From: Bryan Bishop (kanzure@gmail.com)
Date: Sat Jul 19 2008 - 00:29:29 MDT
On Friday 18 July 2008, Matt Mahoney wrote:
> We don't know that self improvement is possible except as an
> evolutionary process. As discussed earlier, we lack even a
> mathematical model of deliberate self improvement, as well as a means
> for an agent to test another agent of greater intelligence. Even in
> simplified environments, we know of no classes of problems which are
> provably hard to solve but easy to verify.
Re: mathematical models of RSI. Since we don't have one of these at the
moment, let's start with something simpler, like recursive
self-augmentation. In other words, the augmentations that are made have
to allow some broader set of functional modifications and or
augmentations than would have been otherwise possible or would have
otherwise occured without the modifications. This is difficult to test
in non-digital organisms because of the issue of copies and the
differences between one organism and its own clone (the 'divergence')
and so this model might not be useful in a biological sense. But, in a
broader digital perspective the ability to change can be measured
thanks to bits and the quantification thereof. Whether or not changing
bits in variable memory or bits in control sectors specifying the
programs, and whether or not this qualifies, is still up for grabs. So
let's just start with recursive augmentation, and the feedback
mechanisms thereof. Where can we go from there?
http://en.wikipedia.org/wiki/Systemantics
http://en.wikipedia.org/wiki/Cybernetics
http://en.wikipedia.org/wiki/12_leverage_points
Partly the issue that you'll have to eventually admit is that
quantifying the augmentation can only occur in one situation, where
you, as the one running the simulations and experiments, already know
the overall set of available augmentations for experimental digital
agents. Then you can run measurements on these agents and see which
ones perform the best and pursue all available augmentations etc.
[whatever your measurement happens to be]. This would be susceptible to
mathematical modeling, I think.
The reality of the situation calls for something different since we are
not privledged with that information in advance. Quite literally,
reality is producing information as quickly as possible and those
augmentations are being constructed on the spot -- if they were already
known, there'd be no issue in the first place. So what it turns into is
more a task of an epistemiological approach to the internal modeling of
augmentation options and the pursuit thereof. Maybe by wearing a system
administrator cap (gown?).
>From here we can get a pragmattic, functional model of augmentation from
within, but it's not using the same approach to strict external
mathematical modeling of RSI.
- Bryan
________________________________________
http://heybryan.org/
Engineers: http://heybryan.org/exp.html
irc.freenode.net #hplusroadmap
"Genius is the ability to escape the human condition;
Humanity is the need to escape." -- Q. Uim
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:03 MDT