KataGo V1.3

[lightvector]

Isn't that up to the GUI you use, and/or your own choice of where to put your stones?

The kind of handicap configuration you post is what KataGo itself would might choose if *it* was the one playing with that many stones. I'm not sure though why your GUI would ask KataGo to place the stones as if KataGo were the one getting the extra stones, rather than letting you do so.

[jlt]

I am using the last version of Lizzie. I unpacked the last version of KataGo in the directory Lizzie\katago, downloaded the weight g170e-b15c192-s1305382144-d335919935.txt.gz in that directory. In the list of engines, I entered the command "katago\katago.exe gtp -model katago\g170e-b15c192-s1305382144-d335919935.txt.gz -config katago\gtp_example.cfg".

That's all. So maybe it's in the configuration of Lizzie, but I couldn't find out where.

[lightvector]

Ah, I see. Apparently the GTP spec indicates that bots should implement a fixed handicap placement method in addition to a free one, according to exact rules defined in the spec that vary with board size oddness/evenness and a few other things. (Yuck. )

KataGo never implemented this, since Chinese rules allow free placement, and until recently KataGo didn't support any other rules, so it wasn't necessary anywhere that KataGo could work. I'll go ahead and add this for next minor release.

The interesting thing is that even if I implement this now, Lizzie won't pick it up automatically because Lizzie, rather than either checking for the availability of the bot's placement methods, or placing the stones itself, appears to be hardcoded to call KataGo's free placement method.

Anyways, as a workaround until Lizzie also itself gets a new version, you can just play all the stones yourself (passing as white), and then just resume play from there?

[jlt]

Anyways, as a workaround until Lizzie also itself gets a new version, you can just play all the stones yourself (passing as white), and then just resume play from there?

Yes, I will do that.

[And]

Lizzie with KataGo "g170 20 block s1.42G" or "g170e 10 block s1.14G" quickly find a winning move:

viewtopic.php?p=244851#p244851

SAI and ELF could not find this move in an hour

[go4thewin]

Someone is running a katago 1.3.1 bot on the KGS server using AWS and getting 1000 playouts per second. The bot is ridiculously strong: https://www.gokgs.com/gameArchives.jsp? ... Accounts=y

Username: katagobot

You can view all the sgfs in the link. This is a fun bot to watch play! Hopefully in the future, these bots make human pros obsolete.
Katagobot uses newest g170 20 block s1.42G net

[ez4u]

See also user "petgo2", which is the latest katago running on an RTX 2080 Ti. Petgo2 is 9d. Up to now katagobot is only "9d?" because it has not yet lost a rated game!

Edit: 45 minutes after posting the above, katagobot finally lost a rated game by 6.5 points, giving 6 stones to a 5d.

[xela]

Someone is running a katago 1.3.1 bot on the KGS server using AWS and getting 1000 playouts per second. The bot is ridiculously strong: https://www.gokgs.com/gameArchives.jsp? ... Accounts=y

Username: katagobot

What's more, katagobot is one of only two players in the KGS top 10 that's achieved their rank by playing humans, rather than by learning to beat up HiraBot on H6. (The other is stakeout at number 10).

[Yakago]

@lightvector

Do you have some additional details on your internal rating tests btw?

On CGOS, some users have put up 800 visits katago bots.

This one I would guess is the first 1.3 release, and it ranks slightly higher than LZ ELFv2 v800, but is worse in the head-2-head matchup.

Another one should be the 2nd release. It has not played that much, and has been taken offline, but had even poorer results against LZ ELF v2

We of course don't know the settings of these, but what do you make of it?

[lightvector]

Do you mean "LZ_05db_ELFv2_p800"? Taking a wild guess based on the name I would guess that this is ELF with 800 playouts, not visits? If so, then it may be getting somewhere around 2x the number of playouts that KataGo is getting. KataGo and LZ-ELF being almost an even match despite that difference seems like KataGo is doing well.

How much compute time that takes depends on hardware and a lot of other details. Some users have reported that LZ-ELF does get almost 2x as many playouts due to LZ having had more GPU optimization work, making it close. Some users have reported a much smaller gap than that already, such as on FP16 tensor core cards, and with a smaller gap KataGo easily outperforms.

(And if you wait a week or so, some more performance improvements should be coming too - working on it...).

[Yakago]

Ah I totally missed the 'p' Thanks for clarifying

[ez4u]

@lightvector
Do you have any plans to release any further updates to the 10b net? On my old hardware it is a joy to fool around with.

I understand you have to prioritize things though.

[go4thewin]

@lightvector

Do you have some additional details on your internal rating tests btw?

On CGOS, some users have put up 800 visits katago bots.

This one I would guess is the first 1.3 release, and it ranks slightly higher than LZ ELFv2 v800, but is worse in the head-2-head matchup.

Another one should be the 2nd release. It has not played that much, and has been taken offline, but had even poorer results against LZ ELF v2

We of course don't know the settings of these, but what do you make of it?

Cgos rankings have katago 1.3.1 s103 at 800 visits at a comparable strength to lz elfv1 at 1600 playouts, so 9d
http://www.yss-aya.com/cgos/19x19/cross ... 3v800.html

Estimates of different bots human rankings (overestimation)
https://github.com/breakwa11/GoAIRatings

So basically you will never in your life beat katabot on ogs in an even game. Which is such an amazing thing for such a young bot

[Gomoto]

I also love the 10b network, even on a 1080ti. Responsivness is everything

[lightvector]

@lightvector
Do you have any plans to release any further updates to the 10b net? On my old hardware it is a joy to fool around with.

I understand you have to prioritize things though.

I already stopped training that one - it seemed mostly to be capped out and not getting noticeably stronger at the time I released it. Perhaps you could gain still another 100-200 Elo with really really really long training times on big net data with low learning rates, but I think overall it's close to as strong as possible at that size given KataGo's current methods.

Glad you like it!

And if you do want to go one step up the ladder of slower but stronger, the 15-block net here https://github.com/lightvector/KataGo/r ... 1.3.1-nets is just shy of the old KataGo (g104) 20 block net in strength. I'm continuing to train it a bit more and it's now just a little past g104 20 block in strength. - I'll release an updated version when it's done - it's slowing down a lot and probably going to cap out soon.