Re: computer-go: abstract info and neural nets

["Tyler Gaede" <tgaede@xxxxxxxxxxx>]

----- Original Message -----
From: <heikki@xxxxxxxxxxxxxxxxx>
To: <computer-go@xxxxxxxxxxxxxxxxx>
Sent: Friday, January 11, 2002 2:09 PM
Subject: Re: computer-go: abstract info and neural nets


> I am sorry, but I do not see much future in just feeding the board
position
> to a neural network, and expecting it to learn anything from it. NNs are
> good at recognising patterns. They can even generalize from the known
> patterns and correctly classify patterns that are a bit off.
Unfortunately,
> there are two problems in trying to use this in evaluating go positions:
>
> 1) There are so many whole-board patterns to learn. The board is huge, and
> without any "intelligent" grouping of stones all you can do is to
recognize
> the overall pattern. There are so very many of them - enough to argue that
> no two players ever play the exactly same game. If you only see your
> patterns once, it is hard to learn from them.

the neural network doesn't have to look at a whole board as a pattern.
smaller patterns (of a 5x5 block, or maybe a 9x9 block) probably come up a
lot more often and would be much more recognizable. in this way, instead of
focusing on whole-board tactics, neural networks could help the "close
combat" sequences.

>
> 2) Even small differences in the position can totally change the value of
> the position. If two large one-eyed groups can be connected, they are very
> valuable. If the connection can not be made, they are a huge win to the
> opponent. The feasibility of this connection can depend on a single stone
> far removed from the position (a ladder breaker, for example).
>

true. even when a neural network only looks at a 5x5 section, one stone can
mean the difference between a huge win and a huge loss. i'm not sure if this
kind of rigidity is well suited for NN's.

>
> Whole-board pattern learning may be useful in the very opening, but
> unfortunately it doesn't help anything to play a good opening if you can
not
> continue with even a modest middle and end game.

once again, you don't need to have a whole board NN. you could have a whole
board NN do the work in the beginning, and then switch to a 9x9 in midgame,
and then 5x5 endgame, with each neural network having different goals.

>
> You may have some luck on very small boards, but I wouldn't bet even on
> that.
>
>
> Of course, I may have misunderstood what you are trying to do, or I may
have
> misunderstood the neural nets - I don't have much experience with them. Or
I
> may be simply wrong. Keep us posted, if you make progress!
>
> Good luck!
>
> Heikki
>
>
>
> --
> Heikki Levanto  LSD - Levanto Software Development   <heikki@xxxxxxxxxxxxxxxxx>
>
>

-tyler