Re: [computer-go] Pattern Matcher

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

From: "Vlad Dumitrescu" <vlad_dumitrescu@xxxxxxxxxxxxxxxxx>
Subject: Re: [computer-go] Pattern Matcher


> You say that your pattern matched 85% of the plays from an unseen game. I
think
> it is a remarkable fact.

I most definately never said that.
I said it matched around 85% of the first 120 moves of a very particular
game, a game that appears to follow standard ways of finishing certain
Joseki. This particular game Mr. Popo vs. GoMonster stood out that way.
Other games have a much worse pro prediction percentage.

> What concerns me is the rest 15%. From 120 moves, this is about 18 moves.
Did
> you compare them with what was actually played? Were they better or worse?

I didn't do anything else.
I immediately started work on a NN addition and I have only one system to my
disposal now.
A process has been running for about a week and I can't interrupt it. It
takes up so much memory that I am unable to run the MoyoGo software, as it
needs huge databases. So I am programming a module at the moment and will do
new experiments (very thoroughly woth a lot of test results) with the
pattern system PLUS neural network ASAP.


> I'd also make a comment. I also was into pattern matching (not at the
level
> discussed here, of course) and I thought it would be a great thing. But
later I
> realized one thing: a program using this technique is doomed to be a step
back
> from human players. It has no creativity, it can't invent new things [*].
This
> isn't meaning it isn't useful - but that it is only a part of the answer,
not
> THE answer[**].

When have I said that was the answer?

I have been busy explaining to eveyone and his dog that my pattern matcher
will have only one use. It will sort moves for a search. I am sick and tired
explaining that my pattern system is a tiny little cog, used to sort move in
search routines and for Joseki/Fuseki opening library. I am very tired of
having to explain that by no means will my program use a pattern matcher to
"do moves". But it seems it is totally useless to explain anything, as
people simply say: "You are lying in fact you intend to make a Go program
that works only with a pattern matcher so you are stupid because it will
never work".

So, whatever.

> [*] Let's assume a perfect pattern matcher that had as base data all games
ever
> played up to 1900 (recorded or not). I'm sure it would be very, very
strong.

No it wouldn't.
It would suck totally.

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