Mike Tintner wrote:
Vladimir: In experience-based learning there are two main problems
relating to
knowledge acquisition: you have to come up with hypotheses and you
have to assess their plausibility. ...you create them based on various
heuristics.
How is this different from narrow AI? It seems like narrow AI - does
Nars have the ability to learn unprogrammed, or invent, totally new
kinds of logic? Or kinds of algebra?
In fact, the definitions of Nars:
"NARS is "intelligent" in the sense that it is adaptive, and works
with insufficient
knowledge and resources.
By "adaptive", we mean that NARS uses its experience (i.e., the
history of its
interaction with the environment) as the guidance of its inference
activities.
For each question, it looks for an answer that is most consistent with
its
experience (under the restriction of available resources)."
define narrow AI systems - which are also "intelligent," "adaptive,"
"work with insufficient knowledge and resources" and learn from
experience. There seems to be nothing in those definitions which is
distinctive to AGI.
With a sufficient knowledge base, which would require learning, NARS
looks as if it could categorize that which it knows about, and make
guesses as to how certain pieces of information are related to other
pieces of information.
An extended version should be "adaptive" in the patterns that it recognizes.
OTOH, I don't recognize any features that would enable it to take
independent action, so I suspect that it would be but one module of a
more complex system.
N.B.: I'm definitely no expert at NARS, I've only read two of the
papers a a few arguments. Features that I didn't notice could well be
present. And they could certainly be in the planning stage.
I'm a bit hesitant about the theoretical framework, as it appears
computationally expensive. Still, implementation doesn't necessarily
follow theory, and theory can jump over the gnarly bits, leaving them
for implementation. It's possible that lazy evaluation and postponed
stability calculations could make things a LOT more efficient. These
probably aren't practical until the database grows to a reasonable size,
however.
But as I understand it, this still wouldn't be an AGI, but merely a
categorizer. (OTOH, I only read two of the papers. These could just be
the papers that cover the categorizer. Plausibly other papers cover
other aspects.)
N.B.: The current version of NARS, as described, only parses a
specialized language covering topics of inheritance of characteristics.
As such, that's all that was covered by the paper I most recently read.
This doesn't appear to be an inherent limitation, as the terminal nodes
are primitive text and, as such, could, in principle, invoke other
routines, or refer to the contents of an image. The program would
neither know nor care.
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=8660244&id_secret=51310341-2108b3
