Which approach to AGI is most promising?
michield
Approaches like Schmidhuebers Goedel machine, proof searcher for
improvements and Bayesian problem solver for execution of minimisation
problems. Data mining can be used for model creation and on impact of
behaviour on the real world. This approach may be sped up so that
training does not have to be as intensive by making special pre-
programmed functions for e.g. vision, complex actions. Also training
can be done in a virtual environment, then the start values of the
network can be used to initialise slower real-world robots. This
approach is good for very complex tasks of great Kolmogorov
complexity.
Approaches like Novamente which model on a concepts level and
parallelise in an entirely different way. I don't think I fully
understand the overview of this approach yet and the relation with the
approach of the Goedel machine.
Approaches that emulate the mind. These are too compute intensive
right now, and maybe the other approaches will lead to a more
optimised form than this can lead to.
Humans solve tasks by combining low Kolmogorov complexity parts to a
whole that is very complex. Also specialised circuitry for vision etc.
Breaking up problem into small parts for a model like the Goedel
machine would be very interesting. How do you break a problem up in
parts? Is the OOPS algorithm essentially something like this (because
it uses programs that it has created before)?
Jim Lewis
with Baum's Hayek machine I have become less enamored with program
search approaches - see my article "Code Cloning". What is your view
on program search approaches? Many approaches seem too theoretical or
too narrow. I agree that Goertzel's Novamente is rather complex for
outsiders to understand well, though Ben appears to be very well
informed. And I agree that neural simulation is impractical currently
and have some comments about "Forward Engineering" on my site.
Where are you from? Do you have a web site? You can see some of my
ideas at rule110.com and I welcome your comments.
michield
The interest in AGI I have is out of personal interest, not
professional.
I have read the book of Goertzel and of Hutter recently, and I have
read several papers of Schmidhueber.
I find his program search algorithms very interesting:
http://www.idsia.ch/~juergen/ contains the following parts,
the Goedel machine http://www.idsia.ch/~juergen/goedelmachine.html
and OOPS: http://www.idsia.ch/~juergen/oops.html
OOPS is easier for me to understand than Goedel machine, it devotes
time to two strategies of program creation:
one is finding ever more complex programs and the other is combining
already existing ones.
With this approach he has been able to solve very large tower of hanoi
problems.
I am from Holland myself, but I am a physicist not an AI expert (very
interested though!).
With Schmidhueber I have had some correspondence on the following
ideas I had:
-speeding up program search with already existing programs by
implementing them in hardware (use lava, a haskell-like way)
-using probability monads in Haskell to make bayesian networks as easy
as logic programming.
-using graphical cards to speed up some parallellisable tasks.
Regards, Michiel
michield
> > Where are you from? Do you have a web site? You can see some of my
> > ideas at rule110.com and I welcome your comments.
I have just checked your website, very interesting! It seems your
approach of re-using code (copying) indeed
has similarities with OOPS using other code to search program space
(half the time is devoted to that approach
initially, but maybe it can bias itself to that approach and less to
the one making increases in complexity).
Very interesting about the Go game. I wonder what happens if you let
OOPS play against your program :-)
The OOPS code is available on the site of Schmidhueber by the way, it
is in plain C.
Have you looked at Haskell as a programming language? You use boo, but
I think Haskell might be
very interesting too (for its probability monads, and ease of making
DSL's).
Jim Lewis
Though I have not yet studied Hutter and Schmidhueber's work, I have
come across various abstracts, comments etc. As I mentioned, after my
experiments with Hayek, I am hesitant about conventional program space
searching as an approach to AGI - even with hardware speed up. It
seems to me that code in a sense is too expensive to "buy". So we have
to "steal" it. Baum seems to have a related view but he thinks we have
to "steal" it from the work of evolution and the resulting DNA. It
seems to me more likely that we must "steal" it from the world.
I used Haskell a few years ago to design a puzzle - see one of my
sites quirkle.com. Perhaps I did not use Haskell long enough to get
comfortable but I found it cumbersome.
Next on my reading list is "Confabulation Theory" by Robert Hecht-
Neilsen. This type of approach seems more promising to me.
Jim