Re: [computer-go] Computer Go hardware

["Frank de Groot" <frank@xxxxxxxxxxxxx>]

> Your approach sounds reasonable.  There was a Japanese group that did a
> program based on learning patterns by finding the significant parts due
> and splitting to form new patterns and growing them.  Do you know of it?

Katsunari works quite like my program.
David Stoutamire's work is much like the theoretical basis for my system.

For the rest, yes I have read a lot of literature on patterns and I settled
for "simple patterns" to avoid explosion and to keep the ability to
generalize.


> Michael Reiss worked on extracting patterns for move selection
> automatically, and got pretty good pro prediction.  I think later he
abandoned it for
> patterns that were selected by a very strong player instead.  Michael, can
you
> comment?

Michael was constrained (by having a commercial product that had to run on
machines with not too much RAM). So human-selected patterns work better in
that case I assume.

My database is now settled to 80 MB and 1024x1024x8 patterns plus their
parameters (value & frequency). The patterns perform a bit better when I
include liberty / chain length info (as Stoutamire also found).

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/