Re: [gnugo-devel] Pattern VA49

[Gunnar Farnebäck]

Emanuele Cisbani wrote:
> I'm discovering how to "debugging" at patterns-level.
>
> <shell>
> $ gnugo.exe --trace -l cisba01.sgf 2>&1 | tail -n 13
>
> Top moves:
> 1. H5  2.44
> 2. J10 2.25
> 3. A1  0.71
> 4. P5  0.19
> 5. Q5  0.19
> 6. F9  0.10
> 7. N7  0.10
> Checking safety of a white move at H5
> Move generation likes H5 with value 2.44
> genmove() recommends H5 with value 2.44
> white (O) move H5
>
> </shell>
>
> Move H5 is clearly wrong.

Indeed.

I'm adding some context for those who haven't looked at the sgf file:

 8 X O O . O O X . X X X X X O X X X O . 8
 7 X X O O O X . X X O O X . O O X X O . 7
 6 . . X O X . X O O O X O O O O O X O . 6
 5 . X . X . X O . O X X X O X . . X O O 5
 4 . . . X X X X O . O X X X X O O O O X 4
 3 X X X O X O O . . O O O X X X O X X X 3
 2 O X O O O . . . . . . O O X . X X . . 2
 1 . O O . . . . . . . . O X X . . . . . 1
   A B C D E F G H J K L M N O P Q R S T

> <shell>
> $ gnugo.exe --trace -l cisba01.sgf --decide-owl H4
> finished examine_position
> owl_attack H4
> Pattern VA49 found at H5 with value 35
> Variation 13: ALIVE (2 or more secure eyes)
> H4 cannot be attacked (14 variations)
> owl_defend H4
> Pattern VA49 found at H5 with value 35
> Variation 13: ALIVE (2 or more secure eyes)
> H4 is alive as it stands
> </shell>
>
> Can I conclude that pattern VA49 has a bug?

Not at all. As it stands the H4 dragon is perfectly alive and safe, 
which is exactly what the owl reading says. It's questionable why VA49 
is shown in these traces but in this case it's involved in the 
calculation of the number of eyes for the dragon (where it does a 
dubious job, but that's an algorithm weakness, not a pattern mistake).

> It's first time I try to debug a pattern.
> Someone can help me?

The real problem is the line "Checking safety of a white move at H5" 
above. That is where it should be detected that playing H5 is a blunder.

I'm afraid it's difficult to find out what's going wrong there without 
single-stepping the code, starting with the call to blunder_size() in 
engine/utils.c.

My guess is that it's not detected as a tactical blunder because H5 can 
reach a sufficient number of liberties (5) and it's not detected as an 
atari_atari blunder because after W H5 - B H3 - W J4 - B G2 - W F2 black 
needs to play J2 which is not found as a tactical attack move on H5. 
Finally it's not detected as an owl blunder because the connectivity 
would need to be reconsidered. If my guess is correct the best fix is to 
recompute dragons in detect_owl_blunder(). I believe we have sufficient 
infrastructure for this, but it would require quite a bit of familiarity 
with the code to implement it.

/Gunnar