[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: Evaluating positions
This is interesting stuff. It's really difficult to break down these
concepts into the "fundamentals", and I'll tell why I think that is
the case.
Any concept we come up with is not "REAL" in an absolute sense. These
are just symbolisms our minds use in an attempt to bring some
structure into our thinking about things.
Most games, in my opinion have only a single real concept, usually
stated by the goal of the game. In Checkers, the goal of the game is
to run the opponent out of moves (NOT to take all his men.) We
pretend that winning tokens in checkers is a "fundamental principle",
because it is a powerful way to THINK about progress towards our goal,
but you can WIN in checkers even if you are behind in tokens!
What we call "fundamental principles" should be reworded for accuracy
to be "reality approximators" or something like this. We count heads
in checkers because we don't know how to determine which side will
realize the final goal without actually calculating this recursively.
You can tell what is fundamental by seeing how a good program write a
program. When the opponent is out of moves, all these silly
heuristics go out the window and the programmer will consider this a
WIN, no matter how many checkers each side has! This is in
recognition of the fact that counting heads was just a guess, not the
real thing.
The only principle in Go, in this sense, boils down to this single
simple principle: "OWN as many squares as possible when the game is
over."
You have asked the perfectly valid question, how should we think about
the approximators? Which things give the approxiamately best
correlation to winning the game? How much do they overlap each other?
Don
From: "Vlad Dumitrescu" <vladdu@xxxxxxxxxxxxxxxxx>
Date: Sun, 17 Jun 2001 21:33:49 +0200
Reply-To: computer-go@xxxxxxxxxxxxxxxxx
Content-Type: multipart/alternative; boundary="----=_NextPart_000_0018_01C0F775.327A8B20"
Content-Length: 3941
This is a multi-part message in MIME format.
------=_NextPart_000_0018_01C0F775.327A8B20
Content-Type: text/plain;
charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
Hi all,
I had a little vacation and because of the discussions recently about =
neural networks and all that, my thoughts wandered in that direction. =
And I tried to discern the basic ingredients that make an evaluation =
function, on a conceptual level. It became a longish list, but then I =
noticed many concepts are related, or are different views over the same =
more general concept.
In the end, the list came down to two items, that are independent as far =
as I can tell (for statical evaluation):
- influence (of a group)
- strength (of a group); this includes life/death
I managed to define all other concepts in the list (fore example aji) =
using the above, and even some dynamic concepts that I thought were more =
fundamental (as sente/gote).=20
It seems like a mirroring of the simplicity of the rules of the game, to =
be able to use only this few concepts to evaluate a position. But there =
might also be things that are escaping me. So I thought to ask all of =
you wiser people: What do you think?
I'm not sure if there are any practical implications. A decent practical =
evaluation function will probably have to account for the 'derivate' =
concepts, because they might be easier to compute.
regards, Vlad
------=_NextPart_000_0018_01C0F775.327A8B20
Content-Type: text/html;
charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<HTML><HEAD>
<META http-equiv=3DContent-Type content=3D"text/html; =
charset=3Diso-8859-1">
<META content=3D"MSHTML 5.50.4616.200" name=3DGENERATOR>
<STYLE></STYLE>
</HEAD>
<BODY bgColor=3D#ffffff>
<DIV><FONT face=3DArial size=3D2>Hi all,</FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2>I had a little vacation and because of =
the=20
discussions recently about neural networks and all that, my thoughts =
wandered in=20
that direction. And I tried to discern the basic ingredients that make =
an=20
evaluation function, on a conceptual level. It became a longish list, =
but then I=20
noticed many concepts are related, or are different views over the same =
more=20
general concept.</FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2>In the end, the list came down to two =
items, that=20
are independent as far as I can tell (for statical =
evaluation):</FONT></DIV>
<DIV><FONT face=3DArial size=3D2> - influence (of a=20
group)</FONT></DIV>
<DIV><FONT face=3DArial size=3D2> - strength (of a =
group); this=20
includes life/death</FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2>I managed to define all other concepts =
in the list=20
(fore example aji) using the above, and even some dynamic concepts that =
I=20
thought were more fundamental (as sente/gote). </FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2>It seems like a mirroring of the =
simplicity of the=20
rules of the game, to be able to use only this few concepts to evaluate =
a=20
position. But there might also be things that are escaping me. So I =
thought to=20
ask all of you wiser people: What do you think?</FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2>I'm not sure if there are any practical =
implications. A decent practical evaluation function will probably =
have to=20
account for the 'derivate' concepts, because they might be easier to=20
compute.</FONT></DIV>
<DIV><FONT face=3DArial size=3D2></FONT> </DIV>
<DIV><FONT face=3DArial size=3D2>regards, =
Vlad</FONT></DIV></BODY></HTML>
------=_NextPart_000_0018_01C0F775.327A8B20--