A new run of KataGo released - strength comparable to ELFv2

[lightvector]

Probably not. The problem is that the larger the board size limit, the more memory the program uses and the worse the performance, even if you don't actually play on the larger board, because some internal datastructures get initialized in order to be able to support those large boards (due to some tricky implementation details in the current implementation).

Also, the neural net is only trained on size 19 and smaller, so it will perform very poorly on such huge sizes, so there is really no point. One could try training on larger boards, but the larger the board the more expensive the training, and boards larger than 19 I don't see as actually that popular, so it doesn't seem worth the cost.

Although huge sizes are quite bad for the net, I think the current best net does okay on sizes in the low 20s despite training only on 19 and smaller. I haven't checked this in detail myself. But if that's the case, I could release a completely separate version supporting up to, say, 25x25. (separate, to avoid the cost for users who don't care about the large board).

[And]

lightvector, Thank you!

KataGo plays very strong, I think it will play weaker on large boards, but still very good.
I can take up testing, buy a 1050ti graphics card. can release a "test" version at least up to 37x37?
but if will be a version 25x25, also great!
about the popularity of large boards - why rarely fly to the moon? not popular or no starship? maybe if will a good program for big boards, will play more? at kgs 37x37 tournaments are sometimes held: https://www.gokgs.com/tournInfo.jsp?id=739

the game on the big boards becomes completely different, so it is so interesting! about the game 37x37:
https://senseis.xmp.net/?LargeBoards

Leela 0.6.2 (2016-06-04) (37x37)
https://sjeng.org/leela.html

mygofriend Adjustable Board Size from 2x2 to 52x52
http://mygofriend.com/home

[And]

The new version of q5Go-1.0 https://github.com/bernds/q5Go/releases/tag/q5go-1.0
viewtopic.php?p=240327#p240327
KataGo works on rectangular boards, for example 13x19, 4x19!

[And]

new KataGo networks, even stronger https://github.com/lightvector/KataGo/i ... -533617837
2 games KataGo 10x128-s1108 - KataGo 10x128-s4073 (gt 610, 5 sec/move). in the first, the new network lost (b), in the second won (w):

[Maharani]

I'm sorry if this has been asked before.

If komi is set to 7, does KataGo understand draws? If not, would it be possible to train KataGo to understand draws in the future?

[And]

50 games KataGo 10x128-s1108 - KataGo 10x128-s4073 (gt 610, 5 sec/move), new network won 42:8 (21 b and 21 w)
in one game, the old network surrendered on move 150:

[And]

50 games KataGo 15x192-s2973 - KataGo 10x128-s4073 (gt 610, 5 sec/move), new network won 32:18 (15 b and 17 w)
look how gracefully the new network won in this game (15x192 surrendered on 101 moves):

[And]

new network 10x128 https://github.com/lightvector/KataGo/f ... del.txt.gz
https://github.com/lightvector/KataGo/i ... -567718559

[And]

50 games 5 sec/move:
KataGo 10x128-s4588 - ELF v2 (gt 610) 37:13 (KataGo b 16:9 and KataGo w 21:4)
8 games the ELF lost due to the ladder.
without these games
KataGo - ELF 29:13 (KataGo b 13:9 and KataGo w 16:4)
two games, won KataGo:

[splee99]

For higher playouts (visits), Katago is still much weaker than ELFv2.

[Maharani]

For higher playouts (visits), Katago is still much weaker than ELFv2.

What amount of playouts are we talking?

[Mike Novack]

And what real time associated with each? Discussing strength in terms of number of playouts if ignoring how much TIME involved for a given amount of hardware does not give a true picture, because go is played with finite time.

Thus telling us that network A is weaker than network B at X playouts each does not mean that network A is weaker than network B on machine Y at time control Z if when running on machine Y network A can do X playouts within Z seconds and network B cannot.

[And]

with the same number of playouts, the elf network thinks about 7 times longer than 10x128.
if I play against a human, then I’m interested in how long he thinks, and it’s not at all interesting how much he weighs, what his pulse, temperature, etc.
or splee99 implies higher playouts, but with time parity?

[Maharani]

Even if we're just talking pure number of playouts (which I do find interesting)... At what point does ELF OpenGo v2 start giving much better results than KataGo?

[splee99]

I was running a match between Katago(B) and ELF(W). Yes I used time parity with 20 seconds per move. The playouts are variable, but ELF used roughly 5000 playouts per move. It looks like ELF was ahead from the beginning to the end. I didn't set any resign threshold but at the end both sides indicated that Katago was lost.

kata-elf.sgf

(2.88 KiB) Downloaded 992 times