RE: computer-go: abstract info and neural nets

[Petersen Kjeld-WFKP1396 <Kjeld.Petersen@xxxxxxxxxxxx>]

I think this is very correct, that humans learns the very smallest structures first  !
But how to implement this in a computer ?

Making a program that, when it finds a particular pattern, extend the view, to
see if the surrounding of this pattern has any meaning here ?

I came up with an idea, to make a database an register all played 4x4 patterns,
and count what was the next move from this pattern.

ex: .=0=empty, 1=black, 2=white, 3=edge
pattern hitcount   
....    0  0  0  0
1...    0 478 0  0 (478 could be anything, but just as an example here)
121. =  0  0  0  0
.1..    0  0  0  0
(Note it maybe that a actual hitcount will find different numbers)
The hitcount of a move only reduces the possible moves.

If the total count of a pattern becomes very large (ex. hitcount=10000 times)
then the pattern in the database is added the possibility to check the surrounding of the pattern.
                                         xxxx
....              ....x          x....   ....     ....
.111              .111x          x.111   .111     .111
122. found, check 122.x but also x122. , 122. and 122.
.111              .111x          x.111   .111     .111
                                                  xxxx

this database will ofcource grow huge, very huge, but what to do instead ?



-----Original Message-----
From: Jeffrey Sorenson [mailto:jws@xxxxxxxxxxxxxxxxx]
Sent: 13. januar 2002 02:34
To: computer-go@xxxxxxxxxxxxxxxxx
Subject: Re: computer-go: abstract info and neural nets


Jeff Sorenson opines:
When a human learns go, it is the very smallest structures that are generally
learned first: the cut with atari- atari is repeated at length until the
beginner learns the idea that just a simple extension from the cut may be
better. The insupportable hane over and over against the opposing wall is
attempted by the tyro until sufficient evidence of its weakness has accumulated
to make some alternative preferable.  After White has played at the center of
Black's three in a row yet again leaving another group eyeless, Black may learn
the
importance of taking that spot.  Most human students must learn the smallest
elements of the go-geometry first, the simple wall and extension, basic life,
the elements commonly referred to in the "proverbs", "at the head of two stones
play hane" and so on.

I heartily concur with most of Hekki's comments on this.  It seems to me that if

an NN can't learn the simplest scale of go-geometry required to live on a very
small board, say 7x7, then progress will remain a very very long way away...
:-)


Good luck!






Heikki Levanto wrote:

> On Sat, Jan 12, 2002 at 03:31:40PM +0100, Erik van der Werf wrote:
> > If it plays locally correct, this means that locally a lot of moves can
> > be ignored (or ordered at the bottom). This is certainly worth
> > something, just think of branchingfactors. Ofcourse you're right that
> > something else is needed to make the global decissions...
>
> Yes, of course NNs can be used for move ordering and pruning, etc. I am not
> trying to argue them to be useless, only that they are not sufficient to
> play the game of go.
>
> Even "locally correct" move depends on the global situation. If a group has
> eyes elsewhere, making the territorially largest endgame move (in sente) is
> locally correct. But if those eyes are lacking, making two one-point eyes is
> certainly preferable, even in gote.
>
> Interesting stuff, all the same!
>
> -Heikki
>
> --
> Heikki Levanto  LSD - Levanto Software Development   <heikki@xxxxxxxxxxxxxxxxx>