Re: computer-go: Computer Go Tournament Program

[Gunnar Farnebäck <gunnar@xxxxxxxxxxxxxx>]

Don Dailey wrote:
> This  sounds great but  how about   adding  a Tromp-Taylor  setting in
> gnu-go?  I have  considered patching this in myself  but would it  get
> incorportated into the distribution if I went to the trouble? 

As it happens this can also be expected in the next version, since I
already have an implementation ready.

But I have no idea why people seem to think that it should be easy to
generate moves which effectively remove all opponent dead
stones. Sure, many times it's just to play on their liberties in some
arbitrary order, but when there are big nakades around quite a bit of
care is needed. Also, a serious implementation should be able to
remove bent-four-in-the-corner shapes. These are not too hard to
statically identify as dead (or more exactly effectively worth one
eye) by eye shape pattern matching, but in order to actually remove
them from the board you must first eliminate all ko threats. The only
way I know of robustly doing this is to first play other moves which
make all (other) own living stones unconditionally (in the strongest
sense) alive and all (other) dead opponent stones unconditionally dead.

So this is what GNU Go does, or at least tries to do. There will be
two new options, --play_out_aftermath which forces the engine to play
on until as much as possible is unconditionally settled, and
--remove_all_dead_stones which plays on until all dead opponent stones
are gone. For the reasons explained above, the latter option
effectively includes the former.

> By  the way,  I do a  lot  of autotesting  between  versions of my own
> program  using TT  rules and  it's simple   and elegant.   I also have
> tested against   older versions  of gnugo  (There  is  little point in
> testing  newer versions since I  only win about 25%  of the games with
> 2.6 version)  and had to go  to a lot of  trouble  to design a scoring
> module that works  with gnu-go's habit of  ending the game whenever it
> feels like it.

Recent versions of GNU Go are reasonably good at (Japanese style)
scoring. I have tested the "--score aftermath" option, or more exactly
the GTP equivalent final_score, on Dave Dyer's test set of 623 scored
professional games. The latest results are

66% correct score
89% off by at most one point
97% off by at most three points
99% off by at most ten points

Most of the misscorings are due to mistakes in the late endgame when
dame is being filled.

/Gunnar