Re: [gnugo-devel] endgame tuning

[Gunnar Farnebäck]

I wrote:
> The most straightforward solution to this problem is to add a
> non-territory pattern, saying that white can't get territory on B5 if
> black moves first. 
> 
> Pattern Nonterritory56
> [...]

Appended is a patch with this pattern plus another pattern to solve
endgame:910.

Regression results:
endgame:910     PASS T2 [T2]
endgame:920     PASS C4 [C4]
nngs2:240       FAIL B14 [F17]
nngs4:560       PASS B13 [B13]
gunnar:34       PASS M5 [M4|M5]
olympiad2004:2  PASS G14 [G14]
olympiad2004:105 PASS T17 [O17|N17|T17]
6 PASS
1 FAIL
Total nodes: 1644306544 3042323 12931177 (+0.014% +0.17% -0.47%)

Breakage analysis:
endgame:910        Good. White T2 gets one more point of territory at R2.
endgame:920        Properly solved. C4 now worth 1.06 points.
nngs2:240          Accidental failure. E16 no longer gets territorial value
                   in determination of the safety of the F16 dragon.
nngs4:560          Somewhat lucky.
gunnar:34          Properly solved (except for a reverse followup
                   still missing). Similar to endgame:920.
olympiad2004:2     Change for good, but pass mostly accidental.
olympiad2004:105   Change for good, but pass accidental.

- influence tuning

/Gunnar

Index: patterns/barriers.db
===================================================================
RCS file: /cvsroot/gnugo/gnugo/patterns/barriers.db,v
retrieving revision 1.66
diff -u -r1.66 barriers.db
--- patterns/barriers.db        13 Sep 2004 10:45:11 -0000      1.66
+++ patterns/barriers.db        22 Sep 2004 15:39:55 -0000
@@ -3415,4 +3415,44 @@
 >non_xterritory(b);
 
 
+Pattern Nonterritory56
+# gf New pattern. (3.7.1)
+# See endgame:920.
+
+...
+X.O
+?O.
+
+:8,t
+
+eac
+XbO
+?Od
+
+;oplay_attack(a,b,c,d,d)
+
+>non_xterritory(e);
+
+
+Pattern Nonterritory57
+# gf New pattern. (3.7.1)
+# See endgame:910.
+
+|..X
+|.XO
+|...
++---
+
+:8,t
+
+|.eX
+|bXO
+|dac
++---
+
+;safe_omove(a) && oplay_attack(a,b,c,d,b)
+
+>non_xterritory(e);
+
+
 # END OF FILE