variableLog2ParamComplexity.h

Go to the documentation of this file.

#ifndef GUM_VARIABLE_LOG2_PARAM_COMPLEXITY_H
#define GUM_VARIABLE_LOG2_PARAM_COMPLEXITY_H

#include <cstddef>
#include <string>
#include <fstream>

#include <agrum/agrum.h>
#include <agrum/core/math/math.h>
#include <agrum/core/hashTable.h>

namespace gum {


  // the CTable cache for log2(C_n^r), with n in {0,...,999} and r in {2,3,4,5}
  extern const double VariableLog2ParamComplexityCTable[4][1000];

  // the size in r of the CTable cache
  constexpr std::size_t VariableLog2ParamComplexityCTableRSize{std::size_t(4)};

  // the size in n of the CTable cache
  constexpr std::size_t VariableLog2ParamComplexityCTableNSize{std::size_t(1000)};


  template < template < typename > class ALLOC = std::allocator >
  class VariableLog2ParamComplexity : private ALLOC< double > {
    public:
    using allocator_type = ALLOC< double >;

    // ########################################################################
    // ########################################################################

    VariableLog2ParamComplexity(const allocator_type& alloc = allocator_type());

    VariableLog2ParamComplexity(const VariableLog2ParamComplexity& from);

    VariableLog2ParamComplexity(const VariableLog2ParamComplexity& from,
                                const allocator_type&              alloc);

    VariableLog2ParamComplexity(VariableLog2ParamComplexity&& from);

    VariableLog2ParamComplexity(VariableLog2ParamComplexity&& from,
                                const allocator_type&         alloc);

    virtual VariableLog2ParamComplexity* clone() const;

    virtual VariableLog2ParamComplexity* clone(const allocator_type& alloc) const;

    virtual ~VariableLog2ParamComplexity();



    // ########################################################################
    // ########################################################################

    VariableLog2ParamComplexity&
       operator=(const VariableLog2ParamComplexity& from);

    VariableLog2ParamComplexity& operator=(VariableLog2ParamComplexity&& from);



    // ########################################################################
    // ########################################################################

    double log2Cnr(const std::size_t r, const double n);

    void CnrToFile(const std::string& filename);

    void useCache(const bool on_off);

    void clearCache();

    allocator_type getAllocator() const;



    private:
    const double __Szpankowski_threshold{VariableLog2ParamComplexityCTableNSize};

    // constants used to speed-up the computation of the Szpankowski
    // approximation.
    // The formula for the approximation given in Silander, Roos,
    // Kontkanen and Myllymaki (2007) "Factorized Normalized Maximum "
    // Likelihood Criterion for Learning Bayesian Network Structures" paper
    // is incorrect. However, the one in Kontkanen, Buntine, Myllymaki,
    // Rissanen and Tirri (2003) "Efficient Computation of Stochastic
    // Complexity" is correct. So we use the latter and simplify it. Thus,
    // the approximation of log2(Cnr) is equal to:
    // 0.5 log2(n) - 0.5 + log2(sqrt(pi)) + (sqrt(2/pi)/3) / sqrt(n) +
    // (3/36 - 4/(9*pi)) / n.
    // So, given the constants below, it is equal to:
    // 0.5 * std::log2 (n) + __cst1 + __cst2 / std::sqrt(n) + __cst3 / n
    const double __cst1 = -0.5 + std::log2(std::sqrt(M_PI));
    const double __cst2 = std::sqrt(2.0 / M_PI) / 3.0;
    const double __cst3 = 3.0 / 36.0 - 4.0 / (9.0 * M_PI);

    // indicates whether we should use a cache or not
    bool __use_cache{true};

    // the cache used, eventually, to store the log2Cnr values
    HashTable< std::pair< std::size_t, double >, double > __cache;
  };

} /* namespace gum */


// always include the template implementation
#include <agrum/core/math/variableLog2ParamComplexity_tpl.h>

#endif /* GUM_VARIABLE_LOG2_PARAM_COMPLEXITY_H */