AI breeding (was: Infeasibility of AI-boxing revisited)

From: Cliff Stabbert (cps46@earthlink.net)
Date: Tue Sep 03 2002 - 16:31:05 MDT


Tuesday, September 3, 2002, 9:31:10 AM, Alfio Puglisi wrote:

AP> I would take this more as an input for future AI-breeding
AP> programs - the ones that will have to choose the most
AP> intelligent AIs from the new release pack. Can we find a
AP> definition of "intelligence" good enough to discriminate
AP> the brights AIs from the cheaters ones in a practical way?

Give them a complex and constantly changing environment. Worked for
us.

We do "cheat" in a number of ways -- there are all sorts of built-in
heuristics that work most the time, but can get fouled up (see, e.g.,
Adelbert Ames' optical illusions). But we're flexible enough to
recognize and adapt to that -- we're not "sphexish":
  http://www.personalityresearch.org/evolutionary/sphexishness.html

By "constantly changing" I don't mean that the environment should
change rapidly or randomly. It's easy to throw out a one-liner like
the above, but designing a virtual environment for AIs to evolve in is
no simple task. Whether it's necessary is a separate question.

--
Cliff


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:40 MDT