Now that evolutionary algorithms are outwitting humans, some researchers want to raise the bar even higher. At Stanford University, for example, professor of biomedical informatics John Koza – yet another Holland protégé – is exploring a closely related field called genetic programming. Evolutionary algorithms have fixed sets of instructions and merely vary the data they manipulate. Genetic programs are more like sexual organisms, capable of improving over time by shuffling bits of code among themselves. The “discoveries” made so far by Koza’s programs range from novel computerized methods for sorting proteins to cutting-edge designs for electronic circuits.
The circuit designs emerged from Koza’s work with Matthew Streeter of Carnegie Mellon University and Martin Keane of Econometrics, a marketing strategy consultancy based in Chicago. Together, the researchers built a program that draws schematic circuit diagrams. Their first challenge was to see whether the genetic approach could derive from scratch circuit designs already patented by past engineers. The program had little trouble generating simple designs that matched those patented in the 1930s and 1940s. Indeed, Koza began referring to the program as an “invention machine” and created a Web page that tracks the latest discoveries by “human competitive” software.
By the time Koza’s group tested the fourth or fifth versions of their program, however, something even more surprising began to happen: the program kicked out circuit designs unpublished anywhere in the patent literature. Two of these designs – a pair of controller circuits that regulate feedback – were so original that Koza and his colleagues have taken out patents on them.
As proud as he is of his software, Koza isn’t about to assign responsibility for the new designs to the program itself. The patents credit Keane, Koza, and Streeter, in that order. But there are a few new pseudophilosophical conundrums lurking here: If something is invented with no human near, is it really an invention? Who is the inventor? And if the invention actually works, does it matter if we don’t understand how?
On that last point, says NASA’s Lohn, “There are two schools of thought. One says I just need something that does X, Y, and Z, and if evolution gives me X, Y, and Z, that’s all I care about. The other school wants to know what’s in there and how it works. We can’t really help those people, because we frequently see evolved designs that are completely unintelligible.”
There’s no need yet for humans to feel jealous of “human competitive” software, says Koza, since the ultimate goal is simply to hand over engineering’s hardest drudge work to computers. He does foresee a time in the near future – perhaps 20 years from now – when genetic algorithms running on ultrafast computers will take over basic design tasks in fields as diverse as electronics and optics. But even then, Koza believes, human and machine intelligence will work in partnership. “We’ve never reached the place where computers have replaced people,” Koza says. “In particular narrow areas, yes – but historically, people have moved on to work on harder problems. I think that will continue to be the case.”
Sam Williams is a freelance technology writer based in Staten Island, NY. He is a frequent contributor to Salon.