OSNews

Of course guys already tried evolutionary optimization (like genetic algorithms) on software configure hardware (FPGAs).

They tried to breed a generation of chips that identify the spoken words "yes" and "no". The chips were initially trained to recognize the average frequency of those words and then several generations of optimization wee run, selection for the best breed that can recognize the two words.

Indeed they got some usable chips.

But those bred chips had some very interesting properties.

1.) The burn patterns only worked on the FPGA chip, which was used during optimization. Buring another FPGA chip with the same burn pattern did not result in a working chip.
Different chips are not exactly identical, obciously the small manufacturing tolerances made a difference!

2.) Analyzing the gate structure brought up surprising structures. Stuff that had no purpose, spiral like structures that were only connected at one end..

The situation is very similiar to breeding of horses.The breeders usually won't have a clue why the horse runs good, the mechanics in the horse, even made of simply parts, is too complex for us too understand.

I think a technology that results in nice hardware, were we have no clue why it works or only vague ideas is certainlya difference to previous hardware and much more common to breeding in agriculture.

In the above example what was happend acutally was that the evolutionary optimization process didn't care about the fact that our chips designs run in a model like situiation we call digital cuircuit logic. In that model we have low and high signal values, were we say low is about 0V, high about 5V and give a bit tolerance. We also have discrete time steps (clock speed).
Thus operate in a discrete model situation.
But the real process runs in good old analog phyiscs.
So the spirals observer were actually structures that made be useful for guiding a wave and reflecting it, perhaps getting some time delay effect or what ever.
Digital curicuitry is not supposed to make use of those physical effects! It should just use electron conduction no wave effects.
But the evolutionary optimization scheme won't care about.
It just selects whatever works.
It optimized not over the space of digital cuircuits, but over the space of alle psossible physical behaviours of the individual chip specimen under optimization!
There was no real reason why the result should be nice, toleratn digittal cuircuit instead of some voodoo efect indivdual chip!

Regarding Undecidability:
That is well established for Turing machines.
The open question is, if there are other computer architectures beyond, that are more powerful.
An indeed quantum computing is one such approach.
Who knows what kind of computational power other analog computers have (like ouir brains).

Regards,
Marc