CParticleFilter.h

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          http://www.mrpt.org/                          |
   |                                                                        |
   | Copyright (c) 2005-2018, Individual contributors, see AUTHORS file     |
   | See: http://www.mrpt.org/Authors - All rights reserved.                |
   | Released under BSD License. See details in http://www.mrpt.org/License |
   +------------------------------------------------------------------------+ */
#pragma once

#include <mrpt/system/COutputLogger.h>
#include <mrpt/config/CLoadableOptions.h>
#include <mrpt/typemeta/TEnumType.h>

namespace mrpt
{
namespace obs
{
class CSensoryFrame;
class CActionCollection;
}  // namespace obs

/** The namespace for Bayesian filtering algorithm: different particle filters
 * and Kalman filter algorithms. \ingroup mrpt_bayes_grp
 */
namespace bayes
{
class CParticleFilterCapable;

/** This class acts as a common interface to the different interfaces (see
 *CParticleFilter::TParticleFilterAlgorithm) any bayes::CParticleFilterCapable
 *class can implement: it is the invoker of particle filter algorithms.
 *   The particle filter is executed on a probability density function (PDF)
 *described by a CParticleFilterCapable object, passed in the constructor or
 *alternatively through the CParticleFilter::executeOn method.<br>
 *
 * For a complete example and further details, see the <a
 *href="http://www.mrpt.org/Particle_Filter_Tutorial" >Particle Filter
 *tutorial</a>.
 *
 *   The basic SIR algorithm (pfStandardProposal) consists of:
 *      - Execute a prediction with the given "action".
 *      - Update the weights of the particles using the likelihood of the
 *"observation".
 *      - Normalize weights.
 *      - Perform resampling if the ESS is below the threshold options.BETA.
 *
 * \ingroup mrpt_bayes_grp
 * \sa mrpt::poses::CPoseParticlesPDF
 */
class CParticleFilter : public mrpt::system::COutputLogger
{
   public:
    /** Defines different types of particle filter algorithms.
     *  The defined SIR implementations are:
     *      - pfStandardProposal: Standard proposal distribution + weights
     *according to likelihood function.
     *      - pfAuxiliaryPFStandard: An auxiliary PF using the standard proposal
     *distribution.
     *      - pfOptimalProposal: Use the optimal proposal distribution (where
     *available!, usually this will perform approximations)
     *      - pfAuxiliaryPFOptimal: Use the optimal proposal and a auxiliary
     *particle filter (see <a
     *href="http://www.mrpt.org/Paper:An_Optimal_Filtering_Algorithm_for_Non-Parametric_Observation_Models_in_Robot_Localization_(ICRA_2008)"
     *>paper</a>).
     *
     * See the theoretical discussion in <a
     *href="http://www.mrpt.org/Resampling_Schemes" >resampling schemes</a>.
     */
    enum TParticleFilterAlgorithm
    {
        pfStandardProposal = 0,
        pfAuxiliaryPFStandard,
        pfOptimalProposal,
        pfAuxiliaryPFOptimal
    };

    /** Defines the different resampling algorithms.
     *  The implemented resampling methods are:
     *      - prMultinomial (Default): Uses standard select with replacement
     *(draws
     *M random uniform numbers)
     *      - prResidual: The residual or "remainder" method.
     *      - prStratified: The stratified resampling, where a uniform sample is
     *drawn for each of M subdivisions of the range (0,1].
     *      - prSystematic: A single uniform sample is drawn in the range
     *(0,1/M].
     *
     * See the theoretical discussion in <a
     *href="http://www.mrpt.org/Resampling_Schemes" >resampling schemes</a>.
     */
    enum TParticleResamplingAlgorithm
    {
        prMultinomial = 0,
        prResidual,
        prStratified,
        prSystematic
    };

    /** The configuration of a particle filter.
     */
    struct TParticleFilterOptions : public mrpt::config::CLoadableOptions
    {
       public:
        // See base docs:
        void loadFromConfigFile(
            const mrpt::config::CConfigFileBase& source,
            const std::string& section) override;
        virtual void saveToConfigFile(
            mrpt::config::CConfigFileBase& target,
            const std::string& section) const override;

        /** A flag that indicates whether the CParticleFilterCapable object
         * should perform adative sample size (default=false). */
        bool adaptiveSampleSize{false};
        /** The resampling of particles will be performed when ESS (in range
         * [0,1]) < BETA (default is 0.5) */
        double BETA{0.5};
        /** The initial number of particles in the filter (it can change only if
         * adaptiveSampleSize=true) (default=1) */
        unsigned int sampleSize{1};

        /** In the algorithm "CParticleFilter::pfAuxiliaryPFOptimal" (and in
         * "CParticleFilter::pfAuxiliaryPFStandard" only if
         * pfAuxFilterStandard_FirstStageWeightsMonteCarlo = true) the number of
         * samples for searching the maximum likelihood value and also to
         * estimate the "first stage weights" (see papers!) (default=100)
         */
        unsigned int pfAuxFilterOptimal_MaximumSearchSamples{100};
        /** An optional step to "smooth" dramatic changes in the observation
         * model to affect the variance of the particle weights, eg
         * weight*=likelihood^powFactor (default=1 = no effects). */
        double powFactor{1};
        /** The PF algorithm to use (default=pfStandardProposal) See
         * TParticleFilterAlgorithm for the posibilities. */
        TParticleFilterAlgorithm PF_algorithm{pfStandardProposal};
        /** The resampling algorithm to use (default=prMultinomial). */
        TParticleResamplingAlgorithm resamplingMethod{prMultinomial};

        /** Only for PF_algorithm=pfAuxiliaryPFOptimal: If a given particle has
         * a max_likelihood (from the a-priori estimate) below the maximum from
         * all the samples - max_loglikelihood_dyn_range, then the particle is
         * directly discarded.
         *  This is done to assure that the rejection sampling doesn't get
         * stuck in an infinite loop trying to get an acceptable sample.
         *  Default = 15 (in logarithmic likelihood)
         */
        double max_loglikelihood_dyn_range{15};

        /** Only for PF_algorithm==pfAuxiliaryPFStandard:
         * If false, the APF will predict the first stage weights just at the
         * mean of the prior of the next time step.
         * If true, these weights will be estimated as described in the papers
         * for the "pfAuxiliaryPFOptimal" method, i.e. through a monte carlo
         * simulation.
         *  In that case, "pfAuxFilterOptimal_MaximumSearchSamples" is the
         * number of MC samples used.
         */
        bool pfAuxFilterStandard_FirstStageWeightsMonteCarlo{false};

        /** (Default=false) In the algorithm
         * "CParticleFilter::pfAuxiliaryPFOptimal", if set to true, do not
         * perform rejection sampling, but just the most-likely (ML) particle
         * found in the preliminary weight-determination stage. */
        bool pfAuxFilterOptimal_MLE{false};
    };

    /** Statistics for being returned from the "execute" method. */
    struct TParticleFilterStats
    {
        double ESS_beforeResample{0};
        double weightsVariance_beforeResample{0};
    };

    /** Default constructor.
     *   After creating the PF object, set the options in
     * CParticleFilter::m_options, then execute steps through
     * CParticleFilter::executeOn.
     */
    CParticleFilter();

    virtual ~CParticleFilter() {}
    /** Executes a complete prediction + update step of the selected particle
     * filtering algorithm.
     *    The member CParticleFilter::m_options must be set before calling this
     * to settle the algorithm parameters.
     *
     * \param obj           The object representing the probability distribution
     * function (PDF) which apply the particle filter algorithm to.
     * \param action        A pointer to an action in the form of a
     * CActionCollection,
     * or nullptr if there is no action.
     * \param observation   A pointer to observations in the form of a
     * CSensoryFrame, or nullptr if there is no observation.
     * \param stats An output structure for gathering statistics of the particle
     * filter execution, or set to nullptr if you do not need it (see
     * CParticleFilter::TParticleFilterStats).
     *
     * \sa CParticleFilterCapable, executeOn
     */
    void executeOn(
        CParticleFilterCapable& obj, const mrpt::obs::CActionCollection* action,
        const mrpt::obs::CSensoryFrame* observation,
        TParticleFilterStats* stats = nullptr);

    /** The options to be used in the PF, must be set before executing any step
     * of the particle filter.
     */
    CParticleFilter::TParticleFilterOptions m_options;

};  // End of class def.

}  // namespace bayes
}  // namespace mrpt

MRPT_ENUM_TYPE_BEGIN(mrpt::bayes::CParticleFilter::TParticleFilterAlgorithm)
MRPT_FILL_ENUM_MEMBER(mrpt::bayes::CParticleFilter, pfStandardProposal);
MRPT_FILL_ENUM_MEMBER(mrpt::bayes::CParticleFilter, pfAuxiliaryPFOptimal);
MRPT_FILL_ENUM_MEMBER(mrpt::bayes::CParticleFilter, pfAuxiliaryPFStandard);
MRPT_FILL_ENUM_MEMBER(mrpt::bayes::CParticleFilter, pfOptimalProposal);
MRPT_ENUM_TYPE_END()

MRPT_ENUM_TYPE_BEGIN(mrpt::bayes::CParticleFilter::TParticleResamplingAlgorithm)
MRPT_FILL_ENUM_MEMBER(mrpt::bayes::CParticleFilter, prMultinomial);
MRPT_FILL_ENUM_MEMBER(mrpt::bayes::CParticleFilter, prResidual);
MRPT_FILL_ENUM_MEMBER(mrpt::bayes::CParticleFilter, prStratified);
MRPT_FILL_ENUM_MEMBER(mrpt::bayes::CParticleFilter, prSystematic);
MRPT_ENUM_TYPE_END()