Re: [computer-go] Pattern Matcher

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 19:49 8-11-2004 +0100, Frank de Groot wrote:
>----- Original Message ----- 
From: "Mark Boon" <tesuji@xxxxxxxxxxxxxxxxx>
>To: "computer-go" <computer-go@xxxxxxxxxxxxxxxxx>
>Sent: Monday, November 08, 2004 7:39 PM
>Subject: RE: [computer-go] Pattern Matcher
>
>
>
>> I didn't say anything like "your shit stinks", it's only your
>> interpretation. I asked you to show the extra value your system has to a
>Go
>> playing program over an ordinary pattern-matcher . To this day you
>haven't.
>
>
>I can't go and prove everything people ask me to prove.
>The discussion was about the worth of a Go module, not about how good my
>software is.

Oh skip the worth discussion. I am not exactly convinced you understand the
value of a working selflearning system.

It would be invaluable to the world. 

Not because you can earn that much in go. As you see you can earn there at
most $1000 from Mark in the go world. And i can assure you that's not a
fake price. He will really pay that out. It would be more invaluable than
any of the algorithms is which we discuss here, as such a learning system
can be applied to other practical forms.

The problem for automatic learners is that it is not easy to do better to
do than good hand coded patterns. They predict at around 80%-95% thanks to
dynamic sorting code when search depth gets bigger.

Where predicting = playing a more or less equal move and not a very bad move.

In this world regurarly great new inventions get done. The reason why
predicting with automatic learned patterns + software a move ordering is so
hard, is because with a few generalized patterns in your database you must
predict a move for a NEW position that is at the board and which never
before happened in life.

That is completely different from what a few learners are trying to do,
such as voice recognition, or even face recognition. 

So hard data on certain games which you didn't use yet for your database
would be good to get back STATISTICAL SIGNIFICANT data.

That's very easy to do.

Pick 200 new games which were not in database.

Try to predict moves 5..105 for the person winning the game.

20000 data points.

We are interested in 2 things. 

a) an outputfile (text) with a move ordering for each position of say the
first 5 moves.

b) the chance the first move was the played move
c) the chance that the played move was among the first 5 moves

That C is more important than you might guess for the efficiency of the
search tree.

There have been 100000 people before you in different areas trying to
achieve a similar thing. 100000 of them wrote a paper that it all worked
real great. 

"Paper supports everything"
   Arturo Ochoa - Caracas, Venezuela



90000 wrote down actually very similar simple things.

>From the other 1000 unique aproaches 900 basically do this:

 while( !success ) {
   a = GetFromBlackBox();
   x = myverycleverformula(a);
   succes = simplesuccessformula(x);
 }

Those 900 write a whole thesis full about MyVeryCleverFormula(). They
forget to explain how to get 'a' from the blackbox. They just assume it is
there. Their problem usually is they write something down on paper instead
making a computer program that works.

>From those 100 interesting approaches you trivially are an unique one. That
already gives you a very privileged spot in the A.I. world as the vaste
majority just works with paper.

However your 100 predecessors didn't do too well so far. So your chance of
success is 0% until you disprove that, which would be great of course.

Until then it is better to keep a bit silent on how good the system works,
as my very simple go program will win from you. Even if you have it a bit
to work and implemented in a go program my go program still win win as it
plays at the squares :)

>You said that a life-death module is more expensive because it takes more
>time to program.
>
>I say nonsense, the value of a module is how many stones it improves the
>play.
>
>Your reply: "Prove to me that your stuff doesn't stink".
>That's not an argument.
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/