Re: [gnugo-devel] Big match with breakin code

[Arend Bayer]

On Fri, 4 Jul 2003, Inge Wallin wrote:

> About a week ago I ran a big match with 1000 games on Arends account
> on "my" Linux cluster.  The objective was to find out about how much
> stronger the breakin code made Gnu go.  It would be interesting to
> hear the result of this match.
>
> Arend?

First the good news -- no crashes :-) (I.e. as far as stability is
concerned, there isn't much point in running more gtp matches.)

The results are: of 500 matches with the break-in code playing white,
it won 254, and 247 out of 500 playing black.

The time penalty was a lot higher than in my tests, s.th. over 20%. This
strongly suggests disabling the break-in code.
I've also replayed a few games with the respective other version, and
got typically 5 differing moves. This certainly is not worth the time
costs.
Of course, one might wonder whether the break-in code is less important in
gnugo-vs-gnugo matches than usual. I am unconvinced of this.

I have observed one statistical oddity, namely an abnormally big number
of _white_ wins by resignation (all gnugo's played with
--resign-allowed). I want to look into this further. (In the few
cases I actually looked at, the resignation was absolutely reasonable.)

Arend