gridsearch()

gridsearch(func: Callable, candidates: ndarray, gridsearch_only: bool = False, label: str = 'parameter', verbose: bool = False) → float | ndarray

Minimize a function by first checking a collection of candidates, then following up with a derivative-free optimization routine.

If the candidates are one-dimensional, meaning func() is a scalar map \(f:\RR_+\to\RR\), the optimization is carried out via scipy.optimize.minimize_scalar() with method='brent'.

Otherwise (\(f:\RR_+^p\to\RR\) for some \(p > 1\)), the optimization is carried out via scipy.optimize.minimize() with method='Nelder-Mead'.

Parameters:

funccallable(candidate) -> float

Function to minimize.

candidateslist of floats or ndarrays

Candidates to check before starting the optimization routine.

gridsearch_onlybool

If True, stop after checking all candidates and do not follow up with optimization.

labelstr

Label for the types of candidates being checked.

verbosebool

If True, print information at each iteration.

Returns:

winnerfloat or ndarray

Optimized candidate.

Raises:

RuntimeError

If the grid search fails because func() returns np.inf for each candidate.

Warns:

OpInfWarning

1) If the candidates are one-dimensional and the grid search winner is the smallest or largest candidate, or 2) If the minimization fails or does not result in a solution that improves upon the grid search.

Notes

The optimization minimizes func(x) by varying x logarithmically. Unless gridsearch_only=True, it is assumed that all entries of the argument x are positive.

If only one scalar candidate is given, the (one-dimensional) optimization sets the bounds to [candidate / 100, candidate * 100].