Abstract
A sense of the nearest live pillbox's ring was added to the nets' senses and proved harmless to every shipped net. Nets imitating the rules with it did no better on the two ring gauntlets than nets without it, and worse on G7. By the rule written before the run it is not shipped; the code is parked in a stash.
I. Intent
Paul: "perhaps the sensor will indicate a ring, as it would indicate and identify the foe."
II. Decision log
| Decision | Cause |
|---|---|
| The block goes last, six numbers | every older net and patch widens with zeros and plays as it did |
| A pillbox "that fires on the tank" is a live enemy's or a neutral one | a neutral pillbox fires on everyone |
| Judged by imitation, five runs a side | one run of each differed by more than the effect sought |
| The no-ring nets' six input columns zeroed after training | weights on an input that never moved stay at their random start |
| Not shipped, parked in a stash | the rule was not met |
III. Method
The gauntlet for the rules and every shipped net, with the block and at LR 145's commit without it. RECORD_COURSE on G7, G8, G18 and G19, 2,000 seeds from 9,100,000; the same file with the block zeroed; treadcore-bc with BC_DEVICE=flex, 30 epochs, batches of 512, rate 0.001; each net on the gauntlet's 200 seeds.
IV. Results
Identical to the seed for every shipped net. Imitation, five runs a side, passed of 200: G7 42.1% with the ring, 52.0% without; G8 85.2% and 83.3%; G18 100% and 100%; G19 40.7% and 44.5%, each run anywhere from 0 to 172.
V. Discussion
The nets' failure on G18 was a lack of experience: imitating the rules where the choice comes up teaches it without the new sense. G19 is the open one, and its spread is the training's, not the senses'.
VI. What is not built
Several training starts with the best kept, for G19.
VII. Paul's decisions
Whether to drop the parked block or keep it for a later trial.