Re: [computer-go] move ordering for gnugo

["Jeffrey Rainy" <jrainy@xxxxxxxxxxxx>]

Hi,

> I think the more relevant question is why anyone would think it could
> work.

I'm am absolutely completly positively aware that such an approach will not
produce a strong go playing program. However, at this point, I'm interested
in finding a statistical model that often give a high value to good moves
and often gives low value to bad moves. (A lot) More will be need before it
can play.

Right now, the best I have is a predictor which place the good human move,
in average, 24th. But with a still high standard deviation of 47. What kind
of value do you think would qualify as "working" ?

> I was certainly hoping for some kind of emergent behavior, where
> little bits of plausible data would reinforce one another, leading to
> a stronger evaluator.   No doubt there are domains where this works,
> but I didn't see any here.

Indeed, in go, something more clever is needed. I believe that the
architecture I choose forces some go-specific knowledge in the evaluation
and that this will help a little. Should it work, I'll be happy to document
it, and hopefully publish something.

> Mostly, hill climbing with some random perturbation and hand tuning;
> comparing the current set of parameters to the previous.  It's very
> easy to get the illusion of progress unless you train on one set of
> data, then evaluate your progress on different data set.

Agreed. I use 150000 games to train, and 150000 others to validate. So far,
I've only used about 100 to validate. In my case it's simply stochastic
gradient descent. I'm using games of players with various strength. At a
later stage I might tune with only expert games, but I think learning is
possible with average games.

Thanks.
Jeffrey.

P.S.: All this input is appreciated, even the negative one. I did not know
of Dave Dyer's draft paper before, and some of it helped a lot.


_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go