Institute For The Future

Evolutionary algorithms, and design vs. understanding

Following on the work we did in the 2007 Ten Year Forecast on the end of science: this week's New Scientist has an article on evolutionary algorithms (sometimes also called genetic algorithms) and debates over their use. Put simply, EAs "mimic the processes of natural selection and random mutation by 'breeding,' selecting and re-breeding possible designs to produce the fittest ones." They might start with two current designs for an antenna, and generate a number of offspring that borrow characteristics from both. The offspring are tested; most fail to work and are discarded, while the survivors are matched up, and the cycle is repeated-- a few thousand times. The result is a new antenna design that is better than any of its ancestors.

Some of the biggest successes with EA (or evolutionary design, or artificial design, take your pick) have been with technologies where the scientific foundations aren't very firm, or don't work as well as you'd expect. Antenna design, for example, has had some notable successes with EA, in part because, as NASA scientist Jason Lohn notes, "Maxwell wrote down the four equations which govern all of wireless communication.... They describe the physics, but the weird thing is, you never use them. In practice, this field is so squirrely, the only way to learn is through trial and error. It's the school of hard knocks."

These methods have been around for a while, but they seem likely to become more widely accessible soon:

 

[Traditionally, EAs have required] ultra-fast computers, both to breed the thousands, or even billions, of generations and to simulate the results to select those offspring that are fit for re-breeding. This has limited their use to a few niche applications.

That is now changing with the availability of ever more powerful computers, the advent of distributed computing "grids", which pool the resources of thousands of PCs, and the emergence of multicore chips, which suit EAs because it's easy to divide up the tasks between cores. As a result, designs can now be evolved in days rather than months or years and EAs are going mainstream.

 

As one evolutionary designer recalled in 2004, "When I started doing this, I was running my simulations on a single Pentium 66 [MHz] PC.... That meant I had to be real careful with how large my problems were and how long it took things to run. Now, you can brute-force things a lot more easily."

So why if these methods work, and are becoming more accessible, why are they controversial? Here's where things get really interesting.

Proponents of EAs say they could replace traditional methods in many fields from designing exotic new types of optical fibre and USB memory sticks to more aesthetic computer-generated art. Critics argue that the technique may lead to designs that can't be properly evaluated since no human understands which trade-offs were made and therefore where failure is likely....

Essentially, some worry that these designs might perform better, but if we can't understand them, we won't know know what hidden costs or disadvantages they carry-- until it's too late. NASA scientist Lohn puts it a different way: he sees EA as forcing people into one of two schools of thought.

 

"One school of thought says you need a black box that does X, Y and Z. If I use evolution to get something that does X, Y and Z, I don't care what's in it as long as it works."

And the other school? "That one says, 'I need to understand what's in there,'" Lohn says. "Those are the people we can't really help, because a lot of times, we don't know what's in there."

 

So ultimately, the question isn't whether these designs work, but whether it's important for us to understand why they work.

Technorati Tags: design, evolution