DiffusionControl / results /furuta /margin_transfer.txt
matCercola18's picture
Add files using upload-large-folder tool
20ec019 verified
Raw History Blame Contribute Delete
2.75 kB
=== margin quick-check: transferred-P certificate rejections (N=30, seed 4242) ===
bank data/furuta_nmpc_bank600.npz (550 entries); K=4; P re-solved fresh at every margin level (0.0 included) via certified_catch_set(margin=.)
reference: traj_gate_nmpc.txt reports 40 certificate-rejections / 4 rollout-failures of 60 plants for the 'nearest' arm (bank-stored P, no margin)
margin=0.0 transfer success 33.3% (Lyapunov decrease fails at a vertex:18, rollout failed:2) | self-consistency (same u*,T,K, own plant): 100.0% of 110
margin=0.1 transfer success 46.7% (Lyapunov decrease fails at a vertex:13, rollout failed:3) | self-consistency (same u*,T,K, own plant): 100.0% of 110
margin=0.3 transfer success 66.7% (Lyapunov decrease fails at a vertex:6, rollout failed:4) | self-consistency (same u*,T,K, own plant): 100.0% of 110
--- before / after (margin=0.0 -> margin=0.3), this 30-plant draw ---
transfer success 33.3% -> 66.7% (+33.4 points)
certificate rejections 18 -> 6 of 30 (the doc's 40/60 baseline is ~20/30-equivalent)
rollout failures 2 -> 4 of 30
self-consistency cost 0.0 points (100.0% -> 100.0% of 110 unique source plants touched)
--- decision ---
margin=0.3 chosen for the v2 bank (step 7): largest transfer gain of the 3 levels tried, at a
measured per-plant cost of 0.0 points (budget was <= ~2 points), on the self-consistency proxy.
CAVEAT on "per-plant success cost": self-consistency replays the ORIGINAL (u*, T) -- solved
against the unmargined region -- against the margined/shrunk P on the SAME plant. It is a lower
bound on the true cost of rebuilding the bank with margin, because a from-scratch NMPC solve
targets the smaller Omega directly and could behave differently (better OR worse) than replaying
an unchanged trajectory into a shrunk target. 100% self-consistency at all 3 levels means margin
never pushed the stored trajectory outside its own (now smaller) certified set on this sample --
it does not by itself prove the v2 NMPC bank build will keep 93% per-plant success once every
entry is re-solved from scratch against the margined objective; that is measured directly by
comparing the v2 bank's own kept-fraction / per-plant success against the v1 (unmargined) bank's,
reported in docs/furuta_generated_trajectory.md.
This 30-plant draw is its own fresh seed-4242 sample (see furuta_margin_check.py), NOT a prefix
of the standard 60-plant eval set, so its 33.3%/66.7% are not directly the same 60 plants as
traj_gate_nmpc.txt's 27%/72% rows; they are read as a relative, paired (same 30 plants, same
neighbours) comparison across margin levels, which is what the decision above rests on.