Datasets:
Hyperparameter sensitivity (phase 5)
Generated by uv run python scripts/07_make_report.py from the files in results/ on 2026-11-15. Do not edit by hand; hand-written notes go in reports/notes/.
Question. How far does the test score move when one setting changes?
Methodology
Descriptive sensitivity: the standard recipe with one setting changed (learning rate / 3, discount 0.995, 3x512 networks, weight decay 1e-5). None of these replaces the frozen recipe.
Metric: success-weighted speed (SWS) and finish rate; intervals are 95 % two-level bootstrap intervals (10,000 resamples of seeds and tracks). See reports/preregistration.md for the binding protocol.
Runs
| Run | Status | Phase | Agent steps | Wall hours | Physics | Commit | Evaluated on |
|---|---|---|---|---|---|---|---|
hp-lr-div3/seed1 |
postponed | 5-hyperparameters | — | — | v1 | — | — |
hp-gamma995/seed1 |
postponed | 5-hyperparameters | — | — | v1 | — | — |
hp-net3x512/seed1 |
postponed | 5-hyperparameters | — | — | v1 | — | — |
hp-wd1e-5/seed1 |
postponed | 5-hyperparameters | — | — | v1 | — | — |
hp-lr-div3/seed2 |
postponed | 6-seed-sensitivity | — | — | v1 | — | — |
hp-gamma995/seed2 |
postponed | 6-seed-sensitivity | — | — | v1 | — | — |
hp-net3x512/seed2 |
postponed | 6-seed-sensitivity | — | — | v1 | — | — |
hp-wd1e-5/seed2 |
postponed | 6-seed-sensitivity | — | — | v1 | — | — |
Configuration (logged in each run.json)
No run of this group has logged its configuration yet (the experiment files are in configs/experiments/).
Results
Scores
| Configuration | Role | Seeds | Test SWS [95 % CI] | Test finished | Val SWS [95 % CI] | Val finished | Train SWS [95 % CI] | Train finished | Own train SWS [95 % CI] | Own train finished | OOR SWS [95 % CI] | OOR finished | Campaign SWS [95 % CI] | Campaign finished | Unfiltered SWS [95 % CI] | Unfiltered finished |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Generalist (PPO) | Main method | 1 | 1.214 [1.177, 1.250] | 81.5 % | 1.236 [1.153, 1.315] | 82.7 % | 1.250 [1.214, 1.283] | 83.9 % | 1.250 [1.214, 1.283] | 83.9 % | 0.875 [0.769, 0.982] | 57.8 % | 1.590 [1.451, 1.783] | 100.0 % | 1.231 [1.147, 1.311] | 82.0 % |
Rows from other groups are shown for comparison. Track counts: Test (unseen generated): 971 tracks; Validation: 197 tracks; Training sample: 1,000 tracks; Own training tracks: 1, 1,000 tracks; Out-of-range test: 187 tracks; Campaign (hand-made): 5 tracks; Unfiltered test: 200 tracks.
Learnings
No hand-written notes yet (add reports/notes/hyperparameters.md).