This looks like just a little bit advanced trivial optimizations for reducing size of high gain antennae. Usually I expect to find such articles in IEEE Transactions on Antennas and Propagation.
What is the reason that it appears on Nature? Just because the genetic methods become the hype (again)?
Note that this is not in the journal Nature, but the considerably lower impact Nature Communications Engineering.
It's so new that it does not yet have an impact factor, but Nature themselves say it's less prestigious than their second-tier journal Nature Communications.
The name is also super weird, because you could well be led to believe the journal is about Communications Engineering, but it is actually Communications on Engineering.
I've needed something like this at least once (but IIRC no more than twice ;) ), so I'm glad to know what to look for next time, thanks for the rabbit hole!
The Science paper launched my PhD topic. I asked a professor what they did for research, they listed a bunch of boring things... saw that and then asked "Did you see that talk about evolving equations? Do you want to work on that?" I was like hell yes!
(Hod gave a talk at our uni about all his research topics)
I wrote a deterministic algorithm for SymReg that outperforms genetic programming.
Genetic optimisation seems like an odd choice here. The fundamental operation of genetic optimisation is crossing over - taking two solutions, and picking half the parameters from one, and half from the other. This makes sense if parameters make independent contributions to fitness. But the case of antenna design, surely that's exactly what they don't? The position of each element relative to the other elements matters enormously. If you had two decent antenna designs, cut both in half, and glued the halves together, i would expect that to be a crappy antenna.
It would be interesting to see an evaluation of genetic optimisation vs more conventional techniques like simplex or BOBYQA, or anything else in NLopt.
GA methods also may converge without expert bias in parameters but typically after more time or generations of mutation, crossover, and selection.
Why would the fitness be lower with mutation, crossover, and selection?
Manual optimization is selection: the - possibly triple-blind - researcher biases the analysis by choosing models, hyperparameters, and parameters. This avoids search cost, but just like gradient descent it can result in unexplored spaces where optima may lurk.
There's already GA Genetic Algorithms deep space antenna design success story.
But there are also dielectric and Rydberg antennae, which don't at all need to be as long as the radio wavelength. How long would that have taken the infinite monkeys GA?
What is the reason that it appears on Nature? Just because the genetic methods become the hype (again)?