From: Ben Goertzel (ben@goertzel.org)
Date: Fri Jun 14 2002 - 08:00:10 MDT
> Hey Ben,
>
> Does that include genetic programming?
Evolutionary programming (in a particular form) is one of the key learning
heuristics of my Novamente AI architecture
> Instead, as is currently becomming pretty vogue,
> allowing the program to write itself might be a faster
> approach, even though it seems 'inefficient' due to
> its randomness.
Novamente, once it gets smart enough, will be able to handle the task of
re-coding itself.
But I think that we need to start off an AGI with some pretty good AI data
structures and algorithms, which it can then use to think about re-coding
itself.
In my view, evolving an intelligently-goal-directed-self-modification system
from scratch is a cool idea, but is more an "origin of life" project than an
AGI project per se. Also a cool project but, in my view, much more
resource-intensive.
> >From my perspective, a new approach to this method
> could make real headway with relatively few lines of
> code, many versions, access to a simple command
> library, and a whole lot of computing power. And I
> don't mean rewriting the entire program, or gen 1
> redesigning itself from some grand plan. Just simple
> changes, line by line, over time.
>
> Are you aware of anyone working in this direction?
Writing simple code that self-modifies itself is no big trick, but also not
very useful. See many references in the domain of "artificial chemistry,"
e.g. some older work by Walter Fontana (Alchemy? was that the name of his
system? I'm not sure...)
I think that starting with fixed-program intelligence, and getting one's
fixed-program intelligence to intelligently self-modify, will be an easier
route than starting with unintelligent self-modification and getting it to
be intelligent....
But I could be wrong, of course -- neither approach has been made to work,
so far!!
> Re: Aaron's "mid-night hackers". I think that he's
> right, especially in light of grid computing.
> Computing power may be a stumbling block not for the
> sake of the AI itself, but because GP requires so darn
> many evolving programs to get any results in a
> reasonable amount of time. The SETI software, though
> not truly advanced grid, is a glimpse of how this
> stuff will become accessable to all of us soon.
I agree, globally distributed computing will be very useful for AGI, once we
have built software that can utilize it in an appropriate way.
I have detailed designs for using globally distributed Net computing to
enhance AGI, but these are on the back burner until my core AGI is more
fully built.
-- Ben G
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:39 MDT