The idea of solving the machine learning conundrum with the techniques of natural selection (an idea called, depending on details, genetic algorithms or GAs, genetic computation, genetic programming, evolutionary computation, and evolutionary algorithms) has for decades been one those ideas that seem wonderfully promising but never quite get out of the laboratory.
The concept is to simulate an artifact in software, vary that artifact in interesting ways, force those variations to compete against each other in a relevant environment, determine the winners, and then create the next population of competitors out of their characteristics.
Suppose for instance you wanted to design a ship hull. Hulls vary along many dimensions (beam, draft, depth, freeboard, length, plus the entire vocabulary of curves and planes.) The first step would be to crank out N rough variations representing different combinations of those dimensions, and time how fast or efficiently or whatever those N hulls sailed though a simulated ocean inside a (very big) computer. As soon as the winners were obvious the bottom fifty would be "culled" and the top fifty would "mate", exchanging their states occupying the various dimensions. A new population would rerun the competition, and a new mating would take place. Over time the most efficient hull design would just emerge, through an artificial equivalent of natural selection.
GAs work, but they have two handicaps. First, they require immense volumes of computation, and while Moore's Law is softening the stringency of that constraint, this is one of those technologies that will always demand more. A hundred years from now the companies relying on evolutionary design will still be demanding the fastest computers available. The other problem is that the software side of the concept is extremely labor intensive. The environment, the replicator, the dimensions, and the variations on those dimensions, all have to be written custom for every single case. Presumably some day someone will figure out a general logic that will make it quicker and therefore cheaper to design useful replicators and relevant environments, but there is no special reason to think this day will come soon.
However, in theory GAs are the ultimate design technology, because they are not prey to the blind spots of human engineers, who need solutions that are orderly and comprehensible so they can be thought about and communicated easily, and are prey to fads, enthusiasms, and bad days. GAs are under no such restriction; often they come up with ideas that no human would have imagined and that no one quite understands at first. Plus a sufficiently powerful GA linked to a 3D printer would be the ultimate consumer device, since it would design and fabricate anything you wanted. It would be an Aladdin's Lamp, a universal tool maker, and the end of capitalism as we know it.
In short, R&D on the technology it has never ceased. This summer speakers at The Genetic and Evolutionary Computation Conference in London showed off memory sticks with an advertised lifetime "up to" (whatever that means) 30 times that of current technology, ultra-high- capacity optical fibers, cochlear implants that optimize to individual patients, and a cancer- biopsy analyzer that diagnosed as well as a human pathologist.
Perhaps the most interesting paper was by John Koza of Stanford, a pioneer in the technology. His team developed an antenna for a client reluctant to pay the license fee to the entity owning the patent of the only technology appropriate to the function.
Koza's algorithms evolved a design that not only was different enough to give the client confidence he would prevail against an infringement suit --- or more accurately, dissuade the patent-holder from even writing a cease letter -- but, as a bonus, actually worked better than the patented technology.
What is interesting about this story is not the technology but the circumstances of the application: using algorithms to dodge the increasingly oppressive patent regimes of the culture. Beating a patent is a lower bar than most R&D projects -- you don't have to be better, just sufficiently different, and genetic algorithms come up with different designs automatically, because they think differently than we do. Plus, patents come in a highly standardized language that might give the engineer writing the GA a head start, perhaps an important head start. Finally, knowing how much money you hope to get away with not paying gives managers a clear idea of what to finance and what not, which is always useful.
The only obvious defense an IP holder has is to buy the best evolutionary computing technology he can and patent its outputs as fast they appear. Something like this might work in some cases. Still, as unpredictable as this technology is, when it does emerge as a practical design option, the collision with the patent system will shake us all.
Join the CIO Australia group on LinkedIn. The group is open to CIOs, IT Directors, COOs, CTOs and senior IT managers.