Hi,
> > I think that the hard problem of AGI is actually the other part:
> >
> > BUILDING A SYSTEM CAPABLE OF SUPPORTING SPECIALIZED INTELLIGENCES THAT
> > COMBINE NARROW-AI HEURISTICS WITH SMALL-N GENERAL INTELLIGENCE
>
> Yes this is part of the problem, the other thing you don't mention
> is the difficulty of trying to solve small-n problems efficiently.
Solving small-n "fully general intelligence" is important. But my opinion
it's an easier problem than the problem of getting a good system of
specialized intelligences running in combination with small-n fully general
intelligence.
We currently have active work going on toward the small-n fully general
intelligence problem, in the Novamente project.
Specifically, we have an approach to creating an algorithm that will
efficiently create a small program approximately achieving any specified
goal.
In short, our approach is:
* Represent programs using directed acyclic graphs whose nodes and links are
typed, and whose nodes contain "elementary functions" drawn from:
arithmetic, boolean and prob. logic, and combinatory logic. [The use of
combinatory logic functions allows us to get looping and recursion without
using variables or cyclic graphs.]
* Encode these program dags in linear strings of functions, using a special
encoding method similar to Gene Expression Programming
* Search program space using a special variation of Pelikan and Goldberg's
Bayesian Optimization Algorithm (BOA), acting on the linear strings [BOA is
a variant of genetic programming that uses Bayes net learning in place of
crossover; on simple problems it's 1-2 orders of magnitude faster than GP]
This is complicated, and at the present rate of progress it may be a year
before it's all implemented and tested (because various limited versions
will be extensively tested on real-world problems along the way).
Anyway, I doubt this is the ONLY good approach, but I think it will be a
workable approach, in terms of giving good average-case functionality for
the general small-n goal-achievement problem.
In short, my feeling is that the small-n fully general intelligence problem
is a problem of "conventional computer science," susceptible to clever
algorithmics. And I think that the CS literature is moving in the right
direction to solve it (for example with technologies like GP, BOA, GEP).
Schmidhuber's OOPS algorithm is another approach to the small-n fully
general intelligence problem, that might well be workable, though my money
is on the probabilistic/evolutionary approach. It could be that a
descendant of OOPS will work, AND our approach will work as an
alternative....
On the other hand, I do not think that the other problem of AGI -- getting
specialized intelligence to work building on small-n fully general
intelligence -- is susceptible to current CS techniques. It's a different
sort of problem, more of a complex-systems type of problem...
I am also skeptical that formal computation theory or alg. info. theory will
be useful for small-n fully general intelligence. This is because I think
that, even for small n, we should be thinking in terms of average case
performance, not worst-case performance. And the CS literature is pretty
weak on average-case performance, generally speaking...
> I would say that there is something very important here that you haven't
> mentioned: The value in having a precisely mathematically defined and
> provably strong definition of what general intelligence actually is.
I think that's a valuable thing to have conceptually.
But I'm not so sure it will be useful for the construction of AGI systems.
I doubt you'll be able to use the definition to derive practical AGI
designs, or to prove theorems about the relative intelligence levels of
different AGI systems. The definition may be able to guide work
conceptually, but perhaps some of us already have a good enough intuitive
notion of what intelligence is, and don't need a formal definition with
attached proofs to guide our work...
Partly, I'm wondering what theorems you intend to prove. You can prove that
arbitrary intelligence is achievable given arbitrary resources. To go
significantly beyond that will be very, very hard given the current array of
mathematical tools, I feel.
> In the words of Charles Kettering,
>
> "A problem well stated is a problem half solved."
Well, there are several different rigorous mathematical statements of the
problem of unifying general relativity and quantum theory. None of them has
led to a practical unification of the two, and there is no agreement on
which rigorous formulation is correct. They are interesting research papers
to read, though...
> The work of Marcus Hutter is, I believe, currently the most
> significant piece of work in this direction to date.
So it would seem, based on what I've read, yeah. I'm grateful to you for
introducing me to his quite interesting work ;-)
But I am not convinced that direction of work is of more than theoretical &
conceptual interest.
-- Ben
-------
To unsubscribe, change your address, or temporarily deactivate your subscription,
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]
