BaytesFilters

Particle Filter module.

abstract type AbstractParticles

Super type for various particle referencing techniques.

Referencing struct: last element in particle history weighted via reference index

Referencing struct: last element in particle history always reference index

struct InitialTrajectory{K<:ParticleKernel, C<:ParticleReferencing, M<:ParticleFilterMemory}

Contains information to obtain reference trajectory for sampling process.

Fields

• kernel::ParticleKernel

  Kernel for particle propagation

• referencing::ParticleReferencing

  Referencing method

• memory::ParticleFilterMemory

  Memory for observed and latent data in PF.

Referencing struct: no updating via reference index

struct Markov{A, B, C} <: ParticleKernel

Markov Kernel for particle propagation.

Fields

• initial::Any

  Initial distribution, function of iter only.

• transition::Any

  Transition distribution, function of full particle trajectory and current iteration count.

• evidence::Any

  Data distribution to weight particles. Function of full data, particle trajectory and current iteration count.

struct ParticleBuffer{A, I, P}

Contains temporary buffer values to avoid allocations during particle propagation.

Fields

• parameter::BaytesCore.ParameterBuffer{A, I} where {A, I}

  Contains buffer values for particles and ancestor for one iteration.

• proposal::Vector

  Proposal trajectory and predicted latent variable.

• prediction::Vector

  Predicted latent and oberved data

• resampled::Vector{Bool}

  Stores boolean if resampled at each iteration.

• ℓobjectiveᵥ::Vector{Float64}

  Stores incremental log target estimates for each iteration.

struct ParticleFilter{A<:Particles, B<:ParticleFilterTune} <: BaytesCore.AbstractAlgorithm

Contains particles, transition kernel and all other relevant tuning information for particle propagation.

Fields

• particles::Particles

  Particle values and kernel.

• tune::ParticleFilterTune

  Tuning configuration for particles.

mutable struct ParticleFilterConfig{A, B, C, D}

Holds information for array structure of data and reference particle.

Fields

• data::BaytesCore.ArrayConfig{A, B} where {A, B}

  Data size and sorting information.

• particle::BaytesCore.ArrayConfig{C, D} where {C, D}

  Particle size and sorting information.

Callable struct to make initializing ParticleFilter sampler easier in sampling library.

Examples

struct ParticleFilterDefault{M<:Union{Nothing, ParticleFilterMemory}, A<:ParameterWeighting, B<:ResamplingMethod, C<:ParticleReferencing, T<:Integer, I<:AbstractInitialization, U<:UpdateBool}

Default arguments for Particle Filter constructor.

Fields

• memory::Union{Nothing, ParticleFilterMemory}

  Memory for particles and data.

• weighting::ParameterWeighting

  Weighting Methods for particles.

• resampling::ResamplingMethod

  Resampling methods for particle trajectories.

• referencing::ParticleReferencing

  Referencing type for last particle at each iteration - either Conditional, Ancestral or Marginal Implementation.

• coverage::Float64

  Coverage of Particles/datapoints.

• threshold::Float64

  ESS threshold for resampling particle trajectories.

• ancestortype::Type{T} where T<:Integer

  Type for ancestors indices

• init::AbstractInitialization

  Method to obtain initial Modelparameter, see 'AbstractInitialization'.

• generated::UpdateBool

  Boolean if generate(_rng, objective) for corresponding model is stored in PF Diagnostics.

Contains information about log-likelihood, expected sample size and proposal trajectory.

Examples

struct ParticleFilterMemory

Contains maximum memory for latent and observed data. Relevant for number of propagation steps from initial particle distribution and for first time particles are weighted with data point.

Fields

• latent::Int64

  Latent variable memory. Number of times that initial particle distribution is used.

• data::Int64

  Observed data memory. First particle weighting taken at this point.

• initial::Int64

  Number of times when sampled from initial distribution. Per default equal to latent.

struct ParticleFilterTune{T<:ModelWrappers.Tagged, A<:ParameterWeighting, B<:ResamplingMethod, C<:ParticleReferencing, D, E, F, G, U<:UpdateBool} <: BaytesCore.AbstractTune

Holds tuning information for Particle Filter.

Fields

• tagged::ModelWrappers.Tagged

  Tagged Model parameter.

• weighting::ParameterWeighting

  Weighting Methods for particles.

• resampling::ResamplingMethod

  Resampling methods for particle trajectories.

• referencing::ParticleReferencing

  Referencing type for last particle at each iteration - either Conditional, Ancestral or Marginal Implementation.

• config::BaytesFilters.ParticleFilterConfig

  Contains data and reference size and sorting.

• chains::BaytesCore.ChainsTune

  Number of particle chains and tuning information.

• memory::ParticleFilterMemory

  Memory for latent and observed data.

• generated::UpdateBool

  Boolean if generated quantities should be generated while sampling

• iter::BaytesCore.Iterator

  Current iteration.

abstract type ParticleKernel <: BaytesCore.AbstractKernel

Available Particle Filter Kernels.

abstract type ParticleReferencing

Super type for various particle referencing techniques.

mutable struct Particles{R, B, I<:Integer} <: AbstractParticles

Contains Particle container for propagation.

Fields

• val::ElasticArrays.ElasticArray{B, 2, 1, Vector{B}} where B

  Particle trajectories, for a discussion about possible shapes for the trajectories.

• ancestor::ElasticArrays.ElasticArray{I, 2, 1, Vector{I}} where I<:Integer

  Saved ancestors of resampling step in pf

• weights::BaytesCore.ParameterWeights

  Particle weights.

• buffer::ParticleBuffer{B, I, R} where {R, B, I<:Integer}

  Necessary buffer values for resampling particles.

• ℓobjective::BaytesCore.Accumulator

  Log likelihood estimate.

struct SemiMarkov{A<:SemiMarkovInitiation, B<:SemiMarkovTransition, C} <: ParticleKernel

Semi-Markov Kernel for particle propagation.

Fields

• initial::SemiMarkovInitiation

  Initial distribution, function of iter only.

• transition::SemiMarkovTransition

  Transition distribution, function of full particle trajectory and current iteration count.

• evidence::Any

  Data distribution to weight particles. Function of full data, particle trajectory and current iteration count.

struct SemiMarkovInitiation{A, B}

Initial distribution for state and duration of semi-Markov kernel.

Fields

• state::Any

  Initial distribution of state variable.

• duration::Any

  Initial distribution of duration variable.

struct SemiMarkovTransition{A, B}

Transition distributions for state and duration of semi-Markov kernel.

Fields

• state::Any

  Transition distribution of state variable.

• duration::Any

  Transition distribution of duration variable.

resize!(particles, reference, Nparticles, Ndata)

Resize particles struct with new Nparticles and Ndata size.

Examples

generate(_rng, algorithm, objective)

Generate statistics for algorithm given model parameter and data.

Examples

generate_showvalues(diagnostics)

Show relevant diagnostic results.

Examples

infer(_rng, kernel, tune, model, data)

Infer type of predictions of kernel.

Examples

infer(_rng, diagnostics, pf, model, data)

Infer ParticleFilter diagnostics type.

Examples

propagate!(_rng, objective, length_forecast)

Propagate forward a single trajectory given ParticleKernel and objective for multiple time steps.

Examples

Propagate particles forward.

Examples

Propagate particle filter forward with new data.

Examples

propagate(_rng, kernel, memory, reference)

Propagate forward a single trajectory given ParticleKernel.

Examples

propose!(_rng, pf, model, data)
propose!(_rng, pf, model, data, proposaltune)

Run particle filter and change reference trajectory with proposal trajectory.

Examples

propose(_rng, kernel, pf, objective)
propose(_rng, kernel, pf, objective, reference)

Run particle filter and do not change reference trajectory.

Examples

resample!(_rng, kernel, particles, tune, reference)

Resample particle ancestors and shuffle current particles dependency.

Examples

results(diagnosticsᵛ, pf, Ndigits, quantiles)

Return summary statistics for PF diagnostics.

Examples

update!(particles, ParticleUpdate, DataUpdate, reference, Nparticles, Ndata)

Update particles struct with new number of particles. Note that kernel and log likelihood are adjusted in another step.

Examples

_checkmemory(f)

Return first iteration at which function is executable.

Examples

_guessmemory(_rng, kernel, reference)

Guess memory of ParticleKernel.

Examples

_try(f, input)

Check if function is exectuable.

Examples

ancestors!(_rng, referencing, resampling, ancestor, iter, Nparticles, weights)

Resampling function for particles, dispatched on ParticleReferencing subtype.

Examples

ancestralweight!(kernel, weights, valₜ, val, iter)

Inplace calculate ancestor weights.

Examples

assign!(buffer, trajectory)

Assign 'trajectory' elements to 'buffer' up until index 'iter'.

Examples

dynamics_propagate(objective)

Setup dynamics for propagate!() step. By default, same as dynamics_propose, but can be extended if faster setup is available.

Examples

dynamics_propose(objective)

Setup dynamics for propose!() step.

Examples

estimate_Nparticles(_rng, pf, objective, variance)

Function that checks for number of particles to achieve target variance of log target estimate.

Examples

get_reference!(_rng, kernel, referencing, ancestor, particles, tune, reference)

Compute new reference index from current reference index.

Examples

get_reference(referencing, model, data, tagged)

Assign reference, depending on whether reference is tracked in Model.

Examples

hasnan(val)

Check if a single value in 'val' has nan

Examples

impute(data, t, kernel, particles, tune, reference, objective)

Extendable function to work with missing data

Examples

infer_generated(_rng, pf, model, data)

Infer type of generated quantities of PF sampler.

Examples

initial!(_rng, kernel, val, iter)

Inplace initiate particle given kernel.

Examples

initial!(_rng, kernel, particles, tune, reference, objective)

Initialize particles.

Examples

initial(_rng, kernel)

Initiate particle with given initial distribution.

Examples

sample_ancestor(_rng, kernel, weights, valₜ, val, iter)

Sample ancestor reference for particle history.

Examples

set_reference!(referencing, val, reference, iter)

Set last particle at current iteration.

Examples

Inplace propagate particle forward given current particle and kernel.

Examples

Propagate particle forward given current particle.

Examples

ℓevidence!(kernel, ℓcontainer, dataₜ, val, iter)

Inplace logpdf evaluation of data given current particle trajectory.

Examples

ℓevidence(kernel, dataₜ, val, iter)

Evaluate data given particle trajectory.

Examples

ℓtransition!(kernel, ℓcontainer, valₜ, val, iter)

Inplace calculate log transtion probability from current particle given trajectory.

Examples

ℓtransition(kernel, valₜ, val, iter)

Calculate log transtion probability from particle given particle history.

Examples

predict(_rng, kernel, trajectory, reference, iter)

Predict new latent and observed data.

Examples