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.cou...@gmail.com> wrote:
> 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
> _______________________________________________
> Computer-go mailing list
> Computer-go@computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go
>
_______________________________________________
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go
