From: J. Andrew Rogers (andrew@ceruleansystems.com)
Date: Tue Mar 30 2004 - 22:39:08 MST
On Mar 28, 2004, at 8:51 PM, Ben Goertzel wrote:
> Then relative simplicity may be defined as
>
> I(f|g) = the length of the shortest self-delimiting program that
> computes f given g as prior knowledge
>
> The prior knowledge g is what I'm referring to as a "background
> knowledge base"
>
> I have developed a model of simplicity in terms of "pattern theory",
> which is related but not identical to algorithmic information theory.
Could you elucidate on your use of "pattern theory"?
I don't recall ever seeing it mentioned on this list by anyone, and I'm
almost surprised to see you mention it here though I suppose I
shouldn't be.
One of the many things in the hierarchy of issues that I felt needed to
be addressed along the way to AGI was a fully functional computational
model that fundamentally operates on "patterns" -- a universal Bayesian
pattern calculus. I ignored it for a long time because there were
other more pressing problems that needed to be solved first (like truly
adequate representation), but over the last year or so I've started to
spend a lot more time on it as I moved past other obstacles and it has
proven to be both difficult and interesting. "Pattern Theory" is a bit
of an overloaded term; there is a lot of vague and/or unrelated stuff
on the 'net that uses that term, which is partially why I'm asking the
question. In fact, "pattern calculus" in the context of computational
models brings up a lot of papers that are a bit less novel inside than
they sometimes sound on the surface.
Most of the hard part for me has been reducing it to the simplest
possible complete model, and it is definitely a work-in-progress in my
own pursuits at this point. (I solve many things by applying a skill
learned through many years of chemical engineering discipline, taking
obscenely complex systems of arbitrary size and systematically reducing
them to something elegant -- chemical engineering in a nutshell. A
general heuristic as a ChemE is that more is better; the more equations
and models you can throw into the pot, the easier systematic reduction
will be.) It isn't theoretically hard, just tricky to think about
sometimes, and I've only recently started putting significant effort
into it.
Completely aside from the AGI issue, I am pretty much of the opinion
that developing a proper pattern calculus (and the representation
system to exploit it) will help address a great many outstanding issues
in basic programming theory, issues such as brittleness.
In my list of things to discuss, but which I hadn't gotten around to.
j. andrew rogers
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:46 MDT