Re: computer-go: perfect play : a new fractal algorithm

[Roland <goprog@xxxxxxx>]

Hello Serge,

Friday, September 29, 2000, 4:59:16 AM, you wrote:


SB> I call this algorithm "fractal go".
I didn't read every detail of your algorithm but I think I got the
main ideas. Basically what you are doing when reducing to a smaller
board is making some kind of pattern recognition on the original board
position to find the best move. You make the reduction to enable the
pattern recognition. The same could probably be done also
by a neural network of fuzzy logic approach. I'm not saying that your
algorithm is bad, on the contrary I think the idea is quite good, but
here comes my critic:
The problem is, that in Go every stone can play an important role, so
we have a lot of subtleties. If we reduce the position to a smaller
one or make a pattern recognition, we can possibly wipe out just that
little stone that makes all the difference, and a move that seems to
work on a reduced position may be nonsense on the original. Take the
ladder as an example. One single stone on the other side of the board
can change the status of the ladder from dead to live or the other way
round. So any reducing or pattern recognition or fuzzy algorithm has
the risk to miss an important subtlety of the original position.

As another example lets look at the following two corner positions:
The only difference is the extra stone in the second position, and
likely in a few reduction steps those two stones will be reduced to
one, and both positions will be computed the same way. Yet as far as I
know, the extra stone changes the control over the corner completely.

. . . . .
. . . . .
. . . . .
. . . O .
. . . . .

. . . . .
. . . . .
. . . O .
. . . O .
. . . . .

--
Best regards,
 Roland