def assemble_HJBi(self, t=None):
""" Assembly matrix discretizing second order terms after diffusion
moved to the explicit operator is subtracted. Whenever amount of
diffusion used to make some row of an explicit operator monotonic
exceeds the amount of natural diffusion at that node we call it
artificial diffusion. In such case, this row of the implicit operator
is multiplied by zero
"""
print('Assembling implicit operators')
self.implicit_matrices = np.empty([self.csize_beta, self.csize_alpha],
dtype=object)
remaining_diffusion = Function(self.V)
max_art_dif = 0.0
max_art_dif_loc = None
diffusion_matrix = self.diag_matrix.copy()
for ind_a, alpha in enumerate(self.ctrset_alpha):
for ind_b, beta in enumerate(self.ctrset_beta):
# TODO: Include assert making sure that diffusion is
# non-negative on all of the internal nodes
diffusion = self.parameters.diffusion(alpha, beta)
if t is not None:
diffusion.t = t
diff = interpolate(diffusion, self.V).vector()
if not np.all(diff >= 0):
raise Exception("Choose non-negative diffusion")
diff_vec = np.array([
diff[i] if i not in self.boundary_nodes_list else 0.0
for i in range(self.dim)
])
artdif = self.explicit_diffusion[ind_b][ind_a] - diff_vec
if np.amax(artdif) > max_art_dif:
max_art_dif = np.amax(artdif)
max_art_dif_loc = self.coords[np.argmax(artdif)]
# discretise second order terms
remaining_diffusion.vector()[:] = np.maximum(
-artdif, [0] * self.dim)
diffusion_matrix.set_diagonal(remaining_diffusion.vector())
Iab = matmult(diffusion_matrix, self.laplacian)
self.implicit_matrices[ind_b][ind_a] = self.timestep_size * \
Iab + self.mass_matrix
self.scipy_implicit_matrices = [[toscipy(mat) for mat in Ialist]
for Ialist in self.implicit_matrices]
print('Checking if the implicit operators satisfy'
' monotonicity conditions')
for matrix_list in self.scipy_implicit_matrices:
for matrix in matrix_list:
implicit_check(matrix)
with open(self.ad_data_path + '/ad.txt', 'a') as f:
time_str = f' at time {t}\n' if t is not None else '\n'
f.write(f'For mesh {self.mesh_name} max value of artificial'
' diffusion coefficient was'
f' {max_art_dif} at {max_art_dif_loc}' + time_str)
def assemble_lumpedmm(self):
""" Assembly lumped mass matrix - plays role of identity matrix
"""
print("Assembling lumped mass matrix")
mass_form = self.w * self.u * dx
self.mass_matrix = assemble(mass_form)
mass_action_form = action(mass_form, Constant(1))
self.MM_terms = assemble(mass_action_form)
self.mass_matrix.zero()
self.mass_matrix.set_diagonal(self.MM_terms)
self.scipy_mass_matrix = toscipy(self.mass_matrix)
def assemble_lumpedmm(self):
""" Assembly lumped mass matrix - equal to 0 on robin boundary since
there is no time derivative there.
"""
print("Assembling lumped mass matrix")
mass_form = self.w * self.u * dx
mass_action_form = action(mass_form, Constant(1))
self.MM_terms = assemble(mass_action_form)
for n in self.robint_nodes_list:
self.MM_terms[n] = 1.0
for n in self.robin_nodes_list:
self.MM_terms[n] = 0.0
self.mass_matrix = assemble(mass_form)
self.mass_matrix.zero()
self.mass_matrix.set_diagonal(self.MM_terms)
self.scipy_mass_matrix = toscipy(self.mass_matrix)
def assemble_HJBe(self, t=None):
""" Assembly explicit operator for every control in the control set,
then apply artificial diffusion to all of them. Note that artificial
diffusion is calculated differently on the boundary nodes.
"""
self.explicit_matrices = np.empty([self.csize_beta, self.csize_alpha],
dtype=object)
self.explicit_diffusion = np.empty([self.csize_beta, self.csize_alpha],
dtype=object)
diag_vector = Vector(self.mesh.mpi_comm(), self.dim)
global_min_timestep = float('inf')
# Create explicit operator for each control in control set
for ind_a, a in enumerate(self.ctrset_alpha):
for ind_b, b in enumerate(self.ctrset_beta):
if (ind_a * self.csize_beta + ind_b) % 10 == 0:
print(f'Assembling explicit matrix under control '
f'{ind_a*self.csize_beta + ind_b} out of '
f'{self.csize_alpha*self.csize_beta}')
# coefficients in the interior
advection_x = self.parameters.adv_x(a, b)
advection_y = self.parameters.adv_y(a, b)
reaction = self.parameters.lin(a, b)
if t is not None:
advection_x.t = t
advection_y.t = t
reaction.t = t
# Discretize PDE (set values on boundary rows to zero)
b = np.array([advection_x, advection_y])
E_int_form = (np.dot(b, grad(self.w)) +
reaction * self.w) * self.u * dx
explicit_matrix = assemble(E_int_form)
# Calculate nodewise minimum diffusion required to
# impose monotonicity
min_diffusion = [0] * explicit_matrix.size(0)
for rows, row_num in getrows(explicit_matrix,
self.laplacian,
ignore=self.boundary_nodes_list):
min_diffusion[row_num] = calc_ad(rows, row_num)
self.explicit_diffusion[ind_b][ind_a] = min_diffusion
diffusion_matrix = apply_ad(self.laplacian, min_diffusion)
explicit_matrix += diffusion_matrix
explicit_matrix.get_diagonal(diag_vector)
current_min_timestep = get_min_timestep(
diag_vector, self.MM_terms, self.boundary_nodes_list)
global_min_timestep = min(current_min_timestep,
global_min_timestep)
self.explicit_matrices[ind_b][ind_a] = toscipy(explicit_matrix)
#######################################################################
min_timesteps = int(self.T / global_min_timestep) + 1
if not self.timesteps or self.timesteps < min_timesteps:
self.timesteps = min_timesteps
try:
filename = (f'meshes/{self.parameters.domain}/'
f'Mvalues-{self.parameters.experiment}.json')
with open(filename, 'r') as f:
min_timesteps_dict = json.load(f)
except FileNotFoundError:
min_timesteps_dict = {}
min_timesteps_dict[self.parameters.mesh_name] = self.timesteps
with open(filename, 'w') as f:
json.dump(min_timesteps_dict, f)
self.timestep_size = self.T / self.timesteps # time step size
self.parameters.calculate_save_interval()
self.explicit_matrices = [[
self.timestep_size * E - self.scipy_mass_matrix for E in Elist
] for Elist in self.explicit_matrices]
print('Checking if the explicit operators satisfy'
' monotonicity conditions')
for matrix_list in self.explicit_matrices:
for matrix in matrix_list:
explicit_check(matrix, self.boundary_nodes_list)
def assemble_HJBi(self, t=None):
""" Assembly matrix discretizing second order terms after diffusion
moved to the explicit operator is subtracted. Whenever amount of
diffusion used to make some row of an explicit operator monotonic
exceeds the amount of natural diffusion at that node we call it
artificial diffusion. In such case, this row of the implicit operator
is multiplied by zero
"""
print('Assembling implicit operators')
self.implicit_matrices = []
remaining_diffusion = Function(self.V)
max_art_dif = 0.0
max_art_dif_loc = None
diffusion_matrix = self.diag_matrix.copy()
for explicit_diffusion, alpha in \
zip(self.explicit_diffusion, self.control_set):
diffusion = self.parameters.diffusion(alpha)
if t is not None:
diffusion.t = t
diff = interpolate(diffusion, self.V).vector()
if not np.all(diff >= 0):
raise Exception("Choose non-negative diffusion")
diff_vec = np.array([
diff[i] if i not in self.boundary_nodes_list else 0.0
for i in range(self.dim)
])
artificial_diffusion = explicit_diffusion - diff_vec
if np.amax(artificial_diffusion) > max_art_dif:
max_art_dif = np.amax(artificial_diffusion)
max_art_dif_loc = self.coords[np.argmax(artificial_diffusion)]
# discretise second order terms
remaining_diffusion.vector()[:] = np.maximum(
-artificial_diffusion, [0] * self.dim)
diffusion_matrix.set_diagonal(remaining_diffusion.vector())
implicit_matrix = matmult(diffusion_matrix, self.laplacian)
for j in self.parameters.regions['Robin']:
self.set_directional_derivative(implicit_matrix,
region=j,
nodes=self.robin_nodes_dict[j],
control=alpha,
time=t)
self.implicit_matrices.append(self.timestep_size *
implicit_matrix + self.mass_matrix)
self.scipy_implicit_matrices = [
toscipy(mat) for mat in self.implicit_matrices
]
print('Checking if the implicit operators satisfy'
' monotonicity conditions')
for implicit_matrix in self.scipy_implicit_matrices:
implicit_check(implicit_matrix)
with open(self.ad_data_path + '/ad.txt', 'a') as f:
time_str = f' at time {t}\n' if t is not None else '\n'
f.write(f'For mesh {self.mesh_name} max value of artificial'
' diffusion coefficient was'
f' {max_art_dif} at {max_art_dif_loc}' + time_str)
def assemble_HJBe(self, t=None):
""" Assembly explicit operator for every control in the control set,
then apply artificial diffusion to all of them. Note that artificial
diffusion is calculated differently on the boundary nodes.
"""
self.explicit_matrices = np.empty(self.control_set_size, dtype=object)
self.explicit_diffusion = np.empty([self.control_set_size, self.dim])
diagonal_vector = Vector(self.mesh.mpi_comm(), self.dim)
global_min_timestep = float('inf')
# Create explicit operator for each control in control set
for k, alpha in enumerate(self.control_set):
if k % 10 == 0:
print(f'Assembling explicit matrix under control {k}'
f' out of {self.control_set_size}')
# coefficients in the interior
advection_x = self.parameters.adv_x(alpha)
advection_y = self.parameters.adv_y(alpha)
reaction = self.parameters.lin(alpha)
if t is not None:
advection_x.t = t
advection_y.t = t
reaction.t = t
# Discretize PDE (set values on boundary rows to zero)
b = np.array([advection_x, advection_y])
E_interior_form = (np.dot(b, grad(self.w)) +
reaction * self.w) * self.u * dx
explicit_matrix = assemble(E_interior_form)
set_zero_rows(explicit_matrix, self.boundary_nodes_list)
# Calculate diffusion necessary to make explicit operator monotone
min_diffusion = np.zeros(explicit_matrix.size(0))
for rows, row_num in getrows(explicit_matrix,
self.laplacian,
ignore=self.boundary_nodes_list):
min_diffusion[row_num] = calc_ad(rows, row_num)
self.explicit_diffusion[k] = min_diffusion
discrete_diffusion = apply_ad(self.laplacian, min_diffusion)
explicit_matrix += discrete_diffusion
for j in self.parameters.regions['RobinTime']:
self.set_directional_derivative(
explicit_matrix,
region=j,
nodes=self.robint_nodes_dict[j],
control=alpha,
time=t)
explicit_matrix.get_diagonal(diagonal_vector)
current_min_timestep = get_min_timestep(
diagonal_vector, self.MM_terms,
self.dirichlet_nodes_list | self.robin_nodes_list)
global_min_timestep = min(current_min_timestep,
global_min_timestep)
self.explicit_matrices[k] = toscipy(explicit_matrix)
#######################################################################
min_timesteps = int(self.T / global_min_timestep) + 1
if not self.timesteps or self.timesteps < min_timesteps:
self.timesteps = min_timesteps
try:
filename = (f'meshes/{self.parameters.domain}/'
f'Mvalues-{self.parameters.experiment}.json')
with open(filename, 'r') as f:
min_timesteps_dict = json.load(f)
except FileNotFoundError:
min_timesteps_dict = {}
min_timesteps_dict[self.parameters.mesh_name] = self.timesteps
with open(filename, 'w') as f:
json.dump(min_timesteps_dict, f)
self.timestep_size = self.T / self.timesteps # time step size
self.parameters.calculate_save_interval()
self.explicit_matrices = self.timestep_size * self.explicit_matrices - \
np.repeat(self.scipy_mass_matrix, self.control_set_size)
print('Checking if the explicit operators satisfy'
' monotonicity conditions')
for explicit_matrix in self.explicit_matrices:
explicit_check(explicit_matrix, self.dirichlet_nodes_list)