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

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

From: "Don Dailey" <drd@xxxxxxxxxxxxxxxxx>
Subject: Re: [computer-go] Pattern matching - example play


> thing.  It became  very clear right away that he  still didn't get the
> concept of  eyes.  When I EXPLAINED it  to him, which only  took a few
> minutes, it was like a light bulb went off in his head.

I already said that the learning system should start with basic concepts
like eyes, connectivity, running space, territory, moyo, aji, sente, shape
etc. But then it's on its own. Trying to put dogma's into the system and
knit them all together is IMO counterproductive. And of course search is
needed.

> I have  to assume that you are  not expert or even  reasonably good at
> any game,

I used to be a strong chess player. No idea about my ELO rating but those
times I played club players that were full of themselves, I often won. I
hardly ever read a book on chess but I played thousands of games since I was
7.


> because if you were,  you would understand that in most well
> established  games there  are decades  or centuries  of  knowledge and
> experience  that  has  accumulated  and  that cannot  just  simply  be
> "gleaned" from playing out games.


Never assume that people are much more stupid than you are.
(Never underestimate you enemy :)

I am fully aware of all the counter-arguments and I consider them all to be
flawed.
Time will tell. I am not planning to learn Go so my system will have to
prove my point.
I am quite sure, health permitting, that I will be able to produce a
dan-level system within 5 years.

> It takes  a real teacher, someone  who can relate to  the problems and
> difficulties of the student, that  can introduce him to various "light
> bulbs" of understanding that might NEVER be figured out on his own.

You do not really understand what automatic learning is.
Concepts like "control the center" and "develop your pawns" and "don't
develop your queen to fast" and whatever are all easily learned
automatically.

You ONLY have to introduce a certain convept and let the system look for it.
Introduce the concept of "piece development" and "area control" and let the
system observe & record how players develop their pieces, and which pieces
at what stage in the game control which parts of the board.

>From that you derive metrics that say: "This is, on average, typical for a
strong game".
You can do that with Go patterns, and with eye formation, and with
connectivity patterns, with territorial development and many other things.
When you record the 8 million most used patterns and their statistics, you
capture a surprizing number of "Go proverbs".

Of course only shape-related, when you record shape patterns.
The interesting part starts when you record different patterns (like eye
shapes) and combine them with "good shape".
Then you can let the system decide what is more important, finishing a
certain potential eye or making a certain shape etc.

And all the time the system build upon huge knowledge libraries and weeks of
number crunching.


> Beleive me, I know.  I had a  good friend who was a chess master, that
> taught  me some chess  concepts that  I never  knew existed  and would
> NEVER have figured  them out just by playing  over master games.

I said that it is perfectly OK to get those concepts from a book, but NOT
"program them in" as details.
So you do not make a heuristic in your evaluation function that assigns a
certain value to the development of the pawns, but you tell the learning
system: "There is such a thing as pawn development, figure it out by
yourself".

Not hard to do. Easy.
Same in Go. Piece of cake.
Just takes time but the process is near-trivial.
Boring as hell in fact, programming for a bank is more exciting.


> Yes, hidden very  deep.  An excellent teacher might  save you years or
> even  a lifetime of  study trying  to understand  one of  those games.

Nah, I think they are simply optimizing simple basic Go concepts like
territory, life, connectivity and a few other things.
Chess is a much messier business.

> Don't get me  wrong, the information is there  somewhere, but it's not
> very accessible.

There is nothing more than the concepts of Go.
It all can be learned in detail when the general concept is known to the
system, when it knows what it should look for.


> the teachers  because they are BOUND  to be biased  about things.  All
> the GO  masters who tutored and  trained other GO  masters should have
> stayed out of the way and let them learn on their own.  This is silly.

I am talking about methods to teach a computer, not to teach a human.
A human is unable to play though 500,000 games several times and
administrate hugely complex details about every position.

For a computer, this is trivial. I am saying my method is best for a
computer, not for wetware.
(Stay with the topic :)

A human is not a von-Neuman machine, but then again neither is most brains'
wetware Turing-complete.

> I think you have this backwards.   In one case, you can learn directly
> from a teacher, in  the other case, you have to learn  from the game a
> teacher played.  How  can you call this more  direct?  I'm not getting
> it.

>From what does the teacher teach?
(from games)

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