# -*- coding: utf-8 -*-
# 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 - API and implementation and/or documentation
# :author: Benoit Pauwels
# OTHER AUTHORS - MACROSCOPIC CHANGES
"""
Post-optimal analysis
*********************
"""
from __future__ import division, unicode_literals
import logging
from numpy import atleast_1d, dot, hstack, ndarray, vstack, zeros_like
from numpy.linalg.linalg import norm
from gemseo.algos.lagrange_multipliers import LagrangeMultipliers
LOGGER = logging.getLogger(__name__)
[docs]class PostOptimalAnalysis(object):
r"""
Post-optimal analysis of a parameterized optimization problem.
Consider the parameterized optimization problem below, whose objective and
constraint functions depend on both the optimization variable :math:`x` and
a parameter :math:`p`.
.. math::
\begin{aligned}
& \text{Minimize} & & f(x,p) \\
& \text{relative to} & & x \\
& \text{subject to} & & \left\{\begin{aligned}
& g(x,p)\le0, \\
& h(x,p)=0, \\
& \ell\le x\le u.
\end{aligned}\right.
\end{aligned}
Denote :math:`x^\ast(p)` a solution of the problem, which depends on
:math:`p`.
The post-optimal analysis consists in computing the following total
derivative:
.. math::
\newcommand{\total}{\mathrm{d}}
\frac{\total f(x^\ast(p),p)}{\total p}(p)
=\frac{\partial f}{\partial p}(x^\ast(p),p)
+\lambda_g^\top\frac{\partial g}{\partial p}(x^\ast(p),p)
+\lambda_h^\top\frac{\partial h}{\partial p}(x^\ast(p),p),
where :math:`\lambda_g` and :math:`\lambda_h` are the Lagrange multipliers
of :math:`x^\ast(p)`.
N.B. the equality above relies on the assumption that
.. math::
\newcommand{\total}{\mathrm{d}}
\lambda_g^\top\frac{\total g(x^\ast(p),p)}{\total p}(p)=0
\text{ and }
\lambda_h^\top\frac{\total h(x^\ast(p),p)}{\total p}(p)=0.
"""
# Dictionary key for term "Lagrange multipliers dot constraints Jacobian"
MULT_DOT_CONSTR_JAC = "mult_dot_constr_jac"
def __init__(self, opt_problem, ineq_tol=None):
"""Constructor.
:param opt_problem: solved optimization problem to be analyzed
:type opt_problem: OptimizationProblem
:param ineq_tol: tolerance to determine active inequality constraints.
If None, its value is fetched in the optimization problem.
"""
self.lagrange_computer = LagrangeMultipliers(opt_problem)
# N.B. at creation LagrangeMultipliers checks the optimization problem
self.opt_problem = opt_problem
# Get the optimal solution
self.x_opt = self.opt_problem.design_space.get_current_x()
# Get the objective name
outvars = self.opt_problem.objective.outvars
if len(outvars) != 1:
raise ValueError("The objective must be single-valued.")
self.outvars = outvars
# Set the tolerance on inequality constraints
if ineq_tol is None:
self.ineq_tol = self.opt_problem.ineq_tolerance
else:
self.ineq_tol = ineq_tol
[docs] def check_validity(self, total_jac, partial_jac, parameters, threshold):
"""Checks whether the assumption for post-optimal validity holds.
:param total_jac: total derivatives of the post-optimal constraints
:type total_jac: dict(dict(ndarray))
:param partial_jac: partial derivatives of the constraints
:type total_jac: dict(dict(ndarray))
:param parameters: names list of the optimization problem parameters
:type parameters: list(str)
:param threshold: tolerance on the validity assumption
:type threshold: number
"""
# Check the Jacobians
func_names = self.opt_problem.get_constraints_names()
self._check_jacobians(total_jac, func_names, parameters)
self._check_jacobians(partial_jac, func_names, parameters)
# Compute the Lagrange multipliers
multipliers = self.lagrange_computer.compute(self.x_opt, self.ineq_tol)
_, mul_ineq = multipliers.get(LagrangeMultipliers.INEQUALITY, ([], []))
_, mul_eq = multipliers.get(LagrangeMultipliers.EQUALITY, ([], []))
# Get the array to validate the inequality constraints
total_ineq_jac = self._get_act_ineq_jac(total_jac, parameters)
partial_ineq_jac = self._get_act_ineq_jac(partial_jac, parameters)
ineq_tot, ineq_part, ineq_corr = self._compute_validity(
total_ineq_jac, partial_ineq_jac, mul_ineq, parameters
)
# Get the array to validate the equality constraints
total_eq_jac = self._get_eq_jac(total_jac, parameters)
partial_eq_jac = self._get_eq_jac(partial_jac, parameters)
eq_tot, eq_part, eq_corr = self._compute_validity(
total_eq_jac, partial_eq_jac, mul_eq, parameters
)
# Compute the error
error_list = [arr for arr in [ineq_tot, eq_tot] if arr is not None]
part_norm_list = [arr for arr in [ineq_part, eq_part] if arr is not None]
if error_list and part_norm_list:
error = norm(vstack(error_list))
part_norm = norm(vstack(part_norm_list))
if part_norm > threshold:
error /= part_norm
else:
error = 0.0
# Assess the validity
valid = error < threshold
if valid:
LOGGER.info("Post-optimality is valid.")
else:
msg = "Post-optimality assumption is wrong by "
msg += str(error * 100.0) + "%."
LOGGER.info(msg)
return valid, ineq_corr, eq_corr
def _compute_validity(self, total_jac, partial_jac, multipliers, parameters):
"""Computes the arrays necessary to the validity check.
:param total_jac: total derivatives of the post-optimal constraints
:type total_jac: dict(dict(ndarray))
:param partial_jac: partial derivatives of the constraints
:type total_jac: dict(dict(ndarray))
:param multipliers: Lagrange multipliers
:type multipliers: ndarray
:param parameters: names list of the optimization problem parameters
:type parameters: list(str)
"""
corrections = dict.fromkeys(parameters, 0.0) # corrections terms
total_prod_list = []
partial_prod_list = []
for input_name in parameters:
total_jac_block = total_jac.get(input_name)
partial_jac_block = partial_jac.get(input_name)
if total_jac_block is not None and partial_jac_block is not None:
total_prod_block = dot(multipliers, total_jac_block)
partial_prod_block = dot(multipliers, partial_jac_block)
total_prod_list.append(total_prod_block)
partial_prod_list.append(partial_prod_block)
# Store the correction term
corrections[input_name] = -total_prod_block
if not self.opt_problem.minimize_objective:
corrections[input_name] *= -1.0
total_prod = hstack(total_prod_list) if total_prod_list else None
partial_prod = hstack(partial_prod_list) if partial_prod_list else None
return total_prod, partial_prod, corrections
[docs] def execute(self, outputs, inputs, functions_jac):
"""Performs the post-optimal analysis.
:param outputs: names list of the outputs to differentiate
:type outputs: list(str)
:param inputs: names list of the inputs w.r.t. which to differentiate
:type inputs: list(str)
:param functions_jac: Jacobians of the optimization functions w.r.t.
the differentiation inputs
:type functions_jac: dict(dict(ndarray))
"""
# Check the outputs
nondifferentiable_outputs = set(outputs) - set(self.outvars)
if nondifferentiable_outputs:
raise ValueError(
"Only the post-optimal Jacobian of "
+ self.outvars[0]
+ " can be computed"
+ ", not the one(s) of "
+ ", ".join(nondifferentiable_outputs)
+ "."
)
# Check the inputs and Jacobians consistency
func_names = self.outvars + self.opt_problem.get_constraints_names()
PostOptimalAnalysis._check_jacobians(functions_jac, func_names, inputs)
# Compute the Lagrange multipliers
self._compute_lagrange_multipliers()
# Compute the Jacobian of the Lagrangian
jac = self.compute_lagrangian_jac(functions_jac, inputs)
return jac
@staticmethod
def _check_jacobians(functions_jac, func_names, inputs):
"""Checks the consistency of the Jacobians with the required inputs.
:param functions_jac: Jacobians of the optimization function w.r.t. the
differentiation inputs
:type functions_jac: dict(dict(ndarray))
:param func_names: names list of the functions differentiated
:type func_names: list(str)
:param inputs: names list of the inputs w.r.t. which to differentiate
:type inputs: list(str)
"""
# Check the consistency of the Jacobians
for output_name in func_names:
jac_out = functions_jac.get(output_name)
if jac_out is None:
raise ValueError("Jacobian of " + output_name + " is missing.")
for input_name in inputs:
jac_block = jac_out.get(input_name)
if jac_block is None:
raise ValueError(
"Jacobian of "
+ output_name
+ " with respect to "
+ input_name
+ " is missing."
)
if not isinstance(jac_block, ndarray):
raise ValueError(
"Jacobian of "
+ output_name
+ " with respect to "
+ input_name
+ " must be of type ndarray."
)
if len(jac_block.shape) != 2:
raise ValueError(
"Jacobian of "
+ output_name
+ " with respect to "
+ input_name
+ " must be a 2-dimensional ndarray."
)
def _compute_lagrange_multipliers(self):
"""Computes the Lagrange multipliers at the solution."""
self.lagrange_computer.compute(self.x_opt, self.ineq_tol)
[docs] def compute_lagrangian_jac(self, functions_jac, inputs):
"""Computes the Jacobian of the Lagrangian.
:param functions_jac: Jacobians of the optimization function w.r.t. the
differentiation inputs
:type functions_jac: dict(dict(ndarray))
:param inputs: names list of the inputs w.r.t. which to differentiate
:type inputs: list(str)
"""
# Get the Lagrange multipliers
multipliers = self.lagrange_computer.lagrange_multipliers
if multipliers is None:
self._compute_lagrange_multipliers()
multipliers = self.lagrange_computer.lagrange_multipliers
_, mul_ineq = multipliers.get(LagrangeMultipliers.INEQUALITY, ([], []))
_, mul_eq = multipliers.get(LagrangeMultipliers.EQUALITY, ([], []))
# Build the Jacobians of the active constraints
act_ineq_jac = self._get_act_ineq_jac(functions_jac, inputs)
eq_jac = self._get_eq_jac(functions_jac, inputs)
jac = {self.outvars[0]: dict(), self.MULT_DOT_CONSTR_JAC: dict()}
for input_name in inputs:
# Contribution of the objective
jac_obj_arr = functions_jac[self.outvars[0]][input_name]
jac_cstr_arr = zeros_like(jac_obj_arr)
# Contributions of the inequality constraints
jac_ineq_arr = act_ineq_jac.get(input_name)
if jac_ineq_arr is not None:
jac_cstr_arr += dot(mul_ineq, jac_ineq_arr)
# Contributions of the equality constraints
jac_eq_arr = eq_jac.get(input_name)
if jac_eq_arr is not None:
jac_cstr_arr += dot(mul_eq, jac_eq_arr)
# Assemble the Jacobian of the Lagrangian
if not self.opt_problem.minimize_objective:
jac_cstr_arr *= -1.0
jac[self.MULT_DOT_CONSTR_JAC][input_name] = jac_cstr_arr
jac[self.outvars[0]][input_name] = jac_obj_arr + jac_cstr_arr
return jac
def _get_act_ineq_jac(self, jacobians, inputs):
"""Builds the Jacobian of the active inequality constraints.
:param jacobians: Jacobians of the inequality constraints w.r.t. the
differentiation inputs
:type jacobians: dict(dict(ndarray))
:param inputs: names list of the differentiation inputs
:type inputs: list(str)
"""
# Get the active constraints
ineq_cstr = self.opt_problem.get_active_ineq_constraints(
self.x_opt, self.ineq_tol
)
# Build the Jacobian
jac_dict = dict()
for input_name in inputs:
jac_input_list = []
for func, act_set in ineq_cstr.items():
if True in act_set:
jac_block = jacobians[func.name][input_name]
jac_block = jac_block[atleast_1d(act_set), :]
jac_input_list.append(jac_block)
if jac_input_list:
jac_input_arr = vstack(jac_input_list)
jac_dict[input_name] = jac_input_arr
return jac_dict
def _get_eq_jac(self, jacobians, inputs):
"""Builds the Jacobian of the equality constraints.
:param jacobians: Jacobians of the equality constraints w.r.t. the
differentiation inputs
:type jacobians: dict(dict(ndarray))
:param inputs: names list of the differentiation inputs
:type inputs: list(str)
"""
eq_cstr = self.opt_problem.get_eq_constraints()
jac_dict = dict()
for input_name in inputs:
jac_input_list = [jacobians[func.name][input_name] for func in eq_cstr]
if jac_input_list:
jac_input_arr = vstack(jac_input_list)
jac_dict[input_name] = jac_input_arr
return jac_dict