I was wondering if a simple system of reason based reasoning could be used
to start an expanding system of knowledge acquisition.  I am not talking
about a human-level AGI program.  I am talking about a very simple, very
artificial system to test the viability and the flexibility of the
reason-based reasoning strategy for general learning.

Reason-based reasoning is just a strategy in which analysis and response to
a situation is based on reasons which the AGI program can access.  In some
ways this makes a great deal of sense and it is almost impossible to
understand why this idea has not gained traction in AI discussions.  In
another sense this method may be a little more complicated than it seems
because it requires the AGI program to integrate knowledge in ways that I
don't fully understand and it can act as an obstruction to making efficient
decisions and reactions.  As our insights become better developed we become
more adept to reacting to the situations for which the insight is relevant
without really thinking of all of the reasons we react the way we do.  This
is part of how habits are formed and as best I can tell, part of the reason
that we can react to situations as quickly as we can is because we can
respond effectively to familiar situations without considering all the
reasons why our reactions should work.  As we are learning, our reactions
have to be tailored with reasons for making decisions, but once we learn to
recognize a situation we seem to react without having to focus on all of
the reasons why we should make one decision or another.  Obviously this
doesn't always work, but it works well enough most of the time to make it
look spectacular from my perspective.  Of course, even with expertise we
are still looking for the reasons we should react in certain ways but our
focus seems to be on a more sophisticated level than it had been at an
earlier stage of learning.

So my question is whether or not reason-based reasoning can be used
effectively in a simplistic system to enable the program to make good
reactions based on what it had learned.  But I do not fully understand how
human beings are able to adeptly recognize and react to complicated
situations.

Analysis and reactions do not only act on some form of output.  They can
govern the analysis and reaction modes as well.  One issue is how much a
reaction to a particular situation should affect a previously learned
analytical or reactive method.  You would not want a system to forget
everything it ever learned in response to a situation but you do want the
program to learn how to improve previously acquired reaction and analytical
methods.  One of the issues that I am aware of is that insights are almost
always tied to the generality level of a subject matter and this idea of a
generality level also applies to analytical and reactive methods as well.  For
example, a general modification of reactive methods might be applied
temporarily at a global level.  This implies that a global reaction might
impact a broad variety of analytical and reactive methods.  This in turn
implies that these methods can be modified by other methods that are not
directly embedded into the reaction.  I can go on and on about this but no
one has yet shown much interest in my thoughts about this issue.

One problem that I do not completely understand is how concepts are
integrated.  Reason-based reasoning will help but it does not explain
everything.  I am thinking about starting with a primitive artificial
language to make the program work a little like a programming method.  However,
with reason-based reasoning that is able to act on recognition and reaction
methods there is no reason why I could not experiment with language
acquisition.

This shows that the idea I am talking about is something that is clearly
different from the old narrower AI methods, like expert systems.  However,
while I think that this idea could work to enable the program to gain
general knowledge, I am not saying that it would be anything near
human-level reasoning.  I am just saying that if a simplistic method might
be able to gain some low level traction for general reasoning in novel ways
then I could have a better base to conduct some experiments into more
complicated problems.  I am not sure if I am going to try this or not but
it certainly seems interesting to me right now.

Jim Bromer



-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968
Powered by Listbox: http://www.listbox.com

Reply via email to