Day of the Robotic Tentacle 199

Actually, when I was back in college, genetic algorithms were the hot topic in one of our VLSI classes that year. What they did was apply the genetic algorthim process to a Field Programmable Gate Array (FPGA [wikipedia.org]) chip to solve a computation. An FPGA is sort of like a giant array of ANDs, ORs, and NOTs with a "control array" of flip-flops that allow you to control the routing of the inputs; you can just load in a new sequence and end up with different outputs using the same hardware. The genetic algorithm comes in by randomly generating the control array sequence. You then compare the output with the target output, then blend the successful solutions together until you have the final solution, all without any hardware design involved.

The story goes, some researchers did this to attempt to reproduce a non-linear equation, I think like a Fourier Transform. The plus side was, they were successfully able to demonstrate that the resultant chip configuration was able to provide the expected results. However, after analyzing the actually solution, the researchers found that the chip was actually creating resonance between different parts of the circuit in such a way that there was no direct path between the input and output signal.

The genetic algorithm had created an analog solution in digital hardware by incorporating the electromagnetic losses and field coupling of the FPGA wiring itself; if they had tried to tweak the "solution" by removing portions of the unused pieces of the FPGA circuit, or even using the same control sequence on a different FPGA, the "solution" would not work.