ec3118004d
geo_push.hpp and replay.hpp cited memllib @ 0a541cc. upstream/main pins
e291192, where the same code had been deliberately redesigned — and because
InterfaceRL was not in the tree (fixed one commit ago), we carried the
superseded version for months. Three changes, all upstream's:
kGeometricPushScale 0.5 -> 1.0 (InterfaceRL.hpp:409)
kNegLRBase 0.5 -> 1.5 (InterfaceRL.hpp:410)
/(1+len) taper deleted (InterfaceRL.tpp:724)
Upstream's own comment on the taper: "a 'no' should clearly move the mapping
away even from a sound already far from the liked region (the taper used to
kill exactly that case)". The direction is already a unit vector, so the
taper only ever shrank the push for exactly the sounds a user is most likely
to be rejecting.
Cold start is folded into the same path. Upstream's useRandom is
`!havePositives || len <= 1e-4`: with nothing liked yet there is no centroid
to push away from, so every dim goes in a random direction. Ours instead
kept the older 0a541cc fallback — train AWAY from the heard action at a
NEGATIVE lr — which was inert whenever the heard action equalled the net's
own output, i.e. in the common case. One path now, and a "no" moves the
mapping before any likes exist (ml_bench E1: 0 -> 2.3e-3). The
GeometricColdStart action is still reported so callers keep their "like a
few sounds first" prompt; only the training changed.
Measured (ml_bench, one dislike at a point):
A4 0.0157 -> 0.0533 at-point displacement (3.4x), so end to end across
this and the RMSProp fix: 5.3e-5 -> 5.3e-2, ~1000x. The gap to the
legacy Diffuse design closes from ~4100x to ~4.2x.
D1 effective_lr 4.7e-4 -> 1.5e-3; 10 presses now reach 0.34, 100 reach
the full intended push.
A5 compounding 0.87 -> 0.96 (a second press at the same spot is no
longer noticeably weaker than the first).
A7 damage_ratio essentially unchanged (0.87-1.57) — the collateral
damage to protected positives scales with the push and is NOT
addressed here; it is the negative-feedback design question.
NOT adopted, deliberately: upstream's per-tick batch retraining over all
live negatives, and its fixed kDislikeLifetimeMs=2500 in place of our
proportional decay. Both need something the core does not have — a per-tick
call site and a millisecond clock inside nisps/ml — so they change
FeedbackControllerCore's interface rather than its constants. Recorded in
ALIGNMENT's deferred-debt entry alongside the existing one-press-one-step
divergence, and filed as its own task.
test_mlp_geo_dislike.cpp: the taper test now pins its ABSENCE (equal
displacement near and far), the cold-start test pins movement where it used
to pin inertness, and a new test covers the random-direction branch.
ALIGNMENT defect 6b resolved. Gates: build-cpp-tests 139 tests / ctest 4/4,
parity-check PASS (WASM rebuilt), lint-cpp clean, firmware slpworkshop
SUCCESS.
126 KiB
Executable file
126 KiB
Executable file