# Copyright 2021 IRT Saint Exupéry, https://www.irt-saintexupery.com
#
# This program is free software; you can redistribute it and/or
# modify it under the terms of the GNU Lesser General Public
# License version 3 as published by the Free Software Foundation.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
# Lesser General Public License for more details.
#
# You should have received a copy of the GNU Lesser General Public License
# along with this program; if not, write to the Free Software Foundation,
# Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
# Contributors:
# INITIAL AUTHORS - initial API and implementation and/or initial
# documentation
# :author: Francois Gallard
# OTHER AUTHORS - MACROSCOPIC CHANGES
"""Quantification of robustness of the optimum to variables perturbations."""
from __future__ import annotations
from typing import Sized
import numpy as np
from numpy.random import multivariate_normal
from gemseo.post.core.hessians import BFGSApprox
from gemseo.post.core.hessians import LSTSQApprox
from gemseo.post.core.hessians import SR1Approx
[docs]class RobustnessQuantifier:
"""classdocs."""
AVAILABLE_APPROXIMATIONS = ["BFGS", "SR1", "LEAST_SQUARES"]
def __init__(self, history, approximation_method: str = "SR1") -> None:
"""
Args:
history: An approximation history.
approximation_method: The name of an approximation method for the Hessian.
""" # noqa: D205, D212, D415
self.history = history
if approximation_method not in self.AVAILABLE_APPROXIMATIONS:
raise ValueError(
f"Unknown hessian approximation method {approximation_method}; "
f"the available ones are: {self.AVAILABLE_APPROXIMATIONS}."
)
if approximation_method == "SR1":
approx_class = SR1Approx
elif approximation_method == "BFGS":
approx_class = BFGSApprox
elif approximation_method == "LEAST_SQUARES":
approx_class = LSTSQApprox
self.approximator = approx_class(history)
self.b_mat = None
self.x_ref = None
self.f_ref = None
self.fgrad_ref = None
[docs] def compute_approximation(
self,
funcname: str,
first_iter: int = 0,
last_iter: int = 0,
b0_mat=None,
at_most_niter: int = -1,
func_index=None,
):
"""Build the BFGS approximation for the Hessian.
Args:
funcname: The name of the function.
first_iter: The index of the first iteration.
last_iter: The index of the last iteration.
b0_mat: The Hessian matrix at the first iteration.
at_most_niter: The maximum number of iterations to take
func_index: The component of the function.
Returns:
An approximation of the Hessian matrix.
"""
self.b_mat, _, x_ref, grad_ref = self.approximator.build_approximation(
funcname=funcname,
first_iter=first_iter,
last_iter=last_iter,
b_mat0=b0_mat,
at_most_niter=at_most_niter,
return_x_grad=True,
func_index=func_index,
)
self.x_ref = x_ref
self.f_ref = self.approximator.f_ref
self.fgrad_ref = grad_ref
return self.b_mat
[docs] def compute_expected_value(self, expect: Sized, cov):
r"""Compute the expected value of the output.
Equal to :math:`0.5\mathbb{E}[e^TBe]`
where :math:`e` is the expected values
and :math:`B` the covariance matrix.
Args:
expect: The expected value of the inputs.
cov: The covariance matrix of the inputs.
Returns:
The expected value of the output.
Raises:
ValueError: When expectation and covariance matrices
have inconsistent shapes or when the Hessian approximation is missing.
"""
n_vars = len(expect)
if cov.shape != (n_vars, n_vars):
raise ValueError("Inconsistent expect and covariance matrices shapes")
if self.b_mat is None:
raise ValueError(
"Build Hessian approximation before computing expected_value_offset"
)
b_approx = 0.5 * self.b_mat
exp_val = np.trace(b_approx @ cov)
delta = expect - self.x_ref
exp_val += delta.T @ (b_approx @ delta)
return exp_val
[docs] def compute_variance(self, expect: Sized, cov):
r"""Compute the variance of the output.
Equal to :math:`0.5\mathbb{E}[e^TBe]`
where :math:`e` is the expected values
and :math:`B` the covariance matrix.
Args:
expect: The expected value of the inputs.
cov: The covariance matrix of the inputs.
Returns:
The variance of the output.
Raises:
ValueError: When expectation and covariance matrices
have inconsistent shapes or when the Hessian approximation is missing.
"""
if self.b_mat is None:
raise ValueError("Build Hessian approximation before computing variance")
n_vars = len(expect)
if cov.shape != (n_vars, n_vars):
raise ValueError("Inconsistent expect and covariance matrices shapes")
b_approx = 0.5 * self.b_mat
mu_cent = expect - self.x_ref
b_approx_cov = b_approx @ cov
v_mat = b_approx_cov @ b_approx_cov
b_approx_mu_cent = b_approx @ mu_cent
v_mat += 4 * b_approx_mu_cent.T @ (cov @ b_approx_mu_cent)
return 2 * np.trace(v_mat)
[docs] def compute_function_approximation(self, x_vars) -> float:
"""Compute a second order approximation of the function.
Args:
x_vars: The point on which the approximation is evaluated.
Returns:
A second order approximation of the function.
"""
if self.b_mat is None or self.x_ref is None:
raise ValueError(
"Build Hessian approximation before computing function approximation"
)
x_l = x_vars - self.x_ref
return 0.5 * x_l.T @ (self.b_mat @ x_l) + self.fgrad_ref.T @ x_l + self.f_ref
[docs] def compute_gradient_approximation(self, x_vars):
"""Computes a first order approximation of the gradient based on the hessian.
Args:
x_vars: The point on which the approximation is evaluated.
"""
if self.b_mat is None or self.fgrad_ref is None:
raise ValueError(
"Build Hessian approximation before computing function approximation"
)
x_l = x_vars - self.x_ref
return self.b_mat @ x_l + self.fgrad_ref
[docs] def montecarlo_average_var(
self, mean: Sized, cov, n_samples: int = 100000, func=None
):
"""Computes the variance and expected value using Monte Carlo approach.
Args:
mean: The mean value.
cov: The covariance matrix.
n_samples: The number of samples for the distribution.
func: If ``None``, the ``compute_function_approximation`` function,
otherwise a user function.
"""
n_dv = len(mean)
if not cov.shape == (n_dv, n_dv):
raise ValueError(
"Covariance matrix dimension " + "incompatible with mean dimensions"
)
ran = multivariate_normal(mean=mean, cov=cov, size=n_samples).T
vals = np.zeros(n_samples)
if func is None:
func = self.compute_function_approximation
for i in range(n_samples):
vals[i] = func(ran[:, i])
average = np.average(vals)
var = np.var(vals)
return average, var
