Re: Defining symbolic AI, what it really is and what it's not?
- From: curt@xxxxxxxx (Curt Welch)
- Date: 25 Apr 2006 21:16:12 GMT
"heiska mikko" <mikk2lnx@xxxxxxxxxxx> wrote:
If we want to know what are the good and bad sides of symbolic AI,
where it should be used and what are its limits, then first it would
be good to have a definition for it. We should not confuse
definitional principles/properties with derived, accidental,
descriptive or version specific. As current and previous versions of
symbolic AI principles don't allow much intelligence, then we should
consider carefully, if it is really impossible or possible to fix
that problem by altering the principles, adding more, extending the
principles with something that may not be considered "symbolic" and/or
using symbolic AI as just one software component alongside with other
kinds of AI and other kinds of software components.
I think that the basic definition should be very short with bare
minimum, and all the rest should be considered version
specific, derived, accidental or descriptive.
Should the definition for "symbolic AI" be this:
"AI that has symbols as input and output, and internally processes
symbols"?
It is not: "AI that processes only symbols, nothing else". Let's
appreciate 'and' more.
Can anybody really say that there(*) are no formulation/version of
symbolic AI that has not all those problems that people talk about?
(*)At least in the non-physical world of mathematical ideas, along the
lines of Plato
What are definitional or derived problems and what are version
specific or accidental problems?
------------
What is that "symbol" in symbol processing?
At least mathematical statements, blueprints and programming
languages should be considered consisting of symbols alongside
numbers.
Natural language and perception have other kind of symbols with
unclear definitions. Sets are replaced with indirect ways of
expressing and processing sets of subsets, that often have connections
to numbers in ways that we are not conscious about. In this context
"number" can mean neuronal axon strength instead of digit string.
In AGI, symbols need to be processed, one way or the other, whether
you like symbolic AI or not.
Mikko Heiska 22:23
Yeah, the last thing you want to do is see symbols as something evil for AI
just because one approach to AI that has the word symbol in its name seems
to have stalled.
The field of symbolic AI is in general, the approach of trying to produce a
thinking machine or a knowledge machine that stores or manipulates
knowledge in ways similar to how humans use and manipulate language. The
world symbol in that case is just an abstraction of the concept of a
natural language word. Humans do very powerful things with language so
it's no surprise that we would attempt to emulate human language behavior
in order to create an intelligent machine.
However, more generally, symbols are just discrete events that can easily
be differentiated from one another by the hardware. All digital computers
are very much symbol processing machines (all based on the symbols 1 and
0). Any solution to AI implemented on a digital computer is going to be a
symbol processing solution. However, this is a far more general concept of
symbol processing than what has been used in the work known as symbolic AI.
Symbol processing however is not the only option. The alternative is
analog processing. There are plenty of pure analog effects in this
universe to create processes out of such as moving objects of all types
from electrons to mountains, field effects like electromagnetism and
gravity, and even time itself can all be seen as non-symbolic systems of
processing and information storage. Some of these might even be required
as part of the solution to general AI simply for performance reasons.
However, the problems of symbolic AI is not because it used symbols. Every
other approach to AI which has been done with digital computers has been
purely symbol based as well. The question to be solved is what should the
symbols represent and how should they be manipulated. Symbolic AI was just
one broad class of attempts at answering that question. In the end, the
solution to AI will almost certainly be symbolic in nature - the only
question is how will the symbols be used and what will they represent.
The approach I like the best currently are temporal pulse solutions. A
pulse network is just another attempt to create symbol manipulation
hardware to explain human behavior. It simply represents information at a
far lower level than what was attempted by classic symbolic AI.
I created a thread here long ago with a title something like "pulses are
symbols" to debate this very point.
--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.
- Follow-Ups:
- References:
- Defining symbolic AI, what it really is and what it's not?
- From: heiska mikko
- Defining symbolic AI, what it really is and what it's not?
- Prev by Date: Re: Crick's answer to lester's conundrum
- Next by Date: Re: Strong AI Thesis (No Chinese room, I promise)
- Previous by thread: Re: Defining symbolic AI, what it really is and what it's not?
- Next by thread: Re: Defining symbolic AI, what it really is and what it's not?
- Index(es):
Relevant Pages
|
|
