Newsgroups: comp.ai.games,comp.ai
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!fas-news.harvard.edu!newspump.wustl.edu!news.ecn.bgu.edu!vixen.cso.uiuc.edu!howland.reston.ans.net!news.nic.surfnet.nl!sun4nl!aie.nl!news
From: geert@sparc.aie.nl (Geert-Jan van Opdorp)
Subject: Re: The game: GO
Sender: news@aie.nl (Usenet News)
Message-ID: <GEERT.95Oct12232108@sparc.aie.nl>
In-Reply-To: y-tony@bu.edu's message of 12 Oct 1995 18:51:01 GMT
Date: Thu, 12 Oct 1995 22:21:08 GMT
Lines: 105
X-Nntp-Posting-Host: sparc.aie.nl
References: <455602$54o@newshound.uidaho.edu> <45i0m7$ob4@hpax.cup.hp.com>
	<45jo2l$j8d@news.bu.edu>
Organization: AI Engineering BV, Amsterdam, Netherlands
Xref: glinda.oz.cs.cmu.edu comp.ai.games:2501 comp.ai:34039

In article <45jo2l$j8d@news.bu.edu> y-tony@bu.edu (Yan Lee) writes:


> 
> Hello,
> 	I am wondering....I only had a limited experience playing Go.  It 
> seems that some of the problems with Go is that many of the aspects are 
> qualitative aspects.  Is there any useful information on the board ( 
> besides position of the pieces ) that can be quantified?  I remember 
> Backgammon using a neural net and once quantitative information was put 
> into the neural net ( ie, pip number, number of blocks, number of open 
> pieces, etc ), the net skyrocketed and can beat the top players.
> 
> 
> Tony
> 

There are quite a number of aspects of the game that can
more or less be quantified. The problem is that most of
them are hard to define formally. Still I believe that
it may be worth to try and to some extent this is
maybe done in some of the programs that are around.
To mention some:

Number of liberties of a group. (this is the only easy one)

Number of moves it would take to capture a group.
(This is almost the same as the liberties above,
but just enough different to be rather difficult already.)

Number of eyes of a group

Number of moves one can ignore before
an eye (or life) is endangered.

Number of moves one can ignore before a connection is threatened.

How many different moves would threaten 
eye, life or connection.

How much influence does a group excert?
This is vital for strategic judgement, but to
assess it all the quantities above must be
estimated.

Some programs implement an 'influence field', based
on relative weakbess and strength of positions, and
the distance to them. Human players do it, and itis essential
since when playing between your own and your opponents
position it is important to strike the right balance.

Number of more or less secure points for each
side. (For example based on that 'field').

And finally, how much time do I want to spend
programming Go, when I could actualy be playing it?

Even apart from the huge problem of formulating
these quantities precisely, how to feed them into
a neural net is also quite unclear. My own experience
is that the smaller a neural net, the better
and this one would be huge, since without
the spatial relations all this quantities are
meaningless. Still, it is worth to think about.
The tuning of the weight of parameters in the 
aforementioned 'field' function seems a
candidate for example.

However, beating top players is a long long
way to go. All the authors of the current
programs are aware of the above, many of
them are strong (amateur) players and
bright programmers and scientists. I'm
sure none of them thinks a professional
will be beaten in the first 20 years,
and that is putting it mildly. I think
David Fotland, the author of one of the leading programs,
mentioned his opinion in this thread. He also
said that current programs are somewhere halfway
between beginners and professional players.
Here you should realize that that means
halfway in 'grades', not in skill. It will take
a talented beginner a year to reach the strength
of these programs; very few ever reach the
professional level.

Now don't let this all discourage you! Go is
the most wonderfull game you will ever learn
and though it may not make you a milionaire,
trying to write a strong program for it
is, precisely because all those difficulties,
one of the most exciting and interresting
AI research projects! 

Go for it!


Geert-Jan


-- 
Geert-Jan van Opdorp
AI-Engineering
Amsterdam, The Netherlands
geert@aie.nl
