Framsticks

This looks alot like "Punctuated Equilibrium" in Evolutionary Theory.
Long periods of genetic stability punctuated in between with rapid
change during shifts in the environment.

Marty Rabens wrote:
> Hello Everyone,
>
> Just some more random observations from an amateur.
>
> It seems that often the best path to evolving creatures for specific
> fitness factors is to start with other, different fitness parameters,
> evolve for a while, then change to the fitness parameters you really
> want, and continue evolving.
>
> For example, I started an experiment with a simple "X" genotype, and had
> Distance as the primary parameter. I think I also had some very small
> weights for a couple other parameters. After running for a few days on
> a 700Mhz Pentium III, the best creatures were getting a distance of a
> little over 300 in their 10000 cycle life span. A couple days later,
> they hit 350, and seemed to have leveled off.
>
> I then started a second experiment, also with an "X" genotype, and had
> Vertical Position and Vertical Velocity as my primary parameters, with a
> small value for Distance. After just running overnight, they were
> getting distances of 350, even though the heaviest parameters were for
> vertical position and vertical velocity, and horizontal distance had a
> much smaller relative weight. I haven't tried it yet, but I believe
> that if I now switched the parameters to have a higher distance weight
> and continued evolution, I'd get creatures that did much greater
> distance than those in my first experiment that started out with
> Distance as the primary parameter.
>
> I think that the evolution in the first experiment was getting pulled
> toward a nearby attractor basin in the solution space, and settled at
> the center of this attractor basin. It got "stuck" at this local
> maximum, but there were other, better maxima elsewhere in the solution
> space (I think this is a common problem in alife and GA work).
>
> In the second experiment, the vertical velocity-oriented parameters
> pulled it in a different direction in the solution space, which actually
> led to a deeper attractor basin in the terms of the distance parameter.
> It's almost as if the creature needed help "getting on it's feet," so it
> could see a better place a little farther off. Kind of like helping
> water up over a small rise, so it can make its way down to a deeper
> place on the other side, deeper than it would have found on it's own.
>
> Anyway, sorry for the long ramble. Just thought I'd share my thoughts
> so far.
>
> --Marty Rabens
>
>

I think that's what exacly Richard Dwakins was trying to write in "Climbing
mount improbable".

> I think that the evolution in the first experiment was getting pulled
> toward a nearby attractor basin in the solution space, and settled at
> the center of this attractor basin. It got "stuck" at this local
> maximum, but there were other, better maxima elsewhere in the solution
> space (I think this is a common problem in alife and GA work).

Yes, although mutations and crossovers are supposed to prevent
this drawback... you can also use speciation!

> In the second experiment, the vertical velocity-oriented parameters
> pulled it in a different direction in the solution space, which actually
> led to a deeper attractor basin in the terms of the distance parameter.

Yes, that is probable. If you think that was a problem, you can
try to use higher mutation rates and lower percentage of
"unchanged" individuals.

> It's almost as if the creature needed help "getting on it's feet," so it
> could see a better place a little farther off. Kind of like helping
> water up over a small rise, so it can make its way down to a deeper
> place on the other side, deeper than it would have found on it's own.

Your metaphors are just perfect!

MacKo