Parallel evaluation (pymoo 0.6.2)

Use a runner when individual evaluations dominate overhead. ElementwiseProblem sets elementwise=True; no elementwise_evaluation argument is needed. Its _evaluate receives one candidate, while the runner schedules multiple candidates. Vectorized Problem evaluates a matrix directly and may be faster for NumPy models.

Threads or processes

from multiprocessing.pool import ThreadPool
from pymoo.algorithms.soo.nonconvex.ga import GA
from pymoo.core.problem import ElementwiseProblem
from pymoo.optimize import minimize
from pymoo.parallelization.starmap import StarmapParallelization

class MyProblem(ElementwiseProblem):
    def __init__(self, **kwargs):
        super().__init__(n_var=3, n_obj=1, xl=-5, xu=5, **kwargs)

    def _evaluate(self, x, out, *args, **kwargs):
        out["F"] = (x**2).sum()

if __name__ == "__main__":
    with ThreadPool(2) as pool:
        problem = MyProblem(elementwise_runner=StarmapParallelization(pool.starmap))
        result = minimize(problem, GA(pop_size=20), ("n_gen", 5), seed=1)

The native test checks serial/thread equality on deterministic evaluations. For CPU-bound Python functions, substitute multiprocessing.get_context("spawn").Pool(2) inside the main guard. Keep worker classes/functions at importable module scope. That process-pool variant is illustrative; it was not executed in this refresh. Threads suit I/O and native code that releases the GIL. Avoid nested oversubscription from BLAS or model runtimes. Context managers release workers on exceptions.

Joblib

from pymoo.parallelization.joblib import JoblibParallelization
runner = JoblibParallelization(n_jobs=2, backend="threading")
problem = MyProblem(elementwise_runner=runner)
result = minimize(problem, GA(pop_size=20), ("n_gen", 5), seed=1)

Install joblib separately. The constructor accepts n_jobs and joblib keyword arguments; it does not take a lambda that manually wraps Parallel. The default backend is joblib's process-based loky. Threading was executed here; distributed and GPU evaluation require separate validation for the user's model and resources.

Optimizer seeds do not automatically seed stochastic external simulations or worker-local RNGs. Supply deterministic per-evaluation seeds when appropriate and record the evaluation protocol. Serialization of an algorithm does not preserve live pools, remote clients, or external simulator state; recreate those resources when resuming a checkpoint.

Sources: Starmap, Joblib, and released 0.6.2 source; reviewed 2026-10-01.