Re: [gnugo-devel] Death by Magic Sword

[Gunnar Farneback]

Michael wrote:
> I've been having GNU Go (3.3.18) play some games against itself, and
> found a case where a non-joseki local move got a higher score than the
> local joseki move:
> 
>   C13: 11.27 - change in territory
>   C13: 10.19 - strategic effect on D16
>   C13: 6.22 - strategic effect on E14
>   C13: 6.16 - strategic effect on D12
>   C13:   6.45 - total followup value, added 6.45 as territorial followup
>   C13: 3.22 - added due to followup (6.45) and reverse followup values (0.00)
> Move generation values C13 to 37.06
>   C14: 5.74 - change in territory
>   C14: 7.17 - strategic effect on D16
>   C14: 5.16 - strategic effect on E14
>   C14:   6.91 - total followup value, added 6.91 as territorial followup
>   C14: 3.46 - added due to followup (6.91) and reverse followup values (0.00)
>   C14:   35.00 - minimum accepted value
> Move generation values C14 to 35.00
> 
> This led to a rather large B group dying later on. W did spend a few too
> many moves killing it, though, while B was getting big fuseki points.

These things happen occasionally. The proper solution would be to
generally improve the evaluation, especially the strategic values of
moves. For a missed joseki there's also a simple solution, namely
marking the bad move as antisuji in the joseki database.

Related is this recently added item in the TODO list:

 * Write a script which plays through the joseki databases and checks
   that the engine really generates a joseki move for all positions in
   the databases. This would also be interesting to run with the
   --nojosekidb option.

This would allow us to systematically find places where antisuji moves
are needed (alternatively some more joseki lines when the proposed
move is also joseki).

> Another interesting move is W92. I think W should have played a more
> severe move such as F8, F6, G6, etc. - this could make a difference of
> 50 points or so. Of course, I am only a lowly 12kyu and suck at
> counting, so I might be wrong. ;)
> 
> As it is the move actually played (H10) was given the same score for
> "strategic effect on C8" as F8, whereas IMHO the effects are very much
> different.

Yes, this is a well-known shortcoming of the engine and one of the
major areas needing work. See for example this FIXME for the
dragon_weakness() function in value_moves.c:

 * FIXME: Important to test more exactly how effective a strategical
 *        attack or defense of a weak dragon is. This can be done by
 *        measuring escape factor and moyo size after the move and
 *        compare with the old values. Also necessary to test whether
 *        an attack or defense of a critical dragon is effective.
 *        Notice that this wouldn't exactly go into this function but
 *        rather where it's called.

> PS Is there any chat place (such as IRC) where GNU Go developers hang
> out?

Stuart wrote:
> I think what you're looking for is NNGS (http://nngs.cosmic.org/) channel 50.

Yes. The activity is somewhat sporadic but that's the place to go, in
particular since GNU Go (generally the most recent development
version) is regularly playing games there.

/Gunnar