Terry Pratchett - The Science of Discworld I

Segments of DNA that code for working proteins are called genes. Segments that don't rejoice in a variety of names. Some of them code for proteins that control when a given gene 'switches on', that is, starts to make proteins: these are known as regulatory (or homeotic) genes. Some bits are colloquially called 'junk DNA', a scientific term meaning 'we don't know what these bits are for'. Some literally minded scientists read this as 'they're not for any­thing', thereby getting the horse of nature neatly aligned with the rear end of the cart of human understanding. Most likely they are a mix of different things: DNA that used to have some function way back in evolution but currently does not (and might possibly be revived if, say, an ancient parasite reappeared), DNA that controls how genes switch their protein manufacturing on and off, DNA that controls those, and so on. Some may actually be genuine junk. And some (so the joke goes) may encode a message like 'It was me, I'm God, I existed all along, ha ha.'

Evolutionary processes do not always direct themselves along paths that are neatly comprehensible to humans. This doesn't mean Darwin was wrong: it means that even when he's right, there may be a surprising absence of narrativium, so that a 'story' that makes perfect sense to evolution may not make sense to humans. We sus­pect that a lot of what you find in living organisms is like that -offering a small advantage at every stage of its evolution, but an advantage in such a complex game is that we can't tell a convincing story about why it's an advantage. To show just how bizarre evolu­tionary processes can be, even in comparatively simple circumstances, we must look not to animals or plants, but to elec­tronic circuits.

Since 1993 an engineer named Adrian Thompson has been evolving circuits. The basic technique, known as 'genetic algo­rithms', is quite widely used in computer science. An algorithm is a specific program, or recipe, to solve a given problem. One way to find algorithms for really tough problems is to 'cross-breed' them and apply natural selection. By 'cross breed' we mean 'mix parts of one algorithm with parts of the other'. Biologists call this 'recom­bination' and each sexual organism, like you, recombines its parents' chromosomes in just this manner. Such a technique, or its result, is called a genetic algorithm. When the method works, it works brilliantly; its main disadvantage is that you can't always give a sensible explanation of how the resulting algorithm accomplishes whatever it does. More of that in a moment: first we must discuss the electronics.

Thompson wondered what would happen if you used the genetic algorithm approach on an electronic circuit. Decide on some task, randomly cross-breed circuits that might or might not solve it, keep the ones that do better than the rest, and repeat for as many generations as it takes.

Most electronic engineers, thinking about such a project, will quickly realize that it's silly to use genuine circuits. Instead, you can simulate the circuits on a computer (since you know exactly how a circuit behaves) and do the whole job more quickly and more cheaply in simulation. Thompson mistrusted this line of argument, though: maybe real circuits 'knew' something that a simulation would miss.

He decided on a task: to distinguish between two input signals of different frequencies, 1 kilohertz and 10 kilohertz, that is, sig­nals that made 1000 vibrations per second and 10,000 vibrations per second. Think of them as sound: a low tone and a high tone. The circuit should accept the tone as input signal, process it in some manner to be determined by its eventual structure, and pro­duce an output signal. For the high tone, the circuit should output a steady zero volts, that is, no output at all, and for the low tone, the circuit should output a steady five volts. (Actually, these prop­erties were not specified at the start: any two different steady signals would have been acceptable. But that's how it ended up.)

It would take forever to build thousands of trial circuits by hand, so he employed a 'field-programmable gate array'. This is a microchip that contains a number of very tiny transistorized 'logic cells', mildly intelligent switches, so to speak, whose connections can be changed by loading new instructions into the chip's config­uration memory.

Those instructions are analogous to an organism's DNA code, and can be cross-bred. That's what Thompson did. He started with an array of one hundred logic cells, and used a computer to ran­domly generate a population of fifty instruction codes. The computer loaded each set into the array, fed in the two tones, looked at the outputs, and tried to find some feature that might help in evolving a decent circuit. To begin with, that feature was anything that didn't look totally random. The 'fittest' individual in the first generation produced a steady five-volt output no matter which tone it heard. The least fit instruction codes were then killed off (deleted), the fit ones were bred (copied and recombined), and the process was repeated.

What's most interesting about the experiment is not the details, but how the system homed in on a solution, and the remarkable nature of that solution. By the 220th generation, the fittest circuit produced outputs that were pretty much the same as the inputs, two waveforms of different frequencies. The same effect could have been obtained with no circuit at all, just a bare wire! The desired steady output signals were not yet in prospect.

By the 650th generation, the output for the low tone was steady, but the high tone still produced a variable output signal. It took until generation 2800 for the circuit to give approximately steady, and different, signals for the two tones; only by generation 4100 did the odd glitch get ironed out, after which point little further evolu­tion occurred.

The strangest thing about the eventual solution was its struc­ture. No human engineer would ever have invented it. Indeed no human engineer would have been able to find a solution with a mere 100 logic cells. The human engineer's solution, though, would have been comprehensible, we would be able to tell a convincing 'story' about why it worked. For example, it would include a 'clock', a cir­cuit that ticks at a constant rate. That would give a baseline to compare the other frequencies against. But you can't make a clock with 100 logic cells. The evolutionary solution didn't bother with a clock. Instead, it routed the input signal through a complicated series of loops. These presumably generated time-delayed and oth­erwise processed versions of the signals, which eventually were combined to produce the steady outputs. Presumably. Thompson described how it functioned like this: 'Really, I don't have the faintest idea how it works.'

Amazingly, further study of the final solution showed that only 32 of its 100 logic cells were actually needed. The rest could be removed from the circuit without affecting its behaviour. At first it looked as if five other logic cells could be removed, they were not connected electrically to the rest, nor to the input or output. However, if these were removed, the circuit ceased to work. Presumably these cells reacted to physical properties of the rest of the circuit other than electrical current, magnetic fields, say. Whatever the reason, Thompson's hunch that a real silicon circuit would have more tricks up its sleeve than a computer simulation turned out to be absolutely right.

The technological justification for Thompson's work is the pos­sibility of evolving highly efficient circuits. But the message for basic evolutionary theory is also important. In effect, it tells us that evolution has no need for narrativium. An evolved solution may 'work' without it being at all clear how it does whatever it does. It may not follow any 'design principle' that makes sense to human beings. Instead, it can follow the emergent logic of Ant Country, which can't be captured in a simple story.