Source code for smt_optim.acquisition_functions.composite_expected_improvement
import numpy as np
[docs]
def positive_part(x: float | np.ndarray) -> float | np.ndarray:
"""Return max(x, 0)."""
return np.maximum(x, 0.0)
# Backward compatibility alias
PositivePart = positive_part
[docs]
def init_bi_obj_composite_ei(state, kwargs):
"""
Initialize the composite expected improvement acquisition function.
"""
phi = kwargs["phi"]
n_expectancy = kwargs.get("n_accuracy", 1000)
seed = kwargs.get("seed", getattr(state, "seed", None))
models = state.obj_models
training_outputs = state.scaled_dataset.export_data([0, 1], 0)
f_min = min(phi(y) for y in training_outputs)
# Use deterministic RNG when seed is provided to ensure smooth acquisition surface
rng = np.random.default_rng(seed) if seed is not None else None
def composite_expected_improvement(
mu: np.ndarray, s: np.ndarray, f_min: float, n_expectancy: int = n_expectancy
) -> float:
"""
Expected Improvement composite acquisition function.
Parameters
----------
mu : np.ndarray
Mean prediction of shape (2,).
s : np.ndarray
Standard deviation prediction of shape (2,).
f_min : float
Best minimum objective value in training data.
n_expectancy : int
Number of Monte-Carlo samples.
Returns
-------
float
Expected Improvement value.
"""
if np.all(s <= 1e-16):
return float(positive_part(f_min - phi(mu)))
if rng is not None:
sample_z = rng.standard_normal((n_expectancy, 2))
else:
sample_z = np.random.normal(size=(n_expectancy, 2))
# Vectorized generation of samples: shape (n_expectancy, 2)
samples = mu + s * sample_z
phi_values = np.fromiter(
(phi(pt) for pt in samples), dtype=float, count=n_expectancy
)
improvements = positive_part(f_min - phi_values)
return float(np.mean(improvements))
def composite_ei(x_pred: np.ndarray) -> float:
x_pred = np.atleast_2d(x_pred)
var0 = max(
1e-16, float(np.asarray(models[0].predict_variances(x_pred)).ravel()[0])
)
var1 = max(
1e-16, float(np.asarray(models[1].predict_variances(x_pred)).ravel()[0])
)
s = np.array([np.sqrt(var0), np.sqrt(var1)])
y = np.array(
[
float(np.asarray(models[0].predict_values(x_pred)).ravel()[0]),
float(np.asarray(models[1].predict_values(x_pred)).ravel()[0]),
]
)
return composite_expected_improvement(y, s, f_min)
return composite_ei