| # Planner-family transfer |
|
|
| Reproduces `tab:planner-transfer`: the same checkpoints scored under six planners, to test whether |
| VIScore depends on the search procedure it was developed with. |
|
|
| ```bash |
| python reproduce/planner/planner_tables.py # the six blocks, as printed in the paper |
| python reproduce/planner/planner_tables.py --latex # LaTeX rows |
| ``` |
|
|
| Verified against the published table: 6 blocks x 7 rows, 0 mismatches. |
|
|
| ## Why this lives in its own folder |
|
|
| Every arm must be evaluated on the **same** checkpoints for the comparison to be paired, so this |
| experiment uses a fixed 36-checkpoint set rather than the pools of `reproduce/tables.py`. All 25 of |
| its training runs also appear in those pools, and 20 of the 36 checkpoints are themselves cells of |
| the development or held-out pool; the remaining 16 are further epochs of the same runs, which widens |
| the quality range any single pool covers. It is a subset of the paper's **runs**, not of its |
| reported **checkpoints**, so these numbers are not comparable to `reproduce/tables.py`'s — the |
| pools differ, and so do the success-rate levels they average over. |
|
|
| A refresh of this experiment on the full 103-checkpoint held-out pool is in progress. It is not what |
| the paper reports, and it is not published here. |
|
|
| ## Files |
|
|
| | File | Contents | |
| |------|----------| |
| | `planner_cells.csv` | 36 checkpoints: planner-independent metrics, one sobriety per probe family, one success rate per planner | |
| | `planner_tables.py` | recomputes the table; no GPU, no checkpoints, no dataset | |
|
|
| ## The probe assignment is a rule, not a choice |
|
|
| Sobriety is the one factor that contains a search, so each planner block needs a probe. The rule was |
| declared before any outcome was seen — **same search family, matched budget fraction**: |
|
|
| | planner | probe | |
| |---------|-------| |
| | CEM, MPPI, iCEM, predictive sampling | mini-CEM (`sobriety_mcem`) | |
| | gradient, 100 starts x 30 steps | mini-AdamW (`sobriety_madam`) | |
| | single-start gradient, 1 start x 100 steps | single-start mini-AdamW (`sobriety_ss`) | |
|
|
| Selecting per planner whichever probe correlated best would manufacture the result: predictive |
| sampling's pooled correlation is +0.51 under the declared probe and +0.80 under the one that would |
| have flattered it. The frozen table carries all three sobriety columns so this is checkable rather |
| than described. |
|
|
| ## Scope |
|
|
| All six planners optimise an action sequence through the learned predictor and replan on a receding |
| horizon — that is, they are model-predictive controllers. Amortized policies (an inverse dynamics |
| model, goal-conditioned behaviour cloning, a learned actor) are a different class and are outside |
| what this table tests: they neither roll the predictor out nor search at deployment, so veracity |
| loses its causal path and sobriety has no estimand at all rather than merely a missing estimator. |
| `visscore`'s `exclude` option exists for that case. |
|
|
