Re: [gnugo-devel] engine/influence.c

[Arend Bayer]

On Sat, 21 Sep 2002, Gunnar Farneback wrote:

> Today we could probably try a more aggressive cutoff.

> I reviewed the patterns using the xterri and oterri helpers and I'm
> fairly certain these could equally well be replaced by checking
> whether the territorial value is larger than 0.95 or something like
> that.
>
> The only potential problem I can see is what happens at vertices where
> both colors have small influence and an early cutoff would cause only
> one of them down to zero. This may or may not be a problem.

The territory valuation should be robust against this (due to the
min_inf_for_territory policy). I have sligthly revised whose_territory()
and whose_moyo_restricted() to take this into account. I get three FAILs
(I am comparing CVS+loop bug fixed with CVS+loop bug fixed+aggressive
cutoff). Overall I think it is OK:

nngs2:
10 unexpected FAIL: Correct '!H6', got 'H6'
H6 is bad yes, but H7 (played by CVS) is even worse.
nngs3:15:
400 unexpected FAIL: Correct 'N13', got 'Q15'
This is caused by a change in the valuation by 0.02, i.e. accidental.
safety.tst:
1 unexpected FAIL: Correct 'R6|S2', got 'O14'
For this we can indeed make the change responsible. O14 is a bit
overvalued, partly due to a higher follow-up value. I think the current
code is overvalueing potential center territory a bit, and this should
be fixed, but preferably in a separate patch.

Runing safety.tst, I now get:
 time   seconds   seconds    calls  ms/call  ms/call  name
  7.76      5.23     5.23    12578     0.42     0.45 compute_primary_domains
  6.51      9.62     4.39  3496496     0.00     0.00  scan_for_patterns
  5.06     13.03     3.41    23255     0.15     0.15 accumulate_influence

I.e. a considerable speed improvement. Maybe one could make the cutoff
level dependent.

Arend

Patch is on top of unrolling_3_9.1a.


diff -u engine/influence.c engine/influence.c
--- engine/influence.c  20 Sep 2002 21:17:56 -0000
+++ engine/influence.c  23 Sep 2002 13:10:26 -0000
@@ -129,13 +129,8 @@
          int a = (arg_di)*(i-m) + (arg_dj)*(j-n); \
          contribution *= (a*a) * b; /* contribution *= cos(phi) */ \
        } \
-       if (contribution <= INFLUENCE_CUTOFF) { \
-          if (permeability < 1.0 || q->w[arg_i][arg_j] != 0.0) \
-           continue; \
-          else { \
-            contribution = 1.01 * INFLUENCE_CUTOFF; \
-         } \
-       } \
+       if (contribution <= INFLUENCE_CUTOFF) \
+         continue; \
        if (q->w[arg_i][arg_j] == 0.0) { \
          q->queuei[queue_end] = (arg_i); \
          q->queuej[queue_end] = (arg_j); \
@@ -1096,8 +1091,7 @@
        accumulate_influence(q, i, j, BLACK);
     }

-  if (q->is_territorial_influence)
-    value_territory(q, m, n, color);
+  value_territory(q, m, n, color);
   segment_influence(q);

   if (((q == &initial_influence || q == &initial_opposite_influence)
@@ -1123,13 +1117,14 @@
 {
   float bi = q->black_influence[m][n];
   float wi = q->white_influence[m][n];
+  float terr = q->territory_value[m][n];

   ASSERT_ON_BOARD2(m, n);

-  if (bi > 0.0 && wi == 0.0)
+  if (bi > 0.0 && wi == 0.0 && terr < -0.95)
     return BLACK;

-  if (wi > 0.0 && bi == 0.0)
+  if (wi > 0.0 && bi == 0.0 && terr > 0.95)
     return WHITE;

   return EMPTY;
@@ -1184,9 +1179,9 @@
   /* default */
   if (territory_color != EMPTY)
     color = territory_color;
-  else if ((bi > 10.0 * wi && bi > 10.0 && wi < 10.0) || bi > 25.0 * wi)
+  else if (bi > 10.0 * wi && bi > 10.0 && wi < 10.0)
     color = BLACK;
-  else if ((wi > 10.0 * bi && wi > 10.0 && bi < 10.0) || wi > 25.0 * bi)
+  else if (wi > 10.0 * bi && wi > 10.0 && bi < 10.0)
     color = WHITE;
   else
     color = EMPTY;