CParticleFilterCapable.h

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          https://www.mrpt.org/                         |
   |                                                                        |
   | Copyright (c) 2005-2020, Individual contributors, see AUTHORS file     |
   | See: https://www.mrpt.org/Authors - All rights reserved.               |
   | Released under BSD License. See: https://www.mrpt.org/License          |
   +------------------------------------------------------------------------+ */
#pragma once

#include <mrpt/bayes/CParticleFilter.h>
#include <cstdint>
#include <cstdlib>
#include <vector>

namespace mrpt::bayes
{
#define INVALID_LIKELIHOOD_VALUE \
    (-1e300)  // An invalid log-likelihood value, used to signal non-initialized
// likelihood variables.

/** 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 <a href="http://www.mrpt.org/Particle_Filter_Tutorial" >Particle
 * Filter tutorial</a> explaining how to use the particle filter-related
 * classes.
 * \sa CParticleFilter, CParticleFilterData
 * \ingroup mrpt_bayes_grp
 */
class CParticleFilterCapable
{
    friend class CParticleFilter;

   private:
    static const unsigned PARTICLE_FILTER_CAPABLE_FAST_DRAW_BINS;

   public:
    CParticleFilterCapable() : m_fastDrawAuxiliary() {}
    /** Virtual destructor
     */
    virtual ~CParticleFilterCapable() = default;
    /** 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.
     * \param index This is the index of the particle its probability is being
     * computed.
     * \param action The value of this is the parameter passed to
     * "prepareFastDrawSample"
     * \param observation The value of this is the parameter passed to
     * "prepareFastDrawSample"
     *  The action and the observation are declared as "void*" for a greater
     * flexibility.
     * \sa prepareFastDrawSample
     */
    using TParticleProbabilityEvaluator = double (*)(
        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.
     * \sa prepareFastDrawSample
     */
    static double defaultEvaluator(
        [[maybe_unused]] const bayes::CParticleFilter::TParticleFilterOptions&
            PF_options,
        [[maybe_unused]] const CParticleFilterCapable* obj, size_t index,
        [[maybe_unused]] const void* action,
        [[maybe_unused]] const void* observation)
    {
        return obj->getW(index);
    }

    /** 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):
     *      - <b>DYNAMIC SAMPLE SIZE=NO</b>: 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.
     *      - <b>DYNAMIC SAMPLE SIZE=YES</b>: 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 <a
     *href="http://www.mrpt.org/Particle_Filters" >Particle Filter
     *tutorial</a>.
     *  Custom supplied "partEvaluator" functions must take into account the
     *previous particle weight, i.e. multiplying the current observation
     *likelihood by the weights.
     * \sa fastDrawSample
     */
    void prepareFastDrawSample(
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options,
        TParticleProbabilityEvaluator partEvaluator = defaultEvaluator,
        const void* action = nullptr, const void* observation = nullptr) 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 <a
     *href="http://www.mrpt.org/Particle_Filters" >Particle Filter
     *tutorial</a>.
     *
     * 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.
     * \sa prepareFastDrawSample
     */
    size_t fastDrawSample(
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options) const;

    /** Access to i'th particle (logarithm) weight, where first one is index 0.
     */
    virtual double getW(size_t i) const = 0;

    /** Modifies i'th particle (logarithm) weight, where first one is index 0.
     */
    virtual void setW(size_t i, double w) = 0;

    /** Get the m_particles count.
     */
    virtual size_t particlesCount() const = 0;

    /** Performs the prediction stage of the Particle Filter.
     *  This method simply selects the appropiate protected method according to
     * the particle filter algorithm to run.
     * \sa
     * prediction_and_update_pfStandardProposal,prediction_and_update_pfAuxiliaryPFStandard,prediction_and_update_pfOptimalProposal,prediction_and_update_pfAuxiliaryPFOptimal
     */
    void prediction_and_update(
        const mrpt::obs::CActionCollection* action,
        const mrpt::obs::CSensoryFrame* observation,
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options);

    /**  Performs the substitution for internal use of resample in particle
     * filter algorithm, don't call it directly.
     *  \param indx The indices of current m_particles to be saved as the new
     * m_particles set.
     */
    virtual void performSubstitution(const std::vector<size_t>& indx) = 0;

    /** Normalize the (logarithmic) weights, such as the maximum weight is zero.
     * \param out_max_log_w If provided, will return with the maximum log_w
     * before normalizing, such as new_weights = old_weights - max_log_w.
     * \return The max/min ratio of weights ("dynamic range")
     */
    virtual double normalizeWeights(double* out_max_log_w = nullptr) = 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.
     */
    virtual double ESS() const = 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"
     * \param[in] out_particle_count The desired number of output particles
     * after resampling; 0 means don't modify the current number.
     * \sa resample
     */
    void performResampling(
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options,
        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
     * \param[in] out_particle_count The desired number of output particles
     * after resampling; 0 means don't modify the current number.
     * \sa performResampling
     */
    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 compute the linear, normalized (the sum the unity)
     * weights from log-weights.
     * \sa performResampling
     */
    static void log2linearWeights(
        const std::vector<double>& in_logWeights,
        std::vector<double>& out_linWeights);

   protected:
    /** Performs the particle filter prediction/update stages for the algorithm
     * "pfStandardProposal" (if not implemented in heritated class, it will
     * raise a 'non-implemented' exception).
     * \sa prediction_and_update
     */
    virtual void prediction_and_update_pfStandardProposal(
        const mrpt::obs::CActionCollection* action,
        const mrpt::obs::CSensoryFrame* observation,
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options);
    /** Performs the particle filter prediction/update stages for the algorithm
     * "pfAuxiliaryPFStandard" (if not implemented in heritated class, it will
     * raise a 'non-implemented' exception).
     * \sa prediction_and_update
     */
    virtual void prediction_and_update_pfAuxiliaryPFStandard(
        const mrpt::obs::CActionCollection* action,
        const mrpt::obs::CSensoryFrame* observation,
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options);
    /** Performs the particle filter prediction/update stages for the algorithm
     * "pfOptimalProposal" (if not implemented in heritated class, it will raise
     * a 'non-implemented' exception).
     * \sa prediction_and_update
     */
    virtual void prediction_and_update_pfOptimalProposal(
        const mrpt::obs::CActionCollection* action,
        const mrpt::obs::CSensoryFrame* observation,
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options);
    /** Performs the particle filter prediction/update stages for the algorithm
     * "pfAuxiliaryPFOptimal" (if not implemented in heritated class, it will
     * raise a 'non-implemented' exception).
     * \sa prediction_and_update
     */
    virtual void prediction_and_update_pfAuxiliaryPFOptimal(
        const mrpt::obs::CActionCollection* action,
        const mrpt::obs::CSensoryFrame* observation,
        const bayes::CParticleFilter::TParticleFilterOptions& PF_options);

    /** Auxiliary vectors, see CParticleFilterCapable::prepareFastDrawSample for
     * more information
     */
    struct TFastDrawAuxVars
    {
        TFastDrawAuxVars() = default;

        std::vector<double> CDF;
        std::vector<uint32_t> CDF_indexes;
        std::vector<double> PDF;
        std::vector<uint32_t> alreadyDrawnIndexes;
        size_t alreadyDrawnNextOne{0};
    };

    /** Auxiliary vectors, see CParticleFilterCapable::prepareFastDrawSample for
     * more information
     */
    mutable TFastDrawAuxVars m_fastDrawAuxiliary;

};  // End of class def.

}  // namespace mrpt::bayes