Fixed bug in BipartiteGraph::eraseEdge and improved documentation
[libdai.git] / include / dai / doc.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) 2008-2009 Joris Mooij [joris dot mooij at libdai dot org]
8 */
9
10
11 /** \file
12 * \brief Contains additional doxygen documentation
13 *
14 * \todo Improve documentation
15 *
16 * \todo Merge COPYING into doxygen documentation
17 * \todo Merge README into doxygen documentation
18 * \todo Document examples, tests and utils
19 *
20 * \todo Adapt (part of the) guidelines in http://www.boost.org/development/requirements.html#Design_and_Programming
21 *
22 * \todo Use "gcc -MM" to generate dependencies for targets: http://make.paulandlesley.org/autodep.html
23 * \todo Investigate whether switching to cmake as cross-platform build system would be a good idea.
24 *
25 * \todo Replace VarSets by SmallSet<size_t> where appropriate, in order to minimize the use of FactorGraph::findVar().
26 *
27 * \idea Disentangle structures. In particular, ensure that graphical properties are not
28 * entangled with probabilistic properties. For example, a FactorGraph contains several
29 * components:
30 * - a BipartiteGraph
31 * - an array of variable labels
32 * - an array of variable state space sizes
33 * - an array of pointers to factor value vectors
34 * In this way, each factor could be implemented differently, e.g., we could have
35 * some sparse factors, some noisy-OR factors, some dense factors, some arbitrary
36 * precision factors, etc.
37 *
38 * \idea Use Boost::uBLAS framework to deal with matrices, especially, with 2D sparse matrices.
39 * See http://www.boost.org/libs/numeric/ublas/doc/matrix_sparse.htm
40 * I read somewhere that boost::uBLAS concentrates more on correct implementation than on performance.
41 *
42 * \idea Introduce naming scheme:
43 * - all Vars should be named v_..., e.g. v_i instead of i
44 * - all VarSets should be named vs_..., e.g. v_i instead of i
45 * - all Factors should be named f_..., e.g. f_I instead of I
46 * - all indices should be named _..., e.g. _k instead of k
47 * - all iterators should be named i_, e.g. i_i is an iterator to i
48 * - all const_iterators should be named ci_, e.g. ci_i is an iterator to i
49 **/
50
51
52 /** \page discussion Discussion of possible improvements
53 * \section discuss_extendedgraphs Extended factorgraphs/regiongraphs
54 *
55 * A FactorGraph and a RegionGraph are often equipped with
56 * additional properties for nodes and edges. The code to initialize those
57 * is often quite similar. Maybe one could abstract this, e.g.:
58 * \code
59 * template <typename Node1Properties, typename Node2Properties, typename EdgeProperties>
60 * class ExtFactorGraph : public FactorGraph {
61 * public:
62 * std::vector<Node1Properties> node1Props;
63 * std::vector<Node2Properties> node2Props;
64 * std::vector<std::vector<EdgeProperties> > edgeProps;
65 * // ...
66 * }
67 * \endcode
68 *
69 * Advantages:
70 * - Less code duplication.
71 * - Easier maintainability.
72 * - Easier to write new inference algorithms.
73 *
74 * Disadvantages:
75 * - Cachability may be worse.
76 * - A problem is the case where there are no properties for either type of nodes or for edges.
77 * Maybe this can be solved using specializations, or using variadac template arguments?
78 * Another possible solution would be to define a "class Empty {}", and add some code
79 * that checks for the typeid, comparing it with Empty, and doing something special in that case
80 * (e.g., not allocating memory).
81 * - The main disadvantage of this approach seems to be that it leads to even more entanglement.
82 * Therefore this is probably a bad idea.
83 *
84 * \section discuss_templates Polymorphism by template parameterization
85 * Instead of polymorphism by inheritance, use polymorphism by template parameterization.
86 * For example, the real reason for introducing the complicated inheritance scheme of dai::InfAlg
87 * was for functions like dai::calcMarginal. Instead, one could use a template function:
88 * \code
89 * template<typename InfAlg>
90 * Factor calcMarginal( const InfAlg &obj, const VarSet &ns, bool reInit );
91 * \endcode
92 * This would assume that the type InfAlg supports certain methods. Ideally, one would use
93 * concepts to define different classes of inference algorithms with different capabilities,
94 * for example the ability to calculate logZ, the ability to calculate marginals, the ability to
95 * calculate bounds, the ability to calculate MAP states, etc. Then, one would use traits
96 * classes in order to be able to query the capabilities of the model. For example, one would be
97 * able to query whether the inference algorithm supports calculation of logZ. Unfortunately,
98 * this is compile-time polymorphism, whereas tests/testdai needs runtime polymorphism.
99 * Therefore this is probably a bad idea.
100 */
101
102
103 /** \mainpage libDAI reference manual
104 * \author Joris Mooij
105 * \version git HEAD
106 * \date October 10, 2008
107 *
108 * \section about About libDAI
109 * libDAI is a free/open source C++ library (licensed under GPLv2+) that provides
110 * implementations of various (approximate) inference methods for discrete
111 * graphical models. libDAI supports arbitrary factor graphs with discrete
112 * variables; this includes discrete Markov Random Fields and Bayesian
113 * Networks.
114 *
115 * The library is targeted at researchers. To be able to use the library, a
116 * good understanding of graphical models is needed.
117 *
118 * \section limitations Limitations
119 * libDAI is not intended to be a complete package for approximate inference.
120 * Instead, it should be considered as an "inference engine", providing
121 * various inference methods. In particular, it contains no GUI, currently
122 * only supports its own file format for input and output (although support
123 * for standard file formats may be added later), and provides very limited
124 * visualization functionalities. The only learning method supported currently
125 * is EM for learning factor parameters.
126 *
127 * \section features Features
128 * Currently, libDAI supports the following (approximate) inference methods:
129 * - Exact inference by brute force enumeration;
130 * - Exact inference by junction-tree methods;
131 * - Mean Field;
132 * - Loopy Belief Propagation [\ref KFL01];
133 * - Tree Expectation Propagation [\ref MiQ04];
134 * - Generalized Belief Propagation [\ref YFW05];
135 * - Double-loop GBP [\ref HAK03];
136 * - Various variants of Loop Corrected Belief Propagation
137 * [\ref MoK07, \ref MoR05];
138 * - Gibbs sampler;
139 * - Conditioned BP [\ref EaG09];
140 *
141 * In addition, libDAI supports parameter learning of conditional probability
142 * tables by Expectation Maximization.
143 *
144 * \section language Why C++?
145 * Because libDAI is implemented in C++, it is very fast compared with
146 * implementations in MatLab (a factor 1000 faster is not uncommon).
147 * libDAI does provide a MatLab interface for easy integration with MatLab.
148 */
149
150
151 /** \example example.cpp
152 */
153
154
155 /** \page quickstart Quick start
156 * An example program illustrating basic usage of libDAI is given in examples/example.cpp.
157 */
158
159
160 /** \page bibliography Bibliography
161 * \section Bibliograpy
162 * \anchor KFL01 \ref KFL01
163 * F. R. Kschischang and B. J. Frey and H.-A. Loeliger (2001):
164 * "Factor Graphs and the Sum-Product Algorithm",
165 * <em>IEEE Transactions on Information Theory</em> 47(2):498-519.
166 * http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=910572
167 *
168 * \anchor MiQ04 \ref MiQ04
169 * T. Minka and Y. Qi (2004):
170 * "Tree-structured Approximations by Expectation Propagation",
171 * <em>Advances in Neural Information Processing Systems</em> (NIPS) 16.
172 * http://books.nips.cc/papers/files/nips16/NIPS2003_AA25.pdf
173 *
174 * \anchor MoR05 \ref MoR05
175 * A. Montanari and T. Rizzo (2005):
176 * "How to Compute Loop Corrections to the Bethe Approximation",
177 * <em>Journal of Statistical Mechanics: Theory and Experiment</em>
178 * 2005(10)-P10011.
179 * http://stacks.iop.org/1742-5468/2005/P10011
180 *
181 * \anchor YFW05 \ref YFW05
182 * J. S. Yedidia and W. T. Freeman and Y. Weiss (2005):
183 * "Constructing Free-Energy Approximations and Generalized Belief Propagation Algorithms",
184 * <em>IEEE Transactions on Information Theory</em>
185 * 51(7):2282-2312.
186 * http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1459044
187 *
188 * \anchor HAK03 \ref HAK03
189 * T. Heskes and C. A. Albers and H. J. Kappen (2003):
190 * "Approximate Inference and Constrained Optimization",
191 * <em>Proceedings of the 19th Annual Conference on Uncertainty in Artificial Intelligence (UAI-03)</em> pp. 313-320.
192 * http://www.snn.ru.nl/reports/Heskes.uai2003.ps.gz
193 *
194 * \anchor MoK07 \ref MoK07
195 * J. M. Mooij and H. J. Kappen (2007):
196 * "Loop Corrections for Approximate Inference on Factor Graphs",
197 * <em>Journal of Machine Learning Research</em> 8:1113-1143.
198 * http://www.jmlr.org/papers/volume8/mooij07a/mooij07a.pdf
199 *
200 * \anchor MoK07b \ref MoK07b
201 * J. M. Mooij and H. J. Kappen (2007):
202 * "Sufficient Conditions for Convergence of the Sum-Product Algorithm",
203 * <em>IEEE Transactions on Information Theory</em> 53(12):4422-4437.
204 * http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=4385778
205 *
206 * \anchor EaG09 \ref EaG09
207 * F. Eaton and Z. Ghahramani (2009):
208 * "Choosing a Variable to Clamp",
209 * <em>Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS 2009)</em> 5:145-152
210 * http://jmlr.csail.mit.edu/proceedings/papers/v5/eaton09a/eaton09a.pdf
211 */
212
213
214 /** \page fileformat libDAI factorgraph file format
215 *
216 * This page describes the .fg fileformat used in libDAI to store factor graphs.
217 * Markov Random Fields are special cases of factor graphs, as are Bayesian
218 * networks. A factor graph can be specified as follows: for each factor, one has
219 * to specify which variables occur in the factor, what their respective
220 * cardinalities (i.e., number of possible values) are, and a table listing all
221 * the values of that factor for all possible configurations of these variables.
222 * A .fg file is not much more than that. It starts with a line containing the
223 * number of factors in that graph, followed by an empty line. Then all factors
224 * are specified, one block for each factor, where the blocks are seperated by
225 * empty lines. Each variable occurring in the factor graph has a unique
226 * identifier, its index (which should be a nonnegative integer). Comment lines
227 * start with #.
228 *
229 * Each block starts with a line containing the number of variables in that
230 * factor. The second line contains the indices of these variables, seperated by
231 * spaces (indices are nonnegative integers and to avoid confusion, it is
232 * suggested to start counting at 0). The third line contains the number of
233 * possible values of each of these variables, also seperated by spaces. Note that
234 * there is some redundancy here, since if a variable appears in more than one
235 * factor, the cardinality of that variable appears several times in the .fg file.
236 * The fourth line contains the number of nonzero entries in the factor table.
237 * The rest of the lines contain these nonzero entries; each entry consists of a
238 * table index, followed by white-space, followed by the value corresponding to
239 * that table index. The most difficult part is getting the indexing right. The
240 * convention that is used is that the left-most variables cycle through their
241 * values the fastest (similar to MATLAB indexing of multidimensional arrays). An
242 * example block describing one factor is:
243 *
244 * 3\n
245 * 4 8 7\n
246 * 3 2 2\n
247 * 11\n
248 * 0 0.1\n
249 * 1 3.5\n
250 * 2 2.8\n
251 * 3 6.3\n
252 * 4 8.4\n
253 * 6 7.4\n
254 * 7 2.4\n
255 * 8 8.9\n
256 * 9 1.3\n
257 * 10 1.6\n
258 * 12 6.4\n
259 * 11 2.6\n
260 *
261 * which corresponds to the following factor:
262 *
263 * \f[
264 * \begin{array}{ccc|c}
265 * x_4 & x_8 & x_7 & \mbox{value}\\
266 * \hline
267 * 0 & 0 & 0 & 0.1\\
268 * 1 & 0 & 0 & 3.5\\
269 * 2 & 0 & 0 & 2.8\\
270 * 0 & 1 & 0 & 6.3\\
271 * 1 & 1 & 0 & 8.4\\
272 * 2 & 1 & 0 & 0.0\\
273 * 0 & 0 & 1 & 7.4\\
274 * 1 & 0 & 1 & 2.4\\
275 * 2 & 0 & 1 & 8.9\\
276 * 0 & 1 & 1 & 1.3\\
277 * 1 & 1 & 1 & 1.6\\
278 * 2 & 1 & 1 & 2.6
279 * \end{array}
280 * \f]
281 *
282 * Note that the value of x_4 changes fastest, followed by that of x_8, and x_7
283 * varies the slowest, corresponding to the second line of the block ("4 8 7").
284 * Further, x_4 can take on three values, and x_8 and x_7 each have two possible
285 * values, as described in the third line of the block ("3 2 2"). The table
286 * contains 11 non-zero entries (all except for the fifth entry). Note that the
287 * eleventh and twelveth entries are interchanged.
288 *
289 * A final note: the internal representation in libDAI of the factor above is
290 * different, because the variables are ordered according to their indices
291 * (i.e., the ordering would be x_4 x_7 x_8) and the values of the table are
292 * stored accordingly, with the variable having the smallest index changing
293 * fastest:
294 *
295 * \f[
296 * \begin{array}{ccc|c}
297 * x_4 & x_7 & x_8 & \mbox{value}\\
298 * \hline
299 * 0 & 0 & 0 & 0.1\\
300 * 1 & 0 & 0 & 3.5\\
301 * 2 & 0 & 0 & 2.8\\
302 * 0 & 1 & 0 & 7.4\\
303 * 1 & 1 & 0 & 2.4\\
304 * 2 & 1 & 0 & 8.9\\
305 * 0 & 0 & 1 & 6.3\\
306 * 1 & 0 & 1 & 8.4\\
307 * 2 & 0 & 1 & 0.0\\
308 * 0 & 1 & 1 & 1.3\\
309 * 1 & 1 & 1 & 1.6\\
310 * 2 & 1 & 1 & 2.6
311 * \end{array}
312 * \f]
313 */
314
315
316 /** \page license License
317 * \verbinclude COPYING
318 */