2012/12/11 terry mcintyre <[email protected]>

>
>
>
> > From: Darren Cook <[email protected]>
> >
> >>  How much [effort] to determine whether there are multiple peaks?
>  <snippage>
>
> >>  Now the tough question: How can this information be used to improve
> move
> > selection?
> >
> > One approach, not at all sophisticated, is better time management: spend
> > less time on normal distributions, more time when the distribution is
> > messy.  (But I wonder if more time will just make the two peaks stand
> > out more?)
>
>
> If "more time" means "more of the same simulations", it might simply lead
> to filling in the details; the peaks exist (hypothetically) because the
> playouts are doing something non-optimal.
>
> I wonder if post-game review can focus on such positions and learn
> something which would optimize the playouts. If such meta-analysis is quick
> enough, it could be done during the game itself.
>

Live-death is a problem of finding a correct sequence of playouts. The
search tree may solve the live-death problem at the top side of tree.
However, the sequence of moves that solves the live-death problem may leads
to a losing result. Then, the search tree will tend to select other moves,
and leave the live-death problem unsolved at the bottom side of tree, which
"thinks" I have opportunities that you may make mistake at the
endgame. This results in many peaks in the distribution.

To improve the algorithms on live-death problems, a self-learning approach
that can learn the sequence of playouts from the search tree is probably
required. Even if the live-death problem is left to the bottom side of
search tree, the playouts should still solve the live-death problem by a
learned sequence of playouts. And the learned sequence of playouts may come
from the search tree itself somewhere. It is still an open problem, and
might be a potential future works.

-- 
Chin-Chang Yang
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to