KataGo Distributed Training and new networks

[And]

"first new net of the new architecture, b18c384nbt, has been uploaded for testing!"
https://katagotraining.org/networks/
https://katagotraining.org/

[And]

LeelaGUI v2.0.6 https://github.com/MAOmao000/LeelaGUI/releases

[And]

the version of katago_avx2 (katago-v1.12.4-eigenavx2-windows-x64) included in the LeelaGUI v2.0.6 package works ~5 times faster than the official version
https://discord.com/channels/4170221623 ... 2965661706

[And]

LeelaGUI v2.0.8 https://github.com/MAOmao000/LeelaGUI/releases

Notable Changes
Supports PCs with less than 2 CPU cores.
Only the Windows 64-bit version will bundle KataGo conjunctioned with "mimalloc".

Other notes
The following dlls are required for KataGo (Windows 64-bit version) using "mimalloc".
mimalloc-override.dll
mimalloc-redirect.dll
msvcp140.dll
msvcp140_1.dll
vcruntime140.dll
vcruntime140_1.dll

[And]

KataGo b4c32-s709 (playouts 100000, Japanese rules) - Zenith Go 7 6d 1:0
katago - tensorRT, GeForce GTX 1650, 2.5-3 min/move

[And]

KataGo b4c32-s709 (playouts 100000, Japanese rules) - CS Zero 6d 1:0

[And]

KataGo b4c32-s709 (playouts 100000, Japanese rules) - CS Zero 7d 1:0

[Amigo]

has anyone held a match with KataGo optimistic-policy?

[And]

policyOptimism = 1, rootPolicyOptimism = 1, b18c384nbt-optimisticv13-s5971M.bin.gz

KataGo (playouts 1, Japanese rules) - original Zenith Go 7 9d H5 0:10
one game:

[PC_Screen]

Test carried out by Wing on the Go AI discord:

Managed to get hold of a 3080ti + 4090 PC plus another 3090 + 4090 PC, ran 3 batches of tests last night.
Networks used: kata1-b18c384nbt-s5832081920-d3223508649.bin.gz (for simplicity I'll refer to it as 583.gz) and b18c384nbt-optimisticv13-s5971M.bin.gz (I'll refer to it as cv13.gz below)
1000 visits per move, komi = 6.5

Batch A 250 games cv13.gz policyOptimism=1 vs 583.gz policyOptimism=0

cv13.gz ahead 145:104 58.23% win rate 57.73 ELO +-44.43 stronger

Batch B 430 games cv13.gz policyOptimism=0 vs 583.gz policyOptimism=0

583.gz ahead 237:191 55.37% win rate 37.47 ELO +-33.65 stronger (2 double pass games)

Batch C 200 games cv13.gz policyOptimism=1 vs cv13.gz policyOptimism=0

cv13.gz policyOptimism=1 ahead 123:77 61.50% win rate 81.37 ELO +-50.31 stronger

Conclusion from the test: policyOptimism = 1 is around a 80 ELO boost over policyOptimism = 0

[And]

Test carried out by Wing on the Go AI discord:

Managed to get hold of a 3080ti + 4090 PC plus another 3090 + 4090 PC, ran 3 batches of tests last night.
Networks used: kata1-b18c384nbt-s5832081920-d3223508649.bin.gz (for simplicity I'll refer to it as 583.gz) and b18c384nbt-optimisticv13-s5971M.bin.gz (I'll refer to it as cv13.gz below)
1000 visits per move, komi = 6.5

Batch A 250 games cv13.gz policyOptimism=1 vs 583.gz policyOptimism=0

cv13.gz ahead 145:104 58.23% win rate 57.73 ELO +-44.43 stronger

Batch B 430 games cv13.gz policyOptimism=0 vs 583.gz policyOptimism=0

583.gz ahead 237:191 55.37% win rate 37.47 ELO +-33.65 stronger (2 double pass games)

Batch C 200 games cv13.gz policyOptimism=1 vs cv13.gz policyOptimism=0

cv13.gz policyOptimism=1 ahead 123:77 61.50% win rate 81.37 ELO +-50.31 stronger

Conclusion from the test: policyOptimism = 1 is around a 80 ELO boost over policyOptimism = 0

There are different numbers of games in these tests. someone may decide that matches were stopped when it was necessary to achieve a certain result

[And]

v13 : b20_s530 97:3 (50:0 v13 W, 47:3 v13 B), policyOptimism = 1, rootPolicyOptimism = 1, 100 visits, chinese
one game:

[And]

KataGo v13 : leelaz b40_287 73:27 (41:9 v13 W, 32:18 v13 B),
v13 policyOptimism = 1, rootPolicyOptimism = 1, 100 visits
leelaz 1000 playouts
one game:

[And]

KataGo s583 : leelaz b40_287 78:22 (46:4 s583 W, 32:18 s583 B),
s583 100 visits
leelaz 1000 playouts
one game:

[And]

KataGo s583 : leelaz b40_287 55:45 (36:14 s583 W, 19:31 s583 B),
s583 50 visits
leelaz 1000 playouts
one game: