The Power of Artificial Evolution

Is anyone else as disappointed as me at the lack of surprises coming from the field of artificial life? My antenna has been tuned for advances, but I haven't seen much since having my mind blown by the amazing personality Karl Sims got out of a bunch of random cubes back in '94.

With that work, he opened a lot of eyes to the power of artificial evolution by computer, and its potential to use Darwinian processes to solve problems and design solutions that are far too complex to be purposefully addressed by mere humans. Machine understanding of language, for instance. I'm not talking about those "voice recognition" menus on the phone, or even machine translation from one language to another. Those challenges are trivial compared to the decades-long attempts to get computers to actually understand human language as well as an infant does, and then to do something with that understanding. It's so complex, it might even be beyond human capacity to engineer solutions in the ways we're used to.

But artificial evolution seemed like an avenue toward solving such problems the way nature does, through massive trial and error and the rewarding of success; a.k.a., survival of the fittest. Think about the logic necessary to parse a sentence like, "What did you bring that book I don't want to be read to out of up for?" – with its five consecutive prepositions (thanks, Steven Pinker). Wouldn't it be great if solutions to such things could be grown, not built?

Karl's work showed the potential, both visually and dramatically. The creatures you see in Karl's movie pack more personality into a couple of hundred polygons than any but the very best animators have ever achieved. And even they would need hundreds of hours of labor to do it

Instead, he let digital natural selection do it for him. He designed a simple way that cubes could be squashed and connected into body parts that rotate in various ways at their connection points. Each such collection has a "genome," a set of numbers that define the squashing and connection that makes them up.

Then he wrote a program that randomly generated thousands of such genomes, kind of like life.

That created thousands of individuals, each of which could be tested for its ability to perform a "fitness test," such as: how fast could it move across a flat surface? Or swim, or hop?

Then came survival of the fittest. The worst performers were killed off, and the best were "bred" to create thousands of new individuals. Each one was randomly generated, but was "related" to its parents. Run that a few times and the creatures get better and better at hopping, swimming, competing.

Partially because they were so visually appealing, these perky polygons raised the possibility that design and creation could rapidly achieve surprising new levels of creativity, because they would no longer be limited by human ability to conceive, plan, engineer, and execute complex processes. We simply needed to conceive the right essential elements, and then design ways for natural selection to do the work for us.

The concept of Evolutionary Design has indeed been embraced in research communities since then, and progress is being made. NASA even has an Evolvable Systems Group, which has produced the one recent product that has demanded much attention, an antenna design that no human would have considered and which performs better than any designed by humans.

I'm not sure why progress has been so slow. The technique works best on problems that center around finding an optimal set of values for a large number of parameters (like the size, shape, and connectivity of Karl's creatures) that lends itself to evolution. That's hard for humans to come up with. Maybe there aren't a lot of problems that fit nicely into that model. There are many examples of evolutionary design projects at Craig Reynold's compendium of links – e.g. including cranes made from Lego – but the page is not exactly up-to-date.

It seemed that videogames could embrace the evolutionary design process in a big way, yielding amazingly complex worlds. They haven't yet, but may soon. The tools are definitely there.

Will Wright, of Sims fame, has been doing some of the best recent work in the field. He is soon releasing Spore, a game that is not designed in the traditional fashion at all. Rather, it is a system in which users begin with simple creatures that they evolve over time based on the creature's ability to survive. That concept promises to go one step further than pure evolutionary design, which doesn't allow for human intervention. Spore, by contrast, takes advantage of human input to shape the direction of the game. Judging from Wright's track record, and this movie of him demoing the game, we could be in for some of the surprises that we've been waiting for.