include/dai/hak.h

   1 /*  This file is part of libDAI - http://www.libdai.org/
   2  *
   3  *  Copyright (c) 2006-2011, The libDAI authors. All rights reserved.
   4  *
   5  *  Use of this source code is governed by a BSD-style license that can be found in the LICENSE file.
   6  */
   7
   8
   9 /// \file
  10 /// \brief Defines class HAK, which implements a variant of Generalized Belief Propagation.
  11 /// \idea Implement more general region graphs and corresponding Generalized Belief Propagation updates as described in [\ref YFW05].
  12 /// \todo Use ClusterGraph instead of a vector<VarSet> for speed.
  13 /// \todo Optimize this code for large factor graphs.
  14 /// \todo Implement GBP parent-child  algorithm.
  15
  16
  17 #ifndef __defined_libdai_hak_h
  18 #define __defined_libdai_hak_h
  19
  20
  21 #include <string>
  22 #include <dai/daialg.h>
  23 #include <dai/regiongraph.h>
  24 #include <dai/enum.h>
  25 #include <dai/properties.h>
  26
  27
  28 namespace dai {
  29
  30
  31 /// Approximate inference algorithm: implementation of single-loop ("Generalized Belief Propagation") and double-loop algorithms by Heskes, Albers and Kappen [\ref HAK03]
  32 class HAK : public DAIAlgRG {
  33     private:
  34         /// Outer region beliefs
  35         std::vector<Factor>                _Qa;
  36         /// Inner region beliefs
  37         std::vector<Factor>                _Qb;
  38         /// Messages from outer to inner regions
  39         std::vector<std::vector<Factor> >  _muab;
  40         /// Messages from inner to outer regions
  41         std::vector<std::vector<Factor> >  _muba;
  42         /// Maximum difference encountered so far
  43         Real _maxdiff;
  44         /// Number of iterations needed
  45         size_t _iters;
  46
  47     public:
  48         /// Parameters for HAK
  49         struct Properties {
  50             /// Enumeration of possible cluster choices
  51             /** The following cluster choices are defined:
  52              *   - MIN minimal clusters, i.e., one outer region for each maximal factor
  53              *   - DELTA one outer region for each variable and its Markov blanket
  54              *   - LOOP one cluster for each loop of length at most \a Properties::loopdepth, and in addition one cluster for each maximal factor
  55              *   - BETHE Bethe approximation (one outer region for each maximal factor, inner regions are single variables)
  56              */
  57             DAI_ENUM(ClustersType,MIN,BETHE,DELTA,LOOP);
  58
  59             /// Enumeration of possible message initializations
  60             DAI_ENUM(InitType,UNIFORM,RANDOM);
  61
  62             /// Verbosity (amount of output sent to stderr)
  63             size_t verbose;
  64
  65             /// Maximum number of iterations
  66             size_t maxiter;
  67
  68             /// Maximum time (in seconds)
  69             double maxtime;
  70
  71             /// Tolerance for convergence test
  72             Real tol;
  73
  74             /// Damping constant (0.0 means no damping, 1.0 is maximum damping)
  75             Real damping;
  76
  77             /// How to choose the outer regions
  78             ClustersType clusters;
  79
  80             /// How to initialize the messages
  81             InitType init;
  82
  83             /// Use single-loop (GBP) or double-loop (HAK)
  84             bool doubleloop;
  85
  86             /// Depth of loops (only relevant for \a clusters == \c ClustersType::LOOP)
  87             size_t loopdepth;
  88         } props;
  89
  90     public:
  91     /// \name Constructors/destructors
  92     //@{
  93         /// Default constructor
  94         HAK() : DAIAlgRG(), _Qa(), _Qb(), _muab(), _muba(), _maxdiff(0.0), _iters(0U), props() {}
  95
  96         /// Construct from FactorGraph \a fg and PropertySet \a opts
  97         /** \param fg Factor graph.
  98          *  \param opts Parameters @see Properties
  99          */
 100         HAK( const FactorGraph &fg, const PropertySet &opts );
 101
 102         /// Construct from RegionGraph \a rg and PropertySet \a opts
 103         HAK( const RegionGraph &rg, const PropertySet &opts );
 104     //@}
 105
 106
 107     /// \name General InfAlg interface
 108     //@{
 109         virtual HAK* clone() const { return new HAK(*this); }
 110         virtual HAK* construct( const FactorGraph &fg, const PropertySet &opts ) const { return new HAK( fg, opts ); }
 111         virtual std::string name() const { return "HAK"; }
 112         virtual Factor belief( const VarSet &vs ) const;
 113         virtual std::vector<Factor> beliefs() const;
 114         virtual Real logZ() const;
 115         virtual void init();
 116         virtual void init( const VarSet &vs );
 117         virtual Real run();
 118         virtual Real maxDiff() const { return _maxdiff; }
 119         virtual size_t Iterations() const { return _iters; }
 120         virtual void setMaxIter( size_t maxiter ) { props.maxiter = maxiter; }
 121         virtual void setProperties( const PropertySet &opts );
 122         virtual PropertySet getProperties() const;
 123         virtual std::string printProperties() const;
 124     //@}
 125
 126
 127     /// \name Additional interface specific for HAK
 128     //@{
 129         /// Returns reference to message from outer region \a alpha to its \a _beta 'th neighboring inner region
 130         Factor & muab( size_t alpha, size_t _beta ) { return _muab[alpha][_beta]; }
 131         /// Returns reference to message the \a _beta 'th neighboring inner region of outer region \a alpha to that outer region
 132         Factor & muba( size_t alpha, size_t _beta ) { return _muba[alpha][_beta]; }
 133         /// Returns belief of outer region \a alpha
 134         const Factor& Qa( size_t alpha ) const { return _Qa[alpha]; };
 135         /// Returns belief of inner region \a beta
 136         const Factor& Qb( size_t beta ) const { return _Qb[beta]; };
 137
 138         /// Runs single-loop algorithm (algorithm 1 in [\ref HAK03])
 139         Real doGBP();
 140         /// Runs double-loop algorithm (as described in section 4.2 of [\ref HAK03]), which always convergences
 141         Real doDoubleLoop();
 142     //@}
 143
 144     private:
 145         /// Helper function for constructors
 146         void construct();
 147         /// Recursive procedure for finding clusters of variables containing loops of length at most \a length
 148         /** \param fg the factor graph
 149          *  \param allcl the clusters found so far
 150          *  \param newcl partial candidate cluster
 151          *  \param root start (and end) point of the loop
 152          *  \param length number of variables that may be added to \a newcl
 153          *  \param vars neighboring variables of \a newcl
 154          *  \return allcl all clusters of variables with loops of length at most \a length passing through root
 155          */
 156         void findLoopClusters( const FactorGraph &fg, std::set<VarSet> &allcl, VarSet newcl, const Var & root, size_t length, VarSet vars );
 157 };
 158
 159
 160 } // end of namespace dai
 161
 162
 163 #endif