KalmanProcessUtils

PiecewiseConstantDecay{FT <: Real, IT <: Int} <: LearningRateScheduler

Piecewise constant decay learning rate scheduler.

Halves the time step periodically with period τ.

Fields

• Δt_init::Real: Initial learning rate

• τ::Int64: Halving time

Constructors

PiecewiseConstantDecay(Δt_init, τ)

defined at /home/runner/work/CalibrateEDMF.jl/CalibrateEDMF.jl/src/KalmanProcessUtils.jl:54.

PiecewiseConstantGrowth{FT <: Real, IT <: Int} <: LearningRateScheduler

Piecewise constant growth learning rate scheduler.

Doubles the time step periodically with period τ.

Fields

• Δt_init::Real: Initial learning rate

• τ::Int64: Doubling time

Constructors

PiecewiseConstantGrowth(Δt_init, τ)

defined at /home/runner/work/CalibrateEDMF.jl/CalibrateEDMF.jl/src/KalmanProcessUtils.jl:76.

Retrieve learning rate Δt from a LearningRateScheduler

generate_ekp(
    ref_stats::ReferenceStatistics,
    process::Process,
    u::Union{Matrix{T}, T} = nothing;
    failure_handler::String = "ignore_failures",
    localizer::LocalizationMethod = NoLocalization(),
    outdir_path::String = pwd(),
    to_file::Bool = true,
    verbose::Bool = true,
) where {T}

Generates, and possible writes to file, an EnsembleKalmanProcess from a parameter ensemble and reference statistics.

Inputs:

• ref_stats :: ReferenceStatistics defining the inverse problem.
• process :: Type of EnsembleKalmanProcess used to evolve the ensemble.
• u :: An ensemble of parameter vectors, used if !isa(process, Unscented).
• failure_handler :: String describing what failure handler to use.
• localizer :: Covariance localization method.
• outdir_path :: Output path.
• to_file :: Whether to write the serialized prior to a JLD2 file.
• verbose :: Whether to use verbose EKP object

Output:

• The generated EnsembleKalmanProcess.

generate_tekp(
    ref_stats::ReferenceStatistics,
    priors::ParameterDistribution,
    process::Process,
    u::Union{Matrix{T}, T} = nothing;
    l2_reg::Union{Dict{String, Vector{R}}, R} = nothing,
    failure_handler::String = "ignore_failures",
    localizer::LocalizationMethod = NoLocalization(),
    outdir_path::String = pwd(),
    to_file::Bool = true,
    verbose::Bool = true,
) where {T, R}

Generates, and possible writes to file, a Tikhonov EnsembleKalmanProcess from a parameter ensemble and reference statistics.

Tikhonov regularization is implemented through output state augmentation with the input parameter vector. The input L2 regularization hyperparameter should be interpreted as the inverse of the variance of our prior belief in the magnitude of the parameters.

Inputs:

• ref_stats :: ReferenceStatistics defining the inverse problem.
• priors :: Parameter priors used for L2 (i.e., Tikhonov) regularization
• process :: Type of EnsembleKalmanProcess used to evolve the ensemble.
• u :: An ensemble of parameter vectors, used if !isa(process, Unscented).
• l2_reg :: L2 regularization hyperparameter driving parameter values toward prior. May be a float (isotropic regularization) or a dictionary of regularizations per parameter.
• failure_handler :: String describing what failure handler to use.
• localizer :: Covariance localization method.
• outdir_path :: Output path.
• to_file :: Whether to write the serialized prior to a JLD2 file.
• verbose :: Whether to use verbose EKP object

Output:

• The generated augmented EnsembleKalmanProcess.

Returns the sparse parameter indices given the sparsity configuration and the number of parameters.

Returns the indices of parameters to be regularized, given the l2 regularization configuration dictionary.