Re: [computer-go] Pattern matching - example play

[Erik van der Werf <E.vanderWerf@xxxxxxxxxxxxx>]

I will not go into the physics further as it is OT, if you really want 
to know more about these issues we can discuss it off line.

Frank de Groot wrote:
>  > Except of course that a bad move against a pro is a good move against
>  > your system ;-)
>
> Yes but that is not due to the fact that I have no "bad examples".
> That is due to the inherent limitations to a "pure good shape" system.
>
> So you *really* have not understood the whole concept I am trying to 
> defend.
It was a joke to point out the relativity of good and bad in this context.


> Yes I *do* have "bad examples" and this is why the pattern system 
> approaches
> the theoretical maximum performance of a "pure pattern" system.
How do you know that it approaches the theoretical maximum (except of 
course that it goes up)? What is the theoretical maximum for pro games?


>  > I strongly doubt it. Even on 9x9 the statistics are not favorable for
>  > tengen.
>
> Which statistics?
> Based on actual games?
Yes, many thousands played by players significantly stronger than 
current computer programs.

> Or computer self-play?
No, computers are still too weak for that.


> BTW my pattern system does not work for 9x9.
> 9x9 and 19x19 are so vastly different at the opening stage that it simply
> doesn't scale down well.
Funny, I had a somewhat similar system a couple of years ago, which was 
trained on 19x19 games. Nevertheless, on 9x9, one of the first versions 
of Magog (which did about 9 ply full-width search) was inferior to my 
static predictors without any search.


> I have never even tried it but I can't imagine that it will perform well.
Try it, you might be surprised :-)

> Let alone that I trust the very sparse 9x9 "Tengen" statistics for 19x19.
Well ofcourse the statistics I found don't prove anything. However, if 
at some point we could prove that the center opening is inferior on a 
small board, such as 9x9, it seems unlikely that it would become better 
when the board size grows further.

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