Coarse Models
Several tools are included for building the variance and discrepancy functions, discussed in [1–3, 5]. These methods require, first, the construction of a coarse scale transition matrix $\tilde{k}$. This typically corresponds to an approximation of the Markov kernel, $K$, on the user defined bins,
\[\tilde{K}_{pq} \approx \mathbb{P}(p\to q),\]
Having, computed this matrix, we next solve for the coarse scale discrepancy and one step mutation variance functions:
\[\begin{gather*} (I - \tilde{K})\tilde{h} = \tilde{f} - \tilde{\mu}(\tilde{f})\\ \tilde{v}^2(p) = \mathrm{Var}_{\tilde{K}_{p,\bullet}}(\tilde{h}) \end{gather*}\]
Serial Methods
WeightedEnsemble.build_coarse_transition_matrix
— Functionbuild_coarse_transition_matrix(mutation!, bin_id, x0_vals, n_bins, n_trials)
Contruct a transition matrix amongst the bins (serial version).
Arguments
mutation!
- an in place mutation functionbin_id
- bin identification functionx0_vals
- an array of starting valuesbin0_vals
- an array of the bins corresponding tox0_vals
n_bins
- total number of binsn_trials
- number of independent trials for each x0 starting value
WeightedEnsemble.build_coarse_poisson
— Functionbuild_coarse_poisson(K̃, f̃)
Construct the solution to the Poisson problem and the 1-step variance approximation on the coarser model given the transition matrix, K̃
, and a coarse scale QoI function, f̃
. This solves it using Julia's linear solver.
Arguments
K̃
- coarse scale transition matrixf̃
- quantity of interest vector on the bin space
WeightedEnsemble.build_coarse_vectors
— Functionbuild_coarse_vectors(n_we_steps, K̃, f̃)
Assemble the conditional expectation and 1- step variance approximations on a coarse model, given the transition matrix, K̃
, and a coarse scale QoI function, f̃
.
Arguments
n_we_steps
- number of WE stepsK̃
- coarse scale transition matrixf̃
- quantity of interest vector on the bin space
While build_coarse_poisson
is appropriate when using WE with steady state problems, build_coarse_vectors
is what should be invoked for finite time horizon problems; see [5].
Multithreaded Methods
TBW
Distributed Parellel Methods
TBW