Noise in the Virtual Hotel

Noise in the Virtual Hotel

Consider the difference between virtual worlds and real worlds. If you want to make a real hotel, you have to put a lot of time, energy, and materials into arranging matters so that the people in adjacent rooms can't overhear each other. When you go to make a virtual hotel, you get that insulation for free. In a virtual hotel, if you want the people in adjacent rooms to be able to overhear each other, you have to add that capacity. You have to add non-insulation. You also have to add shadows, aromas, vibration, dirt, footprints, and wear-and-tear. All these nonfunctional features come for free in the real, concrete world. One feature that may need to be added is collision detection. If you have ever tried to make a computer video game, you probably soon realized that putting shapes in motion on the screen is not enough. The shapes pass right through each other without any effect unless you build collision detection into the update loop (a cycle in your program that incessantly intervenes in whatever the program object is doing and asks, "Am I colliding with anything?").

In his book Le Ton Beau de Marot, Doug Hofstadter (1997) draws attention to the role of what he calls spontaneous intrusions into a creative process. In the real world, almost everything that happens leaves a wake, makes shadows, has an aroma, makes noise, and this provides a bounty of opportunities for spontaneous intrusions. It is also precisely what is in short supply in a virtual world. Indeed one of the chief beauties of virtual worlds, from the point of view of computer modelers, is that quietness: nothing happens except what you provide for, one way or another. This permits you to start with a clean slate and add features to your model one at a time, seeing what the minimal model is that will produce the sought-for effects.

This absence of noise makes computer simulations of evolution extremely limited, since evolution by natural selection feeds on noise, turning fortuitously encountered noise into signal, junk into tools, bugs into features. An early computer simulation of evolution that is still one of the most impressive is Karl Sims's Evolved Virtual Creatures (1994), which can be seen at http://www.karlsims.com/evolved-virtual-creatures.html.

Sims started with random assemblies of virtual jointed blocks with virtual muscles to move the joints, and then let them evolve in their virtual world with its virtual physics. The program automatically selected the assemblages that moved the farthest on their trials for virtual mating, and then repeated the cycle with the resulting offspring. Better and better swimmers, walkers, and jumpers evolved, untouched by any intelligent designer's hand. The far-from-random designs that resulted demonstrated how efficiently (virtual) evolution could sorta discover good design principles and sorta reinvent a striking variety of features found in nature.

It is a spectacular example of getting a lot from a relatively simple model, but it also shows the narrow limits of virtual world evolution. Sims designed the simple "developmental" system that takes whole genomes as inputs and yields novel organisms as outputs, but that process is all backstage, not part of the virtual world being simulated. As a result an accidental bump or shock or collision with a bit of (virtual) debris cannot shorten or lengthen a genome, or change the rules of gene expression. All of that machinery is simply not in the virtual world with its virtual stuff, and hence is immutable. There is no way, for instance, for a Sims creature to evolve a new chromosome. Its whole genetic system is outside the model, not facing natural selection itself but merely conveying genetic information by fiat between generations. (Chapter 51 of the book discusses another example of this phenomenon.)

In a computer model of creativity there should be junk lying around that your creative processes can bump into, noises that your creative processes can't help overhearing. The spontaneous intrusion of that little noise from the next room may tweak what those processes are doing in a way that is serendipitous, or in a way that is destructive, but either way this opens up new possibilities. The exploitation of accidents is the key to creativity, whether what is being made is a new genome, a new behavior, or a new melody.

Let me clarify what I'm not saying. The problem with Sims's evolved creatures is not that they are not made of carbon, or that they contain no protein or hemoglobin. The problem is that they are virtual. And by being virtual, they live in a world many orders of magnitude simpler than the world of biological evolution. The same thing is true of David Cope's EMI (Experiments in Musical Intelligence), a music-composition program. Wonderful as it is, EMI is orders of magnitude simpler than the world of human musical composition. What's delightful about both cases is the discovery of just how much you can get from something so clean, so noise-free, so abstract.

We can imagine improving Cope's EMI, or Sims's work, or any other such project in artificial life or artificial creativity, by adding more and more and more junk, more and more opportunities for collisions, into the world. This would provide more virtual stuff for their sims to cope with, and you never can tell when serendipity will strike, if given an opportunity. But consider how counterintuitive such advice would appear:

No matter what you're modeling, make sure that every phenomenon, every subroutine, everything that happens in that world broadcasts a variety of nonfunctional effects through the world: makes extraneous noises, leaves a wake, sheds dust, causes vibrations, and so forth.

Why? What is all this noise for? It's not for anything; it's just there so that every other process has that noise as a potential source of signal, as something that it might turn, by the alchemy of the creative algorithm, into function, into art, into meaning. Every increment of design in the universe begins with a moment of serendipity, the undesigned intersection of two trajectories yielding something that turns out, retrospectively, to be more than a mere collision. But to the extent that computer modelers follow this advice, they squander the efficiency that makes computers such great tools. So there is a sort of homeostasis here. We can see that, not for any mysterious reason, computer modeling of creativity confronts diminishing returns. In order to get closer and closer to the creativity of a human composer, the model has to become ever-more concrete; it has to model more and more of the incidental collisions that impinge on an embodied composer.