Theory of evolution used to improve machines

When the validity of Darwinian Evolution was debated in a Tennessee courtroom in the famous 1925 Scopes Monkey Trial, it was considered ridiculous by the prosecution that just the random changes between member of a species could ever add up to the huge amount of diversity and the specificity of the organisms we see around us today. The basic mechanism of natural selection was absurd, it was simply impossible that it could happen, let alone that it did. Despite the fact that the prosecution of the Scopes Trail’s intellectual descendants still fight that battle, this argument is rarely used in modern times. We have seen examples of natural selection such as an entire species of moth in Great Britain changing its coloration to blend into trees who have been increasingly, then decreasingly covered in soot over the rise and fall of coal sourced air pollution. Breeding of domesticated animals is simply artificial selection, where animals are selected for their ability to serve humans rather than their ability to survive. However, the most interesting example of selection today is outside of biology altogether; rather, we now use Darwin’s insights to design machines.

Dr. Adrian Thompson of the University of Sussex has been experimenting for years with a special kind of microchip called a Field Programmable Gate Array (FPGA). Most chips used today have a fixed structure that makes the chip perform the same logical operations over and over. To do more complex thinking they simply repeat these operations over and over, making it very important that they run quickly and in a coordinated fashion. A FPGA is very slow and very hot, and has no coordinating clock, but it has an entirely flexible logic structure, an engineer can program the chip to run anyway he or she likes.

Dr. Thompson’s work has a computer load 50 different, entirely random, programs onto a FPGA chip and test them one by one to see which one can discern a particular audio frequency when hooked up to a microphone. Of course, an entirely random program will be almost completely unable to do so, but a (conventional) computer picks which one it thinks identified the audio frequency the best and discards the rest. It then adds a little bit of randomness into this selected program, in 50 slightly different ways. The same audio frequency test is performed on the 50 modifications of the best program from the previous 50, and the best is again selected and the rest are discarded. This process is repeated over and over. 50 modifications, pick the best, and modify it 50 different ways, repeat.
After 200 repetitions only a little progress showed. After 700, the chip would occasionally respond to 1 kilohertz frequency audio signals. After 4000 generations, the progress leveled off, and the chip could identify the audio signal every time it was tested. After only a few hundred more generations and a modification to the selection program, it could pick out individual words when spoken to the microphone, a dramatic increase in complexity. A highly sophisticated program on a very inferior chip, performing a task that is still difficult for highly efficient traditional chips running designed programs arose from complete randomness.

Dr. Thompson’s next step was to figure out how the program that the selection process had chosen worked. Only 37 of the available 100 logic gates were in use on the chip. 5 gates were a complete dead end, unconnected to the rest, but when disabled the program failed completely. When the program was loaded onto another FPGA chip of identical manufacture, the program failed. What Dr. Thompson concluded was that the program was relying on the magnetic waves caused by the miniscule amount of current flowing in the chip to carry information across the gates. Gates, which are designed to operate in a binary ON or OFF position, were being balanced at in-between positions. The effects were so complex and subtle that the slight differences that came from the manufacturing process of the chip were relevant to the chip’s good operation.

From complete randomness, a selection program created a machine that was eminently successful but complex, idiosyncratic, unintuitive and totally reliant on its specific environmental circumstances – just like biological systems we see all around us. Scientists and engineers are working on ways to utilize this process for machines that, although based on conventional components, can redesign elements on the fly to adapt and optimize to harsh and unpredictable conditions. Such designs would require that we give up some degree of the immediate control and understanding over machines that we have now. Such machines could act in unpredictable and contrary ways when we would much rather they didn’t, just like some biological systems that once, among other things, argued in a Tennessee courtroom.