Monday, February 14, 2005

Digital evolution, Genetic Algorithms, and syntheic design

Years ago, sometime in 2002 or 2003 I believe, I found an interesting article (in Wired or some other technical/scientific magazine) about how a group had taken an analog circuit and had given it to a genetic algorithm type program to iterate on it. The program's job was to see if it could produce a better circuit - one with more bandwidth and fewer parts - by doing various (mostly random, if I recall) things. The output of the altered circuit would be checked to make sure it performed the same functions as the original, then the altered circuit would be put back into the program. Alter -> test -> repeat.

The result? The circuit evolved into something that was smaller, and worked better. And engineers don't understand the new circuit. This tells me we are still a ways, perhaps a long ways, from truly understanding electronics.

I wish I could find that original article. The closest thing I've come up with is a site about "Evolving Self-Diagnosing Hardware," done by someone at Sussex.

Tim Gorden has also done some interesting work in this area, and /. just had an article about a digital virus which played dead to avoid "being killed". There is even an Evolvable Hardware conference.

Unfortunately, many of these links are going to go away someday since the people are at college. It would be worth it to save all their papers off line for future reference.

A big thanks to Bigdakine, here is the article I was referring to above:
In February's Scientific American (2003), there is an article written by engineers and computer scientists who used GA's to create novel electronic circuit deisgns. They were able to duplicate or better 15 previously patented designs using GA's.

In the case of the most complicated task, designing a "cubic signal generator", the GA evolved a design which out performs a recently patented design that performs the same task. GA's don't think. They have no cognitive ability. Yet this GA *designed* such a good circuit. Its even more interesting than that. TO quote the authors, "The evolved circuit performs with better accuracy than the designed one, but how it functions is not understood. The evolved circuit is clearly more complicated, but also contains redundant parts, such as the purple transistor that contributes nothing to the functioning." (Page 58, Feb 2003 issue of Sci-Am)
There is another pointer to it, along with a different article worth reading, at C# Corner.

Update2: Radio emerges from electronic soup, Analysis of Unconventional Evolved Electronics, /. link. Oh, and wikipedia.

Update3: Synthetic design may be able to solve other problems too... like how to design a better bridge. But can we determine how much better (via how much additional cost or material it adds or consumes)?


Post a Comment

<< Home