Re: [gnugo-devel] current regression failures

[Gunnar Farneback]

Arend wrote:
> > neurogo:16 (atari_atari?)
> > I don't understand this test case. 
> Neither do I.

And neither did Trevor in a message from September 1:
> 16 failed: Correct 'H17|K17|H15|L16', got 'L1'
> L1 seems fine.  Am I missing something?

Dan wrote in a comment:
> Particularly tests 9, 15 and 16 are clearly wrong. I think test 9
> may have been intended for move 154 at which time the W group
> on the right needs eyes and S10 is the obvious move to secure them.
> 
> It seems clear that test 16 isn't at move 275 since the wrong
> color is to move then.

Considering the list of correct answers, it seems like move 276 is the
only possible alternative. There too it looks like a wasted move. I
think that the idea behind the test may have been to reinforce the
position in order to avoid the trouble later when black had the chance
to live. Since GNU Go is already ahead by almost 50 points this may be
a good idea, but I don't think it's worth regressing over.

Actually GNU Go wants to play L16 at move 276, as an owl attack. I'm
skeptical that black could possibly defend his dragon and the owl
reading is doing a very bad job there, running into the owl node
limit. For the time being I'll just change the move number for
neurogo:16, but I think this should be added as an owl test too.

Back to Arend:
> > strategy2:73 (semeai, moyo?)
> > To me it seems like E7 is very similar to F7. Any reason not to add it
> > to the correct answers?
> I think E7 does not capture the black G11 group: W E7 - B F8 - W H7 - 
> B F7 and F6 and J7 are miai. It took me quite some time to spot this
> when I first came across this problem.

But white doesn't have to play H7. After W E7 - B F8 - W F7, G10 is
connected to the outside (B G8 - W H8 makes H7 and E9 miai) and I
don't think black can make two eyes locally.

> > trevor:381 (atari_atari)
> > This actually passed before due to a bug in the atari_atari code. Now
> > the main problem is with an incorrect reverse followup value.
> Indeed there seems to be a duplication. OTOH, I think E8 is fine and
> should be added to the correct answers.

It's clearly worse for ko threats than D7 and slightly worse than C7,
but it doesn't lose a point like C6 (the game move). I'll add it and
let Trevor remove it later if he doesn't agree.

> > trevora:120 (atari_atari)
> > Various valuation mistakes. However, the previous move at E17 isn't
> > satisfactory and shouldn't be listed as correct. Or am I missing
> > something?
> I come to yet another conclusion :-)
> I think that B8 is the correct move. W can answer E7 with E8 and then D8
> with D9. Compared to D8, this only gives up two points around D8, yet
> B8 secures B7 and A6 (not B5 due to a shortage of liberties, but B5 cannot
> be saved anyway unless I am missing s.th. again).

I think B5 can be saved after B8, but only through a huge ko. After
D8, E8, or E7, there's no defense at all for B5 and ko for B7 and A6.
It seems I agree with Arend that B8 is best. Trevor, what's your
opinion?

> Wow, incident94 is really a funny game!

It was played between GNU Go 2.3.95 (white) and someone called glea on
NNGS. One can notice that between move 68 and 116 there are no less
than 27 ko captures but not a single connection.

> > nicklas1:1106
> > What's so bad about F6? 1 point reverse sente is maybe not the biggest
> > move on the board but it's also far from the worst. This test should
> > either be removed or changed to require the largest move on the board
> > (I guess this would be Q6 or some move around T12).
> Largest is probably the sequence Q6 - R7 - R5 - Q4 - Q4 - S5
> (not Q6 - R7 - R5 - S5? - Q4!). This whole sequence is probably a bit too
> subtle for GNU Go at the moment. Still, we should try to find a way to
> value 1pt sente moves like Q6. Also, 3.3 strategical value of F6 for
> the P8 dragon is somewhat weird.

Okay. I'll change the test to require Q6.

/Gunnar

Index: regression/atari_atari.tst
===================================================================
RCS file: /cvsroot/gnugo/gnugo/regression/atari_atari.tst,v
retrieving revision 1.7
diff -u -r1.7 atari_atari.tst
--- regression/atari_atari.tst  2001/12/04 08:11:05     1.7
+++ regression/atari_atari.tst  2001/12/06 09:15:41
@@ -1,7 +1,9 @@
-# N3 is ineffective
+# N3 is ineffective. N2 and N1 differ in that the latter move gives
+# white the option to defend with a (disadvantageous) ko. L3 is also
+# effective but misses the point and wastes a good ko threat.
 loadsgf games/incident107.sgf 216
 1 combination_attack black
-#? [N1]
+#? [N2|N1|L3]
 
 loadsgf games/atari_atari01.sgf 235
 2 combination_attack black
Index: regression/neurogo.tst
===================================================================
RCS file: /cvsroot/gnugo/gnugo/regression/neurogo.tst,v
retrieving revision 1.16
diff -u -r1.16 neurogo.tst
--- regression/neurogo.tst      2001/11/30 14:42:02     1.16
+++ regression/neurogo.tst      2001/12/06 09:15:41
@@ -76,8 +76,9 @@
 15 gg_genmove white
 #? [A16|G19]
 
-# This test may be at the wrong move number.
-loadsgf games/FSF-neurogo.sgf 275
+# It doesn't look like black can live, so this would only be a
+# reinforcing move. It may not be worth regressing over.
+loadsgf games/FSF-neurogo.sgf 276
 16 gg_genmove white
 #? [H17|K17|H15|L16]
 
Index: regression/nicklas1.tst
===================================================================
RCS file: /cvsroot/gnugo/gnugo/regression/nicklas1.tst,v
retrieving revision 1.19
diff -u -r1.19 nicklas1.tst
--- regression/nicklas1.tst     2001/11/30 14:42:02     1.19
+++ regression/nicklas1.tst     2001/12/06 09:15:41
@@ -121,9 +121,12 @@
 1105 gg_genmove black
 #? [!H14]
 
+# Q6 seems largest but we may want to accept some more moves. The
+# original formulation !F6 isn't very good since a one point reverse
+# sente move isn't all that bad in the position.
 loadsgf games/nicklas/nicklas11.sgf 208
 1106 gg_genmove black
-#? [!F6]*
+#? [Q6]*
 
 #CATEGORY=WASTED_MOVE
 #DESCRIPTION=GnuGo 3.1.8 gets N2 on next move;  R3 benign ko threat loss.
Index: regression/owl.tst
===================================================================
RCS file: /cvsroot/gnugo/gnugo/regression/owl.tst,v
retrieving revision 1.26
diff -u -r1.26 owl.tst
--- regression/owl.tst  2001/11/30 14:42:02     1.26
+++ regression/owl.tst  2001/12/06 09:15:41
@@ -902,6 +902,10 @@
 253 owl_defend C2
 #? [1 B2]
 
+# See also strategy2:100.
+loadsgf games/strategy26.sgf 257
+254 owl_defend O13
+#? [1 (M16|O14|N16|N17)]
 
 ########### end of tests #####################
 
Index: regression/strategy2.tst
===================================================================
RCS file: /cvsroot/gnugo/gnugo/regression/strategy2.tst,v
retrieving revision 1.16
diff -u -r1.16 strategy2.tst
--- regression/strategy2.tst    2001/11/30 14:42:02     1.16
+++ regression/strategy2.tst    2001/12/06 09:15:41
@@ -259,6 +259,7 @@
 99 gg_genmove black
 #? [M13]
 
+# See also owl:254.
 loadsgf games/strategy26.sgf 257
 100 gg_genmove black
 #? [M16]
Index: regression/trevor.tst
===================================================================
RCS file: /cvsroot/gnugo/gnugo/regression/trevor.tst,v
retrieving revision 1.28
diff -u -r1.28 trevor.tst
--- regression/trevor.tst       2001/11/30 14:42:02     1.28
+++ regression/trevor.tst       2001/12/06 09:15:41
@@ -364,9 +364,10 @@
 #CATEGORY=READING
 #DESCRIPTION=C6 is wrong way to protect the cut.  Loses one point.
 #SEVERITY=3
+# Added E8. /gf
 loadsgf games/trevor/trevor_38.sgf 39
 381 gg_genmove black
-#? [D7|C7]
+#? [D7|C7|E8]
 
 #CATEGORY=KO_THREATS
 #DESCRIPTION=E9 is terrible for leaving Ko threats.