From: Bryan Bishop (email@example.com)
Date: Tue Jun 24 2008 - 18:32:57 MDT
On Tuesday 17 June 2008, Matt Mahoney wrote:
> My question is not whether such a thing is possible (I think it is),
> but whether a *non-evolutionary* RSI is possible. The crucial
> difference is whether the parent chooses the fitness function (e.g.
> intelligence), or the environment chooses.
Nonevolutionary RSI is possible, yes. I wrote a bit about this on my
site earlier tonight and am finally checking back to SL4.
Whether or not it will work will require a way to quantify improvement.
There's enough trouble testing two slightly different brain states.
Consider, for example, the poor psychiatrist who assigns medication to
a little boy. The next day after taking the neuropsychopharmaceutical,
the boy is, in some way, significantly different. How do you separate
cause and effect? You could use signal-to-noise sorting if you have a
lot of sample data, but not in this scenario. So, the drug may or may
not be an 'improvement'. My alternative is to look more at
functionality. For example, my brain at the moment can't do partial
differential equations in under 20 seconds. But if I implemented a
plug-and-play module for it, and it started working (w/ unit testing
confirmation), then that's a pretty good bet, isn't it?
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:03 MDT