9 November 2010

Hardware evolution

I have previously written about the idea of evolving solutions to problems, but never have I seen a more beautiful example of this approach than Adrian Thompson’s (1997) investigation into hardware evolution. His paper details an attempt at using selection to find a circuit capable of discriminating between two tones. He isolated a 10 by 10 corner of a reconfigurable chip, such that the behaviour of just 100 of the chip’s 4096 cells was assessed. A genetic algorithm was then used to create different arrangements of the connections between these 100 cells over successive generations. Specifically, a population of 50 individual arrangements existed in each generation and the relative contribution of any one of these arrangements to the next generation was dependent on the extent to which it succeeded in discriminating the tones. After around four thousand generations, an arrangement was found that could discriminate consistently.

The structure of the successful circuit is quite astonishing. A schematic of its arrangement shows that just 21 cells were required to carry out the discrimination and, of these, 5 were special. Like the other 16 they were necessary to ensure that the circuit performed normally, but, bizarrely, there was “no connected path by which they could influence the output” (p.399). In other words, they were contributing to the performance of the circuit in some way other than the direct connections between cells! The means by which these 5 cells exerted their effect isn’t clear, but Thompson suggests that it might be something to do with their analogue output such as radiative coupling or temperature modulation. This is supported by the fact that the chip’s performance was sensitive to ambient temperature, working best in conditions experienced during its evolution.

What Thompson's study tells us is that not only can evolution help us to solve a problem we don’t know how to solve but it is capable of exploiting properties of systems that we are barely even aware of, let alone in a position to fully understand and exploit. Indeed, the physical properties utilised in this circuit are so peripheral to how we normally understand the function of logic circuits, that its behaviour wouldn’t even be captured in a standard simulation. Curiously the selection process can only do this because it has no insight into how or why things work; it simply relies on whether they worked in the past.

References
Thompson, A. (1997). An evolved circuit, intrinsic in silicon, entwined with physics. Evolvable Systems: From Biology to Hardware, 1259, 390-405.
Post a Comment