class mrpt::bayes::CParticleFilterCapable

Python API: mrpt.bayes.CParticleFilterCapable

Overview

This virtual class defines the interface that any particles based PDF class must implement in order to be executed by a mrpt::bayes::CParticleFilter.

See the Particle Filter tutorial explaining how to use the particle filter-related classes.

See also:

CParticleFilter, CParticleFilterData

#include <mrpt/bayes/CParticleFilterCapable.h>

class CParticleFilterCapable
{
public:
    // typedefs

    typedef std::function<double(const bayes::CParticleFilter::TParticleFilterOptions&PF_options, const CParticleFilterCapable*obj, size_t index, const void*action, const void*observation)> TParticleProbabilityEvaluator;

    // structs

    struct TFastDrawAuxVars;

    // construction

    CParticleFilterCapable();

    // methods

    void prepareFastDrawSample(
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options,
        TParticleProbabilityEvaluator partEvaluator = defaultEvaluator,
        const void* action = nullptr,
        const void* observation = nullptr
        ) const;

    size_t fastDrawSample(const bayes::CParticleFilter::TParticleFilterOptions& PF_options) const;
    virtual double getW(size_t i) const = 0;
    virtual void setW(size_t i, double w) = 0;
    virtual size_t particlesCount() const = 0;
    void prediction_and_update(const mrpt::obs::CActionCollection* action, const mrpt::obs::CSensoryFrame* observation, const bayes::CParticleFilter::TParticleFilterOptions& PF_options);
    virtual void performSubstitution(const std::vector<size_t>& indx) = 0;
    virtual double normalizeWeights(double* out_max_log_w = nullptr) = 0;
    virtual double ESS() const = 0;
    void performResampling(const bayes::CParticleFilter::TParticleFilterOptions& PF_options, size_t out_particle_count = 0);

    static double defaultEvaluator(
        ] const bayes::CParticleFilter::TParticleFilterOptions& PF_options,
        ] const CParticleFilterCapable* obj,
        size_t index,
        ] const void* action,
        ] const void* observation
        );

    static void computeResampling(
        CParticleFilter::TParticleResamplingAlgorithm method,
        const std::vector<double>& in_logWeights,
        std::vector<size_t>& out_indexes,
        size_t out_particle_count = 0
        );

    static void log2linearWeights(
        const std::vector<double>& in_logWeights,
        std::vector<double>& out_linWeights
        );

    static std::vector<double> logWeightsToLinear(const std::vector<double>& in_logWeights);
};

// direct descendants

template <class Derived, class particle_list_t>
struct CParticleFilterDataImpl;

Typedefs

typedef std::function<double(const bayes::CParticleFilter::TParticleFilterOptions&PF_options, const CParticleFilterCapable*obj, size_t index, const void*action, const void*observation)> TParticleProbabilityEvaluator

A callback function type for evaluating the probability of m_particles of being selected, used in “fastDrawSample”.

The default evaluator function “defaultEvaluator” simply returns the particle weight.

Parameters:

index	This is the index of the particle its probability is being computed.
action	The value passed to “prepareFastDrawSample” (may be nullptr).
observation	The value passed to “prepareFastDrawSample” (may be nullptr).

See also:

prepareFastDrawSample

Methods

void prepareFastDrawSample(
    const bayes::CParticleFilter::TParticleFilterOptions& PF_options,
    TParticleProbabilityEvaluator partEvaluator = defaultEvaluator,
    const void* action = nullptr,
    const void* observation = nullptr
    ) const

Prepares data structures for calling fastDrawSample method next.

This method must be called once before using “fastDrawSample” (calling this more than once has no effect, but it takes time for nothing!) The behavior depends on the configuration of the PF (see CParticleFilter::TParticleFilterOptions):

• DYNAMIC SAMPLE SIZE=NO : In this case this method fills out an internal array (m_fastDrawAuxiliary.alreadyDrawnIndexes) with the random indexes generated according to the selected resample scheme in TParticleFilterOptions. Those indexes are read sequentially by subsequent calls to fastDrawSample.

• DYNAMIC SAMPLE SIZE=YES : Then:

  • If TParticleFilterOptions.resamplingMethod = prMultinomial, the internal buffers will be filled out (m_fastDrawAuxiliary.CDF, CDF_indexes & PDF) and then fastDrawSample can be called an arbitrary number of times to generate random indexes.

  • For the rest of resampling algorithms, an exception will be raised since they are not appropriate for a dynamic (unknown in advance) number of particles.

The function pointed by “partEvaluator” should take into account the particle filter algorithm selected in “m_PFAlgorithm”. If called without arguments (defaultEvaluator), the default behavior is to draw samples with a probability proportional to their current weights. The action and the observation are declared as “void*” for a greater flexibility. For a more detailed information see the Particle Filter tutorial. Custom supplied “partEvaluator” functions must take into account the previous particle weight, i.e. multiplying the current observation likelihood by the weights.

See also:

fastDrawSample

size_t fastDrawSample(const bayes::CParticleFilter::TParticleFilterOptions& PF_options) const

Draws a random sample from the particle filter, in such a way that each particle has a probability proportional to its weight (in the standard PF algorithm).

This method can be used to generate a variable number of m_particles when resampling: to vary the number of m_particles in the filter. See prepareFastDrawSample for more information, or the Particle Filter tutorial.

NOTES:

• You MUST call “prepareFastDrawSample” ONCE before calling this method. That method must be called after modifying the particle filter (executing one step, resampling, etc…)

• This method returns ONE index for the selected (“drawn”) particle, in the range [0,M-1]

• You do not need to call “normalizeWeights” before calling this.

See also:

prepareFastDrawSample

virtual double getW(size_t i) const = 0

Access to i’th particle (logarithm) weight, where first one is index 0.

virtual void setW(size_t i, double w) = 0

Modifies i’th particle (logarithm) weight, where first one is index 0.

virtual size_t particlesCount() const = 0

Get the m_particles count.

void prediction_and_update(
    const mrpt::obs::CActionCollection* action,
    const mrpt::obs::CSensoryFrame* observation,
    const bayes::CParticleFilter::TParticleFilterOptions& PF_options
    )

Performs the prediction stage of the Particle Filter.

This method simply selects the appropriate protected method according to the particle filter algorithm to run.

See also:

prediction_and_update_pfStandardProposal, prediction_and_update_pfAuxiliaryPFStandard, prediction_and_update_pfOptimalProposal, prediction_and_update_pfAuxiliaryPFOptimal

virtual void performSubstitution(const std::vector<size_t>& indx) = 0

Performs the substitution for internal use of resample in particle filter algorithm, don’t call it directly.

Parameters:

indx	The indices of current m_particles to be saved as the new m_particles set.

virtual double normalizeWeights(double* out_max_log_w = nullptr) = 0

Normalize the (logarithmic) weights, such as the maximum weight is zero.

Parameters:

out_max_log_w

If provided, will return with the maximum log_w before normalizing, such as new_weights = old_weights - max_log_w.

Returns:

The max/min ratio of weights (“dynamic range”)

virtual double ESS() const = 0

Returns the normalized ESS (Estimated Sample Size), in the range [0,1].

Note that you do NOT need to normalize the weights before calling this.

void performResampling(const bayes::CParticleFilter::TParticleFilterOptions& PF_options, size_t out_particle_count = 0)

Performs a resample of the m_particles, using the method selected in the constructor.

After computing the surviving samples, this method internally calls “performSubstitution” to actually perform the particle replacement. This method is called automatically by CParticleFilter::execute, andshould not be invoked manually normally. To just obtaining the sequence of resampled indexes from a sequence of weights, use “resample”

Parameters:

out_particle_count

The desired number of output particles after resampling; 0 means don’t modify the current number.

See also:

resample

static double defaultEvaluator(
    ] const bayes::CParticleFilter::TParticleFilterOptions& PF_options,
    ] const CParticleFilterCapable* obj,
    size_t index,
    ] const void* action,
    ] const void* observation
    )

The default evaluator function, which simply returns the particle weight.

The action and the observation are declared as “void*” for a greater flexibility.

See also:

prepareFastDrawSample

static void computeResampling(
    CParticleFilter::TParticleResamplingAlgorithm method,
    const std::vector<double>& in_logWeights,
    std::vector<size_t>& out_indexes,
    size_t out_particle_count = 0
    )

A static method to perform the computation of the samples resulting from resampling a given set of particles, given their logarithmic weights, and a resampling method.

It returns the sequence of indexes from the resampling. The number of output samples is the same than the input population. This generic method just computes these indexes, to actually perform a resampling in a particle filter object, call performResampling

Parameters:

out_particle_count

The desired number of output particles after resampling; 0 means don’t modify the current number.

See also:

performResampling

static void log2linearWeights(
    const std::vector<double>& in_logWeights,
    std::vector<double>& out_linWeights
    )

A static method to compute the linear, normalized (the sum the unity) weights from log-weights.

See also:

performResampling, logWeightsToLinear

static std::vector<double> logWeightsToLinear(const std::vector<double>& in_logWeights)

Returns the normalized linear weights from log-weights (return-by-value convenience wrapper around log2linearWeights).

See also:

log2linearWeights