[gnugo-devel] move valuation patch

[Gunnar Farnebäck]

This patch removes the CONNECT_MOVE move reason from the ones
considered as automatically safe with respect to territorial
evaluation. The regression result is a single PASS for the targeted
testcase gunnar:72 but also a substantial reduction of 1.9% reading
nodes and 1.5% connection nodes.

- do not trust connection moves to be safe in
  estimate_territorial_value() 

/Gunnar

Index: engine/value_moves.c
===================================================================
RCS file: /cvsroot/gnugo/gnugo/engine/value_moves.c,v
retrieving revision 1.149
diff -u -r1.149 value_moves.c
--- engine/value_moves.c        12 Jun 2005 09:34:14 -0000      1.149
+++ engine/value_moves.c        16 Jun 2005 11:21:01 -0000
@@ -2208,7 +2208,6 @@
       /* This move reason is valued as a strategical value. */
       break;
       
-    case CONNECT_MOVE:
     case CUT_MOVE:
     case EXPAND_MOYO_MOVE:
     case EXPAND_TERRITORY_MOVE:
@@ -2216,6 +2215,15 @@
       does_block = 1;
       break;
 
+    case CONNECT_MOVE:
+      /* This used to always set does_block=1, but there is no
+       * guarantee that a connection move is strategically safe. See
+       * for example gunnar:72.
+       */
+      if (move[pos].move_safety)
+       does_block = 1;
+      break;
+      
     case STRATEGIC_ATTACK_MOVE:
     case STRATEGIC_DEFEND_MOVE:
       /* Do not trust these when we are scoring. Maybe we shouldn't