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

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

From: "Mark Boon" <tesuji@xxxxxxxxxxxxxxxxx>
Subject: RE: [computer-go] Pattern matching - example play



> White is doing better in this game.

You mean Black?


> Now if we look at what Frank's system
> does, which is predict which moves are 'pro' moves

Ahh, that is a big misunderstanding (and also the cause of all negative
remarks and my vicious reaction to them).

The purpose of my pattern module is not at all predicting pro moves.

The purpose of my pattern system is being a Fuseki library and a Joseki
expert.
The Joseki expert should be good enough to play Joseki in every corner and
still choose the best moves.

Now, since I am a scientist, I measure the performance of my pattern system
by the percentage of pro-moves it predicts correctly, since one can expect
that pro's use Fuseki and Joseki.

Even 4d's often do not have a large Joseki library in their heads.

And since one sometimes sees Joseki being developed late in the game (a
"neglected corner"), I measure pro-prediction over the entire game.

So no, this pattern module is not at all intended to have much to do with
"playing Go".
Just because I measure its Joseki/Fuseki performance by predicting pro
moves, it doesn't mean the module's task is to predict pro moves. I measure
how often it conforms to pro moves because I do not play Go myself so I
can't judge how well it plays out a Joseki.

Yes, I am glad it is moreflexible than JUST Joseki & Fuseki but that does
not mean I see that module as anything ELSE but just that. And I said it was
worth a small fortune, because when you look at pro games like mrPopo /
GoMonster, where Joseki is used that stretches for 100 moves (I call it
Joseki because several other pro games have IDENTICAL COPIES of it), my
pattern module basically plays along with almost all those moves, which
means that my goal of building a very good Joseki expert system has
succeeded.

My goa was that when a pro with a huge Joseki repertoire plays my pattern
matcher, that he exclains: "This is incredible, this thing knows all joseki
plus continuations ever played".

Of course against Go++, who doesn't have this knowledge, the system quickly
runs out of patterns to play and is reduced to playing out small patterns
that have little statistically significant meaning, and without a tactical
or territorial component, a pattern system is a interesting curiosity, not a
"Go engine".







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