[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: [gnugo-devel] address@hidden: computer-go: 13x13 C GoTournament 2002
RE: [gnugo-devel] address@hidden: computer-go: 13x13 C GoTournament 2002-03 (12)]
Sun, 2 Feb 2003 07:59:02 +0100
> Increasing the level affects a lot of depth parameters. It would be
> interesting to know which of those pay off best.
I happened to take a look at this portion of code today, and I tried to
figure a bit which ones. I first noticed that all but one of these values
increase linearly with the level : at level 15, owl_node_limit gets raised
to around 7500 (!). I doubt this would pay off, though.
I experimented a bit and discovered something interesting, totally by
inadvertance though. I ran the regression suite nando.tst with
--aa-depth 11 (the value it would get at level 15) and was surprised to
find 3 new passes (great!). Unfortunately, all 3 (number 2, 6 and 26) are
pure owl reading tests (doh!). I'm happy there was nobody in the office to
see my face at that moment...
Of course, it's a bug (patch below). A missing break statement, which
copied the parameter in the owl_distrust_depth global variable.
I have often noticed (I'm not alone I believe) owl reading mistakes due to
the too optimistic "getting deep, looks lively" judgement. I think the
engine would problably benefit from raising this depth parameter.
- fixes missing break statements in main.c
RCS file: /cvsroot/gnugo/gnugo/interface/main.c,v
retrieving revision 1.63
diff -u -r1.63 main.c
--- interface/main.c 27 Jan 2003 18:21:18 -0000 1.63
+++ interface/main.c 2 Feb 2003 06:50:06 -0000
@@ -507,6 +507,7 @@
metamachine = 1;
chinese_rules = 0;
@@ -693,6 +694,7 @@
mandated_aa_depth = atoi(gg_optarg);
mandated_owl_distrust_depth = atoi(gg_optarg);
@@ -1060,7 +1062,7 @@
rotate(bi, bj, &bi, &bj, board_size, orientation);
decide_connection(POS(ai, aj), POS(bi, bj));
- RE: [gnugo-devel] address@hidden: computer-go: 13x13 C GoTournament 2002-03 (12)],
Portela Fernand <=