From: James Rogers (jamesr@best.com)
Date: Sun Apr 15 2001 - 22:12:19 MDT
On 4/15/01 5:07 PM, "Mitchell Porter" <you_have_some_sort_of_bug@yahoo.com>
wrote:
> The main problem with his optimal AI is that
> in its idealized form, it is noncomputable,
> and even its resource-bounded form, it still
> scales horribly (see second comment under
> "Outlook"). This is because it's meant to
> deal with every possible environment, so it
> runs afoul of no-free-lunch theorems. What
> interests me is whether these concepts can
> illuminate 'specialized AI' that works with
> domain-specific representations and algorithms.
I should have mentioned that I have never tried to apply these constructs on
a global basis, as there are obvious problems with computational
intractability. I had already developed an interesting adaptive
partitioning algorithm (something completely outside that paper), that
allows the problems to be computationally tractable in implementation.
Human brains appear to partition from a blank substrate at birth into a
cluster of specialized partitions created as a consequence of the
environment they are subjected to. You create domains as needed on the fly,
adding depth of specialization where necessary. This is a rather
conservative approach to learning (learning only what is forced on you by
the environment), but it creates the type of specialization that allows
intelligence to be feasible and creates areas of intelligence that have
proven utility to the human/AI. I don't know anybody who is a universal
intelligence; humans seem to be large collections of specialization of
varying depth, which judging from the consequences of the mathematics in
question here, grants a strong evolutionary advantage over being a Universal
Intelligence (which *can't* specialize significantly for most intents and
purposes).
My history on this is kind of backwards. I developed an interesting and
unusual adaptive partitioning algorithm for some data mining research, and
while doing related research, tripped across some interesting papers on
Kolmogorov complexity and universal predictor functions. I immediately
noticed the potential relationship between algorithmic complexity and AI,
and after further research, realized that I also had an algorithm that
allowed the AI model I had developed to become tractable. Basically I took
a Universal Intelligence model and derived a functionally equivalent
specialization model that is computationally constrained by
experience/environment.
The theory of Universal Intelligence isn't so valuable because it is a
solution to the problem of AI (although it does give it an excellent
mathematical basis), rather it is valuable because it gives us specific
implementation problems to solve, that when solved, should theoretically
result in a functional AI. Being able to know what needs to be done is a
big step in the right direction.
Cheers,
-James Rogers
jamesr@best.com
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT