From: Dave Dyer <[EMAIL PROTECTED]>
>Considering how monte carlo actually works, I think it's plausible
> to argue that it works best where the distance to endgame is small.
> For a 19x19 board, the playing speed may be only a factor of 4 worse,
> but the effective learning speed for an opening position might be
> exponentially worse. In other words, doing 4x as many playouts won't
> get you to the same quality of play. I'm not aware of any data about
> what the scaling exponent is, but I'll wager 1 is not the answer.
Humans tend to read out various local situations - this group dies, this one
lives, this can be
killed, this can be defended. For endgame moves, there is a method of analysis
- if white
plays first , what is the local effect on the score? If black, what then? Who
gets sente? That
information is cached, and periodically checked - did such-and-such a play
alter the status?
These strategies greatly winnow the search tree. ( I'd be tempted to
dynamically add and alter
callback patterns which would trigger appropriately. )
When it comes to opening moves, it might be that programs need to use opening
books,
joseki patterns, and rules of thumb to narrow the search process. The
evaluator, as some
have suggested, should differ in the opening; instead of playing out to the
bitter end, a
rough map of expectations should suffice. Designing such a mapping function
would
be an interesting machine learning exercise; self-play could tune the results.
A few days ago, I was playing a teaching game with a 5 dan player. At several
points, he used
a form of local null-move analysis, though he didn't call it that. If black
plays X, and white ignores
that play, black follows up with Y - with devastating results. Therefore, white
must reply to X,
unless white has an even bigger threat. Having sente, and a position slightly
altered in his favor,
black then plays Z, which kicks white in the head. But Z before X does not work
so well ... move
order often makes the difference between a very strong and a weak play.
A poster recently mentioned 19x19 games and handicap stones. This would help to
quickly separate
wheat from chaff. If program A could defeat all contending programs more than
half the time with a
two or three stone handicap, we'd take that as clear evidence of superiority.
This could spur the development
of much stronger programs. We know that top human players can give the
strongest current 19x19 programs a
9 stone handicap and win better than half the time. Future programs, evolved to
give current contenders a
large handicap and win, will be a lot closer to beating top human players.
____________________________________________________________________________________
Be a better Globetrotter. Get better travel answers from someone who knows.
Yahoo! Answers - Check it out.
http://answers.yahoo.com/dir/?link=list&sid=396545469_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
