def calculate_bounds(
self,
objectives: Callable,
n_objectives: int,
x0: np.ndarray,
epsilons: np.ndarray,
bounds: Union[np.ndarray, None],
constraints: Optional[Callable],
method: Union[ScalarMethod, str, None],
) -> np.ndarray:
"""
Calculate the new bounds using Epsilon constraint method.
Args:
objectives (np.ndarray): The objective function values for each input vector.
n_objectives (int): Total number of objectives.
x0 (np.ndarray): Initial values for decision variables.
epsilons (np.ndarray): Previous iteration point.
bounds (Union[np.ndarray, None]): Bounds for decision variables.
constraints (Callable): Constraints of the problem.
method (Union[ScalarMethod, str, None]): The optimization method the scalarizer should be minimized with.
Returns:
np.ndarray: New lower bounds for objective functions.
"""
new_lower_bounds: np.ndarray = [None] * n_objectives
# set polish to False
method_e: ScalarMethod = ScalarMethod(
lambda x, _, **y: differential_evolution(x, **y),
method_args={"disp": False, "polish": False, "tol": 0.000001, "popsize": 10, "maxiter": 50000},
use_scipy=True,
)
# solve new lower bounds for each objective
for i in range(n_objectives):
eps = ECM.EpsilonConstraintMethod(
objectives,
i,
# take out the objective to be minimized
np.array([val for ind, val in enumerate(epsilons) if ind != i]),
constraints=constraints,
)
cons_evaluate = eps.evaluate_constraints
scalarized_objective = Scalarizer(objectives, eps)
minimizer = ScalarMinimizer(
scalarized_objective, bounds, constraint_evaluator=cons_evaluate, method=method_e
)
res = minimizer.minimize(x0)
# store objective function values as new lower bounds
new_lower_bounds[i] = objectives(res["x"])[0][i]
return new_lower_bounds
def test_scipy_minimize_cons():
solver = ScalarMinimizer(
simple_problem,
np.array([[0.0, 0.0, 0.0], [1.0, 1.0, 1.0]]).T,
simple_constr,
"scipy_minimize",
)
res = solver.minimize(np.array([0.21, 0.999, 0.001]))
assert not res["success"]
def test_dummy_cons():
method = ScalarMethod(dummy_minimizer)
solver = ScalarMinimizer(simple_problem, np.array([[0, 0, 0], [1, 1, 1]]),
simple_constr, method)
res = solver.minimize(np.array([0.5, 0.5, 0.1]))
assert res["success"]
res = solver.minimize(np.array([0.5, 0.5, 0.5]))
assert not res["success"]
def test_dummy_no_cons():
method = ScalarMethod(dummy_minimizer)
solver = ScalarMinimizer(simple_problem, np.array([[0, 0, 0], [1, 1, 1]]),
None, method)
x0 = np.array([0.5, 0.5, 0.5])
res = solver.minimize(x0)
assert np.array_equal(res["x"], x0)
assert res["success"]
assert (res["message"] ==
"I just retruned the initial guess as the optimal solution.")
def solve_asf(
self,
problem: Union[MOProblem, DiscreteDataProblem],
ref_point: np.ndarray,
method: Optional[ScalarMethod] = None,
):
"""
Solve the achievement scalarizing function
Args:
problem (MOProblem): The problem
ref_point: A reference point
method (Optional[ScalarMethod], optional): A method provided to the scalar minimizer
Returns:
np.ndarray: The decision vector which solves the achievement scalarizing function
"""
asf = SimpleASF(np.ones(ref_point.shape))
if isinstance(problem, MOProblem):
scalarizer = Scalarizer(
lambda x: problem.evaluate(x).objectives,
asf,
scalarizer_args={"reference_point": np.atleast_2d(ref_point)},
)
if problem.n_of_constraints > 0:
_con_eval = lambda x: problem.evaluate(x).constraints.squeeze()
else:
_con_eval = None
solver = ScalarMinimizer(
scalarizer,
problem.get_variable_bounds(),
constraint_evaluator=_con_eval,
method=method,
)
res = solver.minimize(problem.get_variable_upper_bounds() / 2)
if res["success"]:
return res["x"]
else:
raise ParetoNavigatorException(
"Could solve achievement scalarizing function")
else: # Discrete case
# Find closest objective to ref point
scalarizer = DiscreteScalarizer(asf,
{"reference_point": ref_point})
solver = DiscreteMinimizer(scalarizer)
res = solver.minimize(problem.objectives)
return res['x']
def test_scipy_de_cons():
solver = ScalarMinimizer(
simple_problem,
np.array([[0.0, 0.0, 0.0], [1.0, 1.0, 1.0]]).T,
simple_constr,
"scipy_de",
)
res = solver.minimize(None)
assert res["success"]
assert np.all(np.array(res["constr"]) >= 0)
def __init__(self,
problem: MOProblem,
scalar_method: Optional[ScalarMethod] = None):
# check if ideal and nadir are defined
if problem.ideal is None or problem.nadir is None:
# TODO: use same method as defined in scalar_method
ideal, nadir = payoff_table_method(problem)
self._ideal = ideal
self._nadir = nadir
else:
self._ideal = problem.ideal
self._nadir = problem.nadir
self._scalar_method = scalar_method
# generate Pareto optimal starting point
asf = SimpleASF(np.ones(self._ideal.shape))
scalarizer = Scalarizer(
lambda x: problem.evaluate(x).objectives,
asf,
scalarizer_args={"reference_point": np.atleast_2d(self._ideal)},
)
if problem.n_of_constraints > 0:
_con_eval = lambda x: problem.evaluate(x).constraints.squeeze()
else:
_con_eval = None
solver = ScalarMinimizer(
scalarizer,
problem.get_variable_bounds(),
constraint_evaluator=_con_eval,
method=self._scalar_method,
)
# TODO: fix tools to check for scipy methods in general and delete me!
solver._use_scipy = True
res = solver.minimize(problem.get_variable_upper_bounds() / 2)
if res["success"]:
self._current_solution = res["x"]
self._current_objectives = problem.evaluate(
self._current_solution).objectives.squeeze()
self._archive_solutions = []
self._archive_objectives = []
self._state = "classify"
super().__init__(problem)
def solve_asf(
self,
ref_point: np.ndarray,
x0: np.ndarray,
preferential_factors: np.ndarray,
nadir: np.ndarray,
utopian: np.ndarray,
objectives: Callable,
variable_bounds: Optional[np.ndarray] = None,
method: Union[ScalarMethod, str, None] = None,
) -> dict:
"""
Solve Achievement scalarizing function.
Args:
ref_point (np.ndarray): Reference point.
x0 (np.ndarray): Initial values for decision variables.
preferential_factors (np.ndarray): Preferential factors on how much would the Decision Maker wish to improve
the values of each objective function.
nadir (np.ndarray): Nadir vector.
utopian (np.ndarray): Utopian vector.
objectives (np.ndarray): The objective function values for each input vector.
variable_bounds (Optional[np.ndarray)]: Lower and upper bounds of each variable
as a 2D numpy array. If undefined variables, None instead.
method (Union[ScalarMethod, str, None]): The optimization method the scalarizer should be minimized with.
Returns:
dict: A dictionary with at least the following entries: 'x' indicating the optimal variables found,
'fun' the optimal value of the optimized function, and 'success' a boolean indicating whether
the optimization was conducted successfully.
"""
if variable_bounds is None:
# set all bounds as [-inf, inf]
variable_bounds = np.array([[-np.inf, np.inf]] * x0.shape[0])
# scalarize problem using reference point
asf = ReferencePointASF(preferential_factors, nadir, utopian, rho=1e-4)
asf_scalarizer = Scalarizer(
evaluator=objectives,
scalarizer=asf,
scalarizer_args={"reference_point": ref_point})
# minimize
minimizer = ScalarMinimizer(asf_scalarizer,
variable_bounds,
method=method)
return minimizer.minimize(x0)
def calculate_new_solutions(
self,
number_of_solutions: int,
levels: np.ndarray,
improve_inds: np.ndarray,
improve_until_inds: np.ndarray,
acceptable_inds: np.ndarray,
impaire_until_inds: np.ndarray,
free_inds: np.ndarray,
) -> Tuple[NimbusSaveRequest, SimplePlotRequest]:
"""Calcualtes new solutions based on classifications supplied by the decision maker by
solving ASF problems.
Args:
number_of_solutions (int): Number of solutions, should be between 1 and 4.
levels (np.ndarray): Aspiration and upper bounds relevant to the some of the classifications.
improve_inds (np.ndarray): Indices corresponding to the objectives which should be improved.
improve_until_inds (np.ndarray): Like above, but improved until an aspiration level is reached.
acceptable_inds (np.ndarray): Indices of objectives which are acceptable as they are now.
impaire_until_inds (np.ndarray): Indices of objectives which may be impaired until an upper limit is
reached.
free_inds (np.ndarray): Indices of objectives which may change freely.
Returns:
Tuple[NimbusSaveRequest, SimplePlotRequest]: A save request with the newly computed soutions, and
a plot request to visualize said solutions.
"""
results = []
# always computed
asf_1 = MaxOfTwoASF(self._nadir, self._ideal, improve_inds,
improve_until_inds)
def cons_1(
x: np.ndarray,
f_current: np.ndarray = self._current_objectives,
levels: np.ndarray = levels,
improve_until_inds: np.ndarray = improve_until_inds,
improve_inds: np.ndarray = improve_inds,
impaire_until_inds: np.ndarray = impaire_until_inds,
):
f = self._problem.evaluate(x).objectives.squeeze()
res_1 = f_current[improve_inds] - f[improve_inds]
res_2 = f_current[improve_until_inds] - f[improve_until_inds]
res_3 = levels[impaire_until_inds] - f_current[impaire_until_inds]
res = np.hstack((res_1, res_2, res_3))
if self._problem.n_of_constraints > 0:
res_prob = self._problem.evaluate(x).constraints.squeeze()
return np.hstack((res_prob, res))
else:
return res
scalarizer_1 = Scalarizer(
lambda x: self._problem.evaluate(x).objectives,
asf_1,
scalarizer_args={"reference_point": levels},
)
solver_1 = ScalarMinimizer(
scalarizer_1,
self._problem.get_variable_bounds(),
cons_1,
method=self._scalar_method,
)
res_1 = solver_1.minimize(self._current_solution)
results.append(res_1)
if number_of_solutions > 1:
# create the reference point needed in the rest of the ASFs
z_bar = np.zeros(self._problem.n_of_objectives)
z_bar[improve_inds] = self._ideal[improve_inds]
z_bar[improve_until_inds] = levels[improve_until_inds]
z_bar[acceptable_inds] = self._current_objectives[acceptable_inds]
z_bar[impaire_until_inds] = levels[impaire_until_inds]
z_bar[free_inds] = self._nadir[free_inds]
# second ASF
asf_2 = StomASF(self._ideal)
# cons_2 can be used in the rest of the ASF scalarizations, it's not a bug!
if self._problem.n_of_constraints > 0:
cons_2 = lambda x: self._problem.evaluate(
x).constraints.squeeze()
else:
cons_2 = None
scalarizer_2 = Scalarizer(
lambda x: self._problem.evaluate(x).objectives,
asf_2,
scalarizer_args={"reference_point": z_bar},
)
solver_2 = ScalarMinimizer(
scalarizer_2,
self._problem.get_variable_bounds(),
cons_2,
method=self._scalar_method,
)
res_2 = solver_2.minimize(self._current_solution)
results.append(res_2)
if number_of_solutions > 2:
# asf 3
asf_3 = PointMethodASF(self._nadir, self._ideal)
scalarizer_3 = Scalarizer(
lambda x: self._problem.evaluate(x).objectives,
asf_3,
scalarizer_args={"reference_point": z_bar},
)
solver_3 = ScalarMinimizer(
scalarizer_3,
self._problem.get_variable_bounds(),
cons_2,
method=self._scalar_method,
)
res_3 = solver_3.minimize(self._current_solution)
results.append(res_3)
if number_of_solutions > 3:
# asf 4
asf_4 = AugmentedGuessASF(self._nadir, self._ideal, free_inds)
scalarizer_4 = Scalarizer(
lambda x: self._problem.evaluate(x).objectives,
asf_4,
scalarizer_args={"reference_point": z_bar},
)
solver_4 = ScalarMinimizer(
scalarizer_4,
self._problem.get_variable_bounds(),
cons_2,
method=self._scalar_method,
)
res_4 = solver_4.minimize(self._current_solution)
results.append(res_4)
# create the save request
solutions = [res["x"] for res in results]
objectives = [
self._problem.evaluate(x).objectives.squeeze() for x in solutions
]
save_request = NimbusSaveRequest(solutions, objectives)
msg = "Computed new solutions."
plot_request = self.create_plot_request(objectives, msg)
return save_request, plot_request
def compute_intermediate_solutions(
self,
solutions: np.ndarray,
n_desired: int,
) -> Tuple[NimbusSaveRequest, SimplePlotRequest]:
"""Computs intermediate solution between two solutions computed earlier.
Args:
solutions (np.ndarray): The solutions between which the intermediat solutions should
be computed.
n_desired (int): The number of intermediate solutions desired.
Raises:
NimbusException
Returns:
Tuple[NimbusSaveRequest, SimplePlotRequest]: A save request with the compured intermediate
points, and a plot request to visualize said points.
"""
# vector between the two solutions
between = solutions[0] - solutions[1]
norm = np.linalg.norm(between)
between_norm = between / norm
# the plus 2 assumes we are interested only in n_desired points BETWEEN the
# two supplied solutions
step_size = norm / (2 + n_desired)
intermediate_points = np.array([
solutions[1] + i * step_size * between_norm
for i in range(1, n_desired + 1)
])
# project each of the intermediate solutions to the Pareto front
intermediate_solutions = np.zeros(intermediate_points.shape)
intermediate_objectives = np.zeros(
(n_desired, self._problem.n_of_objectives))
asf = PointMethodASF(self._nadir, self._ideal)
for i in range(n_desired):
scalarizer = Scalarizer(
lambda x: self._problem.evaluate(x).objectives,
asf,
scalarizer_args={
"reference_point":
self._problem.evaluate(intermediate_points[i]).objectives
},
)
if self._problem.n_of_constraints > 0:
cons = lambda x: self._problem.evaluate(x).constraints.squeeze(
)
else:
cons = None
solver = ScalarMinimizer(
scalarizer,
self._problem.get_variable_bounds(),
cons,
method=self._scalar_method,
)
res = solver.minimize(self._current_solution)
intermediate_solutions[i] = res["x"]
intermediate_objectives[i] = self._problem.evaluate(
res["x"]).objectives
# create appropiate requests
save_request = NimbusSaveRequest(list(intermediate_solutions),
list(intermediate_objectives))
msg = "Computed intermediate solutions"
plot_request = self.create_plot_request(intermediate_objectives, msg)
return save_request, plot_request
def payoff_table_method_general(
objective_evaluator: Callable[[np.ndarray], np.ndarray],
n_of_objectives: int,
variable_bounds: np.ndarray,
constraint_evaluator: Optional[Callable[[np.ndarray], np.ndarray]] = None,
initial_guess: Optional[np.ndarray] = None,
solver_method: Optional[Union[ScalarMethod, str]] = "scipy_de",
) -> Tuple[np.ndarray, np.ndarray]:
"""Solves a representation for the nadir and ideal points for a
multiobjective minimization problem with objectives defined as the result
of some objective evaluator.
Args:
objective_evaluator (Callable[[np.ndarray], np.ndarray]): The
evaluator which returns the objective values given a set of
variabels.
n_of_objectives (int): Number of objectives returned by calling
objective_evaluator.
variable_bounds (np.ndarray): The lower and upper bounds of the
variables passed as argument to objective_evaluator. Should be a 2D
numpy array with the limits for each variable being on each row. The
first column should contain the lower bounds, and the second column
the upper bounds. Use np.inf to indicate no bounds.
constraint_evaluator (Optional[Callable[[np.ndarray], np.ndarray]], optional):
An evaluator accepting the same arguments as
objective_evaluator, which returns the constraint values of the
multiobjective minimization problem being solved. A negative
constraint value indicates a broken constraint. Defaults to None.
initial_guess (Optional[np.ndarray], optional): The initial guess
used for the variable values while solving the payoff table. The
relevancy of this parameter depends on the solver_method being used.
Defaults to None.
solver_method (Optional[Union[ScalarMethod, str]], optional): The
method to solve the scalarized problems in the payoff table method.
Defaults to "scipy_de", which ignores initial_guess.
Returns:
Tuple[np.ndarray, np.ndarray]: The representations computed using the
payoff table for the ideal and nadir points respectively.
"""
scalarizer = Scalarizer(
objective_evaluator,
weighted_scalarizer,
scalarizer_args={"ws": None},
)
solver = ScalarMinimizer(
scalarizer,
variable_bounds,
constraint_evaluator,
solver_method,
)
ws = np.eye(n_of_objectives)
po_table = np.zeros((n_of_objectives, n_of_objectives))
if initial_guess is None:
initial_guess = variable_bounds[:, 0]
for i in range(n_of_objectives):
scalarizer._scalarizer_args = {"ws": ws[i]}
opt_res = solver.minimize(initial_guess)
if not opt_res["success"]:
print(
"Unsuccessful optimization result encountered while computing a payoff table!"
)
po_table[i] = objective_evaluator(opt_res["x"])
ideal = np.diag(po_table)
nadir = np.max(po_table, axis=0)
return ideal, nadir
def solve_pareto_front_representation_general(
objective_evaluator: Callable[[np.ndarray], np.ndarray],
n_of_objectives: int,
variable_bounds: np.ndarray,
step: Optional[Union[np.ndarray, float]] = 0.1,
eps: Optional[float] = 1e-6,
ideal: Optional[np.ndarray] = None,
nadir: Optional[np.ndarray] = None,
constraint_evaluator: Optional[Callable[[np.ndarray], np.ndarray]] = None,
solver_method: Optional[Union[ScalarMethod, str]] = "scipy_de",
) -> Tuple[np.ndarray, np.ndarray]:
"""Computes a representation of a Pareto efficient front from a
multiobjective minimizatino problem. Does so by generating an evenly spaced
set of reference points (in the objective space), in the space spanned by
the supplied ideal and nadir points. The generated reference points are
then used to formulate achievement scalaraization problems, which when
solved, yield a representation of a Pareto efficient solution. The result
is guaranteed to contain only non-dominated solutions.
Args:
objective_evaluator (Callable[[np.ndarray], np.ndarray]): A vector
valued function returning objective values given an array of decision
variables.
n_of_objectives (int): Numbr of objectives returned by
objective_evaluator.
variable_bounds (np.ndarray): The upper and lower bounds of the
decision variables. Bound for each variable should be on the rows,
with the first column containing lower bounds, and the second column
upper bounds. Use np.inf to indicate no bounds.
step (Optional[Union[np.ndarray, float]], optional): Etiher an float
or an array of floats. If a single float is given, generates
reference points with the objectives having values a step apart
between the ideal and nadir points. If an array of floats is given,
use the steps defined in the array for each objective's values.
Default to 0.1.
eps (Optional[float], optional): An offset to be added to the nadir
value to keep the nadir inside the range when generating reference
points. Defaults to 1e-6.
ideal (Optional[np.ndarray], optional): The ideal point of the
problem being solved. Defaults to None.
nadir (Optional[np.ndarray], optional): The nadir point of the
problem being solved. Defaults to None.
constraint_evaluator (Optional[Callable[[np.ndarray], np.ndarray]], optional):
An evaluator returning values for the constraints defined
for the problem. A negative value for a constraint indicates a breach
of that constraint. Defaults to None.
solver_method (Optional[Union[ScalarMethod, str]], optional): The
method used to minimize the achievement scalarization problems
arising when calculating Pareto efficient solutions. Defaults to
"scipy_de".
Raises:
MCDMUtilityException: Mismatching sizes of the supplied ideal and
nadir points between the step, when step is an array. Or the type of
step is something else than np.ndarray of float.
Returns:
Tuple[np.ndarray, np.ndarray]: A tuple containing representationns of
the Pareto optimal variable values, and the corresponsing objective
values.
Note:
The objective evaluator should be defined such that minimization is
expected in each of the objectives.
"""
if ideal is None or nadir is None:
# compure ideal and nadir using payoff table
ideal, nadir = payoff_table_method_general(
objective_evaluator,
n_of_objectives,
variable_bounds,
constraint_evaluator,
)
# use ASF to (almost) guarantee Pareto optimality.
asf = PointMethodASF(nadir, ideal)
scalarizer = Scalarizer(objective_evaluator,
asf,
scalarizer_args={"reference_point": None})
solver = ScalarMinimizer(scalarizer,
bounds=variable_bounds,
method=solver_method)
# bounds to be used to compute slices
stacked = np.stack((ideal, nadir)).T
lower_slice_b, upper_slice_b = np.min(stacked, axis=1), np.max(stacked,
axis=1)
if type(step) is float:
slices = [
slice(start, stop + eps, step)
for (start, stop) in zip(lower_slice_b, upper_slice_b)
]
elif type(step) is np.ndarray:
if not ideal.shape == nadir.shape == step.shape:
raise MCDMUtilityException(
"The shapes of the supplied step array does not match the "
"shape of the ideal and nadir points.")
slices = [
slice(start, stop + eps, s)
for (start, stop, s) in zip(lower_slice_b, upper_slice_b, step)
]
else:
raise MCDMUtilityException(
"step must be either a numpy array or an float.")
z_mesh = np.mgrid[slices].reshape(len(ideal), -1).T
p_front_objectives = np.zeros(z_mesh.shape)
p_front_variables = np.zeros(
(len(p_front_objectives), len(variable_bounds.squeeze())))
for i, z in enumerate(z_mesh):
scalarizer._scalarizer_args = {"reference_point": z}
res = solver.minimize(None)
if not res["success"]:
print("Non successfull optimization")
p_front_objectives[i] = np.nan
p_front_variables[i] = np.nan
continue
# check for dominance, accept only non-dominated solutions
f_i = objective_evaluator(res["x"])
if not np.all(f_i > p_front_objectives[:i]
[~np.all(np.isnan(p_front_objectives[:i]), axis=1)]):
p_front_objectives[i] = f_i
p_front_variables[i] = res["x"]
elif i < 1:
p_front_objectives[i] = f_i
p_front_variables[i] = res["x"]
else:
p_front_objectives[i] = np.nan
p_front_variables[i] = np.nan
return (
p_front_variables[~np.all(np.isnan(p_front_variables), axis=1)],
p_front_objectives[~np.all(np.isnan(p_front_objectives), axis=1)],
)
res_ind = dminimizer.minimize(non_dominated_points)
assert res_ind == 1
# first point as closest, but invalid
dscalarizer._scalarizer_args = {
"reference_point": np.array([0.2, 0.4, 0.6, 0.8])
}
res_ind = dminimizer.minimize(non_dominated_points)
assert res_ind == 1
# all points invalid
dminimizer._constraint_evaluator = lambda x: x[:, 0] > 1.0
with pytest.raises(ScalarSolverException):
_ = dminimizer.minimize(non_dominated_points)
if __name__ == "__main__":
solver = ScalarMinimizer(
simple_problem,
np.array([[0.0, 0.0, 0.0], [1.0, 1.0, 1.0]]).T,
simple_constr,
"scipy_de",
)
res = solver.minimize(np.array([0.21, 0.999, 0.001]))
print(res)
