[Computer-go] 25x25 experiment

Detlef Schmicker ds2 at physik.de
Wed Apr 29 09:54:38 PDT 2015


I did not test 100k playouts, I was lower than 2k playouts in all my 
tests. No I am not sure we are stronger than Fuego, I should run on 
KGS... (possibly after the 19x19 tournament on sunday)
We won against hirabot (1d if playing with 10k playouts on KGS I think) 
in 19x19 slow bot tournament and played about even with aya and pachi in 
13x13.

The winrates may depend strongly on the opponent: On CGOS with 1k we 
have quite different winrates against aya_1k and fuego_10k or pachi_10k, 
which are about equal strong there.
I am not sure, what to learn from this. For me it means, it might be 
very difficult to compare the playing strength, as CNN has very 
different strengths

Detlef

Am 29.04.2015 um 18:06 schrieb Aja Huang:
> On Wed, Apr 29, 2015 at 4:57 PM, Aja Huang <ajahuang at google.com 
> <mailto:ajahuang at google.com>> wrote:
>
>     Interesting, but I don't really understand your experiments. Our
>     12-layer CNN alone is about KGS 2k, and when integrating it into
>     MCTS our program using very simple playouts is as strong as Fuego
>     (about KGS 1.5d to 2d). Are you sure their CNN is stronger?
>
>
> You might misunderstand the paper. In Section 6, our MCTS program 
> integrated with the 12-layer CNN, running 100k playouts per move, 
> scores 86.7%+-3.5 against the 12-layer CNN alone, which is close to 
> Fuego's win rate described in Section 5. Also, note that our MCTS 
> program's playouts are much faster than Fuego so it can be stronger 
> than Fuego when playing fixed time per move.
>
> Aja
>
>
>
> _______________________________________________
> Computer-go mailing list
> Computer-go at computer-go.org
> http://computer-go.org/mailman/listinfo/computer-go

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20150429/959bff6e/attachment.html>


More information about the Computer-go mailing list