[Computer-go] Crazy Stone is playing on CGOS 9x9

[Rémi Coulom]

Hi,

I trained a neural network for 9x9, and it is playing on CGOS.

The network has 40 layers (20 residual blocks) of 256 units. It is running on a Titan V GPU, with a batch of 64, at about 9k playouts per second.

It is using an opening book that you can browse online there:

https://www.crazy-sensei.com/book/go_9x9/

Each node of the book was searched with 400k playouts.

I will let it play for a few days.

Rémi

[uurtamo .]

great book interface, by the way.

s.

_______________________________________________
Computer-go mailing list
Compu...@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

[Brian Lee]

Agreed on the book interface!

If you click through to the end of a variation, you can see that the evaluation is a minimax'd evaluation, instead of an average over the subtree nodes, even if the overall tree was generated via a UCT algorithm. The values seem a bit more fragile to me as a result...

[Rémi Coulom]

Hi,

Thanks to all the strong bots who joined. Kata is impressive. Does anyone know more about its configuration? Is it a single V100 or many?

I watched some games, and some where spectacular.

A firework of ko fights:

http://www.yss-aya.com/cgos/viewer.cgi?9x9/SGF/2020/05/07/997314.sgf

In this game, Crazy Stone won using a typical Monte Carlo trick:

http://www.yss-aya.com/cgos/viewer.cgi?9x9/SGF/2020/05/07/997390.sgf

On move 27, it sacrificed a stone. According to Crazy Stone, the game would have been a draw had Aya just re-captured it. But Aya took the bait and captured the other stone. Crazy Stone's evaluation became instantly winning after this, the sacrificed stone serving as a threat for the winning ko fight, 18 moves later.

Rémi

[Shawn Ligocki]

Thanks for sharing the games, Rémi!

On Thu, May 7, 2020 at 6:27 AM Rémi Coulom <remi....@gmail.com> wrote:

In this game, Crazy Stone won using a typical Monte Carlo trick:

http://www.yss-aya.com/cgos/viewer.cgi?9x9/SGF/2020/05/07/997390.sgf

On move 27, it sacrificed a stone. According to Crazy Stone, the game would have been a draw had Aya just re-captured it. But Aya took the bait and captured the other stone. Crazy Stone's evaluation became instantly winning after this, the sacrificed stone serving as a threat for the winning ko fight, 18 moves later.

Wow, I did not imagine how that move would be useful later! But the very end is confusing to my human brain, couldn't White move 56 retake the ko and win it? It seems like Black only has one real ko threat left (J4 maybe). But White also has one huge threat left (D3), so it seems like White should win this ko and then be about 4 ahead with komi. Am I missing something?

-Shawn

[David Wu]

Yes, it's fun to see suddenly a little cluster of people running strong 9x9 bots. :)

katab40s37-awsp3 is running on one AWS p3 2xlarge instance. So it's a single V100, with some settings tuned appropriately for good performance on that hardware and board size and time control (mainly, 96 threads), and also some of the score maximization config settings toned down so as to focus more on win/loss. It's running one of the most recent 40 block KataGo nets from the current ongoing training run, a net that isn't released yet but which I'll release soon. But it's only more recent by a couple of weeks, the latest released ones should probably be about the same strength too.

I think it gets somewhere from 5000 to 7000 playouts per second, I forget what the exact numbers were. No opening book, just searching from scratch. A book seems like it would enable a large increase in the effective number of playouts early on (and save time for deeper search out-of-book too), but I haven't worked on book code.

The current run, which has been ongoing for a few months on variously from 37 to 47 V100s spends about 4% of its games on 9x9, which probably amounts to 1%-2% of the total compute since 9x9 games are much shorter than games on larger boards. So presumably a lot of the strength on 9x9 is due to the neural net generalizing its knowledge from other board sizes, since the same net trains on all sizes at the same time.

There's a setting in KataGo ("PDA") that causes it to play a bit more aggressively, in whatever way it has learned from self-play that increases the likelihood of weaker players making a mistake. In 19x19 handicap games, it gives a strong boost in how many handicap stones it's capable of offering to weaker opponents, even helping against players near human pro level. I haven't turned it on here, but maybe I could make a second username with this set to a nonzero value. I'd be curious if it would make KataGo avoid some of the more easily drawn lines of play in favor of ones that would be more challenging for most opponents.

On Thu, May 7, 2020 at 6:27 AM Rémi Coulom <remi....@gmail.com> wrote:

[Rémi Coulom]

If White recaptures the Ko, then Black can play at White's 56, capture the stone, and win by 2 points.

[David Wu]

Having it matter which of the stones you capture there is fascinating. Thanks for the analysis - and thanks for "organizing" this 9x9 testing party. :)

[Martin Mueller]

http://www.yss-aya.com/cgos/viewer.cgi?9x9/SGF/2020/05/07/997479.sgf

Great games! This is my favorite so far. The way black lives inside the super-safe white area is incredible...

Martin

[Ryan Hayward]

Hey Martin,

thanks!  I never realized that the sgf came with a viewer... beautiful :)

yeah, wild wild wild...

looking forward to seeing 6x6 games like this :)

R

_______________________________________________
Computer-go mailing list
Compu...@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

--

Professor Ryan B. Hayward, Computing Science Dept., UAlberta, Edmonton T6G 2E8, Canada 780-492-2285  RyanBHayward  @RyanBHayward Hex-the-full-story

[Rémi Coulom]

And congratulations to rn for beating kata in a very beautiful game:

http://www.yss-aya.com/cgos/viewer.cgi?9x9/SGF/2020/05/08/998312.sgf

I am not strong enough to appreciate all the subtleties, but the complexity looks amazing.

Rémi

[Kensuke Matsuzaki]

Rn says move 21 and move 27 was not good, but I can't understand.
rn.6.3.945 is running on EC2 g4dn.12xlarge, and it's network is 256
channel * 20 resnet blocks.

> And congratulations to rn for beating kata in a very beautiful game:
> http://www.yss-aya.com/cgos/viewer.cgi?9x9/SGF/2020/05/08/998312.sgf
>
> I am not strong enough to appreciate all the subtleties, but the complexity looks amazing.

--
Kensuke Matsuzaki

[Kyle Biedermann]

I looked through the game on one point I found some draw lines at move 13 b5 instead of b6 I tried out the line on OGS against KataGo https://online-go.com/game/23768001 small sample if you have any other variations you want me to try out I could do that.

Kyle Biedermann

Creator of Deep Scholar

[David Wu]

I'm running a new account of KataGo that is set to bias towards aggressive or difficult moves now (the same way it does in 19x19 handicap games), to see what the effect is. Although, it seems like some people have stopped running their bots.  Still maybe it will be interesting for the remaining players, or any others who decide to re-turn-on their bot for a little while. :)

It seems like some fraction of the time, it now opens on 5-5 as black, which is judged as worse than 4-4 in an even game, but presumably is more difficult to play. I suspect it will now start to lose a noticeable number of games now due to overplaying, and there's a good chance it does much worse overall. Even so, I'm curious what will happen, and what the draw rate will be. Suddenly having some 5-5 openings should certainly add some variety to the games.

[Kyle Biedermann]

Sounds fun and interesting experiment, I have noticed the increase in preference to the 4-4 as of lately across the majority of AI. I still prefer the 5-5 opening it seems to hold against Katago at the moment. Maybe i'll test out some things to see if it can find some interesting new moves. Is it still the Kata-bot account on OGS?.

Kyle Biedermann

Creator of Deep Scholar

[uurtamo .]

And this has no book, right? So it should be badly abused by a very good book?

s.

[David Wu]

kata-bot on OGS is intended for human players on OGS and is never guaranteed to be any particular version (certainly not an up-to-date version) nor have any specific fixed settings. You should generally not use it for testing - just download KataGo and run it yourself (in Lizzie, or Sabaki, or whatever) if you want reliable settings.

The "aggressive" version that I mentioned in my last email is running on cgos, not on OGS.

[David Wu]

On Fri, May 8, 2020 at 8:46 PM uurtamo . <uur...@gmail.com> wrote:

And this has no book, right? So it should be badly abused by a very good book?

Maybe! 

But the version that was running before which went something like 48-52-1 (last time I counted it up) against the other top 3 bots that were rated 3300+ also had no book. Granted, the newer one is coded to play a little more excitingly on average, probably so its opening lines might have more flaws. At least, that's what the net should have been trained to do. (katab40s37-pda1)

 

[Rémi Coulom]

The two wins of King against Kata have the same opening.

I made the opening book of Crazy Stone to avoid this problem. Whenever it draws or loses a game, it marks the leaf of the tree as losing, so that it avoids playing it again.