Re: computer-go: Test collections?

[Arend Bayer <arend.bayer@xxxxxx>]

On Sat, 9 Aug 2003, Darren Cook wrote:

> > The GNU Go sources contain regression tests in GTP (Go Text Protocol)
> > form, with correct answers marked in the comments of the GTP files.
>
> Thanks - I'd looked at these and thought it was a pure text format, and so
> hard to edit. Now I see you specify an sgf file and move in that file, so
> really it is just a separation of the test position from the test and result.

We found .gtp files easier to maintain.

> I guess it should be possible to have a conversion utility that puts the gtp
> commands in sgf file comments, and vice versa. (does gnugo already have
> something like this?)

There is interface/gtp_examples/sgf2tst for one way.

> > However, this test suite is developed and maintained only with GNU Go in
> > mind.
>
> Yes - it seems many of the choices return gnu go function return values.

Not if you restrict to reg_genmove tests. (For the other tests see
below.)

> I am looking for a test set of life/death and tactical reading problems,
> that mark chains and say at least "alive/dead/unsettled/ko", and preferably
> with more description of what type of ko. What is the closest to that in GTP?

The problem is that the semantics of such commands are hard to define
uniformly across engines. What is chain that is capturable by tactical
reading? E.g. one engine might decide that strings with 4 liberties
are considered tactically stable once the reading is beyond depth 4, the
other beyond depth 8. Then it doesn't make much sense to give these two
engines identical problems.

Even for life-and-death its not so clear. GNU Go's owl_defend command
returns 0 (fail) if a dragon is locally dead but can live in a semeai.
(Semeais are resolved in a separate module.) Other engine's (or GNU Go
in 2 years from now) might do this differently.

Having said that, I think the closest are the owl_attack/defend commands
(return results mean: 0 - fail, 1 - success, others: ko).

Regards,
Arend