This is probably quibbling over a definition, but:
Jim Bromer wrote:
On Sat, Aug 9, 2008 at 5:35 PM, Charles Hixson
<[EMAIL PROTECTED]> wrote:
Jim Bromer wrote:
As far as I can tell, the idea of making statistical calculation about
what we don't know is only relevant for three conditions.
The accuracy of the calculation is not significant.
The evaluation is near 1 or 0.
The problem of what is not known is clearly within a generalization
category and a measurement of the uncertainty is also made within a
generalization category valid for the other generalization category.
But we can make choices about things that are not known based on opinion.
Could you define "opinion" in an operational manner, i.e. in such a way that
it was specified whether a particular structure in a database satisfied that
or not? Or a particular logical operation?
Otherwise I am forced to consider opinion as a conflation of probability
estimates and desirability evaluations. This doesn't seem consistent with
your assertion (i.e., if you intended opinion to be so defined, you wouldn't
have responded in that way), but I have no other meaning for it.
I could define the difference between opinion and general knowledge
with abstract terms but it is extremely difficult to come up with an
operational principle that could be used to reliably detect opinion.
This is true when dealing with human opinion so why wouldn't it be
true when dealing with AI opinion? Most facts are supported by
opinion and most opinions are supported by some facts, although the
connection may be somewhat difficult to see in some cases.
Your opinion that opinion itself can be defined, 'as a conflation of
probability estimates and desirability evaluations,' avoids the
difficulty of the definition by making it dependent on two concepts
neither of which are necessary and both of which would require some
kind of arbitrary evaluation system for most cases. Opinion can be
derived without probability or an evaluation of desirability. And
opinion is not necessarily dependent on some kind of weighted system
of numerical measurement.
But while I cannot provide an operational definition that is
absolutely reliable for all cases, I can begin to discuss it as if it
were still an open question (as opposed to an arbitrary definition).
Opinion will be mixed with facts in almost all cases. One can only
start to distinguish them by devising standard systems that attempt to
separate and categorize them. This system is going to be imperfect
just as it is in everyday life. This idea of creating standard
methods that can be used for general classes of kinds of things is
significant because it is related to the problem of 'grounding'
opinions or theories onto 'observable events'.
My imaginary AI program would use categorical reasoning but it would
also be able to learn. I would use text-based IO at first. So in
this sense 'grounding' would have to based on textual interactions.
This kind of grounding would be weaker than the grounding that humans
are capable of, but people are limited too, in their own way.
Since opinion and fact seem to be gnarly and intertwined, I feel that
the use of standard methods to examine the problem are necessary. Why
'standard methods'? Because standard methods would be established
only after passing a series of tests to demonstrate the kind of
reliability that would be desirable for the kinds of problems that
they would be applied to. This kind of reliability could be measurable
in some cases, but measurability is not a necessary aspect of
detecting opinion. And another aspect of developing standard methods
is that by relying on highly reliable components and by narrowing the
variations of individual interpretation, these standard methods could
act as a base for methods of grounding. Ironically, this helps to
bond individual opinions from human society about what is fact and
what is not, but this process is helpful as long as it is not
totalitarian.
So an opinion that contains some truth, but cannot attain a standard
of reliability based on the use of established standard methods to
examine similar problems would have to continue to be considered as an
opinion. Of course, a theory might only be considered to be an
opinion after the thinking device is exposed to an alternative theory
that explains some reference data in another way.
This problem is directly related to the greater problem of artificial
judgment the lack of elementary methods that could act as the
'independent variables' to produce higher AI. That means that I think
the problem is AI-Complete (to use an interesting phrase that someone
in the group has used).
Jim Bromer
Well, one point where we disagree is on whether "truth" can actually be
known by anything. I don't think this is possible. So to me that which
is called truth is just something with a VERY high probability, and
which is also consistent with the mental models that one is employing to
describe "what's out there".
I don't think it desirable to segment knowledge into "general knowledge"
and "opinion". To me "general knowledge" describes beliefs about the
world that are presumed to be commonly shared (by some group), and
opinion is "personal" beliefs that are based on not only general
knowledge, but also on the entire rest of one's modeling of the world.
As such I think of them as estimates of how things probably will work
out (did work out? tense is situational) that cannot be expected to be
generally shared...but which may well be shared by some subset of the
class with "general knowledge".
Note that since opinions, by my estimation, are based on modeling
actions, they are less certain that the facts and models that they are
based on. Mistakes in application are possible. Also it's usually true
that heuristics are used in creating opinions, so they are even less
certain.
Note that "facts" are only memories of past experiences, and are
therefore possibly incorrect. Not only is the process of storage
subject to errors, but "facts" aren't the experiences, but are rather
only a compressed (via a lossy compression) rendition of the
experience. I suspect that these features will be necessary in any
intelligent system. Retrieval is also a problem. The needed facts may
not be remembered when any particular cycle of running the model is in
process, and were they present then they might affect the result.
Perfection is not in this universe. (I note that we are probably using
the term "facts" differently. This depends on precisely what you mean
when you say "Most facts are supported by opinion". I would also assert
that ALL opinions are supported by SOME facts. Just not enough, so one
is forced to various means of estimation. Often the means of estimation
used will be that which is most convenient rather than that which is
most accurate.
I don't understand how you can assert that estimation of probability
isn't necessary. One can't even walk across the room without estimating
that one will find a floor under ones feet. Desirability estimates are
necessary because without desirability one wouldn't bother to form an
opinion. Calling the weights numeric is reflecting the implementation.
Other means of estimating probability are possible (e.g., it can be done
with shifting weights), but numbers (or rather their electrical analog)
are what a computer deals with, so a computer model uses numbers (or
rather a pattern of electrical charges that is manipulated in a means
isomorphic to the way that numbers are manipulated). This feels like
extreme quibbling. Analog systems aren't numeric, digital systems are
(up to isomorphism, and excluding errors).
An operational definition would be one that I could use as a programmer
examining the code. I don't expect to reliably be able to tell the
difference between a statement of opinion and a statement of fact or
belief when examining the outputs. (As you say, you can't do it with
people, either.)
As to definitions, either a word can be defined, or it's a meaningless
noise. The definition may not be totally precise (few are outside of
mathematics...and not all of those), but it draws a fuzzy boundary.
Operational just means that you can use it, or theoretically could use
it, to determine whether in a particular instance it applied. In a
program, given the inputs and the outputs, any sufficiently explicit
definition could be made operational, but frequently this would involve
lots of nitpicking over the precise meaning of the words used in the
definition. So it's simpler to just start off by asking for an
operational definition, i.e., one you could use in a test. (Note that
in any particular instance my proposed definition would also require
detailed examination of the code and tracing the flow between inputs and
outputs, so as a matter of practice it would be essentially impossible
to apply it operationally. Still, it could theoretically be so applied,
and that is an important distinction. And one could intentionally
design a system that followed that definition.)
I think I'll stop now, unless you want comments on the remainder of the
post.
-------------------------------------------
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51
Powered by Listbox: http://www.listbox.com
