Randomness and reproducibility#
By their nature, Evolutionary Algorithms contain randomness. At the start of an experiment, a random seed is defined. Then, at the start of each generation, another random seed, called “generation seed”, is set. This enables reproducing the experiment from scratch, or from a certain generation.
This is done by the RNG (Random Number Generation) object, which is a field in Algorithm.
RNG currently supports the random and numpy packages.
Supported methods such as random.choice, numpy.random.randint can be used as usual, without having to manually handle seeds.
Support for additional packages that generate random numbers can be done by extending this class and overriding the set_seed method, for instance:
class TorchRNG(RNG):
@override
def set_seed(self, seed: int) -> None:
super().set_seed(seed)
torch.manual_seed(seed)
algo = SimpleEvolution(
# some args
random_generator=TorchRNG()
)