mikkokotila
commented
6 years ago Just to update, as this is taking more work than I thought. This is slightly expanded from the initial scope also. I think we have to consider three kinds of randomness:
- Pseudo randomness
- Quasi randomness
- True randomness / quantum randomness
Given the emphasis we have on random search, I believe it's important to have strong coverage for all three. The differences in terms of output are notable:
So far I've implemented:
r = Randomizer(self.param_log)
r.sobol() # sobol sequence
r.uniform_crypto() # from secrets
r.uniform_mersenne() # from random
r.latin_matrix() # latin hypercube
r.latin_improved() # improved latin hypercube
r.latin_sudoku() # latin hypercube with sudoku type constraint
r.korobov_matrix() # korobov design matrix
r.halton() # halton sequence
r.quantum() # this is actual quantum mechanics true randomness! :) Based on everything I've studied and see so far, it seems fair to say that not all randomness is created equal. As we get these into use, it will be very interesting to see how things materialize from dataset to dataset. Not all are covered in any literature, and those that are, with very limited datasets.
ruojol
C0NTRIBUTE
It looks like possibly the lowest hanging fruit in terms of providing the user with options for the optimization strategy, and improving the results that come from uniform standards, have to do with quasirandom (low discrepancy) sequences.
The issue with uniform random function i.e. pseudorandom function is "clumping" where luck plays an important role. The goal of the quasirandom approaches (e.g. Latin Hypercube, Sobol, Halton...) is to minimize clumping.
One of the more interesting variations here seems to be Latin hypercube with a sudoku-like constraint. [link]
Basically what we seem to need here is a way to perform a quasirandom shuffle (instead of the pseudorandom one we use now). That library is not available it seems, so it might have to be built. My goal would be to make available for the user several quasirandom shuffle options.