from pathlib import Path
from typing import Literal
import cyipopt
import numpy as np
from bonni.constants import INPUT_FN_TYPE
from bonni.function import MAX_EVAL_MSG_STR, FunctionWrapper
[docs]
def optimize_ipopt(
fn: INPUT_FN_TYPE,
x0: np.ndarray,
bounds: np.ndarray,
max_fn_eval: int | None = None,
max_iterations: int | None = None,
save_path: Path | str | None = None,
tol: float = 1e-8,
direction: Literal["maximize", "minimize"] = "minimize",
bound_contract_ratio: float = 1e-3,
) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
"""
Optimize a black-box function using the Interior Point Optimizer (IPOPT).
This function acts as a convenient wrapper around `cyipopt`, managing the
interface between the objective function (which must provide gradients) and the
solver. It handles bound constraints, optimization direction, and recording
of the optimization history.
Args:
fn (Callable[[np.ndarray], tuple[np.ndarray, np.ndarray]]): The objective function to optimize.
It must accept an input array of shape (num_dims,) and return a tuple `(y, g)`,
where `y` is the scalar objective value and `g` represents gradients of shape (num_dims,).
x0 (np.ndarray): The initial starting point for the optimization. Must have shape (num_dims,).
bounds (np.ndarray): A 2D array of shape `(num_dims, 2)` specifying the (lower, upper)
search space boundaries for each of the `num_dims` input dimensions.
max_fn_eval (int | None, optional): The maximum number of allowed function evaluations.
Note that IPOPT performs line searches, so one iteration results in multiple
function evaluations. Defaults to None.
max_iterations (int | None, optional): The maximum number of solver iterations.
Defaults to None.
save_path (Path | str | None, optional): Directory path to save the results of function
evaluation as an npz-file. Defaults to None.
tol (float, optional): The tolerance for convergence. Defaults to 1e-8.
direction (Literal["maximize", "minimize"], optional): The optimization goal. Defaults to "minimize".
bound_contract_ratio (float, optional): A small epsilon value used to contract the provided bounds
slightly (bounds +/- eps * bound_range). This prevents the interior point method from
starting or evaluating exactly on the boundary, which can cause numerical issues.
Defaults to 1e-3.
Returns:
tuple[np.ndarray, np.ndarray, np.ndarray]: A tuple containing the full history of:
- `xs`: Input parameters (shape `(N, D)`).
- `ys`: Objective values (shape `(N,)`).
- `gs`: Gradients (shape `(N, D)`).
"""
assert bounds.ndim == 2 and bounds.shape[1] == 2, "bounds needs to be array of shape (degree_of_freedom, 2)"
assert x0.ndim == 1 and x0.shape[0] == bounds.shape[0], f"x0 needs to have shape (degree_of_freedom,), but got {x0.shape}"
if direction not in ["maximize", "minimize"]:
raise Exception(f"Invalid direction argument: {direction}")
if save_path is not None:
save_path = Path(save_path)
wrapper = FunctionWrapper(
fn,
np.copy(bounds).astype(float),
negate=(direction=="maximize"),
max_fn_eval=max_fn_eval,
save_path=save_path,
)
options = {}
if max_iterations is not None:
options['maxiter'] = max_iterations
bounds_ipopt = np.copy(bounds).astype(float)
bounds_range = bounds_ipopt[:, 1] - bounds_ipopt[:, 0]
bounds_eps = bounds_range * bound_contract_ratio / 2
bounds_ipopt[:, 0] += bounds_eps
bounds_ipopt[:, 1] -= bounds_eps
try:
cyipopt.minimize_ipopt(
fun=wrapper,
x0=x0,
args=(),
method=None,
jac=True,
bounds=bounds_ipopt,
tol=tol,
options=options,
)
except Exception as e:
# if we are early stopping due, then do nothing. If this is real error, raise it again
if str(e) != MAX_EVAL_MSG_STR:
raise e
return wrapper.get_saved_obs()
