Re: computer-go: Perl Module for next move.

[heikki@xxxxxx]

On Tue, Jun 05, 2001 at 09:52:21PM +0200, Mark Boon wrote:

> I'm very skeptical as well. But not so much because of tactical
> considerations. In general I think in order for a neural network to be able
> to learn Go this way, it will have to learn new technigues and concepts by
> itself. Tactical technigues are just part of that. 

Yes, if the thing would have to learn all from first principles, as seems to
be the case here.

I have been speculating (but never even experimented, I admit) about using a
neural net for global evaluation, not unlike the current prospect, but
feeding it with much more information, like numbers of eyes, liberties,
connections, etc. Some of it for every point on the board, some of it
independently of the board image (number and size of eyeless, one-eyed, and
two-eyed groups, for example). Given enough of relevant information,
estimating the score (or winning probability, which I find more relevant)
should not be all too hard a problem... 

With such an evaluator we should have an estimate for each position, and
normal search algorithms could be useful. Of course, recalculating the
tactical info for each possible move would be quite slow in itself, but with
the speed computers will have before I get this done, it shouldn't matter
too much...

- H

-- 
Heikki Levanto  LSD - Levanto Software Development   <heikki@xxxxxxxxxxxxxxxxx>