class MetropolisHastings(MCMC):
@beartype
def __init__(
self,
pdf_target: Union[Callable, list[Callable]] = None,
log_pdf_target: Union[Callable, list[Callable]] = None,
args_target: tuple = None,
burn_length: Annotated[int, Is[lambda x: x >= 0]] = 0,
jump: int = 1,
dimension: int = None,
seed: list = None,
save_log_pdf: bool = False,
concatenate_chains: bool = True,
n_chains: int = None,
proposal: Distribution = None,
proposal_is_symmetric: bool = False,
random_state: RandomStateType = None,
nsamples: PositiveInteger = None,
nsamples_per_chain: PositiveInteger = None,
):
"""
Metropolis-Hastings algorithm :cite:`MCMC1` :cite:`MCMC2`
:param pdf_target: Target density function from which to draw random samples. Either `pdf_target` or
`log_pdf_target` must be provided (the latter should be preferred for better numerical stability).
If `pdf_target` is a callable, it refers to the joint pdf to sample from, it must take at least one input
**x**, which are the point(s) at which to evaluate the pdf. Within :class:`.MCMC` the pdf_target is evaluated
as:
:code:`p(x) = pdf_target(x, \*args_target)`
where **x** is a :class:`numpy.ndarray of shape :code:`(nsamples, dimension)` and `args_target` are additional
positional arguments that are provided to :class:`.MCMC` via its `args_target` input.
If `pdf_target` is a list of callables, it refers to independent marginals to sample from. The marginal in
dimension :code:`j` is evaluated as:
:code:`p_j(xj) = pdf_target[j](xj, \*args_target[j])` where **x** is a :class:`numpy.ndarray` of shape
:code:`(nsamples, dimension)`
:param log_pdf_target: Logarithm of the target density function from which to draw random samples.
Either `pdf_target` or `log_pdf_target` must be provided (the latter should be preferred for better numerical
stability).
Same comments as for input `pdf_target`.
:param args_target: Positional arguments of the pdf / log-pdf target function. See `pdf_target`
:param burn_length: Length of burn-in - i.e., number of samples at the beginning of the chain to discard (note:
no thinning during burn-in). Default is :math:`0`, no burn-in.
:param jump: Thinning parameter, used to reduce correlation between samples. Setting :code:`jump=n` corresponds
to skipping :code:`n-1` states between accepted states of the chain. Default is :math:`1` (no thinning).
:param dimension: A scalar value defining the dimension of target density function. Either `dimension` and
`n_chains` or `seed` must be provided.
:param seed: Seed of the Markov chain(s), shape :code:`(n_chains, dimension)`.
Default: :code:`zeros(n_chains x dimension)`.
If seed is not provided, both n_chains and dimension must be provided.
:param save_log_pdf: Boolean that indicates whether to save log-pdf values along with the samples.
Default: :any:`False`
:param concatenate_chains: Boolean that indicates whether to concatenate the chains after a run, i.e., samples
are stored as an :class:`numpy.ndarray` of shape :code:`(nsamples * n_chains, dimension)` if :any:`True`,
:code:`(nsamples, n_chains, dimension)` if :any:`False`.
Default: :any:`True`
:param n_chains: The number of Markov chains to generate. Either dimension and `n_chains` or `seed` must be
provided.
:param proposal: Proposal distribution, must have a log_pdf/pdf and rvs method. Default: standard
multivariate normal
:param proposal_is_symmetric: Indicates whether the proposal distribution is symmetric, affects computation of
acceptance probability alpha Default: :any:`False`, set to :any:`True` if default proposal is used
:param random_state: Random seed used to initialize the pseudo-random number generator. Default is
:any:`None`.
:param nsamples: Number of samples to generate.
:param nsamples_per_chain: Number of samples to generate per chain.
"""
self.nsamples = nsamples
self.nsamples_per_chain = nsamples_per_chain
super().__init__(
pdf_target=pdf_target,
log_pdf_target=log_pdf_target,
args_target=args_target,
dimension=dimension,
seed=seed,
burn_length=burn_length,
jump=jump,
save_log_pdf=save_log_pdf,
concatenate_chains=concatenate_chains,
random_state=random_state,
n_chains=n_chains,
)
self.logger = logging.getLogger(__name__)
# Initialize algorithm specific inputs
self.proposal = proposal
self.proposal_is_symmetric = proposal_is_symmetric
if self.proposal is None:
if self.dimension is None:
raise ValueError(
"UQpy: Either input proposal or dimension must be provided."
)
from UQpy.distributions import JointIndependent, Normal
self.proposal = JointIndependent([Normal()] * self.dimension)
self.proposal_is_symmetric = True
else:
self._check_methods_proposal(self.proposal)
self.logger.info("\nUQpy: Initialization of " +
self.__class__.__name__ + " algorithm complete.")
if (nsamples is not None) or (nsamples_per_chain is not None):
self.run(
nsamples=nsamples,
nsamples_per_chain=nsamples_per_chain,
)
def run_one_iteration(self, current_state: np.ndarray,
current_log_pdf: np.ndarray):
"""
Run one iteration of the mcmc chain for MH algorithm, starting at current state -
see :class:`MCMC` class.
"""
# Sample candidate
candidate = current_state + self.proposal.rvs(
nsamples=self.n_chains, random_state=self.random_state)
# Compute log_pdf_target of candidate sample
log_p_candidate = self.evaluate_log_target(candidate)
# Compute acceptance ratio
if self.proposal_is_symmetric: # proposal is symmetric
log_ratios = log_p_candidate - current_log_pdf
else: # If the proposal is non-symmetric, one needs to account for it in computing acceptance ratio
log_proposal_ratio = self.proposal.log_pdf(
candidate -
current_state) - self.proposal.log_pdf(current_state -
candidate)
log_ratios = log_p_candidate - current_log_pdf - log_proposal_ratio
# Compare candidate with current sample and decide or not to keep the candidate (loop over nc chains)
accept_vec = np.zeros(
(self.n_chains, )
) # this vector will be used to compute accept_ratio of each chain
unif_rvs = (Uniform().rvs(nsamples=self.n_chains,
random_state=self.random_state).reshape(
(-1, )))
for nc, (cand, log_p_cand,
r_) in enumerate(zip(candidate, log_p_candidate, log_ratios)):
accept = np.log(unif_rvs[nc]) < r_
if accept:
current_state[nc, :] = cand
current_log_pdf[nc] = log_p_cand
accept_vec[nc] = 1.0
# Update the acceptance rate
self._update_acceptance_rate(accept_vec)
return current_state, current_log_pdf
# %% md # # Create a distribution object, generate samples and evaluate the function at the samples. # %% np.random.seed(1) dist_1 = Uniform(loc=-5.12, scale=10.24) dist_2 = Uniform(loc=-5.12, scale=10.24) marg = [dist_1, dist_2] joint = JointIndependent(marginals=marg) n_samples = 100 x = joint.rvs(n_samples) y = function(x[:, 0], x[:, 1]) # %% md # # Visualize the 2D function. # %% xmin, xmax = -6, 6 ymin, ymax = -6, 6 X1 = np.linspace(xmin, xmax, 50) X2 = np.linspace(ymin, ymax, 50) X1_, X2_ = np.meshgrid(X1, X2) # grid of points f = function(X1_, X2_)
# %%
np.random.seed(1)
dist_1 = Uniform(loc=0, scale=2 * np.pi)
dist_2 = Uniform(loc=0, scale=1)
marg = [dist_1] * 4
marg_1 = [dist_2] * 4
marg.extend(marg_1)
joint = JointIndependent(marginals=marg)
n_samples = 9000
x = joint.rvs(n_samples)
y = function(x)
# %% md
#
# Create an object from the PCE class. Compute PCE coefficients using least squares regression.
# %%
max_degree = 6
polynomial_basis = TotalDegreeBasis(joint, max_degree)
least_squares = LeastSquareRegression()
pce = PolynomialChaosExpansion(polynomial_basis=polynomial_basis,
regression_method=least_squares)
pce.fit(x, y)
# %%
errors = []
# construct PCE surrogate models
for max_degree in range(1, 6):
print('Total degree: ', max_degree)
polynomial_basis = TotalDegreeBasis(joint, max_degree)
print('Size of basis:', polynomial_basis.polynomials_number)
# training data
sampling_coeff = 5
print('Sampling coefficient: ', sampling_coeff)
np.random.seed(42)
n_samples = math.ceil(sampling_coeff * polynomial_basis.polynomials_number)
print('Training data: ', n_samples)
xx = joint.rvs(n_samples)
yy = np.array([analytical_eigenvalues_2d(dim_out, x[0], x[1]) for x in xx])
# fit model
least_squares = LeastSquareRegression()
pce_metamodel = PolynomialChaosExpansion(polynomial_basis=polynomial_basis,
regression_method=least_squares)
pce_metamodel.fit(xx, yy)
# coefficients
# print('PCE coefficients: ', pce.C)
# validation errors
np.random.seed(999)
n_samples = 1000
x_val = joint.rvs(n_samples)
# %% # input distributions dist = Uniform(loc=0, scale=1) marg = [dist] * 3 joint = JointIndependent(marginals=marg) # %% md # # Compute reference mean and variance values using Monte Carlo sampling. # %% # reference moments via Monte Carlo Sampling n_samples_mc = 1000000 xx = joint.rvs(n_samples_mc) yy = function(xx) mean_ref = yy.mean() var_ref = yy.var() # %% md # # Create validation data sets, to be used later to estimate the accuracy of the PCE. # %% # validation data sets n_samples_val = 1000 xx_val = joint.rvs(n_samples_val) yy_val = function(xx_val)
# %% md
#
# We must now compute the PCE coefficients. For that we first need a training sample of input random variable
# realizations and the corresponding model outputs. These two data sets form what is also known as an
# ''experimental design''. In case of adaptive construction of PCE by the best model selection algorithm, size of
# ED is given apriori and the most suitable basis functions are adaptively selected.
# %%
# create training data
sample_size = 500
print('Size of experimental design:', sample_size)
# realizations of random inputs
xx_train = joint.rvs(sample_size)
# corresponding model outputs
yy_train = np.array([ishigami(x) for x in xx_train])
# %% md
#
# We now fit the PCE coefficients by solving a regression problem. Here we opt for the _np.linalg.lstsq_ method,
# which is based on the _dgelsd_ solver of LAPACK. This original PCE class will be used for further selection of
# the best basis functions.
# %%
# fit model
least_squares = LeastSquareRegression()
pce = PolynomialChaosExpansion(polynomial_basis=polynomial_basis,
regression_method=least_squares)
pf_stretch = np.zeros((ntrials, 1))
cov1_stretch = np.zeros((ntrials, 1))
cov2_stretch = np.zeros((ntrials, 1))
m = np.ones(2)
m[0] = 5
m[1] = 125
C = np.eye(2)
C[0, 0] = 1
C[1, 1] = 20**2
for i in range(ntrials):
m1 = PythonModel(model_script='local_Resonance_pfn.py',
model_object_name="RunPythonModel")
model = RunModel(model=m1)
dist = MultivariateNormal(mean=m, cov=C)
xx = dist.rvs(nsamples=1000, random_state=123)
xx1 = dist.rvs(nsamples=100, random_state=123)
sampling = Stretch(dimension=2, n_chains=100, log_pdf_target=dist.log_pdf)
x_ss_stretch = SubsetSimulation(
sampling=sampling,
runmodel_object=model,
conditional_probability=0.1,
nsamples_per_subset=1000,
samples_init=xx,
)
pf_stretch[i] = x_ss_stretch.failure_probability
cov1_stretch[i] = x_ss_stretch.independent_chains_CoV
cov2_stretch[i] = x_ss_stretch.dependent_chains_CoV
print(pf_stretch)
