0242c79a5dcb1a89371a83bf3c5e5dbc83770989
[libdai.git] / include / dai / daialg.h
1 /* This file is part of libDAI - http://www.libdai.org/
2 *
3 * libDAI is licensed under the terms of the GNU General Public License version
4 * 2, or (at your option) any later version. libDAI is distributed without any
5 * warranty. See the file COPYING for more details.
6 *
7 * Copyright (C) 2006-2009 Joris Mooij [joris dot mooij at libdai dot org]
8 * Copyright (C) 2006-2007 Radboud University Nijmegen, The Netherlands
9 */
10
11
12 /// \file
13 /// \brief Defines the general interface for inference methods in libDAI (classes InfAlg, DaiAlg<>, DaiAlgFG and DaiAlgRG).
14
15
16 #ifndef __defined_libdai_daialg_h
17 #define __defined_libdai_daialg_h
18
19
20 #include <string>
21 #include <iostream>
22 #include <vector>
23 #include <dai/factorgraph.h>
24 #include <dai/regiongraph.h>
25 #include <dai/properties.h>
26
27
28 namespace dai {
29
30
31 /// InfAlg is an abstract base class, defining the common interface of all inference algorithms in libDAI.
32 /** \idea General marginalization functions like calcMarginal() now copy a complete InfAlg object. Instead,
33 * it would make more sense that they construct a new object without copying the FactorGraph or RegionGraph.
34 * Or they can simply be made methods of the general InfAlg class.
35 * \idea Use a PropertySet as output of an InfAlg, instead of functions like maxDiff() and Iterations().
36 */
37 class InfAlg {
38 public:
39 /// \name Constructors/destructors
40 //@{
41 /// Virtual destructor (needed because this class contains virtual functions)
42 virtual ~InfAlg() {}
43
44 /// Returns a pointer to a new, cloned copy of \c *this (i.e., virtual copy constructor)
45 virtual InfAlg* clone() const = 0;
46 //@}
47
48 /// \name Queries
49 //@{
50 /// Identifies itself for logging purposes
51 virtual std::string identify() const = 0;
52
53 /// Returns reference to underlying FactorGraph.
54 virtual FactorGraph &fg() = 0;
55
56 /// Returns constant reference to underlying FactorGraph.
57 virtual const FactorGraph &fg() const = 0;
58 //@}
59
60 /// \name Inference interface
61 //@{
62 /// Initializes all data structures of the approximate inference algorithm.
63 /** \note This method should be called at least once before run() is called.
64 */
65 virtual void init() = 0;
66
67 /// Initializes all data structures corresponding to some set of variables.
68 /** This method can be used to do a partial initialization after a part of the factor graph has changed.
69 * Instead of initializing all data structures, it only initializes those involving the variables in \a vs.
70 * \throw NOT_IMPLEMENTED if not implemented/supported
71 */
72 virtual void init( const VarSet &vs ) = 0;
73
74 /// Runs the approximate inference algorithm.
75 /** \note Before run() is called the first time, init() should have been called.
76 */
77 virtual Real run() = 0;
78
79 /// Returns the (approximate) marginal probability distribution of a variable.
80 /** \note Before this method is called, run() should have been called.
81 */
82 virtual Factor belief( const Var &v ) const { return belief( VarSet(v) ); }
83
84 /// Returns the (approximate) marginal probability distribution of a set of variables.
85 /** \note Before this method is called, run() should have been called.
86 * \throw NOT_IMPLEMENTED if not implemented/supported.
87 * \throw BELIEF_NOT_AVAILABLE if the requested belief cannot be calculated with this algorithm.
88 */
89 virtual Factor belief( const VarSet &vs ) const = 0;
90
91 /// Returns the (approximate) marginal probability distribution of the variable with index \a i.
92 /** For some approximate inference algorithms, using beliefV() is preferred to belief() for performance reasons.
93 * \note Before this method is called, run() should have been called.
94 */
95 virtual Factor beliefV( size_t i ) const { return belief( fg().var(i) ); }
96
97 /// Returns the (approximate) marginal probability distribution of the variables on which factor \a I depends.
98 /** For some approximate inference algorithms, using beliefF() is preferred to belief() for performance reasons.
99 * \note Before this method is called, run() should have been called.
100 */
101 virtual Factor beliefF( size_t I ) const { return belief( fg().factor(I).vars() ); }
102
103 /// Returns all beliefs (approximate marginal probability distributions) calculated by the algorithm.
104 /** \note Before this method is called, run() should have been called.
105 */
106 virtual std::vector<Factor> beliefs() const = 0;
107
108 /// Returns the logarithm of the (approximated) partition sum (normalizing constant of the factor graph).
109 /** \note Before this method is called, run() should have been called.
110 * \throw NOT_IMPLEMENTED if not implemented/supported
111 */
112 virtual Real logZ() const = 0;
113
114 /// Returns maximum difference between single variable beliefs in the last iteration.
115 /** \throw NOT_IMPLEMENTED if not implemented/supported
116 */
117 virtual Real maxDiff() const = 0;
118
119 /// Returns number of iterations done (one iteration passes over the complete factorgraph).
120 /** \throw NOT_IMPLEMENTED if not implemented/supported
121 */
122 virtual size_t Iterations() const = 0;
123 //@}
124
125 /// \name Changing the factor graph
126 //@{
127 /// Clamp variable with index \a i to value \a x (i.e. multiply with a Kronecker delta \f$\delta_{x_i, x}\f$)
128 /** If \a backup == \c true, make a backup of all factors that are changed.
129 */
130 virtual void clamp( size_t i, size_t x, bool backup = false ) = 0;
131
132 /// Sets all factors interacting with variable with index \a i to one.
133 /** If \a backup == \c true, make a backup of all factors that are changed.
134 */
135 virtual void makeCavity( size_t i, bool backup = false ) = 0;
136 //@}
137
138 /// \name Backup/restore mechanism for factors
139 //@{
140 /// Make a backup copy of factor \a I
141 /** \throw MULTIPLE_UNDO if a backup already exists
142 */
143 virtual void backupFactor( size_t I ) = 0;
144 /// Make backup copies of all factors involving the variables in \a vs
145 /** \throw MULTIPLE_UNDO if a backup already exists
146 */
147 virtual void backupFactors( const VarSet &vs ) = 0;
148
149 /// Restore factor \a I from its backup copy
150 virtual void restoreFactor( size_t I ) = 0;
151 /// Restore the factors involving the variables in \a vs from their backup copies
152 virtual void restoreFactors( const VarSet &vs ) = 0;
153 //@}
154
155 /// \name Managing parameters
156 //@{
157 /// Set parameters of this inference algorithm.
158 /** The parameters are set according to the PropertySet \a opts.
159 * The values can be stored either as std::string or as the type of the corresponding MF::props member.
160 */
161 virtual void setProperties( const PropertySet &opts ) = 0;
162 /// Returns parameters of this inference algorithm converted into a PropertySet.
163 virtual PropertySet getProperties() const = 0;
164 /// Returns parameters of this inference algorithm formatted as a string in the format "[key1=val1,key2=val2,...,keyn=valn]".
165 virtual std::string printProperties() const = 0;
166 //@}
167 };
168
169
170 /// Combines the abstract base class InfAlg with a graphical model (e.g., a FactorGraph or RegionGraph).
171 /** Inference algorithms in libDAI directly inherit from a DAIAlg, currently either
172 * from a DAIAlg<FactorGraph> or from a DAIAlg<RegionGraph>.
173 *
174 * \tparam GRM Should be castable to FactorGraph
175 * \idea A DAIAlg should not inherit from a FactorGraph or RegionGraph, but should
176 * store a reference to the graphical model object. This prevents needless copying
177 * of (possibly large) data structures. Disadvantage: the caller must not change
178 * the graphical model between calls to the inference algorithm (maybe a smart_ptr
179 * or some locking mechanism would help here?).
180 */
181 template <class GRM>
182 class DAIAlg : public InfAlg, public GRM {
183 public:
184 /// \name Constructors/destructors
185 //@{
186 /// Default constructor
187 DAIAlg() : InfAlg(), GRM() {}
188
189 /// Construct from GRM
190 DAIAlg( const GRM &grm ) : InfAlg(), GRM(grm) {}
191 //@}
192
193 /// \name Queries
194 //@{
195 /// Returns reference to underlying FactorGraph.
196 FactorGraph &fg() { return (FactorGraph &)(*this); }
197
198 /// Returns constant reference to underlying FactorGraph.
199 const FactorGraph &fg() const { return (const FactorGraph &)(*this); }
200 //@}
201
202 /// \name Changing the factor graph
203 //@{
204 /// Clamp variable with index \a i to value \a x (i.e. multiply with a Kronecker delta \f$\delta_{x_i, x}\f$)
205 /** If \a backup == \c true, make a backup of all factors that are changed.
206 */
207 void clamp( size_t i, size_t x, bool backup = false ) { GRM::clamp( i, x, backup ); }
208
209 /// Sets all factors interacting with variable with index \a i to one.
210 /** If \a backup == \c true, make a backup of all factors that are changed.
211 */
212 void makeCavity( size_t i, bool backup = false ) { GRM::makeCavity( i, backup ); }
213 //@}
214
215 /// \name Backup/restore mechanism for factors
216 //@{
217 /// Make a backup copy of factor \a I
218 void backupFactor( size_t I ) { GRM::backupFactor( I ); }
219 /// Make backup copies of all factors involving the variables in \a vs
220 void backupFactors( const VarSet &vs ) { GRM::backupFactors( vs ); }
221
222 /// Restore factor \a I from its backup copy
223 void restoreFactor( size_t I ) { GRM::restoreFactor( I ); }
224 /// Restore the factors involving the variables in \a vs from their backup copies
225 void restoreFactors( const VarSet &vs ) { GRM::restoreFactors( vs ); }
226 //@}
227 };
228
229
230 /// Base class for inference algorithms that operate on a FactorGraph
231 typedef DAIAlg<FactorGraph> DAIAlgFG;
232
233 /// Base class for inference algorithms that operate on a RegionGraph
234 typedef DAIAlg<RegionGraph> DAIAlgRG;
235
236
237 /// Calculates the marginal probability distribution for \a vs using inference algorithm \a obj.
238 /** calcMarginal() works by clamping all variables in \a vs and calculating the partition sum for each clamped state.
239 * Therefore, it can be used in combination with any inference algorithm that can calculate/approximate partition sums.
240 * \param obj instance of inference algorithm to be used
241 * \param vs variables for which the marginal should be calculated
242 * \param reInit should be set to \c true if at least one of the possible clamped states would be invalid (leading to a factor graph with zero partition sum).
243 */
244 Factor calcMarginal( const InfAlg& obj, const VarSet& vs, bool reInit );
245
246 /// Calculates beliefs for all pairs of variables in \a vs using inference algorithm \a obj.
247 /** calcPairBeliefs() works by
248 * - clamping single variables in \a vs and calculating the partition sum and the single variable beliefs for each clamped state, if \a accurate == \c false;
249 * - clamping pairs of variables in \a vs and calculating the partition sum for each clamped state, if \a accurate == \c true.
250 *
251 * Therefore, it can be used in combination with any inference algorithm that can calculate/approximate partition sums (and single variable beliefs, if
252 * \a accurate == \c true).
253 * \param obj instance of inference algorithm to be used
254 * \param vs variables for which the pair beliefs should be calculated
255 * \param reInit should be set to \c true if at least one of the possible clamped states would be invalid (leading to a factor graph with zero partition sum).
256 * \param accurate if \c true, uses a slower but more accurate approximation algorithm
257 */
258 std::vector<Factor> calcPairBeliefs( const InfAlg& obj, const VarSet& vs, bool reInit, bool accurate=false );
259
260
261 } // end of namespace dai
262
263
264 #endif