Experiments · E189
Yes. On the old reward, freezing the pairing scored 41.1, level with the other live signals and 15 above a fixed random gain.
In the log: live or re-drawn? Shuffled gains with the permutation frozen after the first call
book/assets/diagrams/exp/E189.svg).E185 left two readings of "shuffled ≈ own ≫ fixed-random": either what pays is that the gain
moves in time with the head (any walker's state, reassigned once), or that it is
re-drawn on every call (a fresh reassignment each move — a source of noise a fixed draw
lacks). FORAGER_MODULATE=shuffle_frozen: the permutation is drawn once and kept, so each
walker follows one other walker's live state for the whole run. Same MPS family as E185,
tag e189, 200 × 4 × 2.
Predictions. 1. E189 ≈ shuffle (mean within 7 of 40.4 and above 33): what matters is a
live, composition-dependent gain — the head's state is doing work even when it is not the
walker's own. 2. E189 ≈ random (mean below 30): the per-call re-drawing was the ingredient,
and "the head's state" is not the active part; the modulator then reduces to a noise
injector and is described as one. 3. Between 30 and 33: undecidable at two seeds.
E189 result (15:31) — prediction 1 confirmed: live, not re-drawn. Shuffled gains with the permutation frozen after the first call (each walker follows one other walker's live state all run): AUC_Q 41.12 ± 1.44 (seeds 39.7 / 42.6), 88 ± 8 finds, best −138. The MPS family now reads:
own-state gains 39.94 live, own
shuffled, re-drawn 40.40 live, someone else's, reassigned every call
shuffled, frozen 41.12 live, someone else's, reassigned once
fixed random draw 25.53 not live
The three live arms are one number; the dead one is 15 below. So the ingredient is a gain that moves with the head's state as the swarm moves — a composition-dependent, time- varying signal — and it does not need to be the walker's own. E170's "heterogeneity, not the head" becomes: the head's live state is the heterogeneity that pays; which walker reads which cell does not matter. That is a statement about the head as a shared field, not as a per-walker controller — and it is the first result in which removing the head (fixed draw) costs the search something on every seed. Task 21 closed.
This entry is written in 2 separate places in the log, shown here in log order.
E185 left two readings of "shuffled ≈ own ≫ fixed-random": either what pays is that the gain
moves in time with the head (any walker's state, reassigned once), or that it is
re-drawn on every call (a fresh reassignment each move — a source of noise a fixed draw
lacks). FORAGER_MODULATE=shuffle_frozen: the permutation is drawn once and kept, so each
walker follows one other walker's live state for the whole run. Same MPS family as E185,
tag e189, 200 × 4 × 2.
Predictions. 1. E189 ≈ shuffle (mean within 7 of 40.4 and above 33): what matters is a
live, composition-dependent gain — the head's state is doing work even when it is not the
walker's own. 2. E189 ≈ random (mean below 30): the per-call re-drawing was the ingredient,
and "the head's state" is not the active part; the modulator then reduces to a noise
injector and is described as one. 3. Between 30 and 33: undecidable at two seeds.
E189 result (15:31) — prediction 1 confirmed: live, not re-drawn. Shuffled gains with the permutation frozen after the first call (each walker follows one other walker's live state all run): AUC_Q 41.12 ± 1.44 (seeds 39.7 / 42.6), 88 ± 8 finds, best −138. The MPS family now reads:
own-state gains 39.94 live, own
shuffled, re-drawn 40.40 live, someone else's, reassigned every call
shuffled, frozen 41.12 live, someone else's, reassigned once
fixed random draw 25.53 not live
The three live arms are one number; the dead one is 15 below. So the ingredient is a gain that moves with the head's state as the swarm moves — a composition-dependent, time- varying signal — and it does not need to be the walker's own. E170's "heterogeneity, not the head" becomes: the head's live state is the heterogeneity that pays; which walker reads which cell does not matter. That is a statement about the head as a shared field, not as a per-walker controller — and it is the first result in which removing the head (fixed draw) costs the search something on every seed. Task 21 closed.
Built with PRISMWebsite and visualizations made using Claude