def test_constraint_violation_fn(self):
def constr_f(x):
return [x[0] + x[1]]
def constr_f2(x):
return [x[0]**2 + x[1], x[0] - x[1]]
nlc = NonlinearConstraint(constr_f, -np.inf, 1.9)
solver = DifferentialEvolutionSolver(rosen, [(0, 2), (0, 2)],
constraints=(nlc))
cv = solver._constraint_violation_fn([1.0, 1.0])
assert_almost_equal(cv, 0.1)
nlc2 = NonlinearConstraint(constr_f2, -np.inf, 1.8)
solver = DifferentialEvolutionSolver(rosen, [(0, 2), (0, 2)],
constraints=(nlc, nlc2))
# for multiple constraints the constraint violations should
# be concatenated.
cv = solver._constraint_violation_fn([1.2, 1.])
assert_almost_equal(cv, [0.3, 0.64, 0])
cv = solver._constraint_violation_fn([2., 2.])
assert_almost_equal(cv, [2.1, 4.2, 0])
# should accept valid values
cv = solver._constraint_violation_fn([0.5, 0.5])
assert_almost_equal(cv, [0., 0., 0.])
def test_accept_trial(self):
# _accept_trial(self, energy_trial, feasible_trial, cv_trial,
# energy_orig, feasible_orig, cv_orig)
def constr_f(x):
return [x[0] + x[1]]
nlc = NonlinearConstraint(constr_f, -np.inf, 1.9)
solver = DifferentialEvolutionSolver(rosen, [(0, 2), (0, 2)],
constraints=(nlc))
fn = solver._accept_trial
# both solutions are feasible, select lower energy
assert fn(0.1, True, np.array([0.]), 1.0, True, np.array([0.]))
assert (fn(1.0, True, np.array([0.]), 0.1, True,
np.array([0.])) == False)
assert fn(0.1, True, np.array([0.]), 0.1, True, np.array([0.]))
# trial is feasible, original is not
assert fn(9.9, True, np.array([0.]), 1.0, False, np.array([1.]))
# trial and original are infeasible
# cv_trial have to be <= cv_original to be better
assert (fn(0.1, False, np.array([0.5, 0.5]), 1.0, False,
np.array([1., 1.0])))
assert (fn(0.1, False, np.array([0.5, 0.5]), 1.0, False,
np.array([1., 0.50])))
assert (fn(1.0, False, np.array([0.5, 0.5]), 1.0, False,
np.array([1., 0.4])) == False)
def test_L6(self):
# Lampinen ([5]) test problem 6
def f(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
fun = (x[1] - 10)**3 + (x[2] - 20)**3
return fun
def c1(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return [(x[1] - 5)**2 + (x[2] - 5)**2 - 100,
-(x[1] - 6)**2 - (x[2] - 5)**2 + 82.81]
N = NonlinearConstraint(c1, 0, np.inf)
bounds = [(13, 100), (0, 100)]
constraints = (N)
res = differential_evolution(f,
bounds,
strategy='rand1bin',
seed=1234,
constraints=constraints,
tol=1e-7)
x_opt = (14.095, 0.84296)
f_opt = -6961.814744
assert_allclose(f(x_opt), f_opt, atol=1e-6)
assert_allclose(res.fun, f_opt, atol=0.001)
assert_allclose(res.x, x_opt, atol=1e-4)
assert res.success
assert_(np.all(np.array(c1(res.x)) >= 0))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
def test_L9(self):
# Lampinen ([5]) test problem 9
def f(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return x[1]**2 + (x[2] - 1)**2
def c1(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return [x[2] - x[1]**2]
N = NonlinearConstraint(c1, [-.001], [0.001])
bounds = [(-1, 1)] * 2
constraints = (N)
res = differential_evolution(f,
bounds,
strategy='rand1bin',
seed=1234,
constraints=constraints)
x_opt = [np.sqrt(2) / 2, 0.5]
f_opt = 0.75
assert_allclose(f(x_opt), f_opt)
assert_allclose(np.abs(res.x), x_opt, atol=1e-3)
assert_allclose(res.fun, f_opt, atol=1e-3)
assert res.success
assert_(np.all(np.array(c1(res.x)) >= -0.001))
assert_(np.all(np.array(c1(res.x)) <= 0.001))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
def test_constraint_wrapper(self):
lb = np.array([0, 20, 30])
ub = np.array([0.5, np.inf, 70])
x0 = np.array([1, 2, 3])
pc = _ConstraintWrapper(Bounds(lb, ub), x0)
assert (pc.violation(x0) > 0).any()
assert (pc.violation([0.25, 21, 31]) == 0).all()
x0 = np.array([1, 2, 3, 4])
A = np.array([[1, 2, 3, 4], [5, 0, 0, 6], [7, 0, 8, 0]])
pc = _ConstraintWrapper(LinearConstraint(A, -np.inf, 0), x0)
assert (pc.violation(x0) > 0).any()
assert (pc.violation([-10, 2, -10, 4]) == 0).all()
pc = _ConstraintWrapper(LinearConstraint(csr_matrix(A), -np.inf, 0),
x0)
assert (pc.violation(x0) > 0).any()
assert (pc.violation([-10, 2, -10, 4]) == 0).all()
def fun(x):
return A.dot(x)
nonlinear = NonlinearConstraint(fun, -np.inf, 0)
pc = _ConstraintWrapper(nonlinear, [-10, 2, -10, 4])
assert (pc.violation(x0) > 0).any()
assert (pc.violation([-10, 2, -10, 4]) == 0).all()
def test_L5(self):
# Lampinen ([5]) test problem 5
def f(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
fun = (np.sin(2 * np.pi * x[1])**3 * np.sin(2 * np.pi * x[2]) /
(x[1]**3 * (x[1] + x[2])))
return -fun # maximize
def c1(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return [x[1]**2 - x[2] + 1, 1 - x[1] + (x[2] - 4)**2]
N = NonlinearConstraint(c1, -np.inf, 0)
bounds = [(0, 10)] * 2
constraints = (N)
res = differential_evolution(f,
bounds,
strategy='rand1bin',
seed=1234,
constraints=constraints)
x_opt = (1.22797135, 4.24537337)
f_opt = -0.095825
print(res)
assert_allclose(f(x_opt), f_opt, atol=2e-5)
assert_allclose(res.fun, f_opt, atol=1e-4)
assert res.success
assert_(np.all(np.array(c1(res.x)) <= 0))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
def test_prepare_constraint_infeasible_x0():
lb = np.array([0, 20, 30])
ub = np.array([0.5, np.inf, 70])
x0 = np.array([1, 2, 3])
enforce_feasibility = np.array([False, True, True], dtype=bool)
bounds = Bounds(lb, ub, enforce_feasibility)
pytest.raises(ValueError, PreparedConstraint, bounds, x0)
x0 = np.array([1, 2, 3, 4])
A = np.array([[1, 2, 3, 4], [5, 0, 0, 6], [7, 0, 8, 0]])
enforce_feasibility = np.array([True, True, True], dtype=bool)
linear = LinearConstraint(A, -np.inf, 0, enforce_feasibility)
pytest.raises(ValueError, PreparedConstraint, linear, x0)
def fun(x):
return A.dot(x)
def jac(x):
return A
def hess(x, v):
return sps.csr_matrix((4, 4))
nonlinear = NonlinearConstraint(fun, -np.inf, 0, jac, hess,
enforce_feasibility)
pytest.raises(ValueError, PreparedConstraint, nonlinear, x0)
def test_concatenation():
rng = np.random.RandomState(0)
n = 4
x0 = np.random.rand(n)
f1 = x0
J1 = np.eye(n)
lb1 = [-1, -np.inf, -2, 3]
ub1 = [1, np.inf, np.inf, 3]
bounds = Bounds(lb1, ub1, [False, False, True, False])
fun, jac, hess = create_quadratic_function(n, 5, rng)
f2 = fun(x0)
J2 = jac(x0)
lb2 = [-10, 3, -np.inf, -np.inf, -5]
ub2 = [10, 3, np.inf, 5, np.inf]
nonlinear = NonlinearConstraint(fun, lb2, ub2, jac, hess,
[True, False, False, True, False])
for sparse_jacobian in [False, True]:
bounds_prepared = PreparedConstraint(bounds, x0, sparse_jacobian)
nonlinear_prepared = PreparedConstraint(nonlinear, x0, sparse_jacobian)
c1 = CanonicalConstraint.from_PreparedConstraint(bounds_prepared)
c2 = CanonicalConstraint.from_PreparedConstraint(nonlinear_prepared)
c = CanonicalConstraint.concatenate([c1, c2], sparse_jacobian)
assert_equal(c.n_eq, 2)
assert_equal(c.n_ineq, 7)
c_eq, c_ineq = c.fun(x0)
assert_array_equal(c_eq, [f1[3] - lb1[3], f2[1] - lb2[1]])
assert_array_equal(c_ineq, [
lb1[2] - f1[2], f1[0] - ub1[0], lb1[0] - f1[0], f2[3] - ub2[3],
lb2[4] - f2[4], f2[0] - ub2[0], lb2[0] - f2[0]
])
J_eq, J_ineq = c.jac(x0)
if sparse_jacobian:
J_eq = J_eq.toarray()
J_ineq = J_ineq.toarray()
assert_array_equal(J_eq, np.vstack((J1[3], J2[1])))
assert_array_equal(
J_ineq,
np.vstack((-J1[2], J1[0], -J1[0], J2[3], -J2[4], J2[0], -J2[0])))
v_eq = rng.rand(c.n_eq)
v_ineq = rng.rand(c.n_ineq)
v = np.zeros(5)
v[1] = v_eq[1]
v[3] = v_ineq[3]
v[4] = -v_ineq[4]
v[0] = v_ineq[5] - v_ineq[6]
H = c.hess(x0, v_eq, v_ineq).dot(np.eye(n))
assert_array_equal(H, hess(x0, v))
assert_array_equal(c.keep_feasible,
[True, False, False, True, False, True, True])
def test_L4(self):
# Lampinen ([5]) test problem 4
def f(x):
return np.sum(x[:3])
A = np.zeros((4, 9))
A[1, [4, 6]] = 0.0025, 0.0025
A[2, [5, 7, 4]] = 0.0025, 0.0025, -0.0025
A[3, [8, 5]] = 0.01, -0.01
A = A[1:, 1:]
b = np.array([1, 1, 1])
def c1(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return [
x[1] * x[6] - 833.33252 * x[4] - 100 * x[1] + 83333.333,
x[2] * x[7] - 1250 * x[5] - x[2] * x[4] + 1250 * x[4],
x[3] * x[8] - 1250000 - x[3] * x[5] + 2500 * x[5]
]
L = LinearConstraint(A, -np.inf, 1)
N = NonlinearConstraint(c1, 0, np.inf)
bounds = [(100, 10000)] + [(1000, 10000)] * 2 + [(10, 1000)] * 5
constraints = (L, N)
with suppress_warnings() as sup:
sup.filter(UserWarning)
res = differential_evolution(f,
bounds,
strategy='rand1bin',
seed=1234,
constraints=constraints,
popsize=3)
f_opt = 7049.248
x_opt = [
579.306692, 1359.97063, 5109.9707, 182.0177, 295.601172, 217.9823,
286.416528, 395.601172
]
assert_allclose(f(x_opt), f_opt, atol=0.001)
assert_allclose(res.fun, f_opt, atol=0.001)
# selectively use higher tol here for 32-bit
# Windows based on gh-11693
if (platform.system() == 'Windows' and np.dtype(np.intp).itemsize < 8):
assert_allclose(res.x, x_opt, rtol=2.4e-6, atol=0.0035)
else:
assert_allclose(res.x, x_opt, atol=0.002)
assert res.success
assert_(np.all(A @ res.x <= b))
assert_(np.all(np.array(c1(res.x)) >= 0))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
def test_L3(self):
# Lampinen ([5]) test problem 3
def f(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
fun = (x[1]**2 + x[2]**2 + x[1] * x[2] - 14 * x[1] - 16 * x[2] +
(x[3] - 10)**2 + 4 * (x[4] - 5)**2 + (x[5] - 3)**2 + 2 *
(x[6] - 1)**2 + 5 * x[7]**2 + 7 * (x[8] - 11)**2 + 2 *
(x[9] - 10)**2 + (x[10] - 7)**2 + 45)
return fun # maximize
A = np.zeros((4, 11))
A[1, [1, 2, 7, 8]] = -4, -5, 3, -9
A[2, [1, 2, 7, 8]] = -10, 8, 17, -2
A[3, [1, 2, 9, 10]] = 8, -2, -5, 2
A = A[1:, 1:]
b = np.array([-105, 0, -12])
def c1(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return [
3 * x[1] - 6 * x[2] - 12 * (x[9] - 8)**2 + 7 * x[10],
-3 * (x[1] - 2)**2 - 4 * (x[2] - 3)**2 - 2 * x[3]**2 +
7 * x[4] + 120, -x[1]**2 - 2 * (x[2] - 2)**2 +
2 * x[1] * x[2] - 14 * x[5] + 6 * x[6],
-5 * x[1]**2 - 8 * x[2] - (x[3] - 6)**2 + 2 * x[4] + 40, -0.5 *
(x[1] - 8)**2 - 2 * (x[2] - 4)**2 - 3 * x[5]**2 + x[6] + 30
]
L = LinearConstraint(A, b, np.inf)
N = NonlinearConstraint(c1, 0, np.inf)
bounds = [(-10, 10)] * 10
constraints = (L, N)
with suppress_warnings() as sup:
sup.filter(UserWarning)
res = differential_evolution(f,
bounds,
seed=1234,
constraints=constraints,
popsize=3)
x_opt = (2.171996, 2.363683, 8.773926, 5.095984, 0.9906548, 1.430574,
1.321644, 9.828726, 8.280092, 8.375927)
f_opt = 24.3062091
assert_allclose(f(x_opt), f_opt, atol=1e-5)
assert_allclose(res.x, x_opt, atol=1e-6)
assert_allclose(res.fun, f_opt, atol=1e-5)
assert res.success
assert_(np.all(A @ res.x >= b))
assert_(np.all(np.array(c1(res.x)) >= 0))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
def test_L8(self):
def f(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
fun = 3 * x[1] + 0.000001 * x[1]**3 + 2 * x[2] + 0.000002 / 3 * x[
2]**3
return fun
A = np.zeros((3, 5))
A[1, [4, 3]] = 1, -1
A[2, [3, 4]] = 1, -1
A = A[1:, 1:]
b = np.array([-.55, -.55])
def c1(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return [
1000 * np.sin(-x[3] - 0.25) + 1000 * np.sin(-x[4] - 0.25) +
894.8 - x[1], 1000 * np.sin(x[3] - 0.25) +
1000 * np.sin(x[3] - x[4] - 0.25) + 894.8 - x[2],
1000 * np.sin(x[4] - 0.25) +
1000 * np.sin(x[4] - x[3] - 0.25) + 1294.8
]
L = LinearConstraint(A, b, np.inf)
N = NonlinearConstraint(c1, np.full(3, -0.001), np.full(3, 0.001))
bounds = [(0, 1200)] * 2 + [(-.55, .55)] * 2
constraints = (L, N)
with suppress_warnings() as sup:
sup.filter(UserWarning)
# original Lampinen test was with rand1bin, but that takes a
# huge amount of CPU time. Changing strategy to best1bin speeds
# things up a lot
res = differential_evolution(f,
bounds,
strategy='best1bin',
seed=1234,
constraints=constraints,
maxiter=5000)
x_opt = (679.9453, 1026.067, 0.1188764, -0.3962336)
f_opt = 5126.4981
assert_allclose(f(x_opt), f_opt, atol=1e-3)
assert_allclose(res.x[:2], x_opt[:2], atol=2e-3)
assert_allclose(res.x[2:], x_opt[2:], atol=2e-3)
assert_allclose(res.fun, f_opt, atol=2e-2)
assert res.success
assert_(np.all(A @ res.x >= b))
assert_(np.all(np.array(c1(res.x)) >= -0.001))
assert_(np.all(np.array(c1(res.x)) <= 0.001))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
def test_constraint_population_feasibilities(self):
def constr_f(x):
return [x[0] + x[1]]
def constr_f2(x):
return [x[0]**2 + x[1], x[0] - x[1]]
nlc = NonlinearConstraint(constr_f, -np.inf, 1.9)
solver = DifferentialEvolutionSolver(rosen, [(0, 2), (0, 2)],
constraints=(nlc))
# are population feasibilities correct
# [0.5, 0.5] corresponds to scaled values of [1., 1.]
feas, cv = solver._calculate_population_feasibilities(
np.array([[0.5, 0.5], [1., 1.]]))
assert_equal(feas, [False, False])
assert_almost_equal(cv, np.array([[0.1], [2.1]]))
assert cv.shape == (2, 1)
nlc2 = NonlinearConstraint(constr_f2, -np.inf, 1.8)
solver = DifferentialEvolutionSolver(rosen, [(0, 2), (0, 2)],
constraints=(nlc, nlc2))
feas, cv = solver._calculate_population_feasibilities(
np.array([[0.5, 0.5], [0.6, 0.5]]))
assert_equal(feas, [False, False])
assert_almost_equal(cv, np.array([[0.1, 0.2, 0], [0.3, 0.64, 0]]))
feas, cv = solver._calculate_population_feasibilities(
np.array([[0.5, 0.5], [1., 1.]]))
assert_equal(feas, [False, False])
assert_almost_equal(cv, np.array([[0.1, 0.2, 0], [2.1, 4.2, 0]]))
assert cv.shape == (2, 3)
feas, cv = solver._calculate_population_feasibilities(
np.array([[0.25, 0.25], [1., 1.]]))
assert_equal(feas, [True, False])
assert_almost_equal(cv, np.array([[0.0, 0.0, 0.], [2.1, 4.2, 0]]))
assert cv.shape == (2, 3)
def test_constraint_wrapper_violation(self):
def cons_f(x):
return np.array([x[0]**2 + x[1], x[0]**2 - x[1]])
nlc = NonlinearConstraint(cons_f, [-1, -0.8500], [2, 2])
pc = _ConstraintWrapper(nlc, [0.5, 1])
assert np.size(pc.bounds[0]) == 2
assert_array_equal(pc.violation([0.5, 1]), [0., 0.])
assert_almost_equal(pc.violation([0.5, 1.2]), [0., 0.1])
assert_almost_equal(pc.violation([1.2, 1.2]), [0.64, 0])
assert_almost_equal(pc.violation([0.1, -1.2]), [0.19, 0])
assert_almost_equal(pc.violation([0.1, 2]), [0.01, 1.14])
def test_constraint_solve(self):
def constr_f(x):
return np.array([x[0] + x[1]])
nlc = NonlinearConstraint(constr_f, -np.inf, 1.9)
solver = DifferentialEvolutionSolver(rosen, [(0, 2), (0, 2)],
constraints=(nlc))
# trust-constr warns if the constraint function is linear
with warns(UserWarning):
res = solver.solve()
assert constr_f(res.x) <= 1.9
assert res.success
def test_violation():
def cons_f(x):
return np.array([x[0]**2 + x[1], x[0]**2 - x[1]])
nlc = NonlinearConstraint(cons_f, [-1, -0.8500], [2, 2])
pc = PreparedConstraint(nlc, [0.5, 1])
assert_array_equal(pc.violation([0.5, 1]), [0., 0.])
np.testing.assert_almost_equal(pc.violation([0.5, 1.2]), [0., 0.1])
np.testing.assert_almost_equal(pc.violation([1.2, 1.2]), [0.64, 0])
np.testing.assert_almost_equal(pc.violation([0.1, -1.2]), [0.19, 0])
np.testing.assert_almost_equal(pc.violation([0.1, 2]), [0.01, 1.14])
def test_L7(self):
# Lampinen ([5]) test problem 7
def f(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
fun = (5.3578547 * x[3]**2 + 0.8356891 * x[1] * x[5] +
37.293239 * x[1] - 40792.141)
return fun
def c1(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return [
85.334407 + 0.0056858 * x[2] * x[5] + 0.0006262 * x[1] * x[4] -
0.0022053 * x[3] * x[5], 80.51249 + 0.0071317 * x[2] * x[5] +
0.0029955 * x[1] * x[2] + 0.0021813 * x[3]**2,
9.300961 + 0.0047026 * x[3] * x[5] + 0.0012547 * x[1] * x[3] +
0.0019085 * x[3] * x[4]
]
N = NonlinearConstraint(c1, [0, 90, 20], [92, 110, 25])
bounds = [(78, 102), (33, 45)] + [(27, 45)] * 3
constraints = (N)
res = differential_evolution(f,
bounds,
strategy='rand1bin',
seed=1234,
constraints=constraints)
# using our best solution, rather than Lampinen/Koziel. Koziel solution
# doesn't satisfy constraints, Lampinen f_opt just plain wrong.
x_opt = [
78.00000686, 33.00000362, 29.99526064, 44.99999971, 36.77579979
]
f_opt = -30665.537578
assert_allclose(f(x_opt), f_opt)
assert_allclose(res.x, x_opt, atol=1e-3)
assert_allclose(res.fun, f_opt, atol=1e-3)
assert res.success
assert_(np.all(np.array(c1(res.x)) >= np.array([0, 90, 20])))
assert_(np.all(np.array(c1(res.x)) <= np.array([92, 110, 25])))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
def test_initial_constraints_as_canonical():
# rng is only used to generate the coefficients of the quadratic
# function that is used by the nonlinear constraint.
rng = np.random.RandomState(0)
x0 = np.array([0.5, 0.4, 0.3, 0.2])
n = len(x0)
lb1 = [-1, -np.inf, -2, 3]
ub1 = [1, np.inf, np.inf, 3]
bounds = Bounds(lb1, ub1, [False, False, True, False])
fun, jac, hess = create_quadratic_function(n, 5, rng)
lb2 = [-10, 3, -np.inf, -np.inf, -5]
ub2 = [10, 3, np.inf, 5, np.inf]
nonlinear = NonlinearConstraint(fun, lb2, ub2, jac, hess,
[True, False, False, True, False])
for sparse_jacobian in [False, True]:
bounds_prepared = PreparedConstraint(bounds, x0, sparse_jacobian)
nonlinear_prepared = PreparedConstraint(nonlinear, x0, sparse_jacobian)
f1 = bounds_prepared.fun.f
J1 = bounds_prepared.fun.J
f2 = nonlinear_prepared.fun.f
J2 = nonlinear_prepared.fun.J
c_eq, c_ineq, J_eq, J_ineq = initial_constraints_as_canonical(
n, [bounds_prepared, nonlinear_prepared], sparse_jacobian)
assert_array_equal(c_eq, [f1[3] - lb1[3], f2[1] - lb2[1]])
assert_array_equal(c_ineq, [
lb1[2] - f1[2], f1[0] - ub1[0], lb1[0] - f1[0], f2[3] - ub2[3],
lb2[4] - f2[4], f2[0] - ub2[0], lb2[0] - f2[0]
])
if sparse_jacobian:
J1 = J1.toarray()
J2 = J2.toarray()
J_eq = J_eq.toarray()
J_ineq = J_ineq.toarray()
assert_array_equal(J_eq, np.vstack((J1[3], J2[1])))
assert_array_equal(
J_ineq,
np.vstack((-J1[2], J1[0], -J1[0], J2[3], -J2[4], J2[0], -J2[0])))
def test_nonlinear_constraint():
n = 3
m = 5
rng = np.random.RandomState(0)
x0 = rng.rand(n)
fun, jac, hess = create_quadratic_function(n, m, rng)
f = fun(x0)
J = jac(x0)
lb = [-10, 3, -np.inf, -np.inf, -5]
ub = [10, 3, np.inf, 3, np.inf]
user_constraint = NonlinearConstraint(fun, lb, ub, jac, hess,
[True, False, False, True, False])
for sparse_jacobian in [False, True]:
prepared_constraint = PreparedConstraint(user_constraint, x0,
sparse_jacobian)
c = CanonicalConstraint.from_PreparedConstraint(prepared_constraint)
assert_array_equal(c.n_eq, 1)
assert_array_equal(c.n_ineq, 4)
c_eq, c_ineq = c.fun(x0)
assert_array_equal(c_eq, [f[1] - lb[1]])
assert_array_equal(
c_ineq, [f[3] - ub[3], lb[4] - f[4], f[0] - ub[0], lb[0] - f[0]])
J_eq, J_ineq = c.jac(x0)
if sparse_jacobian:
J_eq = J_eq.toarray()
J_ineq = J_ineq.toarray()
assert_array_equal(J_eq, J[1, None])
assert_array_equal(J_ineq, np.vstack((J[3], -J[4], J[0], -J[0])))
v_eq = rng.rand(c.n_eq)
v_ineq = rng.rand(c.n_ineq)
v = np.zeros(m)
v[1] = v_eq[0]
v[3] = v_ineq[0]
v[4] = -v_ineq[1]
v[0] = v_ineq[2] - v_ineq[3]
assert_array_equal(c.hess(x0, v_eq, v_ineq), hess(x0, v))
assert_array_equal(c.keep_feasible, [True, False, True, True])
def test_L2(self):
# Lampinen ([5]) test problem 2
def f(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
fun = ((x[1] - 10)**2 + 5 * (x[2] - 12)**2 + x[3]**4 + 3 *
(x[4] - 11)**2 + 10 * x[5]**6 + 7 * x[6]**2 + x[7]**4 -
4 * x[6] * x[7] - 10 * x[6] - 8 * x[7])
return fun
def c1(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
return [
127 - 2 * x[1]**2 - 3 * x[2]**4 - x[3] - 4 * x[4]**2 -
5 * x[5], 196 - 23 * x[1] - x[2]**2 - 6 * x[6]**2 + 8 * x[7],
282 - 7 * x[1] - 3 * x[2] - 10 * x[3]**2 - x[4] + x[5],
-4 * x[1]**2 - x[2]**2 + 3 * x[1] * x[2] - 2 * x[3]**2 -
5 * x[6] + 11 * x[7]
]
N = NonlinearConstraint(c1, 0, np.inf)
bounds = [(-10, 10)] * 7
constraints = (N)
with suppress_warnings() as sup:
sup.filter(UserWarning)
res = differential_evolution(f,
bounds,
strategy='rand1bin',
seed=1234,
constraints=constraints)
f_opt = 680.6300599487869
x_opt = (2.330499, 1.951372, -0.4775414, 4.365726, -0.6244870,
1.038131, 1.594227)
assert_allclose(f(x_opt), f_opt)
assert_allclose(res.fun, f_opt)
assert_allclose(res.x, x_opt, atol=1e-5)
assert res.success
assert_(np.all(np.array(c1(res.x)) >= 0))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
def test_impossible_constraint(self):
def constr_f(x):
return np.array([x[0] + x[1]])
nlc = NonlinearConstraint(constr_f, -np.inf, -1)
solver = DifferentialEvolutionSolver(rosen, [(0, 2), (0, 2)],
constraints=(nlc),
popsize=3,
seed=1)
# a UserWarning is issued because the 'trust-constr' polishing is
# attempted on the least infeasible solution found.
with warns(UserWarning):
res = solver.solve()
assert res.maxcv > 0
assert not res.success
# test _promote_lowest_energy works when none of the population is
# feasible. In this case the solution with the lowest constraint
# violation should be promoted.
solver = DifferentialEvolutionSolver(rosen, [(0, 2), (0, 2)],
constraints=(nlc),
polish=False)
next(solver)
assert not solver.feasible.all()
assert not np.isfinite(solver.population_energies).all()
# now swap two of the entries in the population
l = 20
cv = solver.constraint_violation[0]
solver.population_energies[[0, l]] = solver.population_energies[[l, 0]]
solver.population[[0, l], :] = solver.population[[l, 0], :]
solver.constraint_violation[[0, l], :] = (
solver.constraint_violation[[l, 0], :])
solver._promote_lowest_energy()
assert_equal(solver.constraint_violation[0], cv)
def test_L1(self):
# Lampinen ([5]) test problem 1
def f(x):
x = np.hstack(([0], x)) # 1-indexed to match reference
fun = np.sum(5 * x[1:5]) - 5 * x[1:5] @ x[1:5] - np.sum(x[5:])
return fun
A = np.zeros((10, 14)) # 1-indexed to match reference
A[1, [1, 2, 10, 11]] = 2, 2, 1, 1
A[2, [1, 10]] = -8, 1
A[3, [4, 5, 10]] = -2, -1, 1
A[4, [1, 3, 10, 11]] = 2, 2, 1, 1
A[5, [2, 11]] = -8, 1
A[6, [6, 7, 11]] = -2, -1, 1
A[7, [2, 3, 11, 12]] = 2, 2, 1, 1
A[8, [3, 12]] = -8, 1
A[9, [8, 9, 12]] = -2, -1, 1
A = A[1:, 1:]
b = np.array([10, 0, 0, 10, 0, 0, 10, 0, 0])
L = LinearConstraint(A, -np.inf, b)
bounds = [(0, 1)] * 9 + [(0, 100)] * 3 + [(0, 1)]
# using a lower popsize to speed the test up
res = differential_evolution(f,
bounds,
strategy='best1bin',
seed=1234,
constraints=(L),
popsize=2)
x_opt = (1, 1, 1, 1, 1, 1, 1, 1, 1, 3, 3, 3, 1)
f_opt = -15
assert_allclose(f(x_opt), f_opt)
assert res.success
assert_allclose(res.x, x_opt, atol=5e-4)
assert_allclose(res.fun, f_opt, atol=5e-3)
assert_(np.all(A @ res.x <= b))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))
# now repeat the same solve, using the same overall constraints,
# but specify half the constraints in terms of LinearConstraint,
# and the other half by NonlinearConstraint
def c1(x):
x = np.hstack(([0], x))
return [2 * x[2] + 2 * x[3] + x[11] + x[12], -8 * x[3] + x[12]]
def c2(x):
x = np.hstack(([0], x))
return -2 * x[8] - x[9] + x[12]
L = LinearConstraint(A[:5, :], -np.inf, b[:5])
L2 = LinearConstraint(A[5:6, :], -np.inf, b[5:6])
N = NonlinearConstraint(c1, -np.inf, b[6:8])
N2 = NonlinearConstraint(c2, -np.inf, b[8:9])
constraints = (L, N, L2, N2)
with suppress_warnings() as sup:
sup.filter(UserWarning)
res = differential_evolution(f,
bounds,
strategy='rand1bin',
seed=1234,
constraints=constraints,
popsize=2)
assert_allclose(res.x, x_opt, atol=5e-4)
assert_allclose(res.fun, f_opt, atol=5e-3)
assert_(np.all(A @ res.x <= b))
assert_(np.all(res.x >= np.array(bounds)[:, 0]))
assert_(np.all(res.x <= np.array(bounds)[:, 1]))