Leela-zero: Recommendations 52% above the winning rate are pass

[a270709570]

This will speed up the efficiency of chess and improve efficiency：
The winning rate of the 400 game is 54%, and the probability of chess promotion is 93.95%；
The winning rate of the 400 game is 53%, and the probability of chess promotion is 87.49%；
The winning rate of the 400 game is 52%, and the probability of chess promotion is 77.34%；
The winning rate of the 400 game is 51%, and the probability of chess promotion is 63.68%

[odeint]

Groundhog Day...
https://github.com/gcp/leela-zero/issues/1734
https://github.com/gcp/leela-zero/issues/1524
https://github.com/gcp/leela-zero/issues/906
https://github.com/gcp/leela-zero/issues/676
https://github.com/gcp/leela-zero/issues/543
https://github.com/gcp/leela-zero/issues/420
...and many more.

[l1t1]

only the proved stronger weights can touch the PM, like two elfs, c9d7, but there are a few exception, like dc01, or a mistake made by the machine like b1c3

[wonderingabout]

as said many times already, the 55%/400 games gating avoids a lot of hassle, especially when the project is held by a community made of many different people, see how leela chess zero project struggles because of the elo yoyo, bringing the community unstable everytime there is a drop

it's ok if it's slow, slowly but surely
this loss of speed was expected when moving from 20b to 40b, yet people wanted to promote to 40b, so all what is left now is to wait

also, it is not true that no progress is made
indeed, no new network is promoted, but a lot of selfplay games are played
see for example between lz 172 and lz 173, we waited 10 days for a new promotion, but lz 173 gained 4-5% winrate (at 400 games) against the same 40b network
so again, no promotion does not mean no progress

if you want to decrease gating to for example 53%, you need to increase the number of games to arround 800 games to keep the same significancy
seeing as slow the 40b is to train, i'd rather take these 800 games in selfplay rather than waste them in unsuccessful match games

[l1t1]

https://cloudygo.com/v3-9x9/eval-graphs