def test02a_sensitivity_gradient():
"""
Test whether the sensitivity and adjoint calculations give the same gradient
"""
# Mesh
mesh = Mesh1D(resolution=(100, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(1 - x) < 1e-10)
# Element
Q = QuadFE(mesh.dim(), 'Q2')
dh = DofHandler(mesh, Q)
dh.distribute_dofs()
n_dofs = dh.n_dofs()
phi = Basis(dh, 'u')
# Covariance
cov = Covariance(dh, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
d = len(lmd)
# Coarse field (single sample)
d0 = 2
z0 = np.random.randn(d0, 1)
q0 = sample_q0(V, lmd, d0, z0)
q0_fn = Nodal(data=q0, basis=phi)
# State
J0, u0 = sample_qoi(q0, dh, return_state=True)
u0_fn = Nodal(data=u0, basis=phi)
# Compute gradient using sensitivity
dJs = np.zeros(n_dofs)
for i in range(n_dofs):
# Define perturbation
dq = np.zeros(n_dofs)
dq[i] = 1
dq_fn = Nodal(data=dq, basis=phi)
# Compute gradient using sensitivity
dJs[i] = dJdq_sen(q0_fn, u0_fn, dq_fn)
dJs_fn = Nodal(data=dJs, basis=phi)
plot = Plot()
plot.line(dJs_fn)
# Compute gradient using adjoint method
dJa = dJdq_adj(q0_fn, u0_fn)
dJa_fn = Nodal(data=dJa, basis=phi)
print(dJa)
plot.line(dJa_fn)
def test02_variance():
"""
Compute the variance of J(q) for different mesh refinement levels
and compare with MC estimates.
"""
l_max = 8
for i_res in np.arange(2, l_max):
# Computational mesh
mesh = Mesh1D(resolution=(2**i_res, ))
# Element
element = QuadFE(mesh.dim(), 'DQ0')
dofhandler = DofHandler(mesh, element)
dofhandler.distribute_dofs()
# Linear Functional
mesh.mark_region('integrate',
lambda x: x >= 0.75,
entity_type='cell',
strict_containment=False)
phi = Basis(dofhandler)
assembler = Assembler(Form(4, test=phi, flag='integrate'))
assembler.assemble()
L = assembler.get_vector()
# Define Gaussian random field
C = Covariance(dofhandler, name='gaussian', parameters={'l': 0.05})
C.compute_eig_decomp()
eta = GaussianField(dofhandler.n_dofs(), K=C)
eta.update_support()
n_samples = 100000
J_paths = L.dot(eta.sample(n_samples=n_samples))
var_mc = np.var(J_paths)
lmd, V = C.get_eig_decomp()
LV = L.dot(V)
var_an = LV.dot(np.diag(lmd).dot(LV.transpose()))
print(var_mc, var_an)
#
phi_0 = Basis(dQ0)
phi_1 = Basis(dQ1)
phix_1 = Basis(dQ1, 'vx')
phiy_1 = Basis(dQ1, 'vy')
phi_2 = Basis(dQ2)
phix_2 = Basis(dQ2, 'vx')
phiy_2 = Basis(dQ2, 'vy')
#
# Define Random field
#
cov = Covariance(dQ0, name='gaussian', parameters={'l': 0.01})
cov.compute_eig_decomp()
q = GaussianField(dQ0.n_dofs(), K=cov)
# Sample Random field
n_samples = 100
eq = Nodal(basis=phi_0, data=np.exp(q.sample(n_samples)))
plot.contour(eq, n_sample=25)
#
# Compute state
#
# Define weak form
state = [[
Form(eq, test=phix_1, trial=phix_1),
# Define piecewise constant elements
#
element = QuadFE(mesh.dim(), 'DQ0')
dofhandler = DofHandler(mesh, element)
dofhandler.distribute_dofs()
# Get projection matrix
P = projection_matrix(dofhandler, None, 0)
fig, ax = plt.subplots(1, 1)
for i in range(l_max):
CC = Covariance(dofhandler,
subforest_flag=i,
name='gaussian',
parameters={'l': 0.01})
CC.compute_eig_decomp()
d, V = CC.get_eig_decomp()
print(d)
lmd = np.arange(len(d))
ax.semilogy(lmd, d, '.-', label='level=%d' % i)
plt.legend()
plt.show()
#
# Define random field on the fine mesh
#
C = Covariance(dofhandler, name='gaussian', parameters={'l': 0.05})
C.compute_eig_decomp()
eta = GaussianField(dofhandler.n_dofs(), K=C)
eta.update_support()
def experiment01_problem():
"""
Illustrate the problem: Plot sample paths of the input q, of the output,
and histogram of the QoI.
"""
#
# Computational Mesh
#
mesh = Mesh1D(resolution=(100, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
#
# Element
#
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
#
# Basis
#
phi = Basis(dQ1, 'v')
phi_x = Basis(dQ1, 'vx')
#
# Covariance
#
cov = Covariance(dQ1, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
d = len(lmd)
#
# Sample and plot full dimensional parameter and solution
#
n_samples = 20000
z = np.random.randn(d, n_samples)
q = sample_q0(V, lmd, d, z)
# Define finite element function
qfn = Nodal(data=q, basis=phi)
problem = [[Form(qfn, test=phi_x, trial=phi_x),
Form(1, test=phi)],
[Form(qfn, test=phi_x, dmu='dv', flag='right')]]
# Define assembler
assembler = Assembler(problem)
# Incorporate Dirichlet conditions
assembler.add_dirichlet('left', 0)
assembler.add_dirichlet('right', 1)
comment.tic('assembly')
# Assemble system
assembler.assemble()
comment.toc()
comment.tic('solver')
ufn = Nodal(basis=phi, data=None)
J = np.zeros(n_samples)
for i in range(n_samples):
# Solve system
u = assembler.solve(i_problem=0, i_matrix=i, i_vector=0)
# Compute quantity of interest
J[i] = u.dot(assembler.get_vector(1, i))
# Update sample paths
ufn.add_samples(u)
comment.toc()
#
# Plots
#
"""
# Formatting
plt.rc('text', usetex=True)
# Figure sizes
fs2 = (3,2)
fs1 = (4,3)
plot = Plot(quickview=False)
plot_kwargs = {'color':'k', 'linewidth':0.05}
#
# Plot qfn
#
# Figure
fig = plt.figure(figsize=fs2)
ax = fig.add_subplot(111)
ax = plot.line(qfn, axis=ax,
i_sample=np.arange(100),
plot_kwargs=plot_kwargs)
ax.set_xlabel(r'$x$')
ax.set_ylabel(r'$q$')
plt.tight_layout()
fig.savefig('fig/ex02_gauss_qfn.eps')
plt.close()
#
# Plot ufn
#
fig = plt.figure(figsize=fs2)
ax = fig.add_subplot(111)
ax = plot.line(ufn, axis=ax,
i_sample=np.arange(100),
plot_kwargs=plot_kwargs)
ax.set_xlabel(r'$x$')
ax.set_ylabel(r'$u$')
plt.tight_layout()
fig.savefig('fig/ex02_gauss_ufn.eps')
plt.close()
"""
# Formatting
plt.rc('text', usetex=True)
# Figure sizes
fs2 = (3, 2)
fs1 = (4, 3)
fig = plt.figure(figsize=fs2)
ax = fig.add_subplot(111)
plt.hist(J, bins=100, density=True)
ax.set_xlabel(r'$J(u)$')
plt.tight_layout()
fig.savefig('fig/ex02_gauss_jhist.eps')
def test06_linearization():
"""
Compute samples on fine grid via the linearization
"""
plot = Plot()
#
# Computational mesh
#
mesh = Mesh1D(resolution=(100, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
#
# Element
#
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
nx = dQ1.n_dofs()
x = dQ1.get_dof_vertices()
Q3 = QuadFE(mesh.dim(), 'Q3')
dQ3 = DofHandler(mesh, Q3)
dQ3.distribute_dofs()
#
# Basis
#
phi = Basis(dQ1, 'u')
phi_x = Basis(dQ1, 'ux')
psi = Basis(dQ3, 'u')
psi_x = Basis(dQ3, 'ux')
#
# Covariance
#
cov = Covariance(dQ1, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
d = len(lmd)
# Fix coarse truncation level
d0 = 10
#
# Build Sparse Grid
#
grid = TasmanianSG.TasmanianSparseGrid()
dimensions = d0
outputs = 1
depth = 2
type = 'level'
rule = 'gauss-hermite'
grid.makeGlobalGrid(dimensions, outputs, depth, type, rule)
# Sample Points
zzSG = grid.getPoints()
zSG = np.sqrt(2) * zzSG # transform to N(0,1)
# Quadrature Weights
wSG = grid.getQuadratureWeights()
wSG /= np.sqrt(np.pi)**d0 # normalize weights
# Number of grid points
n0 = grid.getNumPoints()
#
# Sample low dimensional input parameter
#
q0 = sample_q0(V, lmd, d0, zSG.T)
J0 = sample_qoi(q0, dQ1)
#
# Sample conditional expectation
#
# Pick a single coarse sample to check
i0 = np.random.randint(0, high=n0)
# Sample fine, conditional on coarse
n_samples = 1
z1 = np.random.randn(d - d0, n_samples)
q = sample_q_given_q0(q0[:, i0], V, lmd, d0, z1)
# Perturbation
log_qref = np.log(q0[:, i0])
dlog_q = np.log(q.ravel()) - log_qref
dlog_qfn = Nodal(data=dlog_q, basis=phi)
# Perturbed q
n_eps = 12 # Number of refinements
epsilons = [10**(-l) for l in range(n_eps)]
log_qper = np.empty((nx, n_eps))
for i in range(n_eps):
log_qper[:, i] = log_qref + epsilons[i] * dlog_q
"""
plt.plot(x, log_qref, label='ref')
for i in range(n_eps):
plt.plot(x, log_qper[:,i],label='%d'%(i))
"""
assert np.allclose(log_qper[:, 0], np.log(q.ravel()))
plt.legend()
plt.show()
# Define finite element function
exp_qref = Nodal(data=q0[:, i0], basis=phi)
exp_qper = Nodal(data=np.exp(log_qper), basis=phi)
#
# PDEs
#
# 1. State Equation
state_eqn = [Form(exp_qref, test=phi_x, trial=phi_x), Form(1, test=phi)]
state_dbc = {'left': 0, 'right': 1}
# 2. Perturbed Equation
perturbed_eqn = [
Form(exp_qper, test=phi_x, trial=phi_x),
Form(1, test=phi)
]
perturbed_dbc = {'left': 0, 'right': 1}
# 3. Adjoint Equation
adjoint_eqn = [Form(exp_qref, test=psi_x, trial=psi_x), Form(0, test=psi)]
adjoint_dbc = {'left': 0, 'right': -1}
# Combine
eqns = [state_eqn, perturbed_eqn, adjoint_eqn]
bcs = [state_dbc, perturbed_dbc, adjoint_dbc]
#
# Assembly
#
assembler = Assembler(eqns, n_gauss=(6, 36))
# Boundary conditions
for i, bc in zip(range(3), bcs):
for loc, val in bc.items():
assembler.add_dirichlet(loc, val, i_problem=i)
# Assemble
assembler.assemble()
#
# Solve
#
# Solve state
ur = assembler.solve(i_problem=0)
u_ref = Nodal(data=ur, basis=phi)
ux_ref = Nodal(data=ur, basis=phi_x)
# Solve perturbed state
u_per = Nodal(basis=phi)
for i in range(n_eps):
# FEM solution
up = assembler.solve(i_problem=1, i_matrix=i)
u_per.add_samples(up)
plt.plot(x, up - ur)
plt.show()
ux_per = Nodal(data=u_per.data(), basis=phi_x)
# Solve adjoint equation
v = assembler.solve(i_problem=2)
v_adj = Nodal(data=v, basis=psi)
vx_adj = Nodal(data=v, basis=psi_x)
#
# Sensitivity
#
# Sensitivity Equation
ker_sen = Kernel(f=[exp_qref, dlog_qfn, ux_ref],
F=lambda eq, dq, ux: -eq * dq * ux)
sensitivity_eqn = [
Form(exp_qref, test=phi_x, trial=phi_x),
Form(ker_sen, test=phi_x)
]
sensitivity_dbc = {'left': 0, 'right': 0}
assembler = Assembler(sensitivity_eqn, n_gauss=(6, 36))
for loc in sensitivity_dbc:
assembler.add_dirichlet(loc, sensitivity_dbc[loc])
assembler.assemble()
s = assembler.solve()
sx = Nodal(data=s, basis=phi_x)
plt.plot(x, s)
plt.show()
#
# Quantities of Interest
#
# Reference
k_ref = Kernel(f=[exp_qref, ux_ref], F=lambda eq, ux: eq * ux)
ref_qoi = [Form(k_ref, dmu='dv', flag='right')]
# Perturbed
k_per = Kernel(f=[exp_qper, ux_per], F=lambda eq, ux: eq * ux)
per_qoi = [Form(k_per, dmu='dv', flag='right')]
# Adjoint
k_adj = Kernel(f=[exp_qref, dlog_qfn, ux_ref, vx_adj],
F=lambda eq, dq, ux, vx: -eq * dq * ux * vx)
adj_qoi = [Form(k_adj)]
# Sensitivity
k_sens = Kernel(f=[exp_qref, dlog_qfn, ux_ref, sx],
F=lambda eq, dq, ux, sx: eq * dq * ux + eq * sx)
sens_qoi = Form(k_sens, dmu='dv', flag='right')
qois = [ref_qoi, per_qoi, adj_qoi, sens_qoi]
# Assemble
assembler = Assembler(qois, mesh=mesh)
assembler.assemble()
# Evaluate
J_ref = assembler.get_scalar(0)
J_per = []
for i in range(n_eps):
J_per.append(assembler.get_scalar(1, i))
# Finite difference approximation
dJ = []
for eps, J_p in zip(epsilons, J_per):
dJ.append((J_p - J_ref) / eps)
# Adjoint differential
dJ_adj = assembler.get_scalar(2)
# Sensitivity differential
dJ_sen = assembler.get_scalar(3)
print(dJ_adj)
print(dJ_sen)
print(dJ)
"""
def test01_finite_elements():
"""
Test accuracy of the finite element approximation
"""
#
# Construct reference solution
#
plot = Plot(quickview=False)
# Mesh
mesh = Mesh1D(resolution=(2**11, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
# Element
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
# Basis
phi = Basis(dQ1, 'v')
phi_x = Basis(dQ1, 'vx')
#
# Covariance
#
cov = Covariance(dQ1, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
d = len(lmd)
#
# Sample and plot full dimensional parameter and solution
#
n_samples = 1
z = np.random.randn(d, n_samples)
q_ref = sample_q0(V, lmd, d, z)
print(q_ref.shape)
# Define finite element function
q_ref_fn = Nodal(data=q_ref, basis=phi)
problem = [[Form(q_ref_fn, test=phi_x, trial=phi_x),
Form(1, test=phi)],
[Form(q_ref_fn, test=phi_x, dmu='dv', flag='right')]]
# Define assembler
assembler = Assembler(problem)
# Incorporate Dirichlet conditions
assembler.add_dirichlet('left', 0)
assembler.add_dirichlet('right', 1)
# Assemble system
assembler.assemble()
# Solve system
u_ref = assembler.solve()
# Compute quantity of interest
J_ref = u_ref.dot(assembler.get_vector(1))
# Plot
fig = plt.figure(figsize=(6, 4))
ax = fig.add_subplot(111)
u_ref_fn = Nodal(basis=phi, data=u_ref)
ax = plot.line(u_ref_fn, axis=ax)
n_levels = 10
J = np.zeros(n_levels)
for l in range(10):
comment.comment('level: %d' % (l))
#
# Mesh
#
mesh = Mesh1D(resolution=(2**l, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
#
# Element
#
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
#
# Basis
#
phi = Basis(dQ1, 'v')
phi_x = Basis(dQ1, 'vx')
# Define problem
problem = [[
Form(q_ref_fn, test=phi_x, trial=phi_x),
Form(1, test=phi)
], [Form(q_ref_fn, test=phi_x, dmu='dv', flag='right')]]
assembler = Assembler(problem)
# Incorporate Dirichlet conditions
assembler.add_dirichlet('left', 0)
assembler.add_dirichlet('right', 1)
assembler.assemble()
A = assembler.get_matrix()
print('A shape', A.shape)
u = assembler.solve()
J[l] = u.dot(assembler.get_vector(1))
print(u.shape)
print(phi.n_dofs())
ufn = Nodal(basis=phi, data=u)
ax = plot.line(ufn, axis=ax)
plt.show()
#
# Plots
#
# Formatting
plt.rc('text', usetex=True)
# Figure sizes
fs2 = (3, 2)
fs1 = (4, 3)
print(J_ref)
print(J)
#
# Plot truncation error for mean and variance of J
#
fig = plt.figure(figsize=fs2)
ax = fig.add_subplot(111)
err = np.array([np.abs(J[i] - J_ref) for i in range(n_levels)])
h = np.array([2**(-l) for l in range(n_levels)])
plt.loglog(h, err, '.-')
ax.set_xlabel(r'$h$')
ax.set_ylabel(r'$|J-J^h|$')
plt.tight_layout()
fig.savefig('fig/ex02_gauss_fem_error.eps')
def experiment06_sensitivity_stats():
"""
Compute the sensitivities
"""
comment = Verbose()
comment.comment('Computing statistics for the sensitivity dJ_dq')
#
# Computational mesh
#
mesh = Mesh1D(resolution=(100, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
#
# Element
#
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
n_dofs = dQ1.n_dofs()
phi = Basis(dQ1, 'u')
#
# Covariance
#
cov = Covariance(dQ1, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
d = len(lmd)
# Fix coarse truncation level
d0 = 10
#
# Build Sparse Grid
#
grid = TasmanianSG.TasmanianSparseGrid()
dimensions = d0
outputs = 1
depth = 4
type = 'level'
rule = 'gauss-hermite'
grid.makeGlobalGrid(dimensions, outputs, depth, type, rule)
# Sample Points
zzSG = grid.getPoints()
zSG = np.sqrt(2) * zzSG # transform to N(0,1)
# Quadrature Weights
wSG = grid.getQuadratureWeights()
wSG /= np.sqrt(np.pi)**d0 # normalize weights
# Number of grid points
n0 = grid.getNumPoints()
comment.comment('Element DOFs: {0}'.format(n_dofs))
comment.comment('Sparse Grid Size: {0}'.format(n0))
#
# Sample low dimensional input parameter
#
comment.tic('Sampling reference')
q0 = sample_q0(V, lmd, d0, zSG.T)
J0, u0 = sample_qoi(q0, dQ1, return_state=True)
comment.toc()
comment.tic('Sampling gradient')
dJdq = np.zeros((n_dofs, n0))
for i in range(n0):
# Sample input and state
q = Nodal(data=q0[:, i], basis=phi)
u = Nodal(data=u0[:, i], basis=phi)
# Compute gradient using adjoint approach
dJdq[:, i] = dJdq_adj(q, u)
comment.toc()
# Compute sparse grid mean and variance
E_dJ = np.dot(dJdq, wSG)
V_dJ = np.dot(dJdq**2, wSG) - E_dJ**2
E_dJ = Nodal(data=E_dJ, basis=phi)
V_dJ = Nodal(data=V_dJ, basis=phi)
fig, ax = plt.subplots(nrows=1, ncols=2)
plot = Plot(quickview=False)
ax[0] = plot.line(E_dJ, axis=ax[0])
ax[1] = plot.line(V_dJ, axis=ax[1])
plt.show()
def experiment05_conditioning():
"""
Obtain an estimate of J using sparse grids on the coarse scale and MC as a
correction.
REMARKS: This takes very long, especially since the convergence rate of the
conditional samples is low.
"""
#
# Computational mesh
#
mesh = Mesh1D(resolution=(100, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
#
# Element
#
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
#
# Covariance
#
cov = Covariance(dQ1, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
d = len(lmd)
# Fix coarse truncation level
d0 = 10
#
# Build Sparse Grid
#
grid = TasmanianSG.TasmanianSparseGrid()
dimensions = d0
outputs = 1
depth = 2
type = 'level'
rule = 'gauss-hermite'
grid.makeGlobalGrid(dimensions, outputs, depth, type, rule)
# Sample Points
zzSG = grid.getPoints()
zSG = np.sqrt(2) * zzSG # transform to N(0,1)
# Quadrature Weights
wSG = grid.getQuadratureWeights()
wSG /= np.sqrt(np.pi)**d0 # normalize weights
# Number of grid points
n0 = grid.getNumPoints()
#
# Sample low dimensional input parameter
#
q0 = sample_q0(V, lmd, d0, zSG.T)
J0 = sample_qoi(q0, dQ1)
# Compute sparse grid mean and variance
EJ0 = np.sum(wSG * J0)
VJ0 = np.sum(wSG * (J0**2)) - EJ0**2
J = np.load('data/j_mc.npy')
mean_ref = np.mean(J)
var_ref = np.var(J)
# Record errors
mean_err = [np.abs(EJ0 - mean_ref)]
var_err = [np.abs(VJ0 - var_ref)]
for n_samples in [10, 100, 1000]:
mean_Jg0 = 0
var_Jg0 = 0
for i in range(n0):
z = np.random.randn(d - d0, n_samples)
qg0 = sample_q_given_q0(q0[:, i], V, lmd, d0, z)
Jg0 = sample_qoi(qg0, dQ1)
mean_Jg0 += wSG[i] * np.mean(Jg0)
mean_err.append(np.abs(mean_Jg0 - mean_ref))
# Formatting
plt.rc('text', usetex=True)
# Figure sizes
fs2 = (3, 2)
fs1 = (4, 3)
fig = plt.figure(figsize=fs2)
ax = fig.add_subplot(111)
ax.semilogy([0, 10, 100, 1000], mean_err, '.-')
ax.set_xlabel(r'$n$')
ax.set_ylabel(r'$\mathrm{Error}$')
fig.tight_layout()
fig.savefig('fig/ex02_gauss_hyb_mean_err.eps')
"""
#
# Plot conditional variances
#
fig = plt.figure(figsize=fs2)
ax = fig.add_subplot(111)
ax.hist(varJg,bins=30, density=True)
ax.set_xlabel(r'$\sigma_{J|q_0}^2$')
fig.tight_layout()
fig.savefig('fig/ex02_gauss_cond_var.eps')
"""
"""
def experiment04_sparse_grid():
"""
Test sparse grid
"""
#
# Computational mesh
#
mesh = Mesh1D(resolution=(100, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
#
# Element
#
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
#
# Covariance
#
cov = Covariance(dQ1, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
# Truncation levels
truncation_levels = [1, 5, 10, 20]
# Formatting
plt.rc('text', usetex=True)
# Set figure and axis
fs2 = (3, 2)
fs1 = (4, 3)
# For mean
fig1 = plt.figure(figsize=fs1)
ax1 = fig1.add_subplot(111)
# For variance
fig2 = plt.figure(figsize=fs1)
ax2 = fig2.add_subplot(111)
for d0 in truncation_levels:
J = []
mean = []
var = []
n = []
for depth in range(5):
#
# Construct Sparse Grid
#
grid = TasmanianSG.TasmanianSparseGrid()
dimensions = d0
outputs = 1
type = 'level'
rule = 'gauss-hermite'
grid.makeGlobalGrid(dimensions, outputs, depth, type, rule)
# Get Sample Points
zzSG = grid.getPoints()
zSG = np.sqrt(2) * zzSG # transform to N(0,1)
wSG = grid.getQuadratureWeights()
wSG /= np.sqrt(np.pi)**d0 # normalize weights
n0 = grid.getNumPoints()
n.append(n0)
#
# Sample input parameter
#
q0 = sample_q0(V, lmd, d0, zSG.T)
J = sample_qoi(q0, dQ1)
EJ = np.sum(wSG * J)
VJ = np.sum(wSG * (J**2)) - EJ**2
mean.append(EJ)
var.append(VJ)
J_mc = np.load('data/j_%d_mc.npy' % (d0))
# Compute mean and variance
mean_mc = np.mean(J_mc)
var_mc = np.var(J_mc)
# Plot mean error
mean_err = [np.abs(mean[i] - mean_mc) for i in range(5)]
ax1.loglog(n, mean_err, '.-.', label=r'$k=%d$' % (d0))
ax1.set_xlabel(r'$n$')
ax1.set_ylabel(r'$\mathrm{Error}$')
ax1.legend()
fig1.tight_layout()
# Plot variance error
var_err = [np.abs(var[i] - var_mc) for i in range(5)]
ax2.loglog(n, var_err, '.-.', label=r'k=%d' % (d0))
ax2.set_xlabel(r'$n$')
ax2.set_ylabel(r'$\mathrm{Error}$')
ax2.legend()
fig2.tight_layout()
fig1.savefig('fig/ex02_gauss_sg_mean_error.eps')
fig2.savefig('fig/ex02_gauss_sg_var_error.eps')
def experiment03_truncation():
"""
Investigate the error in truncation level
"""
generate = False
mesh = Mesh1D(resolution=(100, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
#
# Element
#
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
#
# Basis
#
phi = Basis(dQ1, 'v')
phi_x = Basis(dQ1, 'vx')
#
# Covariance
#
cov = Covariance(dQ1, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
d = len(lmd)
# Truncation levels
truncation_levels = [1, 5, 10, 20, 50]
n_samples = 1000000
if generate:
n_batches = 1000
batch_size = n_samples // n_batches
for d0 in truncation_levels:
comment.tic('d = %d' % (d0))
J = np.empty(n_samples)
for i in range(n_batches):
# Print progress
#print('.',end='')
# Sample diffusion coefficient
z = np.random.randn(d0, batch_size)
q = sample_q0(V, lmd, d0, z)
# Evaluate quantity of interest
J[(i) * batch_size:(i + 1) * batch_size] = sample_qoi(q, dQ1)
# Save current update to file
np.save('./data/j_%d_mc.npy' % (d0), J)
comment.toc()
#
# Compute estimates and errors
#
n_levels = len(truncation_levels)
mean = []
var = []
for d0 in truncation_levels:
J = np.load('data/j_%d_mc.npy' % (d0))
# Compute mean and variance
mean.append(np.mean(J))
var.append(np.var(J))
# Load reference
J = np.load('data/j_mc.npy')
mean_ref = np.mean(J)
var_ref = np.var(J)
#truncation_levels.append(101)
err_mean = [np.abs(mean[i] - mean_ref) for i in range(n_levels)]
err_var = [np.abs(var[i] - var_ref) for i in range(n_levels)]
#
# Plots
#
# Formatting
plt.rc('text', usetex=True)
# Figure sizes
fs2 = (3, 2)
fs1 = (4, 3)
#
# Plot truncation error for mean and variance of J
#
fig = plt.figure(figsize=fs2)
ax = fig.add_subplot(111)
plt.semilogy(truncation_levels, err_mean, '.-', label='mean')
plt.semilogy(truncation_levels, err_var, '.--', label='variance')
plt.legend()
ax.set_xlabel(r'$k$')
ax.set_ylabel(r'$\mathrm{Error}$')
plt.tight_layout()
fig.savefig('fig/ex02_gauss_trunc_error.eps')
#
# Plot estimated mean and variance
#
fig = plt.figure(figsize=fs2)
ax = fig.add_subplot(111)
truncation_levels.append(101)
mean.append(mean_ref)
var.append(var_ref)
plt.plot(truncation_levels, mean, 'k.-', label='mean')
plt.plot(truncation_levels, var, 'k.--', label='variance')
plt.legend()
ax.set_xlabel(r'$k$')
plt.tight_layout()
fig.savefig('fig/ex02_gauss_trunc_stats.eps')
def experiment02_reference():
"""
Convergence rate of MC
"""
generate = False
#
# Computational Mesh
#
mesh = Mesh1D(resolution=(100, ))
mesh.mark_region('left', lambda x: np.abs(x) < 1e-10)
mesh.mark_region('right', lambda x: np.abs(x - 1) < 1e-10)
#
# Element
#
Q1 = QuadFE(mesh.dim(), 'Q1')
dQ1 = DofHandler(mesh, Q1)
dQ1.distribute_dofs()
#
# Covariance
#
cov = Covariance(dQ1, name='gaussian', parameters={'l': 0.05})
cov.compute_eig_decomp()
lmd, V = cov.get_eig_decomp()
d = len(lmd)
#
# Generate random sample for J
#
n_samples = 1000000
if generate:
n_batches = 1000
batch_size = n_samples // n_batches
J = np.empty(n_samples)
for i in range(n_batches):
# Sample diffusion coefficient
z = np.random.randn(d, n_samples // n_batches)
q = sample_q0(V, lmd, d, z)
# Evaluate quantity of interest
J[(i) * batch_size:(i + 1) * batch_size] = sample_qoi(q, dQ1)
# Save current update to file
np.save('./data/j_mc.npy', J)
#
# Process data
#
# Load MC samples
J = np.load('data/j_mc.npy')
# Compute sample mean and variance of J
EX = np.mean(J)
VarX = np.var(J)
print(EX, VarX)