Bayesian quantification of parameter uncertainty¶
I. Estimating the posterior pdf of the coefficient parameter field in an elliptic PDE¶
In this example we tackle the problem of quantifying the uncertainty in the solution of an inverse problem governed by an elliptic PDE via the Bayesian inference framework. Hence, we state the inverse problem as a problem of statistical inference over the space of uncertain parameters, which are to be inferred from data and a physical model. The resulting solution to the statistical inverse problem is a posterior distribution that assigns to any candidate set of parameter fields, our belief (expressed as a probability) that a member of this candidate set is the “true” parameter field that gave rise to the observed data.
Bayes’s Theorem¶
The posterior probability distribution combines the prior pdf \(\mu_{\text{prior}}(m)\) over the parameter space, which encodes any knowledge or assumptions about the parameter space that we may wish to impose before the data are considered, with a likelihood pdf \(\pi_{\text{like}}(\data \; | \; m)\), which explicitly represents the probability that a given parameter \(m\) might give rise to the observed data \(\data \in \mathbb{R}^{n_t}\), namely:
Note that infinite-dimensional analog of Bayes’s formula requires the use Radon-Nikodym derivatives instead of probability density functions.
The prior¶
We consider a Gaussian prior with mean \(m_{\rm prior}\) and covariance \(\prcov\), \(\mu_{\rm prior} \sim \mathcal{N}({m}_{\rm prior}, \prcov)\). The covariance is given by the discretization of the inverse of differential operator \(\mathcal{A}^{-2} = (-\gamma \Delta + \delta I)^{-2}\), where \(\gamma\), \(\delta > 0\) control the correlation length and the variance of the prior operator. This choice of prior ensures that it is a trace-class operator, guaranteeing bounded pointwise variance and a well-posed infinite-dimensional Bayesian inverse problem.
The likelihood¶
Here \({\bf f}\) is the parameter-to-observable map that takes a parameter \(m\) and maps it to the space observation vector \(\data\).
In this application, \({\bf f}\) consists in the composition of a PDE solve (to compute the state \(u\)) and a pointwise observation of the state \(u\) to extract the observation vector \(\data\).
The posterior¶
The Laplace approximation to the posterior: \(\nu \sim \mathcal{N}({\map},\bf \postcov)\)¶
The mean of the Laplace approximation posterior distribution, \({\map}\), is the parameter maximizing the posterior, and is known as the maximum a posteriori (MAP) point. It can be found by minimizing the negative log of the posterior, which amounts to solving a deterministic inverse problem) with appropriately weighted norms,
The posterior covariance matrix is then given by the inverse of the Hessian matrix of \(\mathcal{J}\) at \(\map\), namely
provided that \(\Hmisfit(\map)\) is positive semidefinite.
The generalized eigenvalue problem¶
In what follows we denote with \(\matHmis, \Gpost, \Gprior \in \mathbb{R}^{n\times n}\) the matrices stemming from the discretization of the operators \(\Hmisfit(\map)\), \(\postcov\), \(\prcov\) with respect to the unweighted Euclidean inner product. Then we considered the symmetric generalized eigenvalue problem
where \({\matrix \Lambda} = \diag(\lambda_i) \in \mathbb{R}^{n\times n}\) contains the generalized eigenvalues and the columns of \({\matrix V}\in \mathbb R^{n\times n}\) the generalized eigenvectors such that \({\matrix V}^T \Gprior^{-1} {\matrix V} = {\matrix I}\).
Randomized eigensolvers to construct the approximate spectral decomposition¶
When the generalized eigenvalues \(\{\lambda_i\}\) decay rapidly, we can extract a low-rank approximation of \(\matHmis\) by retaining only the \(r\) largest eigenvalues and corresponding eigenvectors,
Here, \(\Vr \in \mathbb{R}^{n\times r}\) contains only the \(r\) generalized eigenvectors of \(\matHmis\) that correspond to the \(r\) largest eigenvalues, which are assembled into the diagonal matrix \({\matrix{\Lambda}}_r = \diag (\lambda_i) \in \mathbb{R}^{r \times r}\).
The approximate posterior covariance¶
Using the Sherman–Morrison–Woodbury formula, we write
where \({\matrix{D}}_r :=\diag(\lambda_i/(\lambda_i+1)) \in \mathbb{R}^{r\times r}\). The last term in this expression captures the error due to truncation in terms of the discarded eigenvalues; this provides a criterion for truncating the spectrum, namely that \(r\) is chosen such that \(\lambda_r\) is small relative to 1.
Therefore we can approximate the posterior covariance as
Drawing samples from a Gaussian distribution with covariance $Gpost$¶
Let \({\bf x}\) be a sample for the prior distribution, i.e. \({\bf x} \sim \mathcal{N}({\bf 0}, \Gprior)\), then, using the low rank approximation of the posterior covariance, we compute a sample \({\bf v} \sim \mathcal{N}({\bf 0}, \Gpost)\) as
Full posterior sampling via Markov chain Monte Carlo (MCMC)¶
The posterior can be fully explored by using MCMC algorithms, the most popular method for sampling from a probability distribution. In this example, some of the advanced MCMC algorithms are considered and compared in terms of efficiency and accuracy.
The preconditioned Crank-Nicolson algorithm (pCN)¶
The pCN algorithm is perhaps the simplest MCMC method that is well-defined in the infinite dimensional setting ensuring a mixing rates independent of the dimension of the discretized parameter space.
The algorithm proceeds as follows (see [Cotter et al. (2013)] [Pinski et al. (2015)] for the details):
Given \(m^{(k)}\), propose \(v^{(k+1)} = m_{\rm prop} + \sqrt{1 - \beta^2}(m^{(k)} - m_{\rm prop}) + \beta \xi^{(k)}, \quad \xi^{(k)} \sim \mathcal{N}( 0, \mathcal{C}_{\rm prop} )\)
Set \(m^{(k+1)} = v^{(k+1)}\) with probability \(a(m^{(k)}, v^{(k+1)}) = \min \left(1, \frac{\mu_{\text{post}}(v^{(k+1)}) q(v^{(k+1)}, m^{(k)})}{\mu_{\text{post}}(m^{(k)}) q(m^{(k)}, v^{(k+1)})} \right)\)
where \(q(m,v) \sim \mathcal{N}\left( m_{\rm prop} + \sqrt{1 - \beta^2}(m - m_{\rm prop}), \beta^2 \mathcal{C}_{\rm prop} \right)\) with proposal mean \(m_{\rm prop}\) and covariance \(\mathcal{C}_{\rm prop}\) and \(\beta\) is a parameter controlling the step length of the proposal.
The preconditioned Metropolis adjusted Langevin algorithm (MALA)¶
The MALA algorithm is built on two mechanisms: the overdamped Langevin diffusion to propose a move and the Metropolis–Hastings algorithm to accept or reject the proposal move [Roberts and Tweedie (1996)].
The preconditioned MALA algorithm is described as follows:
Given \(m^{(k)}\), propose \(v^{(k+1)} = m^{(k)} + \tau \mathcal{A}_{\rm prop} \nabla \log \mu_{\text{post}} (m^{(k)}) + \sqrt{2 \tau \mathcal{A}_{\rm prop}} \xi^{(k)}, \quad \xi^{(k)} \sim \mathcal{N}( 0, \mathcal{I})\)
Set \(m^{(k+1)} = v^{(k+1)}\) with probability \(a(m^{(k)}, v^{(k+1)}) = \min \left(1, \frac{\mu_{\text{post}}(v^{(k+1)}) q(v^{(k+1)}, m^{(k)})}{\mu_{\text{post}}(m^{(k)}) q(m^{(k)}, v^{(k+1)})} \right)\)
where \(q(m,v) \sim \mathcal{N}\left( m + \tau \mathcal{A}_{\rm prop} \nabla \log \mu_{\text{post}} (m), 2 \tau \mathcal{A}_{\rm prop} \right)\) with a proposal covariance \(\mathcal{A}_{\rm prop}\) and \(\tau\) is a step size.
The Delayed Rejection (DR)¶
The basic idea of the delayed rejection is to use a sequence of stages in each iteration. Unlike the basic Metropolis-Hastings algorithm, if a candidate is rejected, a new move is proposed. The acceptance rate for the new proposal move is adjusted so that the stationary distribution is preserved. For the details, see [Mira (2001)].
This tutorial shows¶
Definition of the component of an inverse problem (the forward problem, the prior, and the misfit functional) using hIPPYlib
Computation of the maximum a posterior MAP point using inexact Newton-CG algorithm
Low-rank based approximation of the posterior covariance under the Laplace Approximation
Sampling from the prior distribution and Laplace Approximation using hIPPYlib
Construction of a MUQ workgraph using a PDE model defined in hIPPYlib
Exploring the full posterior using the MCMC methods implemented in MUQ
Convergence diagnostics of MCMC simulation results and their comparison
Mathematical tools used¶
Finite element method
Derivation of gradient and Hessian via the adjoint method
Inexact Newton-CG
Randomized eigensolvers
Bayes’ formula
MCMC methods
List of software used¶
hIPPYlib, MUQ and their interfaces are the main software framework in this tutorial. Additional tools used are:
FEniCS, A parallel finite element element library for the discretization of partial differential equations
PETSc, A set of data structures and routines for scalable and efficient linear algebra operations and solvers
Numpy, A python package for linear algebra
Matplotlib, A python package for visualizing the results
References¶
Cotter, S. L., Roberts, G. O., Stuart, A. M., & White, D. (2013). MCMC methods for functions: modifying old algorithms to make them faster. Statistical Science, 424-446.
Pinski, F. J., Simpson, G., Stuart, A. M., & Weber, H. (2015). Algorithms for Kullback–Leibler approximation of probability measures in infinite dimensions. SIAM Journal on Scientific Computing, 37(6), A2733-A2757.
Roberts, G. O., & Tweedie, R. L. (1996). Exponential convergence of Langevin distributions and their discrete approximations. Bernoulli, 2(4), 341-363.
Mira, A. (2001). On Metropolis-Hastings algorithms with delayed rejection. Metron, 59(3-4), 231-241.
II. hIPPYlib-MUQ integration¶
The main objective of this example is to illustrate the interface between hIPPYlib and MUQ.
We make use of hIPPYlib to
Define the forward model, prior distribution, and likelihood function
Compute the MAP point by solving a deterministic inverse problem
Construct the Laplace Approximation to the posterior distribution with a low-rank based approximation of the covariace operator.
The main classes and functions of hIPPYlib employed in this example are
hippylib::PDEVariationalProblem: forward, adjoint and incremental problems solvers and their derivatives evaluationshippylib::BiLaplacianPrior: a biLaplacian Gaussian prior modelhippylib::GaussianLRPosterior: the low rank Gaussian approximation of the posterior (used for generating starting points of MCMC simulations)
MUQ is used to sample from the posterior by implementing MCMC methods with various kernels and proposals.
The main classes and functions used here are
muq.Modeling::PyModPiece: an abstract interface for defining vector-valued modelsmuq.Modeling::PyGaussianBase: an abstract interface for implementing Gaussian distributionsmuq.Modeling::WorkGraph: a graph or a frame of connectedmuq.Modeling::PyModPiece(ormuq.Modeling::WorkPiece) classesmuq.SamplingAlgorithms::CrankNicolsonProposal: the pCN proposalmuq.SamplingAlgorithms::MALAProposal: the MALA proposalmuq.SamplingAlgorithms::MHKernel: the Metropolis-Hastings transition kernelmuq.SamplingAlgorithms::DRKernel: the delayed rejection kernelmuq.SamplingAlgorithms::SingleChainMCMC: a single chain MCMC sampler
To interface hIPPYlib and MUQ for this example, hippylib2muq provides the following classes:
hippylib2muq::Param2LogLikelihood: a child ofmuq::PyModPiecewhich wrapshippylib::PDEVariationalProblemandhippylib:PointwiseStateObservation(solving the forward problem, mapping from parameters to log likelihood and evaluating its derivative)hippylib2muq::BiLaplaceGaussian: a child ofmuq.Modeling::PyGaussianBasewhich wrapshippylib::BiLaplacianPriorhippylib2muq::LAPosteriorGaussian: a child ofmuq.Modeling::PyGaussianBasewhich wrapshippylib::GaussianLRPosterior
III. Implementation¶
1. Load modules¶
from __future__ import absolute_import, division, print_function
import math
import matplotlib.pyplot as plt
%matplotlib inline
import muq.Modeling_ as mm
import muq.SamplingAlgorithms as ms
import dolfin as dl
import hippylib as hp
import hippylib2muq as hm
import numpy as np
import logging
logging.getLogger('FFC').setLevel(logging.WARNING)
logging.getLogger('UFL').setLevel(logging.WARNING)
dl.set_log_active(False)
np.random.seed(seed=1)
2. Generate the true parameter¶
This function generates a random field with a prescribed anisotropic covariance function.
def true_model(prior):
noise = dl.Vector()
prior.init_vector(noise, "noise")
hp.parRandom.normal(1., noise)
mtrue = dl.Vector()
prior.init_vector(mtrue, 0)
prior.sample(noise, mtrue)
return mtrue
3. Set up the mesh and finite element spaces¶
We compute a two dimensional mesh of a unit square with nx by ny elements. We define a P2 finite element space for the state and adjoint variable and P1 for the parameter.
ndim = 2
nx = 32
ny = 32
mesh = dl.UnitSquareMesh(nx, ny)
Vh2 = dl.FunctionSpace(mesh, 'Lagrange', 2)
Vh1 = dl.FunctionSpace(mesh, 'Lagrange', 1)
Vh = [Vh2, Vh1, Vh2]
print("Number of dofs: STATE={0}, PARAMETER={1}, ADJOINT={2}".format(
Vh[hp.STATE].dim(), Vh[hp.PARAMETER].dim(), Vh[hp.ADJOINT].dim()) )
Number of dofs: STATE=4225, PARAMETER=1089, ADJOINT=4225
4. Set up the forward problem¶
Let \(\Omega\) be the unit square in \(\mathbb{R}^2\), and \(\Gamma_D\), \(\Gamma_N\) be the Dirichlet and Neumann portitions of the boundary \(\partial \Omega\) (that is \(\Gamma_D \cup \Gamma_N = \partial \Omega\), \(\Gamma_D \cap \Gamma_N = \emptyset\)). The forward problem reads
where \(u \in \mathcal{V}\) is the state variable, and \(m \in \mathcal{M}\) is the uncertain parameter. Here \(\Gamma_D\) corresponds to the top and bottom sides of the unit square, and \(\Gamma_N\) corresponds to the left and right sides. We also let \(f = 0\), and \(u_D = 1\) on the top boundary and \(u_D = 0\) on the bottom boundary.
To set up the forward problem we use the hp::PDEVariationalProblem class, which requires the following inputs
the finite element spaces for the state, parameter, and adjoint variables
Vhthe pde in weak form
pde_varfthe boundary conditions
bcfor the forward problem andbc0for the adjoint and incremental problems.
The hp::PDEVariationalProblem class offer the following functionality:
solving the forward/adjoint and incremental problems
evaluate first and second partial derivative of the forward problem with respect to the state, parameter, and adojnt variables.
def u_boundary(x, on_boundary):
return on_boundary and ( x[1] < dl.DOLFIN_EPS or x[1] > 1.0 - dl.DOLFIN_EPS)
u_bdr = dl.Expression("x[1]", degree=1)
u_bdr0 = dl.Constant(0.0)
bc = dl.DirichletBC(Vh[hp.STATE], u_bdr, u_boundary)
bc0 = dl.DirichletBC(Vh[hp.STATE], u_bdr0, u_boundary)
f = dl.Constant(0.0)
def pde_varf(u,m,p):
return dl.exp(m)*dl.inner(dl.nabla_grad(u), dl.nabla_grad(p))*dl.dx - f*p*dl.dx
pde = hp.PDEVariationalProblem(Vh, pde_varf, bc, bc0, is_fwd_linear=True)
5. Set up the prior¶
To obtain the synthetic true parameter \(m_{\rm true}\) we generate a realization from the prior distribution.
Here we assume a Gaussian prior, \(\mu_{\rm prior} \sim \mathcal{N}(0, \prcov)\)
with zero mean and covariance matrix \(\prcov = \mathcal{A}^{-2}\),
which is implemented by hp::BiLaplacianPrior class that provides methods to apply the regularization (precision) operator to a vector or to apply the prior covariance operator.
gamma = .1
delta = .5
theta0 = 2.
theta1 = .5
alpha = math.pi/4
anis_diff = dl.CompiledExpression(hp.ExpressionModule.AnisTensor2D(), degree = 1)
anis_diff.set(theta0, theta1, alpha)
prior = hp.BiLaplacianPrior(Vh[hp.PARAMETER], gamma, delta, anis_diff, robin_bc=True)
print("Prior regularization: (delta_x - gamma*Laplacian)^order: "
"delta={0}, gamma={1}, order={2}".format(delta, gamma,2))
mtrue = true_model(prior)
objs = [dl.Function(Vh[hp.PARAMETER],mtrue), dl.Function(Vh[hp.PARAMETER],prior.mean)]
mytitles = ["True parameter", "Prior mean"]
hp.nb.multi1_plot(objs, mytitles)
plt.show()
Prior regularization: (delta_x - gamma*Laplacian)^order: delta=0.5, gamma=0.1, order=2
6. Set up the likelihood and generate synthetic observations¶
To setup the observation operator \(\mathcal{B}: \mathcal{V} \mapsto \mathbb{R}^{n_t}\),
we generate \(n_t\) (ntargets in the code below) random locations where to evaluate the value of the state.
Under the assumption of Gaussian additive noise, the likelihood function \(\pi_{\rm like}\) has the form
where \(u(m)\) denotes the solution of the forward model at a given parameter \(m\).
The class hp::PointwiseStateObservation implements the evaluation of the
log-likelihood function and of its partial derivatives w.r.t. the state
\(u\) and parameter \(m\).
To generate the synthetic observation, we first solve the forward problem using
the true parameter \(m_{\rm true}\). Synthetic observations are obtained by perturbing the state variable at the observation points with a random Gaussian noise.
rel_noise is the signal to noise ratio.
ntargets = 300
rel_noise = 0.005
targets = np.random.uniform(0.05, 0.95, [ntargets, ndim])
print("Number of observation points: {0}".format(ntargets))
misfit = hp.PointwiseStateObservation(Vh[hp.STATE], targets)
utrue = pde.generate_state()
x = [utrue, mtrue, None]
pde.solveFwd(x[hp.STATE], x)
misfit.B.mult(x[hp.STATE], misfit.d)
MAX = misfit.d.norm("linf")
noise_std_dev = rel_noise * MAX
hp.parRandom.normal_perturb(noise_std_dev, misfit.d)
misfit.noise_variance = noise_std_dev*noise_std_dev
model = hp.Model(pde, prior, misfit)
vmax = max( utrue.max(), misfit.d.max() )
vmin = min( utrue.min(), misfit.d.min() )
plt.figure(figsize=(15,5))
hp.nb.plot(dl.Function(Vh[hp.STATE], utrue), mytitle="True state",
subplot_loc=121, vmin=vmin, vmax=vmax, cmap="jet")
hp.nb.plot_pts(targets, misfit.d, mytitle="Observations",
subplot_loc=122, vmin=vmin, vmax=vmax, cmap="jet")
plt.show()
Number of observation points: 300
7. Compute the MAP point¶
We used the globalized Newtown-CG method to compute the MAP point.
m = prior.mean.copy()
solver = hp.ReducedSpaceNewtonCG(model)
solver.parameters["rel_tolerance"] = 1e-6
solver.parameters["abs_tolerance"] = 1e-12
solver.parameters["max_iter"] = 25
solver.parameters["GN_iter"] = 5
solver.parameters["globalization"] = "LS"
solver.parameters["LS"]["c_armijo"] = 1e-4
x = solver.solve([None, m, None])
if solver.converged:
print( "\nConverged in ", solver.it, " iterations.")
else:
print( "\nNot Converged")
print( "Termination reason: ", solver.termination_reasons[solver.reason] )
print( "Final gradient norm: ", solver.final_grad_norm )
print( "Final cost: ", solver.final_cost )
plt.figure(figsize=(18,4))
mtrue_min = dl.Function(Vh[hp.PARAMETER],mtrue).vector().min()
mtrue_max = dl.Function(Vh[hp.PARAMETER],mtrue).vector().max()
hp.nb.plot(dl.Function(Vh[hp.PARAMETER],mtrue), subplot_loc=131, mytitle="True parameter",
vmin=mtrue_min, vmax=mtrue_max)
hp.nb.plot(dl.Function(Vh[hp.PARAMETER], x[hp.PARAMETER]), subplot_loc=132,mytitle="MAP",
vmin=mtrue_min, vmax=mtrue_max)
hp.nb.plot(dl.Function(Vh[hp.STATE], x[hp.STATE]), subplot_loc=133,mytitle="Recovered state", cmap="jet")
plt.show()
It cg_it cost misfit reg (g,dm) ||g||L2 alpha tolcg
1 2 6.95e+03 6.94e+03 9.30e-01 -7.91e+04 1.78e+05 1.00e+00 5.00e-01
2 3 2.81e+03 2.81e+03 1.36e+00 -8.27e+03 5.53e+04 1.00e+00 5.00e-01
3 4 9.25e+02 9.21e+02 3.58e+00 -3.70e+03 2.53e+04 1.00e+00 3.77e-01
4 10 3.39e+02 3.30e+02 9.32e+00 -1.30e+03 9.45e+03 1.00e+00 2.30e-01
5 1 2.71e+02 2.61e+02 9.32e+00 -1.37e+02 1.41e+04 1.00e+00 2.81e-01
6 13 1.73e+02 1.56e+02 1.63e+01 -1.96e+02 3.73e+03 1.00e+00 1.45e-01
7 16 1.44e+02 1.20e+02 2.40e+01 -5.63e+01 1.78e+03 1.00e+00 1.00e-01
8 12 1.41e+02 1.16e+02 2.45e+01 -6.92e+00 1.14e+03 1.00e+00 8.01e-02
9 43 1.34e+02 9.87e+01 3.50e+01 -1.47e+01 8.67e+02 1.00e+00 6.98e-02
10 3 1.34e+02 9.86e+01 3.50e+01 -1.31e-01 3.77e+02 1.00e+00 4.60e-02
11 42 1.34e+02 9.85e+01 3.51e+01 -8.26e-02 8.90e+01 1.00e+00 2.24e-02
12 59 1.34e+02 9.85e+01 3.51e+01 -7.70e-04 8.86e+00 1.00e+00 7.06e-03
Converged in 12 iterations.
Termination reason: Norm of the gradient less than tolerance
Final gradient norm: 0.10248108935885225
Final cost: 133.60199155509807
8. Compute the low-rank based Laplace approximation of the posterior (LA-posterior)¶
We used the double pass algorithm to compute a low-rank decomposition of the Hessian Misfit. In particular, we solve
The effective rank of the Hessian misfit is the number of eigenvalues above the red line (\(y=1\)). The effective rank is independent of the mesh size.
model.setPointForHessianEvaluations(x, gauss_newton_approx=False)
Hmisfit = hp.ReducedHessian(model, misfit_only=True)
k = 100
p = 20
print( "Single/Double Pass Algorithm. Requested eigenvectors: "\
"{0}; Oversampling {1}.".format(k,p) )
Omega = hp.MultiVector(x[hp.PARAMETER], k+p)
hp.parRandom.normal(1., Omega)
lmbda, V = hp.doublePassG(Hmisfit, prior.R, prior.Rsolver, Omega, k)
nu = hp.GaussianLRPosterior(prior, lmbda, V)
nu.mean = x[hp.PARAMETER]
plt.plot(range(0,k), lmbda, 'b*', range(0,k+1), np.ones(k+1), '-r')
plt.yscale('log')
plt.xlabel('number')
plt.ylabel('eigenvalue')
plt.show()
Single/Double Pass Algorithm. Requested eigenvectors: 100; Oversampling 20.
9. Drawing samples from the prior distribution and Laplace Approximation¶
nsamples = 3
noise = dl.Vector()
nu.init_vector(noise,"noise")
s_prior = dl.Function(Vh[hp.PARAMETER], name="sample_prior")
s_post = dl.Function(Vh[hp.PARAMETER], name="sample_post")
post_pw_variance, pr_pw_variance, corr_pw_variance =\
nu.pointwise_variance(method="Exact")
pr_max = 2.5*math.sqrt( pr_pw_variance.max() ) + prior.mean.max()
pr_min = -2.5*math.sqrt( pr_pw_variance.max() ) + prior.mean.min()
ps_max = 2.5*math.sqrt( post_pw_variance.max() ) + nu.mean.max()
ps_min = -2.5*math.sqrt( post_pw_variance.max() ) + nu.mean.min()
fig = plt.figure(figsize=(18,8))
for i in range(nsamples):
hp.parRandom.normal(1., noise)
nu.sample(noise, s_prior.vector(), s_post.vector())
impr = hp.nb.plot(s_prior, subplot_loc=231+i, vmin=pr_min, vmax=pr_max, colorbar=None)
imps = hp.nb.plot(s_post, subplot_loc=234+i, vmin=ps_min, vmax=ps_max, colorbar=None)
fig.tight_layout()
fig.subplots_adjust(left=0.15, right=0.8)
pos_impr = impr.axes.get_position().get_points()
pos_imps = imps.axes.get_position().get_points()
height_im = impr.axes.get_position().size[1]
cbaxes_pr = fig.add_axes([pos_impr[1,0]+0.01, pos_impr[0,1], 0.01, height_im])
cbaxes_ps = fig.add_axes([pos_imps[1,0]+0.01, pos_imps[0,1], 0.01, height_im])
fig.colorbar(impr, cbaxes_pr)
fig.colorbar(imps, cbaxes_ps)
fig.text(0.15, pos_impr[0,1]+0.125, 'Prior samples', fontsize=20, rotation=90)
fig.text(0.15, pos_imps[0,1]+0.1, 'Laplace samples', fontsize=20, rotation=90)
plt.show()
10 Define a quantify of interest¶
As a quantity of interest, we consider the log of the flux through the bottom boundary:
where the state variable \(u\) denotes the pressure, and \(\mathbf{n}\) is the unit normal vector to \(\Gamma_b\) (the bottom boundary of the domain).
class FluxQOI(object):
def __init__(self, Vh, dsGamma):
self.Vh = Vh
self.dsGamma = dsGamma
self.n = dl.Constant((0.,1.))
self.u = None
self.m = None
self.L = {}
def form(self, x):
return dl.exp(x[hp.PARAMETER])*dl.dot( dl.grad(x[hp.STATE]), self.n)*self.dsGamma
def eval(self, x):
u = hp.vector2Function(x[hp.STATE], self.Vh[hp.STATE])
m = hp.vector2Function(x[hp.PARAMETER], self.Vh[hp.PARAMETER])
return np.log( dl.assemble(self.form([u,m])) )
class GammaBottom(dl.SubDomain):
def inside(self, x, on_boundary):
return ( abs(x[1]) < dl.DOLFIN_EPS )
GC = GammaBottom()
marker = dl.MeshFunction("size_t", mesh, 1)
marker.set_all(0)
GC.mark(marker, 1)
dss = dl.Measure("ds", subdomain_data=marker)
qoi = FluxQOI(Vh,dss(1))
11. Exploring the posterior using MCMC methods¶
Define the parameter-to-observable map in MUQ¶
Overall, we want a mapping from parameter coefficients vector to the log target, \(J(m) = - \tfrac{1}{2} \parallel {\bf f}(m) - \data \parallel^{2}_{{\bf \Gamma}_{\text{noise}}^{-1}} \! - \tfrac{1}{2}\parallel m - m_{\rm prior} \parallel^{2}_{\prcov^{-1}}\). To do so, we generate a MUQ WorkGraph of connected ModPieces.
# a place holder ModPiece for the parameters
idparam = mm.IdentityOperator(Vh[hp.PARAMETER].dim())
# log Gaussian Prior ModPiece
gaussprior = hm.BiLaplaceGaussian(prior)
log_gaussprior = gaussprior.AsDensity()
# parameter to log likelihood Modpiece
param2loglikelihood = hm.Param2LogLikelihood(model)
# log target ModPiece
log_target = mm.DensityProduct(2)
workgraph = mm.WorkGraph()
# Identity operator for the parameters
workgraph.AddNode(idparam, 'Identity')
# Prior model
workgraph.AddNode(log_gaussprior, "Log_prior")
# Likelihood model
workgraph.AddNode(param2loglikelihood, "Log_likelihood")
# Posterior
workgraph.AddNode(log_target, "Log_target")
workgraph.AddEdge("Identity", 0, "Log_prior", 0)
workgraph.AddEdge("Log_prior", 0, "Log_target", 0)
workgraph.AddEdge("Identity", 0, "Log_likelihood", 0)
workgraph.AddEdge("Log_likelihood", 0, "Log_target", 1)
workgraph.Visualize("workgraph.png")
# Construct the problem
problem = ms.SamplingProblem(workgraph.CreateModPiece("Log_target"))
Set up MCMC methods¶
We run five different MCMC methods:
pCN: Metropolis-Hastings kernel + pCN proposal with \(m_{\rm prop} = m_{\rm prior} = 0\) and \(\mathcal{C}_{\rm prop} = \prcov\)
MALA: Metropolis-Hastings kernel + MALA proposal with \(\mathcal{A}_{\rm prop} = \prcov\)
h-pCN: Metropolis-Hastings kernel + pCN proposal with \(m_{\rm prop} = \map\) and \(\mathcal{C}_{\rm prop} = \postcov\)
h-MALA: Metropolis-Hastings kernel + MALA proposal with \(\mathcal{A}_{\rm prop} = \postcov\)
DR (h-pCN/h-MALA): Delayed rejection kernel + two stage proposals (h-pCN proposal as first stage and h-MALA proposal as second stage)
where \(\postcov\) is the covariance of the LA-posterior.
We set the value of parameters (\(\beta\) for pCN and \(\tau\) for MALA) such that the acceptance rates are about 20-35% and 50-60% for pCN and MALA, respectively.
# Construct options for MH kernel and MCMC sampler
options = dict()
options['NumSamples'] = 22000 # Number of MCMC steps to take
options['BurnIn'] = 2000 # Number of steps to throw away as burn in
options['PrintLevel'] = 0 # in {0,1,2,3} Verbosity of the output
method_list = dict()
# pCN
opts = options.copy()
opts.update( {'Beta':0.005} )
gauss_pcn = hm.BiLaplaceGaussian(prior)
prop = ms.CrankNicolsonProposal(opts, problem, gauss_pcn)
kern = ms.MHKernel(opts, problem, prop)
sampler = ms.SingleChainMCMC(opts, [kern])
method_list['pCN'] = {'Options': opts, 'Sampler': sampler}
# MALA
opts = options.copy()
opts.update( {'StepSize':0.000006} )
gauss_mala = hm.BiLaplaceGaussian(prior, use_zero_mean=True)
prop = ms.MALAProposal(opts, problem, gauss_mala)
kern = ms.MHKernel(opts, problem, prop)
sampler = ms.SingleChainMCMC(opts, [kern])
method_list['MALA'] = {'Options': opts, 'Sampler': sampler}
# h-pCN
opts = options.copy()
opts.update( {'Beta':0.55} )
gauss_hpcn = hm.LAPosteriorGaussian(nu)
prop = ms.CrankNicolsonProposal(opts, problem, gauss_hpcn)
kern = ms.MHKernel(opts, problem, prop)
sampler = ms.SingleChainMCMC(opts, [kern])
method_list['h-pCN'] = {'Options': opts, 'Sampler': sampler}
# h-MALA
opts = options.copy()
opts.update( {'StepSize':0.1} )
gauss_hmala = hm.LAPosteriorGaussian(nu, use_zero_mean=True)
prop = ms.MALAProposal(opts, problem, gauss_hmala)
kern = ms.MHKernel(opts, problem, prop)
sampler = ms.SingleChainMCMC(opts, [kern])
method_list['h-MALA'] = {'Options': opts, 'Sampler': sampler}
# DR (h-pCN/h-MALA)
opts = options.copy()
opts.update( {'Beta':1.0, 'StepSize':0.1} )
gauss_dr1 = hm.LAPosteriorGaussian(nu)
gauss_dr2 = hm.LAPosteriorGaussian(nu, use_zero_mean=True)
prop1 = ms.CrankNicolsonProposal(opts, problem, gauss_dr1)
prop2 = ms.MALAProposal(opts, problem, gauss_dr2)
kern = ms.DRKernel( opts, problem, [prop1, prop2], [1.0, 1.0] )
sampler = ms.SingleChainMCMC(opts, [kern])
method_list['DR (h-pCN/h-MALA)'] = {'Options': opts, 'Sampler': sampler}
hm.print_methodDict(method_list)
Method Kernel Proposal Beta or Step-size
----------------------------------------------------------
pCN mh pcn 5.0e-03
MALA mh mala 6.0e-06
h-pCN mh pcn 5.5e-01
h-MALA mh mala 1.0e-01
DR (h-pCN/h-MALA) dr pcn 1.0e+00
mala 1.0e-01
Run MCMC methods¶
# Generate starting sample vector for all the MCMC simulations
noise = dl.Vector()
nu.init_vector(noise, "noise")
hp.parRandom.normal(1., noise)
pr_s = model.generate_vector(hp.PARAMETER)
post_s = model.generate_vector(hp.PARAMETER)
nu.sample(noise, pr_s, post_s, add_mean=True)
x0 = hm.dlVector2npArray(post_s)
# Implement MCMC simulations
for mName, method in method_list.items():
# Run the MCMC sampler
sampler = method['Sampler']
samps = sampler.Run([x0])
# Save the computed results
method['Samples'] = samps
method['ElapsedTime'] = sampler.TotalTime()
kernel = sampler.Kernels()[0]
if "AcceptanceRate" in dir(kernel):
method['AcceptRate'] = kernel.AcceptanceRate()
elif "AcceptanceRates" in dir(kernel):
method['AcceptRate'] = kernel.AcceptanceRates()
print("Drawn ", options['NumSamples'] - options['BurnIn'] + 1,
"MCMC samples using", mName)
print("\n")
print("Parameter space dimension:", Vh[hp.PARAMETER].dim())
print("Number of samples:", options['NumSamples'] - options['BurnIn'] + 1)
# Keep track of the quantity of interest
qoi_dataset = hm.track_qoiTracer(pde, qoi, method_list)
hm.print_qoiResult(method_list, qoi_dataset)
hm.plot_qoiResult(method_list, qoi_dataset, max_lag=300)
Drawn 20001 MCMC samples using pCN
Drawn 20001 MCMC samples using MALA
Drawn 20001 MCMC samples using h-pCN
Drawn 20001 MCMC samples using h-MALA
Drawn 20001 MCMC samples using DR (h-pCN/h-MALA)
Parameter space dimension: 1089
Number of samples: 20001
===================================================
Summary of convergence diagnostics (single chain)
===================================================
Method E[QOI] AR ESS ES/min
-----------------------------------------------
pCN -1.006 0.252 8.8 0.6
MALA -0.844 0.550 5.7 0.1
h-pCN 0.491 0.258 191.1 12.8
h-MALA 0.553 0.543 305.3 5.4
DR (h-pCN/h-MALA) 0.505 0.582 550.8 7.8
Copyright © 2020, Army Corps of Engineers, Massachusetts Institute of Technology, University of California–Merced, The University of Texas at Austin, Washington University in St. Louis
All Rights reserved.
Acknowledgment: This work is supported by the National Science Foundation under grants ACI-1550487, ACI-1550547, and ACI-1550593.