Re: computer-go: life and death

[David Mechner <mechner@xxxxxxxxxxx>]

David Fotland wrote:
> Without some kind of automatic knowledge acquisition, it seems that every
> new problem will need some new knowledge, and the program can never get
> very strong at life and death.

This was precisely the reason we trained on odd numbered problems and tested
on even; to look at generalization. The numbers Tim reported in his thesis
shouldn't be interpreted as any kind of snapshot or limit on the ability of
the program. The only point was to say something like "after being 'trained'
on 100 problems, the program could solve X% of new problems of the same
difficulty".

The key question is whether our algorithm and knowledge representation
scheme will support significant generalization, so that as knowledge is
added, the program will approach some reasonably high level of general life
and death expertise. The results we had were very encouraging, in that
respect.
To get a better feeling for where that upper limit would be would require a
lot
more time spent actually entering patterns and running tests.

-David Mechner