3 ms·
Genetic optimisation seems like an odd choice here. The fundamental operation of genetic optimisation is crossing over - taking two solutions, and picking half
by twic 2y ago
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.
- westurner 2y agoGA 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? /?hnlog: antenna, Rydberg, dielectric: https://westurner.github.io/hnlog/#comment-38054784 https://westurner.github.io/hnlog/#comment-38054784