Xref: utzoo rec.games.go:938 comp.ai:3963 rec.games.board:2045
Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!ncar!ames!oliveb!sun!hanami!landman
From: landman%hanami@Sun.COM (Howard A. Landman)
Newsgroups: rec.games.go,comp.ai,rec.games.board
Subject: Re: Computer Go Challenge
Keywords: go, computer games, ai
Message-ID: <100234@sun.Eng.Sun.COM>
Date: 20 Apr 89 23:46:14 GMT
References: <3724@sdsu.UUCP>
Sender: news@sun.Eng.Sun.COM
Reply-To: landman@sun.UUCP (Howard A. Landman)
Organization: Sun Microsystems, Mountain View
Lines: 38

In article <3724@sdsu.UUCP> wintermute@tr.wrld (Wintermute) writes:
>Now that Taiwan's Acer Group and the Ing Chang-Ki Wei-Ch'i Educational
>Foundation have offered a prize of 40 million Taiwanese dollars
>(about $1.3 million US) to anyone who can by the year 2000 program a
>version of the game that can beat a professional Go player, are there
>any takers out there?

Where have you been?  Why do you think Bruce Wilcox, David Fotland, Lynn Beus,
and I were in Taipei last November?  (Not to mention a host of Europeans and
Japanese and others).

>Looks like it's time to start using some AI techniques.)  :-D

No, that's been tried already (Wilcox, Nemesis).  It seems to be being left
in the dust by other approaches.  My territory estimator (Poka) is nearly
as strong as Wilcox's entire program, for example, and neither of them have
a prayer against the better programs like Codan.  These were around 12 kyu
last November, and the general expectation is that November 1989 will see
programs around 9 to 10 kyu battling it out for the top prize.

Current programs are running on <2 MIP machines and take about an hour per
game.  In 2000 they'll be running on 100 MIP machines, and have 7 hours or
more per game.  So we should be aiming at about 4 MIP-hours per move, that is,
an algorithm that can productively use 2 hours of Mac II time (or 20 minutes
of SparcStation 1 time) average per move.  Expanding the depth and width of
tactical search is one way to use more time, but I don't think it will give
the maximal improvement.  Pro class programs will have to do EVERYTHING well;
openings, tactics, life & death, positional judgement, and endgame.

On the other hand, some studies of my pro game database indicate that pros
routinely make hundreds of points of errors per game; so the possibility
that a program might someday beat them is quite real.  And the exploding
development of a solid mathematical theory of the endgame, based on Conway
and Berlekamp's work, promises programs that play rapid and accurate yose.
It's a very exciting time in computer Go this year.

	Howard A. Landman
	landman@hanami.sun.com