Source code for jetgp.hyperparameter_optimizers.powell

import numpy as np
from scipy.optimize import minimize



def powell(func, lb, ub, **kwargs):
    """
    Powell's method with bounds and multi-start.

    Parameters
    ----------
    func : callable
        Function to minimize.
    lb, ub : array-like
        Lower and upper bounds.
    kwargs : dict
        Optional arguments:
        - x0 : initial guess for first restart
        - n_restart_optimizer : number of random restarts (default=10)
        - debug : bool, print intermediate results (default=False)
        - Any Powell-specific options: maxiter, xtol, ftol, disp
    """
    x0 = kwargs.pop("x0", None)
    n_restart_optimizer = kwargs.pop("n_restart_optimizer", 10)
    debug = kwargs.pop("debug", False)

    # Extract Powell-specific options
    options = {}
    for key in ["maxiter", "xtol", "ftol", "disp"]:
        if key in kwargs:
            options[key] = kwargs.pop(key)

    lb = np.array(lb)
    ub = np.array(ub)
    best_x = None
    best_val = np.inf

    for i in range(n_restart_optimizer):
        if x0 is not None and i == 0:
            x_init = np.array(x0)
        else:
            x_init = np.random.uniform(lb, ub)

        res = minimize(
            func,
            x_init,
            method="Powell",
            bounds=list(zip(lb, ub)),
            options=options,
            **kwargs  # extra kwargs forwarded
        )

        if res.fun < best_val:
            best_val = res.fun
            best_x = res.x

        if debug:
            print(f"[Powell] Restart {i+1}/{n_restart_optimizer} -> best_val={best_val}")

    return best_x, best_val