[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[gnugo-devel] Rép : experiments with "cosmic" patch
From: |
Stéphane Nicolet |
Subject: |
[gnugo-devel] Rép : experiments with "cosmic" patch |
Date: |
Tue, 10 Jun 2003 18:54:59 +0200 |
> Hi Stephane,
>
> Btw, did you make further experiments with your "cosmic" patch?
> As I wrote most of the current territory evaluation code, I am of
> course curious about any successes you had with that.
> Arend
Hi Aren, hi all,
I have indeed tried various influence settings during the last month,
but as you may guess, the experiments are slow since I haven't figured
a better way to measure my "progress" in the evaluation function than
tomake long twogtp matches against non-patched versions...
Btw, have you tried some semi-automatic mean to estimate quality of
evaluation, like for instance using some measure of correlation
betweenGNUGO moves and professional moves in a fixed set of games ?
Anyway, here is what worked best for my "cosmic" patch :
1) I try to find some influence values that gives an exciting style
of play, principaly by trial and errors and visual inspection (ie
looking at one or two games against normal GnuGo, and choosing the
styles I like best): the resulting evaluation function are often a
little bit "unreasonable", in the sense that they play a center-
oriented fuseki, but don't know how to make it work later (the
usual problem... :-) )
This usually produces a version that is not really stronger that
the default GnuGo (it wins typically 50%-55% of the games), but
gives fighting games (scores like W+70.0 or B+70.0 are not rare)
2) So I tried to tame it down by making it use the cosmic values only
every other moves... (that's the reason of the ((movenum / 2) % 2)
test in the patch below).
And, to my great surprise, it seems to have worked somehow !
In a sense, alternating between the "slightly unreasonable" and
the "slightly conservative" evaluation functions uses the strong
points of both :-)
Anyway, below are two patches against the current CVS (for influence.h
and influence.h) that I made from my current version. It won 54-26
yesterday in a 80 games match (40 with Black, 40 with White) against
the current CVS version (both programs using --owl-threats, but not
--experimental-breaking)
I would of course be interested if you found the patches worth
testing, and honored if they made it to the CVS.
Usage : * set COSMIC_GO to 0 or 1 in influence.h
* even when COSMIC_GO is set to 1, it is possible to
get the default (3.3.20) influence values by setting
cosmic_importance to 0.0 in init_influence(), so it
should be very easy to introduce a --experimental switch
to test the changes, and clean the code by removing
all the #else parts in the
#if COSMIC_GNUGO [..] #else [..] #end
compilation conditionnals.
See you,
Stephane.
PS : Should the patch be accepted, I hereby assign the copyright of
the changes to the Free Sofware Fondation.
inf.h.patch
Description: application/text
inf.c.patch
Description: Binary data
[Prev in Thread] |
Current Thread |
[Next in Thread] |
- [gnugo-devel] Rép : experiments with "cosmic" patch,
Stéphane Nicolet <=