CRejectionSamplingCapable.h

Go to the documentation of this file.

/* +------------------------------------------------------------------------+
   |                     Mobile Robot Programming Toolkit (MRPT)            |
   |                          https://www.mrpt.org/                         |
   |                                                                        |
   | Copyright (c) 2005-2019, 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/CProbabilityParticle.h>
#include <mrpt/random.h>

namespace mrpt
{
/// \ingroup mrpt_bayes_grp
namespace bayes
{
/** A base class for implementing rejection sampling in a generic state space.
 *   See the main method CRejectionSamplingCapable::rejectionSampling
 *  To use this class, create your own class as a child of this one and
 * implement the desired
 *   virtual methods, and add any required internal data.
 * \ingroup mrpt_bayes_grp
 */
template <
    class TStateSpace, mrpt::bayes::particle_storage_mode STORAGE =
                           mrpt::bayes::particle_storage_mode::POINTER>
class CRejectionSamplingCapable
{
   public:
    using TParticle = CProbabilityParticle<TStateSpace, STORAGE>;

    /** Virtual destructor
     */
    virtual ~CRejectionSamplingCapable() = default;
    /** Generates a set of N independent samples via rejection sampling.
     * \param desiredSamples The number of desired samples to generate
     * \param outSamples The output samples.
     * \param timeoutTrials The maximum number of rejection trials for each
     * generated sample (i.e. the maximum number of iterations). This can be
     * used to set a limit to the time complexity of the algorithm for difficult
     * probability densities.
     *  All will have equal importance weights (a property of rejection
     * sampling), although those samples
     *   generated at timeout will have a different importance weights.
     */
    void rejectionSampling(
        size_t desiredSamples, std::vector<TParticle>& outSamples,
        size_t timeoutTrials = 1000)
    {
        MRPT_START

        TStateSpace x;
        typename std::vector<TParticle>::iterator it;

        // Set output size:
        if (outSamples.size() != desiredSamples)
        {
            // Free old memory:
            outSamples.clear();

            // Reserve new memory:
            outSamples.resize(desiredSamples);
            for (it = outSamples.begin(); it != outSamples.end(); ++it)
                it->d.reset(new TStateSpace);
        }

        // Rejection sampling loop:
        double acceptanceProb;
        for (it = outSamples.begin(); it != outSamples.end(); ++it)
        {
            size_t timeoutCount = 0;
            double bestLik = -1e250;
            TStateSpace bestVal;
            do
            {
                RS_drawFromProposal(*it->d);
                acceptanceProb = RS_observationLikelihood(*it->d);
                ASSERT_(acceptanceProb >= 0 && acceptanceProb <= 1);
                if (acceptanceProb > bestLik)
                {
                    bestLik = acceptanceProb;
                    bestVal = *it->d;
                }
            } while (acceptanceProb <
                         mrpt::random::getRandomGenerator().drawUniform(
                             0.0, 0.999) &&
                     (++timeoutCount) < timeoutTrials);

            // Save weights:
            if (timeoutCount >= timeoutTrials)
            {
                it->log_w = log(bestLik);
                *it->d = bestVal;
            }
            else
            {
                it->log_w = 0;  // log(1.0);
            }
        }  // end for it

        MRPT_END
    }

   protected:
    /** Generates one sample, drawing from some proposal distribution.
     */
    virtual void RS_drawFromProposal(TStateSpace& outSample) = 0;

    /** Returns the NORMALIZED observation likelihood (linear, not
     * exponential!!!) at a given point of the state space (values in the range
     * [0,1]).
     */
    virtual double RS_observationLikelihood(const TStateSpace& x) = 0;

};  // End of class def.

}  // namespace bayes
}  // namespace mrpt