Quoting terry mcintyre <[EMAIL PROTECTED]>:
Filling suc an eye does not require an "extremely strong" opponent;
I am rated a mere 9 kyu AGA, and use this method often.
My opponents also use it every chance they get. When teaching 20 kyu
players, these known-dead shapes are right at the top of the list.
I am not sure you misunderstood me but if you did here is my clarification:
I did not say that the opponent has to be strong to fill the eye, I
pointed out that the opponent had be extremely strong to *force* a
position where *my program has to fill the eye* in order to win.
However, several otherwise strong programs appear to have a blind
spot regarding these situations.
I agree. And I believe the reasons are twofold. One reason is that it
is quite messy to implement. Most programs are doing very well with
the simple 3x3 pattern paradigm, but nakade problems do not fit in to
the existing framework. So one has to program it from scratch and deal
with either hundreds of patterns and special cases or use some messy
algorithm designed for the problem and hope that it is not inefficient
in the end. I guess the faster the playouts are the more reluctant the
programmers is to it.
Which leads to the second reason. It is not certain that such
improvements always leads to better play in general. The program might
become scared of invasions early if it knows "too well how invading
stones are killed" which is disastrous for the opening in 9x9 where
this is a strategic decision that cannot be delayed in many positions.
This was my impression from experimenting with attacking/defending
simple eyeshapes with my older program Viking.
By the way a really hard nut in playouts are seki because it interacts
with the nakade problem. Bent4 is just an extra messy special case of
this. If the program randomly plays all possible moves you are at
least sure of getting a "signal" everytime you play the right move in
a nakade situation. But if you do this it will never see the sekis
because the nakades go suicidal in all playouts.
I recently added a really nice little rule to Valkyria that forbids it
to fill in false eyes in situations where the false eye is a true eye
because of a seki with a small nakade. The program does not see the
seki but since all bad moves are pruned it behaves as if it did know
about the seki. Previously the group would die 100% of the time. This
changed the evaluation a lot for the positions leading up to the seki
in the game I found the problem.
This is an ideal case where a few lines of code eliminated a severe
weakness of program. Unfortunately these kind of special cases are
rare so the improvement in playing strength is not likely to be
measurable. But a lot of little changes like that makes a difference I
hope.
-Magnus
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
