Not exactly the same topic, but I feel like this is a good context to bring up https://extremelearning.com.au/unreasonable-effectiveness-of-quasirandom-sequence

Not exactly the same topic, but I feel like this is a good context to bring up https://extremelearning.com.au/unreasonable-effectiveness-of-quasirandom-sequences/ https://extremelearning.com.au/unreasonable-effectiveness-of.... It provides (in some sense) maximally un-aliased parameter choices when trying to cover a (continuous) parameter space. It generalizes to arbitrary dimensions, doesn't require you to choose the number of samples/experiments beforehand, and actually expands trivially to parameters representing classes (including possibly weighting the classes). I think it would be interesting to see how using this sequence fares in such a fractional factorial analysis, i.e. how close to optimal the (e.g. binarized) pseudorandom parameter choices are for different numbers of experiments!