Petr Baudis wrote:

> I don't quite understand. I think most programs do not
> disallow eye-filling in the tree stage, only in the MC
> simulation stage. So your simulations will get misevaluated,
> but given enough time to reach final position, tree will
> always consider evaluating the eye-filling move too.

Of course. I was thinking "upwards" from the end towards
the root as in Álvaro's induction reasoning and thought that
if the nodes are never evaluated correctly by the simulations,
they will never propagate correct info. But as you state, if
you grow the tree to the infinite, why play simulations at all?
So infinite MCTS does not need simulations and gets the result
correct. Now, why still call it MC if the stochastic part
is unnecessary?

Hideki Kato wrote:

> If you are talking about the article Subject "Re: 19x19 Study.
> Nakade is difficult", Message-Id:<47A2240E.2090508 at dybot.com>,
> some strong programs already have managed that, though I don't
> know the detail.

Hi Hideki. If I remember correctly, the Nakade problem was about
playing or not playing self-atari moves. If done, the program
"understood" nakade but was blind to seki detection, if not,
the nakade problem emerged. Please, anyone correct me if I am wrong
on this, since I think it is a very important aspect of MCTS. Also
feel free to describe what you do to decide when self-atari has to
be played and when not. Finding the correct policy about self atari
solves the nakade problem while still "understanding" seki.

AFAIK nobody has experimented with filling own eyes as the
position was totally artificial and I don't think a program can be
forced into that mistake. That was just hypothetical arguing.


Jacques.

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

Reply via email to