The Genius Behind the Ingenious — October 17th, 2008 by Biologic Staff
When evolutionary biologist Andrew Parker strolls through the vast collection of once-living specimens on display at the Natural History Museum in London, he sees “a treasure trove of brilliant design” [1]. If he’s right about that, then the current fascination with living designs among engineers should come as no surprise. The hot new field of biomimetics was born out of this fascination, fueled by the irresistible thought of translating some of these brilliant designs into lucrative technologies.
But are engineers really even needed for this? What if these technological advances could be had ‘on the cheap’, without any design expertise? A recent National Geographic article put it this way: If “every species, even those that have gone extinct, is a success story, optimized by millions of years of natural selection”, then “why not learn from what evolution has wrought?” [1] Indeed, why not?
In terms of time, of course, millions of years isn’t cheap. But if time is really all the evolutionary process requires—no R&D, no design plans, no genius—shouldn’t we be able to make it cheap by harnessing its power in a much more time-efficient way? For example, what if instead of waiting many generations for the effect of a real mutation to run its evolutionary course, we could simply let that course play itself out at ultra high-speed in a virtual world confined to a rack of computers?
Perhaps the end result, after many rounds of this, really would be impressive designs without the hard work of designing. Think of it this way: Neither spontaneous mutation nor natural selection has the slightest understanding of anything. So, if a mindless pair like that can come up with things like molecular motors, and adaptive camouflage, and nervous systems, and compound eyes, why should we invest so heavily in the laborious tools of education and understanding?
Actually, as you may have guessed, attempts to harness the principles of evolution on computers have been underway for many years now. The field dedicated to this undertaking is known as evolutionary computing, and the results are not altogether encouraging for evolutionary biology.
It’s not that evolutionary design has failed on computers—far from it. One of the most celebrated successes, for example, is a NASA antenna that looks like a bent paper clip. [2] It may not be much to look at, but this odd little design works better than any known alternative, which is why NASA has deployed it in space.
Designs like this are achieved by first generating a number of virtual prototypes on computers, where design parameters are assigned more-or-less arbitrary values, and then using a mathematical model to calculate how good each of these initial prototypes is. Mirroring survival of the fittest, the bad prototypes are discarded and the better ones are replicated to fill out a virtual population. To simulate mutation, the prototypes are then altered by changing their parameter values slightly.
This spawns a large collection of second-generation prototypes. The process of culling, replicating, and mutating is then repeated until an acceptably good design solution results. Numerous variations to this procedure can be applied, but this describes the general approach used in evolutionary computing. And in many interesting cases, like the NASA antenna, it works.
How impressively it works, though, depends on what you were expecting. You can’t fault the NASA engineers for choosing the automated evolutionary approach when you consider the alternative—a pair of needle-nose pliers, half a ton of paper clips, and a whole lot of wrist strain. But if you really saw evolutionary computing as a high-speed version of the process that produced all the jaw-dropping designs of biology, well… you ought to be more than a little disappointed.
Equally sobering is the likelihood that this striking disparity—between the stunning things attributed to evolution and the modest things we get by harnessing it—will persist.
Two major limitations to evolutionary processes seem to assure this. First, it turns out that if you want these processes to go anywhere, you really do need to master the design principles specific to your objective. You’d better believe the NASA team did their homework for the task they were tackling—they knew what materials to use, they knew the range of dimensions to explore, they knew what kind of geometric space to explore, and they knew how to model the performance of any prototype within those specifications. So the software they used was intelligently pre-configured for this particular design task and no other.
It’s not at all obvious how spontaneous mutation and natural selection could pull something like that off in a much more general way sans knowledge. One tempting possibility is that the broad scope of fitness can come to the rescue here. Since fitness can potentially be improved in any number of ways, natural selection might seem like an all-purpose substitute for knowledge-based guidance. But ironically, this very open-endedness is what makes fitness problematic. If the head of NASA had said, “Forget about antennas! From now on I want you to optimize our success instead!”, it’s hard to see how anything would have come of it. Clearly, someone has to do some careful thinking about what success looks like if anything interesting is going to happen.
A similar need for insight has been formally proven in the area of algorithmic optimization. Specifically, landmark work in the mid 1990s shows that if there is no information supplied to guide the search for a better design, then no design procedure done on a computer, including any evolutionary search, consistently outperforms random guessing. [4, 5] Remarkably, then, unguided evolutionary searches cannot generate the information needed for new designs any more than random guessing can. [6, 7] We are led, rather, back to the important insight of information-theory pioneer Leon Brillouin: “The [computing] machine does not create any new information, but it performs a very valuable transformation of known information” [8].
So, the first major limitation is that evolutionary searches need to be intelligently configured in a problem-specific way if they are to outperform random trail and error. The second is that, for all but the simplest design problems, random trial and error is a nonstarter.
In fact, even with the input of intelligent insight, the evolutionary approach is most suited to these simple exceptions. Antenna design is one example. If antennas required complex designs and precision crafting in order to work, you wouldn’t see coat hangers doing the job in a makeshift way. Compare this with something more sophisticated. If you lost your flash drive, would you expect to be able to rig one up the way people rig antennas?
In the case of antennas, it’s easy to find prototypes that work, and nothing can go horribly wrong if you tweak one of those prototypes. That makes the task of zeroing in on a good design relatively easy. Flash drives, on the other hand, are a different story—no one expects to make a working prototype of one of those by accident, or to improve an engineered version by accident.
The problem, of course, is that the designs of biology look more like flash drives (on steroids) than coat-hanger antennas. Parker definitely got that one right.
So, in light of this, we think it’s time to turn the logic around. If it was reasonable to think that selection ought to be as powerful in virtual worlds as it is in the real world (and we think it was), then the considerable limitations we now see demonstrated (even proven) in virtual worlds should perhaps make us re-think the extravagant claims we make about selection in the real world.
Maybe the intuitions that enable us to recognize brilliant design should be kept in mind when we try to explain it.
[1] http://ngm.nationalgeographic.com/2008/04/biomimetics/tom-mueller-text
[2] http://www.nasa.gov/centers/ames/news/releases/2004/04_55AR.html
[3] NASA photo (no copyright); dragonfly eye photo by David L. Green, obtained from Wikimedia Commons under GNU Free Documentation License.
[4] Cullen Schaffer, (1994) A Conservation Law for Generalization Performance, in Machine Learning: Proceedings of the Eleventh International Conference, H Hirsh and W W Cohen, eds., p259-265 (Morgan Kaufmann, San Francisco).
[5] David Wolpert and William G Macready (1997) No free lunch theorems for optimization, IEEE Transactions on Evolutionary Computation 1(1): 67-82.
[6] William A Dembski (2006) No Free Lunch: Why Specified Complexity Cannot Be Purchased without Intelligence, Rowman & Littlefield.
[8] Leon Brillouin (1956) Science and Information Theory (Academic Press, New York).