106 0.02995029283600202 0.6490640704444625 0.06841802866768291 0.1214012118617519 0.060802973848973256 0.11003724780693744 0.13306226874927982 23.00734453583346 0.12037037037037036 0.8796296296296297 0.14285714285714285 1.7904514064841919 0.5512582556357004
|
| 5 |
+
mo-hopper-v5 Hopper a0_shared_rm_bestfit_reg A0 shared-RM | best-fit regularizer A0 best-fit 108 10 0.28302688677523125 0.04194348895699695 0.4693332928326949 0.15098973217737274 0.18630640605746365 0.14356402795291998 0.15948502252130886 0.21302123139875911 65.82639839635476 0.10925925925925926 0.8907407407407407 0.1 2.1057115038844603 0.6751959840523587
|
| 6 |
+
mo-hopper-v5 Hopper a0_shared_rm_common_reg A0 shared-RM | common regularizer A0 common 108 10 0.2672943013917985 0.04325573549884323 0.6651734444573562 0.09345708865023382 0.39787914306555766 0.07360730908508228 0.38256205322597425 0.4107679066372294 148.85433059882135 0.0925925925925926 0.9074074074074074 0.1 1.8352256299378398 0.5505008921034467
|
| 7 |
+
mo-hopper-v5 Hopper a1_independent_rm_bestfit_reg A1 independent-RM | best-fit regularizer A1 best-fit 108 10 0.2711590081634288 0.024429337895498903 0.420291248440295 0.1412071267997361 0.14913224027686614 0.13562172257130287 0.12398970623049761 0.17445997833390267 54.99807706442982 0.11296296296296297 0.8870370370370371 0.1 1.8767907062376388 0.5763117341593688
|
| 8 |
+
mo-hopper-v5 Hopper a1_independent_rm_common_reg A1 independent-RM | common regularizer A1 common 108 10 0.4766182346578902 0.05729680004240387 0.7187980360872858 0.053849318395952174 0.24217980142939552 0.06942559622721753 0.2291052529368976 0.25561998785359846 50.812114144820484 0.11296296296296297 0.8870370370370371 0.1 2.623613160106653 0.9196921123071499
|
| 9 |
+
mo-humanoid-v5 Humanoid a0_shared_rm_bestfit_reg A0 shared-RM | best-fit regularizer A0 best-fit 108 7 0.3985597060619826 0.12860948289783283 0.5750442325749019 0.13495307026561104 0.17648452651291927 0.07789661602608795 0.16151668373558165 0.191455420287224 44.280574234835726 0.14417989417989419 0.8558201058201059 0.14285714285714285 1.147786280508578 0.10618144040167028
|
| 10 |
+
mo-humanoid-v5 Humanoid a0_shared_rm_common_reg A0 shared-RM | common regularizer A0 common 108 7 0.3997987914334205 0.12741644130936816 0.5770504563188036 0.14105717663231876 0.17725166488538308 0.07743234920425987 0.16310060271262647 0.191983797803901 44.33521778539474 0.1388888888888889 0.8611111111111112 0.14285714285714285 1.1034183044747756 0.0745032753665026
|
| 11 |
+
mo-humanoid-v5 Humanoid a1_independent_rm_common_bestfit_reg A1 independent-RM | common = best-fit regularizer A1 common=best-fit 108 7 0.7089261009216262 0.04177690821926761 0.8574814944597835 0.05136996148722909 0.1485553935381573 0.05130151789757354 0.13893305859544058 0.1584119067015577 20.954989997551312 0.1335978835978836 0.8664021164021165 0.14285714285714285 1.2669676231620144 0.19136751153392373
|