Benjamin Goertzel wrote:
2) The definition clearly says at least something about how to measure
this degree of intelligence (rather than just handwaving about the
possibility that there might be different degrees), and
This is the shortcoming of the optimization-based approach to defining
intelligence, at present.
Let's say we define intelligence as "the ability to solve complex
optimization
problems"
and let's say we define the intensity of a pattern P in an entity X as the
degree to which P compresses X,
and let's say we define the complexity of an objective function f as
"the total intensity of the set of patterns in the graph of f,
subtracting off
for overlap"
(ignoring for compactness the subtleties of subtracting off for overlap
in this context)
Then, this is all dandy mathematically speaking, but how do we actually
calculate all the patterns in the graph of a complex real-world objective
function, let alone a whole bunch of such?
But this just falls straight into a glorified version of the same
hidden-subjectivity trap that the behaviorists dug themselves into:
- When you try to cash out that compression function, I claim, you
will end up in a situation where the system's real world behavior
depends on exactly which 'patterns' it chooses to go hunting for, and
how it deploys them. The devil is in the details that you do not
specify here, so any decision about whether this formalism really is
coextensive with commonsense intelligence is pure speculation.
- Now, you can certainly get around that criticism I just leveled,
by taking the Hutter route and postulating systems that have infinite
amounts of time and resources to decide what to do: but postulating
such an infinite-resources definition of intelligence is not a
definition that any scientific endeavor has ever done before. That is
not a definition: it is a mathematical fantasy.
Note my previous clarification, BTW: I only care about attempts to
build formal definitions. So maybe if that is not what you care about
(see your comments below), we might not be disagreeing.
Richard Loosemore
Pragmatically, we can posit a particular pattern-recognition system S, and
talk about "all the patterns in the graph of f that S can recognize."
If we let S = a typical human, then I suggest that the above definition of
intelligence captures a lot of the commonsense human language notion
of "intelligence."
However, letting S=a typical human, is obviously not the only choice
one could make.
According to some other assumed pattern recognition system, different
judgments of intelligence might be made.
I do not claim that this approach captures all aspects of the common
language notion of "intelligence" -- and I'm not that interested in trying
to precisely formalize natural language concepts, either. That seems like
a dead-end pursuit, as someone already noted.
My suggestion is that this is the right sort of conceptualization of
"intelligence" to use for guiding AGI research.
-----
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=231415&user_secret=fabd7936
