def test_vertex_rectangular(m = 5): np.random.seed(0) dom = BoxDomain(-np.ones(m), np.ones(m)) Xhat = dom.sample(10) X0 = dom.sample(100) L = np.ones((1, m)) check_vertex_sample(dom, Xhat, X0, L = L)
def test_vertex_eq(m = 5): np.random.seed(0) dom = BoxDomain(-np.ones(m), np.ones(m)) dom = dom.add_constraints(A_eq = np.ones(m), b_eq = [0]) Xhat = dom.sample(10) X0 = dom.sample(5) check_vertex_sample(dom, Xhat, X0)
def test_vertex_weight(m = 5): np.random.seed(0) dom = BoxDomain(-np.ones(m), np.ones(m)) Xhat = dom.sample(10) X0 = dom.sample(100) L = np.diag(1./np.arange(1,m+1)) check_vertex_sample(dom, Xhat, X0, L = L)
def test_vertex_low_rank_nonrandomize(m = 5): np.random.seed(0) dom = BoxDomain(-np.ones(m), np.ones(m)) Xhat = dom.sample(10) X0 = dom.sample(100) L = np.diag(1./np.arange(1,m+1)) L[0,0] = 0. check_vertex_sample(dom, Xhat, X0, L = L, randomize = False)
def test_sample_grid(m=3):
lb = 0 * np.ones(m)
ub = 1 * np.ones(m)
dom = BoxDomain(lb, ub)
X = dom.sample_grid(5)
assert len(X) == 5**m
def test_func(): client = Client(processes = False) dom = BoxDomain(-1,1) fun = Function(func, dom, dask_client = client) X = dom.sample(5) res = fun.eval_async(X) for r, x in zip(res, X): print(r.result()) assert np.isclose(x, r.result())
def test_isinside(m=5):
np.random.seed(0)
dom = BoxDomain(-10 * np.ones(m), -5 * np.ones(m))
X = dom.sample(10)
Xg = dom.sample_grid(2)
hull = ConvexHullDomain(Xg)
assert np.all(hull.isinside(X))
def test_fill_distance(m = 5): dom = BoxDomain(-np.ones(m), np.ones(m)) Xhat = dom.sample(10) X0 = dom.sample(1e2) x = seq_maximin_sample(dom, Xhat, X0 = X0) d = np.min(cdist(x.reshape(1,-1), Xhat)) d2 = fill_distance_estimate(dom, Xhat, X0 = X0) assert np.isclose(d, d2)
def test_box(m=10):
dom = BoxDomain(-np.ones(m), np.ones(m))
assert len(dom) == m
x = dom.corner(np.ones(m))
assert np.all(np.isclose(x, np.ones(m)))
A = np.ones((1, m))
b = np.zeros(1)
dom2 = dom.add_constraints(A_eq=A, b_eq=b)
x = dom2.sample()
assert np.isclose(np.dot(A, x), b)
def no_test_lipschitz_sample(N = 5, m = 3): dom = BoxDomain(-np.ones(m), np.ones(m)) # Add an inequality constraint so some combinations aren't feasible dom = dom.add_constraints(A = np.ones((1,m)), b = np.ones(1)) Ls = [np.random.randn(1,m) for j in range(2)] # Limit the number of iterations to reduce computational cost X = psdr.lipschitz_sample(dom, N, Ls, maxiter = 3, jiggle = False) print(X) assert np.all(dom.isinside(X)) # Verify that each point is distinct in projections for L in Ls: y = L.dot(X.T).T print(y) assert np.min(pdist(y)) > 0, "points not distinct in projection"
def test_return_grad2(m=3): A = np.random.randn(m, m) A += A.T B = np.random.randn(m, m) B += B.T def func(x, return_grad = False): fx = [0.5*x.dot(A.dot(x)), 0.5*x.dot(B.dot(x))] if return_grad: grad = [A.dot(x), B.dot(x)] return fx, grad else: return fx dom = BoxDomain(-2*np.ones(m), 2*np.ones(m)) fun = Function(func, dom, return_grad = True) X = fun.domain.sample(4) fX, grads = fun(X, return_grad = True) assert fX.shape == (len(X), 2) assert grads.shape == (len(X), 2, m) for x, grad in zip(X, grads): assert np.all(np.isclose(grad, fun.grad(x)))
def build_otl_circuit_domain():
# Parameters
# R_b1, R_b2, R_f, R_c1, R_c2, beta
lb = np.array([50, 25, 0.5, 1.2, 0.25, 50])
ub = np.array([150, 70, 3, 2.5, 1.2, 300])
return BoxDomain(lb, ub)
def build_borehole_domain():
# Parameters
# r_w, r, T_u, H_u, T_l, H_l, L, K_w
lb = np.array([0.05, 100, 63070, 990, 63.1, 700, 1120, 9855])
ub = np.array([0.15, 50e3, 115600, 1110, 116, 820, 1680, 12045])
return BoxDomain(lb, ub)
def test_lambda():
dom = BoxDomain(-1,1)
def f(x):
return x
#f = lambda x: x
print('about to start client')
# We use a threaded version for sanity
# https://github.com/dask/distributed/issues/2515
client = Client(processes = False)
print(client)
fun = Function(f, dom, dask_client = client)
x = dom.sample(1)
res = fun.eval_async(x)
print(x, res.result())
assert np.isclose(x, res.result())
def build_oas_design_domain(n_cp=3):
# Twist
domain_twist = BoxDomain(-1 * np.ones(n_cp),
1 * np.ones(n_cp),
names=['twist %d' % (i, ) for i in range(1, 4)])
# Thick
domain_thick = BoxDomain(
0.005 * np.ones(n_cp),
0.05 * np.ones(n_cp),
names=['thickness %d' % (i, ) for i in range(1, 4)])
# Root Chord
domain_root_chord = BoxDomain(0.7, 1.3, names=['root chord'])
# Taper ratio
domain_taper_ratio = BoxDomain(0.75, 1.25, names=['taper ratio'])
return TensorProductDomain(
[domain_twist, domain_thick, domain_root_chord, domain_taper_ratio])
def build_beam_domain():
# Note the book has an invalid range for height "(20mm > b > 250 mm)" and breadth "(10 mm > b > 50mm)"
# Here we follow the dimensions implied in the corresponding matlab code
# b = x(1)*0.045 +0.005
# h = x(2)*0.23 + 0.02
# I assume these definitions are in meters
domain = BoxDomain([5e-3, 0.02], [50e-3, 0.25],
names=['breadth (m)', 'height (m)'])
return domain
def build_golinski_design_domain():
return BoxDomain([2.6, 0.7, 7.3, 7.3, 2.9, 5.0],
[3.6, 0.8, 8.3, 8.3, 3.9, 5.5],
names=[
"width of gear face", "teeth module",
"shaft 1 length between bearings",
"shaft 2 length between bearings",
"diameter of shaft 1", "diameter of shaft 2"
])
def test_mult_output(M= 10,m = 5):
dom = BoxDomain(-np.ones(m), np.ones(m))
X = dom.sample(M)
a = np.random.randn(m)
b = np.random.randn(m)
def fun_a(X):
return a.dot(X.T)
def fun_b(X):
return b.dot(X.T)
def fun(X):
return np.vstack([X.dot(a), X.dot(b)]).T
print("Single function with multiple outputs")
for vectorized in [True, False]:
myfun = Function(fun, dom, vectorized = vectorized)
print(fun(X))
print("vectorized", vectorized)
print(myfun(X).shape)
assert myfun(X).shape == (M, 2)
print(myfun(X[0]).shape)
assert myfun(X[0]).shape == (2,)
fX = fun(X)
for i, x in enumerate(X):
assert np.all(np.isclose(fX[i], fun(x)))
print("Two functions with a single output each")
for vectorized in [True, False]:
myfun = Function([fun_a, fun_b], dom, vectorized = vectorized)
print(fun(X))
print("vectorized", vectorized)
print(myfun(X).shape)
assert myfun(X).shape == (M, 2)
print(myfun(X[0]).shape)
assert myfun(X[0]).shape == (2,)
fX = fun(X)
for i, x in enumerate(X):
assert np.all(np.isclose(fX[i], fun(x)))
def test_cheb(m=5):
lb = -np.ones(m)
ub = np.ones(m)
dom = BoxDomain(lb, ub)
A = np.ones((1, m))
b = np.zeros(1, )
dom2 = dom.add_constraints(A, b)
center, radius = dom2.chebyshev_center()
print center
print radius
assert dom2.isinside(center), "Center must be inside"
for i in range(100):
p = np.random.randn(m)
p /= np.linalg.norm(p)
assert dom2.extent(center, p) >= radius, "violated radius assumption"
def test_constraints(m=3):
np.random.seed(0)
dom = BoxDomain(-1 * np.ones(m), np.ones(m))
# Lower pyramid portion
dom_con = dom.add_constraints(A=np.ones((1, m)), b=np.ones(1))
# Convex hull describes the same space as dom_con
X = dom.sample_grid(2)
hull = ConvexHullDomain(X, A=dom_con.A, b=dom_con.b)
# Check that the same points are inside
X = dom.sample(100)
assert np.all(hull.isinside(X) == dom_con.isinside(X))
# Check sampling
X = hull.sample(100)
assert np.all(dom_con.isinside(X))
def test_vertex_full(m = 2):
np.random.seed(0)
dom = BoxDomain(-np.ones(m), np.ones(m))
Xhat = dom.sample(10)
#check_vertex(dom, Xhat)
# Check with degenerate points
Xhat[1] = Xhat[0]
#check_vertex(dom, Xhat)
# Check with a Lipschitz matrix
np.random.seed(0)
Xhat = dom.sample(5)
#L = np.random.randn(m,m)
L = np.diag(np.arange(1, m+1))
print(L)
print("Checking with a Lipschitz matrix")
check_vertex(dom, Xhat, L = L)
def test_bad_scaling():
# TODO: This fails with ub = 1e7
# This is mainly due to solver tolerances
lb = [-1, 1e7]
ub = [1, 2e7]
dom1 = BoxDomain(lb=lb, ub=ub)
dom2 = LinQuadDomain(lb=lb, ub=ub, verbose=True)
# Check quality of solution
p = np.ones(len(dom1))
# this calls an algebraic formula
x1 = dom1.corner(p)
# whereas this calls a linear program
x2 = dom2.corner(p)
for x1_, x2_, lb_, ub_ in zip(x1, x2, dom2.lb, dom2.ub):
print("x1:%+15.15e x2:%+15.15e delta:%+15.15e; lb: %+5.2e ub: %+5.2e" %
(x1_, x2_, np.abs(x1_ - x2_), lb_, ub_))
assert np.all(np.isclose(x1, x2, rtol=1e-4, atol=1e-4))
def build_robot_arm_domain():
# Parameters
# theta 1-4, L 1-4
lb = np.array([0, 0, 0, 0, 0, 0, 0, 0])
ub = np.array([2 * np.pi, 2 * np.pi, 2 * np.pi, 2 * np.pi, 1, 1, 1, 1])
return BoxDomain(lb,
ub,
names=[
'theta_1', 'theta_2', 'theta_3', 'theta_4', 'L_1',
'L_2', 'L_3', 'L_4'
])
def test_closest_point(m=5):
Ls = [np.eye(m)]
ys = [np.zeros(m)]
rhos = [1]
dom1 = LinQuadDomain(Ls=Ls, ys=ys, rhos=rhos)
dom2 = BoxDomain([2], [3])
dom = dom1 * dom2
x0 = np.zeros(m + 1)
x = dom.closest_point(x0)
assert np.all(np.isclose(x[0:m], 0))
assert np.all(np.isclose(x[-1], 2))
def build_hartmann_domain():
# Ranges are taken from GCSW17, Table 2
# The second parameter 'fluid density' does not appear in either function
return BoxDomain(
[0.05, 0.5, 0.5, 0.1],
[0.2, 3, 3, 1],
names=[
'fluid viscosity', # mu
'applied pressure gradient', # d p_0/ dx
'resistivity', # eta
'applied magnetic field' # B_0
])
def test_initial_sample(m = 10):
dom = BoxDomain(-np.ones(m), np.ones(m))
L1 = np.random.randn(1,m)
L2 = np.random.randn(2,m)
L3 = np.random.randn(3,m)
Nsamp = 100
for L in [L1, L2, L3]:
# Standard uniform sampling
X1 = dom.sample(Nsamp)
LX1 = L.dot(X1.T).T
d1 = pdist(LX1)
# initial sample algorithm
X2 = initial_sample(dom, L, Nsamp = Nsamp)
assert np.all(dom.isinside(X2))
LX2 = L.dot(X2.T).T
d2 = pdist(LX2)
print("uniform sampling mean distance", np.mean(d1), 'min', np.min(d1))
print("initial sampling mean distance", np.mean(d2), 'min', np.min(d2))
assert np.mean(d2) > np.mean(d1), "Initial sampling ineffective"
def test_bad_scaling():
# TODO: This fails with ub = 1e7
# This is mainly due to solver tolerances
lb = [-1, 1e7]
ub = [1, 2e7]
dom1 = BoxDomain(lb=lb, ub=ub)
dom2 = LinQuadDomain(lb=lb, ub=ub)
# Check quality of solution
p = np.ones(len(dom1))
x1 = dom1.corner(p)
x2 = dom2.corner(p,
verbose=True,
solver='ECOS',
abstol=4e-10,
reltol=1e-14,
feastol=1e-14,
max_iters=500)
for x1_, x2_, lb_, ub_ in zip(x1, x2, dom2.lb, dom2.ub):
print "x1:%+15.15e x2:%+15.15e delta:%+15.15e; lb: %+5.2e ub: %+5.2e" % (
x1_, x2_, np.abs(x1_ - x2_), lb_, ub_)
assert np.all(np.isclose(x1, x2))
def test_point(m=3):
lb = 0 * np.ones(m)
ub = 1 * np.ones(m)
dom = BoxDomain(lb, ub)
assert dom.is_point == False
dom = BoxDomain(ub, ub)
assert dom.is_point == True
dom = BoxDomain(lb, ub)
dom = dom.add_constraints(A_eq=np.ones((1, m)), b_eq=[0])
assert dom.is_point == True
def build_borehole_domain():
r""" Constructs a deterministic domain associated with the borehole function
Returns
-------
dom: BoxDomain
Domain associated with the borehole function
"""
# Parameters
# r_w, r, T_u, H_u, T_l, H_l, L, K_w
lb = np.array([0.05, 100, 63070, 990, 63.1, 700, 1120, 9855])
ub = np.array([0.15, 50e3, 115600, 1110, 116, 820, 1680, 12045])
return BoxDomain(
lb, ub, names=['r_w', 'r', 'T_u', 'H_u', 'T_l', 'H_l', 'L', 'K_w'])
def test_gp_fit(m=3, M=100):
""" check """
dom = BoxDomain(-np.ones(m), np.ones(m))
a = np.ones(m)
b = np.ones(m)
b[0] = 0
f = lambda x: np.sin(x.dot(a)) + x.dot(b)**2
fun = Function(f, dom)
X = dom.sample(M)
fX = f(X)
for structure in ['const', 'diag', 'tril']:
for degree in [None, 0, 1]:
gp = GaussianProcess(structure=structure, degree=degree)
gp.fit(X, fX)
print(gp.L)
I = ~np.isclose(gp(X), fX)
print(fX[I])
print(gp(X[I]))
assert np.all(np.isclose(
gp(X), fX, atol=1e-5)), "we should interpolate samples"
_, cov = gp.eval(X, return_cov=True)
assert np.all(np.isclose(
cov, 0, atol=1e-3)), "Covariance should be small at samples"