# BO loop
n_iters = 10
for iteration in range(n_iters):
# Fit GP model
botorch.fit_gpytorch_model(mll=mll_fct)
# Define the acquisition function
acq_fct = botorch.acquisition.ExpectedImprovement(model=model,
best_f=best_f[-1],
maximize=False)
# Get new candidate
new_x = joint_optimize_manifold(acq_fct,
sphere_manifold,
solver,
q=1,
num_restarts=5,
raw_samples=100,
bounds=bounds,
equality_constraints=eq_constraints)
# Get new observation
new_y = test_function(new_x)[0]
# Update training points
x_data = torch.cat((x_data, new_x))
y_data = torch.cat((y_data, new_y))
# Update best observation
new_best_f, index = y_data.min(0)
best_x.append(x_data[index])
best_f.append(new_best_f)
inequality_constraints = [
max_eigenvalue_constraint, min_eigenvalue_constraint
]
# Define the acquisition function
acq_fct = botorch.acquisition.ExpectedImprovement(model=latent_model,
best_f=best_f[-1],
maximize=False)
# Get new candidate
new_x_projected = joint_optimize_manifold(
acq_fct,
latent_spd_manifold,
acquisition_solver,
q=1,
num_restarts=5,
raw_samples=100,
bounds=bounds,
pre_processing_manifold=vector_to_symmetric_matrix_mandel_torch,
post_processing_manifold=symmetric_matrix_to_vector_mandel_torch,
approx_hessian=True,
inequality_constraints=inequality_constraints)
# To matrix
new_x_projected = vector_to_symmetric_matrix_mandel_torch(
new_x_projected)
# Projection back onto the original SPD manifold
new_x = projection_from_nested_spd_to_spd(
new_x_projected, projection_matrix, projection_complement_matrix,
bottom_spd_matrix, contraction_matrix)
# To vector
# BO loop
n_iters = 25
for iteration in range(n_iters):
# Fit GP model
botorch.fit_gpytorch_model(mll=mll_fct)
# Define the acquisition function
acq_fct = botorch.acquisition.ExpectedImprovement(model=model,
best_f=best_f[-1],
maximize=False)
# Get new candidate
new_x = joint_optimize_manifold(acq_fct,
sphere_manifold,
solver,
q=1,
num_restarts=5,
raw_samples=100,
bounds=bounds)
# Get new observation
new_y = test_function(new_x)[0]
# Update training points
x_data = torch.cat((x_data, new_x))
y_data = torch.cat((y_data, new_y))
# Update best observation
new_best_f, index = y_data.min(0)
best_x.append(x_data[index])
best_f.append(new_best_f)