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

["Mark Boon" <tesujisoftware@xxxxxxxxx>]

> -----Original Message-----
> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of Frank de Groot
> Sent: Monday, November 29, 2004 18:32
> To: computer-go
> Subject: Re: [computer-go] Pattern matching - example play
>
> None of the Go programmers have ever given any arguments as to
> why you can't
> extract Go knowledge from game records, and be better than "manual" Go
> knowledge. I have given a working example for "shape" and theoretical
> examples for effieicntly terminating search branches for L&D based on
> harvested knowlegde, and a good explanation on how "connectivity"
> knowledge
> could be harvested as well.

I do agree that some form of automatic knowledge acquisition would be ideal
and possibly even required to reach the next level in computer-Go. This was
the prediction I made in my master-thesis in 1991. I do have a different
opinion to how I think this will be realised in playing programs. I think Go
knowledge is required to understand the required fundamentals of a Go
program, not necessarily for patterns as I think there's even a possibility
of acquiring those without any input from a strong player or pro-games
whatsoever.

> What are your arguments?
> Keep on repeating that "if you don't have Go knowledge you can't
> program it
> either" is not the point.
> The point is that it is possible to EXTRACT this knowlege
> automatically and
> BETTER.
> At least that is my hunch.
>
> Any reasons why this would not be the case (spare me the religious
> convictions).
> Of course only mean, by "no Go knowledge", no Go *playing* knowledge.
>
> I *do* think it is important to read a lot about stuff, like statical
> evaluation of Semeai etc. etc.
> But apart from knowing the Go literature and using that to extract Go
> knowledge or build algorithms, why would you need to be a good Go player?
> You might say: "To judge where it makes a mistake and to tune
> certain parts"
> but why is that impossible to judge much better and scientifically by
> letting it predict pro moves and analize where it goes wrong? If you can
> build separate modules and tests them using a scientific method and then
> assemble those modules, why would you need to be a Go player for that?

And you ask here if your influence results are any good? Why? Maybe because
this is exactly the kind of knowledge you can't acquire from pro-games?

Did you try to learn influence based on statistical information on how
likely a point will become territory in a pro-game? That might actually be
very interesting to see if that gives anything useful.




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