[Top][All Lists]

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

Re: [gnugo-devel] Fwd: [computer-go] April KGS on-line computer Go tourn

From: Evan Daniel
Subject: Re: [gnugo-devel] Fwd: [computer-go] April KGS on-line computer Go tournament
Date: Wed, 3 May 2006 15:50:16 -0400

On 5/3/06, David G Doshay <address@hidden> wrote:
To be specific, SlugGo_l3 does 12 way branching at the top level, and
then linear lookahead 3 more levels. SlugGo_88 does 8 way branching
at each of the top 2 levels (64 CPUs doing lookahead) with the same
maximum depth, and does slightly worse against almost all strong
opponents, but most notable is the fact that the advantage l3 shows
against GNU Go is completely erased. While this at first confused us,
we came to realize that this is because with second-level branching
SlugGo evaluated the mistakes that GNU Go would make as unlikely to
happen because other choices were considered to be better, whereas
with l3 GNU Go's mistakes were taken as the only likely response and
thus were accepted and taken advantage of. It is not clear why 88
played slightly weaker against other programs, where we had expected
that the additional move choices might give us a better evaluation of
a number of possible responses. But it is clear that the primary
reason that the cgos ELO score of l3 is higher than that of 88 is
that l3 takes proper advantage of GNU Go mistakes, while 88 out-
thinks itself and rates those mistakes as unlikely. SlugGo_88 uses
the same evaluation function for choosing the first level 8 choices
as for the second level 8, and it is not clear to us why an
evaluation function that shows some improvement when applied at the
top level would be weaker against non-GNU Go opponents when applied
to their possible responses. This is something we will investigate

I have observed this same effect in metamachine games with GoFigure --
even numbers of branched plies do worse than odd numbers.  I would
expect something like 8-4-2 to play better than 8-8.  I think the
problem basically arises from which side you "force" to make mistakes
that GNU Go thinks are good moves first.  With 2 branching plies,
SlugGo is the player making mistakes first.

I'm sorry I don't have a more rigorous description for you, I haven't
conducted any formal investigations.


reply via email to

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