[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] influence patch
From: |
Gunnar Farneback |
Subject: |
Re: [gnugo-devel] influence patch |
Date: |
Tue, 24 Sep 2002 18:36:01 +0200 |
User-agent: |
EMH/1.14.1 SEMI/1.14.3 (Ushinoya) FLIM/1.14.2 (Yagi-Nishiguchi) APEL/10.3 Emacs/20.7 (sparc-sun-solaris2.7) (with unibyte mode) |
I wrote:
> It seems we are currently systematically undervaluing moves actually
> taking stones off the board. This includes, but is not limited to, ko
> captures. More precisely a captured stone counts for two points (in
> most cases) when evaluating territory, but when the move takes stones
> off the board, only one point is counted.
>
> The appended patch corrects this. I haven't run the regressions yet.
The regression delta is 3 passes (strategy:10, endgame:501,
trevorc:560) and 3 failures (trevorb:270, trevord:340, nngs3:510). The
failed test cases are all acceptable.
trevorb:270 Two moves misvalued, one corrected by the patch.
trevord:340 Territorial valuation of F8 improved, inappropriate shape values
(empty triangle) cause the failure.
nngs3:510 Completely bogus valuation, inappropriate shape value (ponnuki).
> This one-liner solves reading:172 with only a marginal increase in the
> node count for reading.tst. I'd say it's well worth it.
>
> - break_chain2_moves() revised
3 passes (reading:172, nngs:500,510) and 1 failure (nicklas2:701). The
failed test case is acceptable.
I have added both patches to CVS.
/Gunnar