From: Don Dailey <[EMAIL PROTECTED]>
On Wed, 2008-08-13 at 14:38 +0200, Magnus Persson wrote:
>> Here is my take on joseki and fuseki in computer programs.
>> My older program Viking, had a quite nice patternmatching feature
>> which matched the entire board or smaller parts of it towards a
>> database of 50k games or so. It makes it play nice but as far as I
>> could tell it had no impact on the strength of the program.
>> With 9x9 I have used many systems learned or handmade, but it all
>> boils down to that as been said earlier. It only works for a program
>> that does not change, since it overfits its own strengths and
>> weaknesses.
>Yes, even in chess it seems to be best if you match the book to the program.
>In go it is perhaps not good to even try with a rapidly changing program.
So far, all experience with programs and joseki on 19x19 boards has been with
kyu-level programs, since heretofore there have been no dan-level programs on
the 19x19 board. The proverb "learn joseki, lose three stones" may apply to
programs, for much the same reason - the program does not know how to exploit
bad moves, nor how to preserve the advantages of playing good moves.
An approach used for beginners is to learn simpler joseki which are easier to
get right, and learn how to deal with common overplays. A common strategy by
stronger players is to throw in an overplay which will baffle the beginner,
causing him to make an inferior move. These overplays won't ever appear in
professional game records, nor in most joseki books; they seem to be passed
along by a secret fraternity.
Many joseki depend upon an exact calculation of liberties and tesuji; in the
recent demo game between Mogo and Myungwan Kim, Mogo won a capturing race by
exactly one liberty. Any deviation from correct play would have cost Mogo a
considerable number of points. It's hard to be sure with so few sample games,
but I guess that Mogo with a ten-minute clock would have failed to find the
correct line of play. Hence, any experiments with smaller resources available
might have found knowledge of that joseki to be of little or negative value.
With the current scarcity of supercomputers devoted to playing Go, it may be
impossible to take proper advantage of a pro-level opening book - now. The
takeaway for humans and computers seems to be to learn those joseki which you
can understand and use well at your current level of skill. Can playouts as
presently implemented efficiently exploit joseki knowledge? Can they be
designed to do so?
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
