Re: (meta-)meta-design (was: Coding a Transhuman AI 2.0a)

From: Martin Ling (martin@nodezero.org.uk)
Date: Tue May 23 2000 - 04:53:52 MDT


On Tue, May 23, 2000 at 05:09:01AM -0400, Dan Fabulich wrote:
> Martin Ling, missing the point, replied:
>
> > On Tue, May 23, 2000 at 12:19:15AM -0400, Dan Fabulich wrote:
> > > Matt Gingel noted briefly:
> > >
> > > > In the general case, evolution is a graph search. Each possible
> > > > individual is a node, and mutations spell out the transition rules. By
> > > > manipulating the space of transition rules, we transform the search
> > > > space - we make some points closer together and some further
> > > > apart. Perhaps more importantly, local maxima can be eradicated
> > > > since we change the set reachable nodes. (This analysis still
> > > > holds if we consider crossover operators, every node becoming a
> > > > population.)
> > >
> > > Here's a good example of how someone might use GAs: they might be
> > > interested in finding out what the most aerodynamic design for a
> > > vehicle is. So they'd design a multi-dimensional space allowing for a
> > > variety of different things to be changed in the design of the craft.
> > > Then, they'd set the GA to work.
> > >
> > > When it got to work, it would generate a random sampling of designs,
> > > CHECK TO SEE HOW AERODYNAMIC THEY WERE, and prioritize the designs
> > > which were the most aerodynamic in the next run, which would amount to
> > > random variations, breedings, etc. from the previous generation.
> > >
> > > So. GA. Trying to generate good heuristics. It generates a random
> > > sampling of heuristics (or sets of heuristics). Then it checks to see
> > > which heuristics are best, and prioritizes the best heuristics in the
> > > next generation.
> > >
> > > But how do you figure out whether a heuristic is good or not? Well,
> > > heuristics are rules for action. You USE the heuristic, and you see
> > > how well off you wind up. (Using some definition of "well off" which
> > > has nothing to do with motivations, I suppose.) In fact, just one
> > > trial run really isn't enough... you need to try using the heuristic
> > > many many times in order to even out statistical hiccups.
> >
> > ...so you have a second genetic algorithm, trying out different
> > heuristics on the first. And then, if neccessary, one above that.
> >
> > Has this been tried? Meta-design? Meta-meta-design?
>
> OK, *now* I'm frustrated. [I left my whole post here so you can
> reread it.]

...which I now have. Apologies, I had only just got out of bed. I was
thinking of my own experiments (neural network and simulated two-legged
walking robot, trying to achieve balance for stable movement) in which
my main problem has been implementing evaluative heuristics below the
level of how far forward it got in a given time without falling over.

I think what I was vaguely getting at is that multiple levels of GA can be
put in place until you do get something you could set clear limits on.

The extreme example would be a simulated attempt-at-human AI, where you
ran the multiple-level GAs and decided everything on what mark your
virtual subject gets on his English exam.

Martin

-- 
+--------------------------------------------------------+
| Martin J. Ling              Tel: +44 (0)20 8863 2948   |
| martin@nodezero.org.uk      Fax: +44 (0)20 8248 4025   |
| http://www.nodezero.org.uk  Mobile: +44 (0)7940 482675 |
+--------------------------------------------------------+



This archive was generated by hypermail 2b29 : Thu Jul 27 2000 - 14:11:33 MDT