Parallel run

[fab6]

Hi,
is it possible to run the actual calculation (e.g. ackley or rastrigin function or any other more time consuming calculation) in parallel, e.g. using scoop or dask?

Thank you in advance!
Best
Fab

[ha-mano]

Hi Fab,

Yes, that is possible. cmaes uses an ask-and-tell interface, so a common pattern is:

1. call ask() population_size times,
2. evaluate those candidates in parallel (e.g., SCOOP, Dask, multiprocessing),
3. pass all (x, value) pairs to tell() once.

For example, here is examples/lra_cma.py adapted to run objective evaluations in parallel with SCOOP:

import numpy as np
from cmaes import CMA
from scoop import futures


def rastrigin(x):
    dim = len(x)
    if dim < 2:
        raise ValueError("dimension must be greater one")
    return 10 * dim + sum(x**2 - 10 * np.cos(2 * np.pi * x))


if __name__ == "__main__":
    dim = 40
    optimizer = CMA(mean=3 * np.ones(dim), sigma=2.0, seed=10, lr_adapt=True)

    for generation in range(50000):
        xs = [optimizer.ask() for _ in range(optimizer.population_size)]
        values = list(futures.map(rastrigin, xs))
        if generation % 500 == 0:
            print(f"#{generation} {values}")
        optimizer.tell(list(zip(xs, values)))

        if optimizer.should_stop():
            break