KataGo v1.4

[And]

here http://rusgo.org/2020/04/29/leela-zero-vs-katago/ talking about 40x384 network. Is she learning or has she failed?

[Uberdude]

Kim Yoonyoung 8p (https://www.twitch.tv/kimyoonyoung) is playing a variable handicap series against KataGo running on beefy V100 hardware. She lost 2 games on 2 stones and 6.5 reverse komi and just now one on 3 stones, the "this group won't die" died after a very nice not working ladder KataGo played out to get some helper stones in sente. 2nd 3 stone starting now.

Edit: she won the 2nd on 3 stones by 8.5 points.

[And]

can these games be downloaded in sgf?

[johnsmith]

can these games be downloaded in sgf?

Here are the games: https://online-go.com/player/787607/

[Vargo]

KG 1.4.2 g170-b20c256x2-s4384473088-d968438914
v.
KG 1.4.2 g170e-b20c256x2-s5055114240-d1149032340 (new experimental net from https://github.com/lightvector/KataGo/r ... 1.4.2-nets)

30 games, 1s/move, gpu : 1x GTX1080, gogui-twogtp 1.5.1

The new net wins 19-11 (25 games by resignation, 5 games W+0.5 or B+0.5, no error, no duplicate game)

1.jpg (210.97 KiB) Viewed 11715 times

2.jpg (542.3 KiB) Viewed 11715 times

PS. In another 30 game test between these two nets, the score was 16-14, but I had forgotten to use -alternate, so the new net was 30 times W...

[And]

would now be made 30 times black, there would be 60 games

[Vargo]

KG 1.4.2 g170-b20c256x2-s4384473088-d968438914
v.
KG 1.4.2 g170e-b20c256x2-s5055114240-d1149032340 (new experimental net from https://github.com/lightvector/KataGo/r ... 1.4.2-nets)

After 60 more games (total of 90 games, each side 45 times B, 1s/move, 1x GTX1080)
The new net wins 50-40 (55.6%)

---------------------------------------

30 games between
KG 1.4.2 g170-b40c256x2-s4679779328-d1149909226 (new experimental)
and
KG 1.4.2 g170-b40c256x2-s3708042240-d967973220

The new net wins 16-14 , 22 games by resignation and 8 games(!) W+0.5 or B+0.5

1s/move, 1x GTX1080, GoGui-TwoGtp 1.5.1, no error, no duplicate game

[Vargo]

KG 1.4.2 g170-b30c320x2-s3530176512-d968463914
v.
KG 1.4.2 g170-b30c320x2-s4432082944-d1149895217 (new experimental net)

50 games, 1s/move, 1x GTX1080
The experimental net wins 28-22 (42 games by resignation, 7 games B+0.5 or W+0.5 and 1 game by 1.5)

30_1.jpg (198.54 KiB) Viewed 11600 times

30_2.jpg (787.19 KiB) Viewed 11600 times

[Vargo]

new experimental nets 20b, 30b, 40b (from https://github.com/lightvector/KataGo/r ... 1.4.2-nets)
50 game tests, TwoGtp 1.5.1, 1s/move, gpu : 1x GTX 1080

g170e-b20c256x2-s5055114240-d1149032340 v. g170-b30c320x2-s4432082944-d1149895217
20b wins 31-19 (62%) No error, no duplicate, 43 games by resignation, 7 games B+0.5 or W+0.5

g170e-b20c256x2-s5055114240-d1149032340 v. g170-b40c256x2-s4679779328-d1149909226
20b wins 29-21 (58%) No error, no duplicate, 41 games by resignation, 9 games B+0.5 or W+0.5

g170-b30c320x2-s4432082944-d1149895217 v. g170-b40c256x2-s4679779328-d1149909226
30b wins 26-24 (52%) No error, no duplicate, 45 games by resignation, 5 games B+0.5 or W+0.5

On this PC, with 1x GTX1080
1s/move corresponds to maxVisits =~ 700-800 for g170e-b20c256x2-s5055114240-d1149032340

2030.jpg (237.23 KiB) Viewed 11509 times

2040.jpg (226.73 KiB) Viewed 11509 times

3040.jpg (196.68 KiB) Viewed 11509 times

[go4thewin]

Is there anyway to play the nonzero version of katago on ogs? Thanks for the test matches!

[lightvector]

Is there anyway to play the nonzero version of katago on ogs? Thanks for the test matches!

"kata-bot" on OGS is now running one of the latest nets.

You're likely not going to notice any difference, however (except maybe if you try to play mi-yuting's flying dagger, but even then I think kata now often wants to dodge that joseki on its own). The "non-zero" training doesn't teach it to play any moves or shapes or any style of play that it wouldn't already consider good on its own or that it wasn't already playing. It just helps it to find a few of the blind spots where it would consider a move good if only it searched it deeply enough but didn't actually do so, and a human player came along and demonstrated that move. Of course, plenty of embarrassing blind spots remain.

[johnsmith]

@lightvector which positions besides MYT have been fed to Kata in this most recent net?

[lightvector]

@lightvector which positions besides MYT have been fed to Kata in this most recent net?

https://github.com/lightvector/KataGo/issues/240

[johnsmith]

@lightvector which positions besides MYT have been fed to Kata in this most recent net?

https://github.com/lightvector/KataGo/issues/240

Nice! Would that mean that some other human josekis, like taisha, have been hinted and investigated as well (if they ever occured in those games)?

[lightvector]

Maybe, a little. But I wouldn't count on there being a large effect for any particular joseki. Maybe just some more general familiarity with the kinds of midgame positions and patterns and shapes that occur in human games that would be less likely in a bot vs bot self-play.