WC Archive

Evolution is a pretty amazing process. The combination of internal change (mutation) and environmental pressure (fitness) can have pretty dramatic results, given enough time. And when you do it in a computer, "enough time" can be surprisingly brief.

Evolutionary design is a computerized creative process which relies on the same notions of natural selection and mutation that underlie biological evolution. Take a large number of individuals, each slightly different. Introduce some mutation, either by randomizing small bits or mixing elements from individuals (the electronic version of sexual reproduction). Check the resulting generation against the goal -- how well do the various designs accomplish the needed task? Get rid of some of the designs that do very poorly, add more of the designs that do fairly well. Now repeat the process. Many thousands of times.

We've talked about it here in brief, and it's one of the more powerful techniques underlying the biomimicry concept: you're not just copying nature's results, you're copying how nature comes by its results.

NASA's 2004 conference on evolvable hardware just finished up, and it turns out that NASA is doing some of the most interesting work around with evolutionary design. The Evolvable Systems Group researches techniques for engineering hardware for NASA missions without explicit blueprints. Antenna design consumes a great deal of their attention, as some of the interactions between components in the antenna frame and with the spacecraft itself can be very difficult to model. The antennas that the evolutionary designs come up with often don't really look like traditional devices (see above), but that's okay: it's how they work that counts.

(Antennas have been the focus for evolutionary design researchers for awhile now; Derek Linden, who did some work with NASA on this prior to the ESG's efforts, was working on antennas back in 1997 (PDF), and has suggested that the evolved antenna he patented in 1999 may have been the first patent issued to a nonhuman designer.)

As funky as these antenna designs are, I think that the most intriguing research the ESG is now doing is with coevolutionary algorithms. (The ESG site links to a paper on coevolutionary design the group did for the 2002 IEEE Congress on Evolutionary Computation here -- PDF. It's heavy going, but interesting.) Whereas traditional evolutionary design is a strictly Darwinian, competitive process, coevolutionary design integrates cooperative aspects as well, making for a richer, more complex evolutionary environment -- and one which better mimics reality.