Not even one batch. It was specific to that exact one chip it was evolved on. Trying to move it to another chip of the same model would produce unreliable results.
There is actually a whole lot of variance between individual silicon chips, even two chips right next to each other on the wafer will preform slightly differently. They will all meet the spec on the datasheet, but datasheets always specify ranges, not exact values.
If I recall the original article, I believe it even went a step further. While running on the same chip it evolved on, if you unplugged the lamp that was in the closest outlet the chip the chip stopped working. It was really fascinating how environmentally specific it evolved.
That said, it seems like it would be very doable to first evolve a chip with the functionality you need in a single environment, then slowly vary parameters to evolve it to be more robust.
Or vice versa begin evolving the algorithm using a fitness function that is the average performance across 5 very different chips to ensure some robustness is built in from the beginning.
> slowly vary parameters to evolve it to be more robust
Injecting noise and other constraints (like forcing it place circuits in different parts of the device) are totally valid when it needs to evolve in-place.
For the most part, I think it would be better to run in a simulator where it can evolve against an abstract model, then it couldn't overfit to the specific device and environment. This doesn't work if the best simulator of the system is the system itself.
Agreed, but to some degree relying on a simulator means it can no longer evolve truly novel solutions. No simulator would have accurately simulated many of the effects it was leveraging in the lab. Essentially you would only generate efficient uses of concepts we already know how to model / engineer well. Using the setup he used, it can generate things we don't understand well and can take advantage of. Or begin to study better.
Yeah, if you took it outside the temperature envelope of the lab it failed. I guess thermal expansion?
There were also a bunch of cells that had inputs, but no outputs. When you disconnected them... the circuit stopped working. Shades of "magic" and "more magic".
I've never worked with it, but I've had a fascination with GA/GP ever since this paper/the Tierra paper. I do wonder why it's such an attractive technique - simulated annealing or hill climbing just don't have the same appeal. It's the biological metaphor, I think.
There is actually a whole lot of variance between individual silicon chips, even two chips right next to each other on the wafer will preform slightly differently. They will all meet the spec on the datasheet, but datasheets always specify ranges, not exact values.