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

[heikki@xxxxxx]

On Mon, Jun 04, 2001 at 02:55:04PM -0700, Daniel Bump wrote:
>
> As to how often you get a wrong score, Heikki Levanto has some
> results about that. Around April 25 he played 100 games between
> 2.7.231 and 2.6 and found 8 games where they reported different
> scores. Either 2.6 or 2.7.231 gave a wrong score, and as I recall
> it was 2.7.231 in many cases. This percentage is too high.

This was in games between two gnugo versions, which both had a reasonably
good (and similar!) idea of the ending point of the game. I believe things
could be much worse if you just have two random neural nets playing, since
they can "decide" to stop the game at any point. It is next to impossible to
calculate a valid score on a random position like that, if you play with
Japanese rules.

For the NN experiment, I would warmly recommend using chinese rules, and
requiring all prisoners to be captured before declaring the game over - that
is, all stones left on the board are to be counted alive. This sounds like a
big change in the rules, but actually it can be shown that the optimal play
is the same under both rulesets for a long way into the game, only in the
endgame can there be some slight differences...

- Heikki



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