[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [gnugo-devel] testing gnugo

From: Gunnar Farneback
Subject: Re: [gnugo-devel] testing gnugo
Date: Thu, 22 Aug 2002 21:25:06 +0200
User-agent: EMH/1.14.1 SEMI/1.14.3 (Ushinoya) FLIM/1.14.2 (Yagi-Nishiguchi) APEL/10.3 Emacs/20.7 (sparc-sun-solaris2.7) (with unibyte mode)

John wrote:
> It appears that gnugo is mostly tested on 19x19 boards.

Yes, but not exclusively. Of the game records in the regression
directory about 70% are 19x19, 25% 9x9, and 5% 13x13.

> I wonder if 9x9 games are not sufficient to expose most of the
> weaknesses of gnugo.

Most but not all. In particular the large scale fuseki, which is
currently a serious weak spot, isn't covered. Problems related to life
and death, semeai, and interaction between weak groups certainly are
common also on small boards.

> Certainly in play against humans, gnugo can blunder in 9x9 at least
> as badly as in 19x19. It's funny that in 4 handicap games, gnugo
> playing white will pass after only a few moves, deciding that it
> can't make a living group with white,

I wouldn't exactly say it's wrong about that, but it wouldn't hurt to
add code to force it to struggle for a while longer either.

> while gnugo playing black loses easily. I attach the 2 games
> I played against gnugo 3.2

I don't think most people find it all that easy. :-)

> I think making gnugo strong enough never to lose 4 handicap 9x9 games
> would be an exciting goal to work towards...


One thing you can do if you want to help is to create test cases for
critical mistakes by GNU Go. For example, in the first game it seems
like move 26 at B3 instead of D3 would secure life and make the game
very difficult for white. Expressed in our regression test format this
would be

loadsgf tromp1.sgf 26
1 gg_genmove black
#? [B3]

(This format is documented in more detail in the file
doc/regression.texi or in the regression chapter of the formatted

Of course, if you want to jump into the code and also find solutions
to the mistakes you're most welcome to do so.


reply via email to

[Prev in Thread] Current Thread [Next in Thread]