RE: Designing AGI

From: Ben Goertzel (
Date: Tue Oct 25 2005 - 19:30:21 MDT

> Ben Goertzel wrote:
> > I think that the emergent adult knowledge structures and the learning
> > mechanisms need to be considered together.
> I agree, but you also need to add 'inference mechanisms' and 'goal
> system'. Even if you've got a working learning mechanism building
> descriptive representations, that's still not enough for AGI unless
> those representations can be processed efficiently by an inferential
> search mechanism that can produce well-calibrated, nonobvious and
> goal-relevant predictions. Representations are also critical to
> accurately defining goals, and if you surrender the ability to specify
> representational structure and/or allow them to be mutated in an
> unknown, uncontrolled fashion by learning mechanisms, the design will
> be incompatible with FAI (or any narrow, purposeful application). I'm
> sure you're well aware of the former point Ben (and maybe the later
> one), I'm just making this clear for other readers.
> Beyond that there's planning, tractable sensory processing, final
> stage action generation, realtime processing, reflection and numerous
> other required cognitive competencies. These things need to be
> addressed in a design, and will not just 'emerge' given a mechanism
> for learning representations, but they're not quite as critical and
> interlinked as the four problems identified abouve.

Yes, I agree with these comments, Michael.

I wasn't trying to give a complete
statement of my approach to AGI, merely pointing out what I saw as the
conceptual flaw in Loosemore's statement, using his own language.

The Novamente design addresses all these issues (and more) explicitly,
and does not expect them to "just emerge." However, it DOES expect
specific representations for specific items of knowledge -- and
specific procedures carrying out specific goals -- to emerge within
the framework that addresses these issues. I.e. the architecture
exists to provide a high-level framework in which medium-level
structures and processes can emerge from lower-level learning
mechanisms. (to simplify things just a bit.)

> > Thinking only about the learning mechanism and not about the
> final emergent
> > structures is not going to get you to AGI very quickly. Nor is thinking
> > only about the final emergent structures and not about the learning
> > mechanisms. In short, both traditional cog sci and complex systems
> > thinking are needed, and need to be carefully coordinated.
> I disagree with this because I don't think you need 'Complex systems'
> theory to describe a functional learning system. I did at one time, but
> that was before I knew how to make normative reasoning tractable. I'm
> sure my views on 'emergence' in AGI are pretty clear by now.

Well I am not sure whether or not we really disagree here on anything
substantive. It may be that we are just using words in different ways
due to our different historical biaes.

I didn't say "complex systems theory" btw I said "complex systems thinking."

This is because there isn't really any general "complex systems theory"
out there.... there are theories of particular classes of complex systems,
and then there is mathematical theory relevant to complex systems (e.g.
dynamical systems theory) and then there is complex systems *philosophy*
(which I do find useful for guiding my thinking in many cases)

I find that dynamical systems theory ideas are useful for understanding
some aspects of Novamente, such as the dynamics via which complex
representations may "spontaneously" arise over time out of the combined
actions of a lot of little inference steps carried out in pursuit of
a lot of little goals...

> > We can let different learning mechanisms generate adult systems NOW,
> > it just takes a long time and a lot of work.
> Though of couse generating a whole system from scratch (i.e. by
> evolutionary methods) is a different task from just generating
> representations.

Yes. Alife is one approach to AI, but IMO nowhere near the most
efficient approach...

-- Ben G

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:52 MDT