KataGo V1.3

[go4thewin]

That may be a problem with zen,I don't know. But the much bigger problem from what I've read and experienced personally, is if you keep playing it over and over again you get too strong and the people you usually play against don't enjoy games against you anymore.

[And]

I want to play with KataGo! the question is not whether I will win or not. with or without a handicap, but at the same level. in which case is more interesting?

[Maharani]

in which case is more interesting?

try and see?

[Sneegurd]

I updated Katago and its model file and again it is not working with Sabaki 0.50.1.

Code: Select all

gtp -model C:\Users\Michel\Programme\Drago\Engines\Katago\model.txt.gz -config C:\Users\Michel\Programme\Drago\Engines\Katago\gtp.cfg

gtp config has been created with genconfig flag. model.txt.gz is a renamed g170-b30c320x2-s1287828224-d525929064 gz file.

Sabaki says:

"Connection failed"
"GTP engine can't be synced to current state".

Thx guys.

[Vargo]

1.jpg (40.02 KiB) Viewed 11129 times

This works with Sabaki 0.50.1 if your config file and your network are in the same folder as katago.exe

maybe the problem is your model name finishing with .txt.gz

[Uberdude]

Video with live comments of Cho Yeonwoo 1p playing KataGo on 3 stones:

[Sneegurd]

[removed my useless comment]

[lightvector]

Version 1.3.4 is released! Some changes to make configuration easier, more solid play when receiving handicaps, particularly at low playouts, added optional hack to avoid one opening joseki that both Kata sometimes misplays and that also leads to lack of opening diversity vs some bots since it fills a quarter of the board.

And as usual, continued improvements in the networks as training proceeds.

https://github.com/lightvector/KataGo/releases

[Yakago]

So you simply made it avoid the 'dagger'/'Flying knife' pattern? with the option on of course.

What led to this change ?

[Vargo]

50 game test : KG 1.3.4 g170e-b20c256x2-s3354994176-d716845198 v. LZ#270
twogtp 1.5.1, 1600 visits for LZ and for KG (~time parity), all games by resignation, no error, no duplicate game
KataGo wins 42-8 = 84% (!)

details : (KG always appears as W, because of the command -alternate)

1.jpg (269.22 KiB) Viewed 10645 times

2.jpg (634.47 KiB) Viewed 10645 times

[lightvector]

So you simply made it avoid the 'dagger'/'Flying knife' pattern? with the option on of course.

What led to this change ?

Well, it felt a bit weird to simultaneously have some people in chat lamenting about this issue, yet also simultaneously the majority of them voting to keep the training still "zero"(*) instead of doing the things that make the self-play learning correct the issue. So I did this to try to make both sides happy.

Anyways, if "zero"-like training remains unlucky and never picks it up, I'm pretty confident that it will be "solved" once adding simple ways of leveraging outside data or human and breaking fully from zero. Still, it would be nice to have a way to better improve self-play exploration by itself. Such issues aren't not Go-specific as far as I can tell, from a very broad perspective, you have the same kinds of exploration issues problems in AlphaStar / OpenAI5 / other research that continue to be an unsolved problem for general machine learning.

(*) KataGo is still half-"zero" in the sense that it uses some Go-specific input features and training targets, as well as some optimizations to improve selfplay speed (e.g. Benson's algorithm, minor details around passing), but 100% of the data is via self-play with no use of outside data, and also not even leveraging any other bots either (e.g. ELF or other external weights).

[Vargo]

50 game test :

KG 1.3.4 with best 20b (g170e-b20c256x2-s3354994176-d716845198)
v.
KG 1.3.3 with former best 20b (g170e-b20c256x2-s2971705856-d633407024)

twogtp 1.5.1, 1600 visits for both, all games by resignation, no error, no duplicate game

KataGo 1.3.4 wins 33-17 = 66%

details : (KG 1.3.4 always appears as W, because of the command -alternate)

k133134.jpg (348.33 KiB) Viewed 10535 times

k133134b.jpg (697.52 KiB) Viewed 10535 times

[johnsmith]

Katago is skyrocketing compared to other engines... and it even show us the score and knows ladders.

[go4thewin]

Katago is skyrocketing compared to other engines... and it even show us the score and knows ladders.

johnsmith, being 6dan, are you able to beat katago setting maximum playouts to 12 (and 1 thread)? i'm trying to figure out how strong it is, thanks!

[johnsmith]

Katago is skyrocketing compared to other engines... and it even show us the score and knows ladders.

johnsmith, being 6dan, are you able to beat katago setting maximum playouts to 12 (and 1 thread)? i'm trying to figure out how strong it is, thanks!

I've just tried one 5sec blitz game with those settings and lost by more than 10pts... the bot's direction of playing is rock solid even with 12 playouts. But I was able to trick it on 2 places, like when it didn't connect on a second line, so I cut and pushed a couple of times until it realized it's been caught So the score was even greater before that. So it might be a gamble with simple tactical combos, even though Kata's feeling is better than mine. I had a feeling of a human opponent who could make such simple errors when playing blitz and trying to grab points on every move.

From the Deepmind paper, Alpha zero had a rating of around 3200 when using only one playout. So I'm sure Kata is there somewhere too.
I'd also like to add this page https://github.com/breakwa11/GoAIRatings#go-ai-ratings
There's an approximate rating of various LZ networks, so that you could try Kata against them. After you had found out the rating against which Kata scores 50%, using the same rating system you could find corresponding human here https://www.goratings.org/en/