[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: [gnugo-devel] nasty shape values
From: |
Portela Fernand |
Subject: |
RE: [gnugo-devel] nasty shape values |
Date: |
Fri, 28 Feb 2003 20:56:20 +0100 |
Gunnar wrote:
> Please submit such tests.
Although it was not really necessary, I've been recently reminded that
reviewing patches is a large amount of work for Dan, Arend and you (thank
you all btw). Since my recent test cases submissions proved to be somewhat
imperfect, I was just holding this case in a (soon to be submitted) batch,
until I find the time to review them as carefully as I can.
> As an experiment I tried to significantly limit the effects of shape
> values to a maximum change of 0.95 points (...)
This looks a bit too drastic to me, though it's definitely interesting to
measure the impact of such a change. IMO shape values are important,
and we maybe only need to try to fix their (sometimes) nasty effects.
I just made another experiment with the appended patch (not for CVS either,
just for discussion), which solves the test case I have and plays H8
(instead of M8) in Paul's game. AFAICT, it is otherwise completely neutral
on regressions. Inge's idea was to remove all large shape values, my
experiment is a bit more general, since it limits any shape boost to a
maximum of +40%.
/nando
Index: engine/value_moves.c
===================================================================
RCS file: /cvsroot/gnugo/gnugo/engine/value_moves.c,v
retrieving revision 1.82
diff -u -r1.82 value_moves.c
--- engine/value_moves.c 13 Feb 2003 20:09:42 -0000 1.82
+++ engine/value_moves.c 28 Feb 2003 17:40:06 -0000
@@ -1084,6 +1084,8 @@
exponent += move[pos].numpos_shape - 1;
if (move[pos].numneg_shape > 1)
exponent -= move[pos].numneg_shape - 1;
+ if (exponent > 7)
+ exponent = 7;
return pow(1.05, exponent);
}