Re: [gnugo-devel] GNU Go tried to cheat jimm on NNGS

[Gunnar Farneback]

Dan wrote:
> Here is a game in which jimm [7k]* played black against
> GNU Go. At the end of the game, GNU Go removed the
> string at P12. So jimm undid his final pass and
> connected at N5.
> 
> The final move in the game record is this connection.
> So the position at move 230 was the original final
> position.
> 
> Running gnugo -L230 -T on this position doesn't show
> anything strange so I don't know what happened.

The owl display shows the A18 dragon in green, which is incorrect
since that dragon is dead. This is unfortunate, but not beyond the
capabilities of the aftermath scoring to resolve. By playing on until
all dead stones are captured (or at least unconditionally settled), it
correctly finds that the A18 dragon is indeed dead. But the problem
wasn't with this dragon but with the P12 string (and actually also
M7). So what happened?

To find this out you need to either issue the "final_score" GTP
command after loading at move 230 or remove the last black move from
the game record and run with the "--score aftermath" option. In both
cases it's necessary to add "-t" to see what happens. This is

1. Black finds that A9 saves the A13 string (which is not inessential
since A18 is thought to be alive) and captures A10 tactically with ko.
2. White similarly finds that these strings would be saved by
recapturing at A8 but since this would be an illegal ko capture it
instead generates a ko threat at M5.
3. Black incorrectly chooses to kill (as it thinks) the entire A18
dragon with B11 instead of answering the white ko threat.
4. White follows up his ko threat with M6.
[...]
When the dust settles, the P12 and M7 strings have been captured and
are thus assumed to have been dead at the game end position, just like
the A18 and A10 strings. 

There are at least two ways to solve this.

1. The best solution is of course to make sure that the A18 dragon is
recognized as dead in the first place. The problem here is that the
optics code isn't sophisticated enough to understand the significance
of the cutting stone at C10, thinking that the eyeshape is .XXX. and
therefore at least seki. (Some exceptions from that rule in the corner
are known though.)

2. Stop white from making a ko threat in step 2 above. This may have
some potential problems.

/Gunnar