On Wed, Oct 14, 2009 at 05:12:58PM +0200, Heikki Levanto wrote:
> On Wed, Oct 14, 2009 at 03:34:59PM +0300, Petri Pitkanen wrote:
> > Neural network tend to work well in those cases where evaluation function is
> > smooth, like backgammon. Even inbackgammon neural networks do give good
> > results if situation has possibility of sudden equity changes like deep
> > backgames and deep anchor games. Top backgammon programs 3-ply search on top
> > neural network to handle these problems.
> >
> > I do not know wher neural nets would fit well, perhaps finding invasion
> > spots?
>
> I have been speculating about a NN evaluation function for go, feeding it a
> lot of preprocessed information about the position, like number of strings
> with 1,2,3,4, or more liberties, number of stones in same, number of separate
> groups, number of "obviously" dead stones, strings, and groups, number of
> points "clearly" controlled by each player, etc, etc. This should be possible
> to train from existing games where we know the result (in the beginning it is
> 50-50, in the end one or the other has won 100-0. Assume some simple function
> in between).
Interesting idea, but I think MCTS programs already handle quite fine
the part where there is some simple function in between. Most of the
problematic games seem to be when MCTS is chasing and trying to kill a
single group, and the question is whether it succeeds; IOW, I think the
problem is tactical rather than strategical in MCTS (I have few of my
own ideas to try to deal with that, non-related to NN ;-).
I like Magnus' ideas, though the 3x3 problem seems to be a bit of a
special case of what Remi did, and I don't really have an idea how to
well train the "in-UCT NN".
Another idea I received was actually not related to computers _playing_
Go, but maybe that's all the more interesting in current times - a NN
that could identify the player's "playing style" by looking at few of
her games and deciding which well-known player they look like (e.g.
Takemiya, Cho Chikun, Lee Sedol, TheCaptain ;-). Of course it is quite
an interesting problem what set of features to choose...
Thank you all for your ideas so far!
--
Petr "Pasky" Baudis
A lot of people have my books on their bookshelves.
That's the problem, they need to read them. -- Don Knuth
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
