[gnugo-devel] readconnect patch

[Gunnar Farnebäck]

This patch makes a small change in the readconnect code. Regression
results:

nngs:1280       PASS D13 [D13]
gunnar:54       pass (fails in cvs)
2 PASS (1 PASS, 1 pass)
Total nodes: 1643843760 3037500 12966372 (-0.015% +0.0068% -0.2%)

Both passes are very good and solves mysterious-looking mistakes in
the breakin analysis.

- check_self_atari() in readconnect.c revised

/Gunnar

Index: engine/readconnect.c
===================================================================
RCS file: /cvsroot/gnugo/gnugo/engine/readconnect.c,v
retrieving revision 1.83
diff -u -r1.83 readconnect.c
--- engine/readconnect.c        24 Aug 2004 14:39:53 -0000      1.83
+++ engine/readconnect.c        22 Sep 2004 15:39:55 -0000
@@ -4259,11 +4259,25 @@
    * I leave the code in for a while. /gf
    *
    * Code reactivated, see nando:31. /nn
+   *
+   * Added requirement that no additional stones are sacrificed in the
+   * self atari. /gf
+   *
+   * FIXME: Add a function in board.c to check how big the string
+   *        becomes when playing a move and use for the isolated stone
+   *        test below.
    */
   if (approxlib(pos, color_to_move, 1, &lib) >= 1
       && approxlib(lib, OTHER_COLOR(color_to_move), 3, NULL) <= 2
-      && ladder_capturable(lib, OTHER_COLOR(color_to_move)))
-    return 1;
+      && ladder_capturable(lib, OTHER_COLOR(color_to_move))) {
+    int k;
+    for (k = 0; k < 4; k++) {
+      if (board[pos + delta[k]] == color_to_move)
+       break;
+    }
+    if (k == 4)
+      return 1;
+  }
 #endif
 
   return 0;