Re: [gnugo-devel] life and death problems

[Evan Berggren Daniel]

On Sat, 23 Aug 2003, Gunnar Farneback wrote:

> Evan wrote:
> > I have created a test suite of life and death problems for gnugo from the
> > problem book "The Art of Capturing Stones" (published in the US by
> > Yutopian, book Y25, available from yutopian.com/go/).  I have entered most
> > of the first 32 problems (under the stones problems), with one test case
> > per move in the main variation.  Gnugo gets 61 out of 178 test cases
> > correct; however, some of these passes are accidental, as later moves in
> > the problem are missed.  Gnugo gets only 1 problem completely correct.
>
> Those problems are not easy and quite atypical with respect to eye
> analysis, so I'm somewhat surprised it doesn't get an even lower
> number. Maybe some of the test cases it gets right are mostly tactical
> reading? Which problem does it get completely right?
>
> On a related note I have entered problems from
> http://www.hitachi.co.jp/Sp/tsumego/igohpe.htm, so far up to number
> 244. Out of 589 owl test cases it passes 207, and out of 102 semeai
> tests it passes 51.

If you only count passes where it gets all later moves of the problem
correct too, the number is much lower.  The problem that it gets
completely correct is as follows:

----+
..X.|
O.X.|
OX..|
OOX.|
OXX.|
OXOO|
OO..|
..O.|

Black to play and live.

In general, it seems to do better when trying to live than when trying to
kill with these prolems; I think this is because the tactical analysis
does a fine job understanding how to capture something using under the
stones techniques, and therefore owl can find the lunch it needs.

Also, I think that in order to solve a lot of these problems it will need
a better understanding of cutting points in eyespaces.

Evan Daniel