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 */

9 /** \file

10 * \brief Contains additional doxygen documentation

11 *

12 * \idea Adapt (part of the) guidelines in http://www.boost.org/development/requirements.html#Design_and_Programming

13 *

14 * \idea Use "gcc -MM" to generate dependencies for targets: http://make.paulandlesley.org/autodep.html

15 *

16 * \idea Disentangle structures. In particular, ensure that graphical properties are not

17 * entangled with probabilistic properties. For example, a FactorGraph contains several components:

18 * - a BipartiteGraph

19 * - an array of variable labels

20 * - an array of variable state space sizes

21 * - an array of pointers to factor value vectors

22 * In this way, each factor could be implemented differently, e.g., we could have

23 * some sparse factors, some noisy-OR factors, some dense factors, some arbitrary

24 * precision factors, etcetera.

25 *

26 * \idea Use boost::uBLAS framework to deal with matrices, especially, with 2D sparse matrices.

27 * See http://www.boost.org/libs/numeric/ublas/doc/matrix_sparse.htm

28 * However: I read somewhere that boost::uBLAS concentrates more on correct implementation than on performance.

29 **/

32 /** \mainpage Reference manual for libDAI - A free/open source C++ library for Discrete Approximate Inference methods

33 * \author Joris Mooij (with contributions of Frederik Eaton)

34 * \version git HEAD

35 * \date July 12, 2011 - or later

36 *

37 * <hr size="1">

38 * \section about About libDAI

39 * libDAI is a free/open source C++ library that provides implementations of

40 * various (approximate) inference methods for discrete graphical models. libDAI

41 * supports arbitrary factor graphs with discrete variables; this includes

42 * discrete Markov Random Fields and Bayesian Networks.

43 *

44 * The library is targeted at researchers. To be able to use the library, a

45 * good understanding of graphical models is needed.

46 *

47 * The best way to use libDAI is by writing C++ code that invokes the library;

48 * in addition, part of the functionality is accessibly by using the

49 * - command line interface

50 * - (limited) MatLab interface

51 * - (experimental) python interface

52 * - (experimental) octave interface.

53 *

54 * libDAI can be used to implement novel (approximate) inference algorithms

55 * and to easily compare the accuracy and performance with existing algorithms

56 * that have been implemented already.

57 *

58 * A solver using libDAI was amongst the three winners of the UAI 2010 Approximate

59 * Inference Challenge (see http://www.cs.huji.ac.il/project/UAI10/ for more

60 * information). The full source code is provided as part of the library.

61 *

62 * \section features Features

63 * Currently, libDAI supports the following (approximate) inference methods:

64 * - Exact inference by brute force enumeration;

65 * - Exact inference by junction-tree methods;

66 * - Mean Field;

67 * - Loopy Belief Propagation [\ref KFL01];

68 * - Fractional Belief Propagation [\ref WiH03];

69 * - Tree-Reweighted Belief Propagation [\ref WJW03];

70 * - Tree Expectation Propagation [\ref MiQ04];

71 * - Generalized Belief Propagation [\ref YFW05];

72 * - Double-loop GBP [\ref HAK03];

73 * - Various variants of Loop Corrected Belief Propagation

74 * [\ref MoK07, \ref MoR05];

75 * - Gibbs sampler;

76 * - Conditioned Belief Propagation [\ref EaG09];

77 * - Decimation algorithm.

78 *

79 * These inference methods can be used to calculate partition sums, marginals

80 * over subsets of variables, and MAP states (the joint state of variables that

81 * has maximum probability).

82 *

83 * In addition, libDAI supports parameter learning of conditional probability

84 * tables by Expectation Maximization.

85 *

86 * \section limitations Limitations

87 * libDAI is not intended to be a complete package for approximate inference.

88 * Instead, it should be considered as an "inference engine", providing

89 * various inference methods. In particular, it contains no GUI, currently

90 * only supports its own file format for input and output (although support

91 * for standard file formats may be added later), and provides very limited

92 * visualization functionalities. The only learning method supported currently

93 * is Expectation Maximization (or Maximum Likelihood if no data is missing)

94 * for learning factor parameters.

95 *

96 * \section rationale Rationale

97 *

98 * In my opinion, the lack of open source "reference" implementations hampers

99 * progress in research on approximate inference. Methods differ widely in terms

100 * of quality and performance characteristics, which also depend in different

101 * ways on various properties of the graphical models. Finding the best

102 * approximate inference method for a particular application therefore often

103 * requires empirical comparisons. However, implementing and debugging these

104 * methods takes a lot of time which could otherwise be spent on research. I hope

105 * that this code will aid researchers to be able to easily compare various

106 * (existing as well as new) approximate inference methods, in this way

107 * accelerating research and stimulating real-world applications of approximate

108 * inference.

109 *

110 * \section language Language

111 * Because libDAI is implemented in C++, it is very fast compared with

112 * implementations in MatLab (a factor 1000 faster is not uncommon).

113 * libDAI does provide a (limited) MatLab interface for easy integration with MatLab.

114 * It also provides a command line interface and experimental python and octave

115 * interfaces (thanks to Patrick Pletscher).

116 *

117 * \section compatibility Compatibility

118 *

119 * The code has been developed under Debian GNU/Linux with the GCC compiler suite.

120 * libDAI compiles successfully with g++ versions 3.4 up to 4.6.

121 *

122 * libDAI has also been successfully compiled with MS Visual Studio 2008 under Windows

123 * (but not all build targets are supported yet) and with Cygwin under Windows.

124 *

125 * Finally, libDAI has been compiled successfully on MacOS X.

126 *

127 * \section download Downloading libDAI

128 * The libDAI sources and documentation can be downloaded from the libDAI website:

129 * http://www.libdai.org.

130 *

131 * \section support Mailing list

132 * The Google group "libDAI" (http://groups.google.com/group/libdai)

133 * can be used for getting support and discussing development issues.

134 */

137 /** \page license License

138 * <hr size="1">

139 * \section license-license License

140 *

141 * libDAI is free software; you can redistribute it and/or modify it under the

142 * terms of the BSD 2-clause license (also known as the FreeBSD license), which

143 * can be found in the accompanying LICENSE file.

144 *

145 * [Note: up to and including version 0.2.7, libDAI was licensed under the GNU

146 * General Public License (GPL) version 2 or higher.]

147 *

148 * <hr size="1">

149 * \section license-freebsd libDAI license (similar to FreeBSD license)

150 *

151 * \verbinclude LICENSE

152 */

155 /** \page citations Citing libDAI

156 * <hr size="1">

157 * \section citations-citations Citing libDAI

158 *

159 * If you write a scientific paper describing research that made substantive use

160 * of this library, please cite the following paper describing libDAI:\n

161 *

162 * Joris M. Mooij;\n

163 * libDAI: A free & open source C++ library for Discrete Approximate Inference in graphical models;\n

164 * Journal of Machine Learning Research, 11(Aug):2169-2173, 2010.\n

165 *

166 * In BiBTeX format (for your convenience):\n

167 *

168 * <pre>

169 * \@article{Mooij_libDAI_10,

170 * author = {Joris M. Mooij},

171 * title = {lib{DAI}: A Free and Open Source {C++} Library for Discrete Approximate Inference in Graphical Models},

172 * journal = {Journal of Machine Learning Research},

173 * year = 2010,

174 * month = Aug,

175 * volume = 11,

176 * pages = {2169-2173},

177 * url = "http://www.jmlr.org/papers/volume11/mooij10a/mooij10a.pdf"

178 * }</pre>

179 *

180 * Moreover, as a personal note, I would appreciate it to be informed about any

181 * publications using libDAI at joris dot mooij at libdai dot org.

182 */

185 /** \page authors Authors

186 * \section authors-authors People who contributed to libDAI

187 *

188 * \verbinclude AUTHORS

189 */

192 /** \page build Building libDAI

193 * <hr size="1">

194 * \section build-unix Building libDAI under UNIX variants (Linux / Cygwin / Mac OS X)

195 *

196 * \subsection build-unix-preparations Preparations

197 *

198 * You need:

199 * - a recent version of gcc (at least version 3.4)

200 * - GNU make

201 * - recent boost C++ libraries (at least version 1.37; however,

202 * version 1.37 shipped with Ubuntu 9.04 is known not to work)

203 * - GMP library (or the Windows port called MPIR)

204 * - doxygen (only for building the documentation)

205 * - graphviz (only for using some of the libDAI command line utilities)

206 * - CImg library (only for building the image segmentation example)

207 *

208 * On Debian/Ubuntu, you can easily install the required packages with a single command:

209 * <pre> apt-get install g++ make doxygen graphviz libboost-dev libboost-graph-dev libboost-program-options-dev libboost-test-dev libgmp-dev cimg-dev</pre>

210 * (root permissions needed).

211 *

212 * On Mac OS X (10.4 is known to work), these packages can be installed easily via MacPorts.

213 * If MacPorts is not already installed, install it according to the instructions at http://www.macports.org/.

214 * Then, a simple

215 * <pre> sudo port install gmake boost gmp doxygen graphviz</pre>

216 * should be enough to install everything that is needed.

217 *

218 * On Cygwin, the prebuilt Cygwin package boost-1.33.1-x is known not to work.

219 * You can however obtain the latest boost version (you need at least 1.37.0)

220 * from http://www.boost.org/ and build it as described in the next subsection.

221 *

222 * \subsubsection build-unix-boost Building boost under Cygwin

223 *

224 * - Download the latest boost libraries from http://www.boost.org

225 * - Build the required boost libraries using:

226 * <pre>

227 * ./bootstrap.sh --with-libraries=program_options,math,graph,test --prefix=/boost_root/

228 * ./bjam</pre>

229 * - In order to use dynamic linking, the boost .dll's should be somewhere in the path.

230 * This can be achieved by a command like:

231 * <pre>

232 * export PATH=$PATH:/boost_root/stage/lib</pre>

233 *

234 *

235 * \subsection build-unix-libdai Building libDAI

236 *

237 * To build the libDAI source, first copy a template Makefile.* to Makefile.conf

238 * (for example, copy Makefile.LINUX to Makefile.conf if you use GNU/Linux).

239 * Then, edit the Makefile.conf template to adapt it to your local setup. In case

240 * you want to use Boost libraries which are installed in non-standard locations,

241 * you have to tell the compiler and linker about their locations (using the

242 * -I, -L flags for GCC; also you may need to set the LD_LIBRARY_PATH environment

243 * variable correctly before running libDAI binaries). Platform independent build

244 * options can be set in Makefile.ALL. Finally, run

245 * <pre> make</pre>

246 * The build includes a regression test, which may take a while to complete.

247 *

248 * If the build is successful, you can test the example program:

249 * <pre> examples/example tests/alarm.fg</pre>

250 * or the more extensive test program:

251 * <pre> tests/testdai --aliases tests/aliases.conf --filename tests/alarm.fg --methods JTREE_HUGIN BP_SEQMAX</pre>

252 *

253 *

254 * <hr size="1">

255 * \section build-windows Building libDAI under Windows

256 *

257 * \subsection build-windows-preparations Preparations

258 *

259 * You need:

260 * - A recent version of MicroSoft Visual Studio (2008 is known to work)

261 * - recent boost C++ libraries (version 1.37 or higher)

262 * - GMP or MPIR library

263 * - GNU make (can be obtained from http://gnuwin32.sourceforge.net)

264 * - CImg library (only for building the image segmentation example)

265 *

266 * For the regression test, you need:

267 * - GNU diff, GNU sed (can be obtained from http://gnuwin32.sourceforge.net)

268 *

269 * \subsubsection build-windows-boost Building boost under Windows

270 *

271 * Because building boost under Windows is tricky, I provide some guidance here.

272 *

273 * - Download the boost zip file from http://www.boost.org/users/download

274 * and unpack it somewhere.

275 * - Download the bjam executable from http://www.boost.org/users/download

276 * and unpack it somewhere else.

277 * - Download Boost.Build (v2) from http://www.boost.org/docs/tools/build/index.html

278 * and unpack it yet somewhere else.

279 * - Edit the file \c boost-build.jam in the main boost directory to change the

280 * \c BOOST_BUILD directory to the place where you put Boost.Build (use UNIX

281 * / instead of Windows \ in pathnames).

282 * - Copy the \c bjam.exe executable into the main boost directory.

283 * Now if you issue <tt>"bjam --version"</tt> you should get a version and no errors.

284 * Issueing <tt>"bjam --show-libraries"</tt> will show the libraries that will be built.

285 * - The following command builds the boost libraries that are relevant for libDAI:

286 * <pre>

287 * bjam --with-graph --with-math --with-program_options --with-test link=static runtime-link=shared</pre>

288 *

289 * \subsection build-windows-gmp Building GMP or MPIR under Windows

290 *

291 * Information about how to build GPR or MPIR under Windows can be found on the internet.

292 * The user has to update Makefile.WINDOWS in order to link with the GPR/MPIR libraries.

293 *

294 * \subsection build-windows-libdai Building libDAI

295 *

296 * To build the source, copy Makefile.WINDOWS to Makefile.conf. Then, edit

297 * Makefile.conf to adapt it to your local setup. Platform independent

298 * build options can be set in Makefile.ALL. Finally, run (from the command line)

299 * <pre> make</pre>

300 * The build includes a regression test, which may take a while to complete.

301 *

302 * If the build is successful, you can test the example program:

303 * <pre> examples\\example tests\\alarm.fg</pre>

304 * or the more extensive test program:

305 * <pre> tests\\testdai --aliases tests\\aliases.conf --filename tests\\alarm.fg --methods JTREE_HUGIN BP_SEQMAX</pre>

306 *

307 *

308 * <hr size="1">

309 * \section build-matlab Building the libDAI MatLab interface

310 *

311 * You need:

312 * - MatLab

313 * - The platform-dependent requirements described above

314 *

315 * First, you need to build the libDAI source as described above for your

316 * platform. By default, the MatLab interface is disabled, so before compiling the

317 * source, you have to enable it in Makefile.ALL by setting

318 * <pre> WITH_MATLAB=true</pre>

319 * Also, you have to configure the MatLab-specific parts of

320 * Makefile.conf to match your system (in particular, the Makefile variables ME,

321 * MATLABDIR and MEX). The MEX file extension depends on your platform; for a

322 * 64-bit linux x86_64 system this would be "ME=.mexa64", for a 32-bit linux x86

323 * system "ME=.mexglx". If you are unsure about your MEX file

324 * extension: it needs to be the same as what the MatLab command "mexext" returns.

325 * The required MEX files are built by issuing

326 * <pre> make</pre>

327 * from the command line. The MatLab interface is much less powerful than using

328 * libDAI from C++. There are two reasons for this: (i) it is boring to write MEX

329 * files; (ii) the large performance penalty paid when large data structures (like

330 * factor graphs) have to be converted between their native C++ data structure to

331 * something that MatLab understands.

332 *

333 * A simple example of how to use the MatLab interface is the following (entered

334 * at the MatLab prompt), which performs exact inference by the junction tree

335 * algorithm and approximate inference by belief propagation on the ALARM network:

336 * <pre> cd path_to_libdai/matlab

337 * [psi] = dai_readfg ('../tests/alarm.fg');

338 * [logZ,q,md,qv,qf] = dai (psi, 'JTREE', '[updates=HUGIN,verbose=0]')

339 * [logZ,q,md,qv,qf] = dai (psi, 'BP', '[updates=SEQMAX,tol=1e-9,maxiter=10000,logdomain=0]')</pre>

340 * where "path_to_libdai" has to be replaced with the directory in which libDAI

341 * was installed. For other algorithms and some default parameters, see the file

342 * tests/aliases.conf.

343 *

344 * <hr size="1">

345 * \section build-doxygen Building the documentation

346 *

347 * Install doxygen, graphviz and a TeX distribution and use

348 * <pre> make doc</pre>

349 * to build the documentation. If the documentation is not clear enough, feel free

350 * to send me an email (or even better, to improve the documentation and send a patch!).

351 * The documentation can also be browsed online at http://www.libdai.org.

352 */

355 /** \page changelog Change Log

356 * \verbinclude ChangeLog

357 */

360 /** \page terminology Terminology and conventions

361 *

362 * \section terminology-graphicalmodels Graphical models

363 *

364 * Commonly used graphical models are Bayesian networks and Markov random fields.

365 * In libDAI, both types of graphical models are represented by a slightly more

366 * general type of graphical model: a factor graph [\ref KFL01].

367 *

368 * An example of a Bayesian network is:

369 * \dot

370 * digraph bayesnet {

371 * size="1,1";

372 * x0 [label="0"];

373 * x1 [label="1"];

374 * x2 [label="2"];

375 * x3 [label="3"];

376 * x4 [label="4"];

377 * x0 -> x1;

378 * x0 -> x2;

379 * x1 -> x3;

380 * x1 -> x4;

381 * x2 -> x4;

382 * }

383 * \enddot

384 * The probability distribution of a Bayesian network factorizes as:

385 * \f[ P(\mathbf{x}) = \prod_{i\in\mathcal{V}} P(x_i \,|\, x_{\mathrm{pa}(i)}) \f]

386 * where \f$\mathrm{pa}(i)\f$ are the parents of node \a i in a DAG.

387 *

388 * The same probability distribution can be represented as a Markov random field:

389 * \dot

390 * graph mrf {

391 * size="1.5,1.5";

392 * x0 [label="0"];

393 * x1 [label="1"];

394 * x2 [label="2"];

395 * x3 [label="3"];

396 * x4 [label="4"];

397 * x0 -- x1;

398 * x0 -- x2;

399 * x1 -- x2;

400 * x1 -- x3;

401 * x1 -- x4;

402 * x2 -- x4;

403 * }

404 * \enddot

405 *

406 * The probability distribution of a Markov random field factorizes as:

407 * \f[ P(\mathbf{x}) = \frac{1}{Z} \prod_{C\in\mathcal{C}} \psi_C(x_C) \f]

408 * where \f$ \mathcal{C} \f$ are the cliques of an undirected graph,

409 * \f$ \psi_C(x_C) \f$ are "potentials" or "compatibility functions", and

410 * \f$ Z \f$ is the partition sum which properly normalizes the probability

411 * distribution.

412 *

413 * Finally, the same probability distribution can be represented as a factor graph:

414 * \dot

415 * graph factorgraph {

416 * size="1.8,1";

417 * x0 [label="0"];

418 * x1 [label="1"];

419 * x2 [label="2"];

420 * x3 [label="3"];

421 * x4 [label="4"];

422 * f01 [shape="box",label=""];

423 * f02 [shape="box",label=""];

424 * f13 [shape="box",label=""];

425 * f124 [shape="box",label=""];

426 * x0 -- f01;

427 * x1 -- f01;

428 * x0 -- f02;

429 * x2 -- f02;

430 * x1 -- f13;

431 * x3 -- f13;

432 * x1 -- f124;

433 * x2 -- f124;

434 * x4 -- f124;

435 * }

436 * \enddot

437 *

438 * The probability distribution of a factor graph factorizes as:

439 * \f[ P(\mathbf{x}) = \frac{1}{Z} \prod_{I\in \mathcal{F}} f_I(x_I) \f]

440 * where \f$ \mathcal{F} \f$ are the factor nodes of a factor graph (a

441 * bipartite graph consisting of variable nodes and factor nodes),

442 * \f$ f_I(x_I) \f$ are the factors, and \f$ Z \f$ is the partition sum

443 * which properly normalizes the probability distribution.

444 *

445 * Looking at the expressions for the joint probability distributions,

446 * it is obvious that Bayesian networks and Markov random fields can

447 * both be easily represented as factor graphs. Factor graphs most

448 * naturally express the factorization structure of a probability

449 * distribution, and hence are a convenient representation for approximate

450 * inference algorithms, which all try to exploit this factorization.

451 * This is why libDAI uses a factor graph as representation of a

452 * graphical model, implemented in the dai::FactorGraph class.

453 *

454 * \section terminology-inference Inference tasks

455 *

456 * Given a factor graph, specified by the variable nodes \f$\{x_i\}_{i\in\mathcal{V}}\f$

457 * the factor nodes \f$ \mathcal{F} \f$, the graph structure, and the factors

458 * \f$\{f_I(x_I)\}_{I\in\mathcal{F}}\f$, the following tasks are important:

459 *

460 * - Calculating the partition sum:

461 * \f[ Z = \sum_{\mathbf{x}_{\mathcal{V}}} \prod_{I \in \mathcal{F}} f_I(x_I) \f]

462 * - Calculating the marginal distribution of a subset of variables

463 * \f$\{x_i\}_{i\in A}\f$:

464 * \f[ P(\mathbf{x}_{A}) = \frac{1}{Z} \sum_{\mathbf{x}_{\mathcal{V}\setminus A}} \prod_{I \in \mathcal{F}} f_I(x_I) \f]

465 * - Calculating the MAP state which has the maximum probability mass:

466 * \f[ \mathrm{argmax}_{\mathbf{x}}\,\prod_{I\in\mathcal{F}} f_I(x_I) \f]

467 *

468 * libDAI offers several inference algorithms, which solve (a subset of) these tasks either

469 * approximately or exactly, for factor graphs with discrete variables. The following

470 * algorithms are implemented:

471 *

472 * Exact inference:

473 * - Brute force enumeration: dai::ExactInf

474 * - Junction-tree method: dai::JTree

475 *

476 * Approximate inference:

477 * - Mean Field: dai::MF

478 * - (Loopy) Belief Propagation: dai::BP [\ref KFL01]

479 * - Fractional Belief Propagation: dai::FBP [\ref WiH03]

480 * - Tree-Reweighted Belief Propagation: dai::TRWBP [\ref WJW03]

481 * - Tree Expectation Propagation: dai::TreeEP [\ref MiQ04]

482 * - Generalized Belief Propagation: dai::HAK [\ref YFW05]

483 * - Double-loop GBP: dai::HAK [\ref HAK03]

484 * - Loop Corrected Belief Propagation: dai::MR [\ref MoR05] and dai::LC [\ref MoK07]

485 * - Gibbs sampling: dai::Gibbs

486 * - Conditioned Belief Propagation: dai::CBP [\ref EaG09]

487 * - Decimation algorithm: dai::DecMAP

488 *

489 * Not all inference tasks are implemented by each method: calculating MAP states

490 * is only possible with dai::JTree, dai::BP and dai::DECMAP; calculating partition sums is

491 * not possible with dai::MR, dai::LC and dai::Gibbs.

492 *

493 * \section terminology-learning Parameter learning

494 *

495 * In addition, libDAI supports parameter learning of conditional probability

496 * tables by Expectation Maximization (or Maximum Likelihood, if there is no

497 * missing data). This is implemented in dai::EMAlg.

498 *

499 * \section terminology-variables-states Variables and states

500 *

501 * Linear states are a concept that is used often in libDAI, for example for storing

502 * and accessing factors, which are functions mapping from states of a set of variables

503 * to the real numbers. Internally, a factor is stored as an array, and the array index

504 * of an entry corresponds with the linear state of the set of variables. Below we will

505 * define variables, states and linear states of (sets of) variables.

506 *

507 * \subsection terminology-variables Variables

508 *

509 * Each (random) \a variable has a unique identifier, its \a label (which has

510 * a non-negative integer value). If two variables have the same

511 * label, they are considered as identical. A variable can take on a finite

512 * number of different values or \a states.

513 *

514 * We use the following notational conventions. The discrete

515 * random variable with label \f$l\f$ is denoted as \f$x_l\f$, and the number

516 * of possible values of this variable as \f$S_{x_l}\f$ or simply \f$S_l\f$.

517 * The set of possible values of variable \f$x_l\f$ is denoted

518 * \f$X_l := \{0,1,\dots,S_l-1\}\f$ and called its \a state \a space.

519 *

520 * \subsection terminology-variable-sets Sets of variables and the canonical ordering

521 *

522 * Let \f$A := \{x_{l_1},x_{l_2},\dots,x_{l_n}\}\f$ be a set of variables.

523 *

524 * The \a canonical \a ordering of the variables in \a A is induced by their labels.

525 * That is: if \f$l_1 < l_2\f$, then \f$x_{l_1}\f$ occurs before \f$x_{l_2}\f$ in the

526 * canonical ordering. Below, we will assume that \f$(l_i)_{i=1}^n\f$ is

527 * ordered according to the canonical ordering, i.e., \f$l_1 < l_2 < \dots < l_n\f$.

528 *

529 * \subsection terminology-variable-states States and linear states of sets of variables

530 *

531 * A \a state of the variables in \a A refers to a joint assignment of the

532 * variables, or in other words, to an element of the Cartesian product

533 * \f$ \prod_{i=1}^n X_{l_i}\f$ of the state spaces of the variables in \a A.

534 * Note that a state can also be interpreted as a mapping from variables (or

535 * variable labels) to the natural numbers, which assigns to a variable (or its

536 * label) the corresponding state of the variable.

537 *

538 * A state of \a n variables can be represented as an n-tuple of

539 * non-negative integers: \f$(s_1,s_2,\dots,s_n)\f$ corresponds to the

540 * joint assignment \f$x_{l_1} = s_1, \dots, x_{l_n} = s_n\f$.

541 * Alternatively, a state can be represented compactly as one non-negative integer;

542 * this representation is called a \a linear \a state. The linear state

543 * \a s corresponding to the state \f$(s_1,s_2,\dots,s_n)\f$ would be:

544 * \f[

545 * s := \sum_{i=1}^n s_i \prod_{j=1}^{i-1} S_{l_j}

546 * = s_1 + s_2 S_{l_1} + s_3 S_{l_1} S_{l_2} + \dots + s_n S_{l_1} \cdots S_{l_{n-1}}.

547 * \f]

548 *

549 * Vice versa, given a linear state \a s for the variables \a A, the

550 * corresponding state \f$s_i\f$ of the \a i 'th variable \f$x_{l_i}\f$ (according to

551 * the canonical ordering of the variables in \a A) is given by

552 * \f[

553 * s_i = \left\lfloor\frac{s \mbox { mod } \prod_{j=1}^i S_{l_j}}{\prod_{j=1}^{i-1} S_{l_j}}\right\rfloor.

554 * \f]

555 *

556 * Finally, the \a number \a of \a states of the set of variables \a A is simply the

557 * number of different joint assignments of the variables, that is, \f$\prod_{i=1}^n S_{l_i}\f$.

558 */

561 /** \page fileformats libDAI file formats

562 *

563 * \section fileformats-factorgraph Factor graph (.fg) file format

564 *

565 * This section describes the .fg file format used in libDAI to store factor graphs.

566 * Markov Random Fields are special cases of factor graphs, as are Bayesian

567 * networks. A factor graph can be specified as follows: for each factor, one has

568 * to specify which variables occur in the factor, what their respective

569 * cardinalities (i.e., number of possible values) are, and a table listing all

570 * the values of that factor for all possible configurations of these variables.

571 *

572 * A .fg file is not much more than that. It starts with a line containing the

573 * number of factors in that graph, followed by an empty line. Then all factors

574 * are specified, using one block for each factor, where the blocks are seperated

575 * by empty lines. Each variable occurring in the factor graph has a unique

576 * identifier, its label (which should be a nonnegative integer). Comment lines

577 * which start with # are ignored.

578 *

579 * \subsection fileformats-factorgraph-factor Factor block format

580 *

581 * Each block describing a factor starts with a line containing the number of

582 * variables in that factor. The second line contains the labels of these

583 * variables, seperated by spaces (labels are nonnegative integers and to avoid

584 * confusion, it is suggested to start counting at 0). The third line contains

585 * the number of possible values of each of these variables, also seperated by

586 * spaces. Note that there is some redundancy here, since if a variable appears

587 * in more than one factor, the cardinality of that variable appears several

588 * times in the .fg file; obviously, these cardinalities should be consistent.

589 * The fourth line contains the number of nonzero entries

590 * in the factor table. The rest of the lines contain these nonzero entries;

591 * each line consists of a table index, followed by white-space, followed by the

592 * value corresponding to that table index. The most difficult part is getting

593 * the indexing right. The convention that is used is that the left-most

594 * variables cycle through their values the fastest (similar to MatLab indexing

595 * of multidimensional arrays).

596 *

597 * \subsubsection fileformats-factorgraph-factor-example Example

598 *

599 * An example block describing one factor is:

600 *

601 * <pre>

602 * 3

603 * 4 8 7

604 * 3 2 2

605 * 11

606 * 0 0.1

607 * 1 3.5

608 * 2 2.8

609 * 3 6.3

610 * 4 8.4

611 * 6 7.4

612 * 7 2.4

613 * 8 8.9

614 * 9 1.3

615 * 10 1.6

616 * 11 2.6

617 * </pre>

618 *

619 * which corresponds to the following factor:

620 *

621 * \f[

622 * \begin{array}{ccc|c}

623 * x_4 & x_8 & x_7 & \mbox{value}\\

624 * \hline

625 * 0 & 0 & 0 & 0.1\\

626 * 1 & 0 & 0 & 3.5\\

627 * 2 & 0 & 0 & 2.8\\

628 * 0 & 1 & 0 & 6.3\\

629 * 1 & 1 & 0 & 8.4\\

630 * 2 & 1 & 0 & 0.0\\

631 * 0 & 0 & 1 & 7.4\\

632 * 1 & 0 & 1 & 2.4\\

633 * 2 & 0 & 1 & 8.9\\

634 * 0 & 1 & 1 & 1.3\\

635 * 1 & 1 & 1 & 1.6\\

636 * 2 & 1 & 1 & 2.6

637 * \end{array}

638 * \f]

639 *

640 * Note that the value of \f$x_4\f$ changes fastest, followed by that of \f$x_8\f$, and \f$x_7\f$

641 * varies the slowest, corresponding to the second line of the block ("4 8 7").

642 * Further, \f$x_4\f$ can take on three values, and \f$x_8\f$ and \f$x_7\f$ each have two possible

643 * values, as described in the third line of the block ("3 2 2"). The table

644 * contains 11 non-zero entries (all except for the fifth entry). Note that the

645 * eleventh and twelveth entries are interchanged.

646 *

647 * A final note: the internal representation in libDAI of the factor above is

648 * different, because the variables are ordered according to their indices

649 * (i.e., the ordering would be \f$x_4 x_7 x_8\f$) and the values of the table are

650 * stored accordingly, with the variable having the smallest index changing

651 * fastest:

652 *

653 * \f[

654 * \begin{array}{ccc|c}

655 * x_4 & x_7 & x_8 & \mbox{value}\\

656 * \hline

657 * 0 & 0 & 0 & 0.1\\

658 * 1 & 0 & 0 & 3.5\\

659 * 2 & 0 & 0 & 2.8\\

660 * 0 & 1 & 0 & 7.4\\

661 * 1 & 1 & 0 & 2.4\\

662 * 2 & 1 & 0 & 8.9\\

663 * 0 & 0 & 1 & 6.3\\

664 * 1 & 0 & 1 & 8.4\\

665 * 2 & 0 & 1 & 0.0\\

666 * 0 & 1 & 1 & 1.3\\

667 * 1 & 1 & 1 & 1.6\\

668 * 2 & 1 & 1 & 2.6

669 * \end{array}

670 * \f]

671 *

672 *

673 * \section fileformats-evidence Evidence (.tab) file format

674 *

675 * This section describes the .tab fileformat used in libDAI to store "evidence",

676 * i.e., a data set consisting of multiple samples, where each sample is the

677 * observed joint state of some variables.

678 *

679 * A .tab file is a tabular data file, consisting of a header line, followed by

680 * an empty line, followed by the data points, with one line for each data point.

681 * Each line (apart from the empty one) should have the same number of columns,

682 * where columns are separated by one tab character. Each column corresponds to

683 * a variable. The header line consists of the variable labels (corresponding to

684 * dai::Var::label()). The other lines are observed joint states of the variables, i.e.,

685 * each line corresponds to a joint observation of the variables, and each column

686 * of a line contains the state of the variable associated with that column.

687 * Missing data is handled simply by having two consecutive tab characters,

688 * without any characters in between.

689 *

690 * \subsection fileformats-evidence-example Example

691 *

692 * <pre>

693 * 1 3 2

694 *

695 * 0 0 1

696 * 1 0 1

697 * 1 1

698 * </pre>

699 *

700 * This would correspond to a data set consisting of three observations concerning

701 * the variables with labels 1, 3 and 2; the first observation being

702 * \f$x_1 = 0, x_3 = 0, x_2 = 1\f$, the second observation being

703 * \f$x_1 = 1, x_3 = 0, x_2 = 1\f$, and the third observation being

704 * \f$x_1 = 1, x_2 = 1\f$ (where the state of \f$x_3\f$ is missing).

705 *

706 * \section fileformats-emalg Expectation Maximization (.em) file format

707 *

708 * This section describes the file format of .em files, which are used

709 * to specify a particular EM algorithm. The .em files are complementary

710 * to .fg files; in other words, an .em file without a corresponding .fg

711 * file is useless. Furthermore, one also needs a corresponding .tab file

712 * containing the data used for parameter learning.

713 *

714 * An .em file starts with a line specifying the number of maximization steps,

715 * followed by an empty line. Then, each maximization step is described in a

716 * block, which should satisfy the format described in the next subsection.

717 *

718 * \subsection fileformats-emalg-maximizationstep Maximization Step block format

719 *

720 * A maximization step block of an .em file starts with a single line

721 * describing the number of shared parameters blocks that will follow.

722 * Then, each shared parameters block follows, in the format described in

723 * the next subsection.

724 *

725 * \subsection fileformats-emalg-sharedparameters Shared parameters block format

726 *

727 * A shared parameters block of an .em file starts with a single line

728 * consisting of the name of a ParameterEstimation subclass

729 * and its parameters in the format of a PropertySet. For example:

730 * <pre> CondProbEstimation [target_dim=2,total_dim=4,pseudo_count=1]</pre>

731 * The next line contains the number of factors that share their parameters.

732 * Then, each of these factors is specified on separate lines (possibly

733 * seperated by empty lines), where each line consists of several fields

734 * seperated by a space or a tab character. The first field contains

735 * the index of the factor in the factor graph. The following fields should

736 * contain the variable labels of the variables on which that factor depends,

737 * in a specific ordering. This ordering can be different from the canonical

738 * ordering of the variables used internally in libDAI (which would be sorted

739 * ascendingly according to the variable labels). The ordering of the variables

740 * specifies the implicit ordering of the shared parameters: when iterating

741 * over all shared parameters, the corresponding index of the first variable

742 * changes fastest (in the inner loop), and the corresponding index of the

743 * last variable changes slowest (in the outer loop). By choosing the right

744 * ordering, it is possible to let different factors (depending on different

745 * variables) share parameters in parameter learning using EM. This convention

746 * is similar to the convention used in factor blocks in a factor graph .fg

747 * file (see \ref fileformats-factorgraph-factor).

748 *

749 * \section fileformats-aliases Aliases file format

750 *

751 * An aliases file is basically a list of "macros" and the strings that they

752 * should be substituted with.

753 *

754 * Each line of the aliases file can be either empty, contain a comment

755 * (if the first character is a '#') or contain an alias. In the latter case,

756 * the line should contain a colon; the part before the colon contains the

757 * name of the alias, the part after the colon the string that it should be

758 * substituted with. Any whitespace before and after the colon is ignored.

759 *

760 * For example, the following line would define the alias \c BP_SEQFIX

761 * as a shorthand for "BP[updates=SEQFIX,tol=1e-9,maxiter=10000,logdomain=0]":

762 * <pre>

763 * BP_SEQFIX: BP[updates=SEQFIX,tol=1e-9,maxiter=10000,logdomain=0]

764 * </pre>

765 *

766 * Aliases files can be used to store default options for algorithms.

767 */

769 /** \page bibliography Bibliography

770 * \anchor EaG09 \ref EaG09

771 * F. Eaton and Z. Ghahramani (2009):

772 * "Choosing a Variable to Clamp",

773 * <em>Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS 2009)</em> 5:145-152,

774 * http://jmlr.csail.mit.edu/proceedings/papers/v5/eaton09a/eaton09a.pdf

775 *

776 * \anchor EMK06 \ref EMK06

777 * G. Elidan and I. McGraw and D. Koller (2006):

778 * "Residual Belief Propagation: Informed Scheduling for Asynchronous Message Passing",

779 * <em>Proceedings of the 22nd Annual Conference on Uncertainty in Artificial Intelligence (UAI-06)</em>,

780 * http://uai.sis.pitt.edu/papers/06/UAI2006_0091.pdf

781 *

782 * \anchor HAK03 \ref HAK03

783 * T. Heskes and C. A. Albers and H. J. Kappen (2003):

784 * "Approximate Inference and Constrained Optimization",

785 * <em>Proceedings of the 19th Annual Conference on Uncertainty in Artificial Intelligence (UAI-03)</em> pp. 313-320,

786 * http://www.snn.ru.nl/reports/Heskes.uai2003.ps.gz

787 *

788 * \anchor KFL01 \ref KFL01

789 * F. R. Kschischang and B. J. Frey and H.-A. Loeliger (2001):

790 * "Factor Graphs and the Sum-Product Algorithm",

791 * <em>IEEE Transactions on Information Theory</em> 47(2):498-519,

792 * http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=910572

793 *

794 * \anchor KoF09 \ref KoF09

795 * D. Koller and N. Friedman (2009):

796 * <em>Probabilistic Graphical Models - Principles and Techniques</em>,

797 * The MIT Press, Cambridge, Massachusetts, London, England.

799 * \anchor Min05 \ref Min05

800 * T. Minka (2005):

801 * "Divergence measures and message passing",

802 * <em>MicroSoft Research Technical Report</em> MSR-TR-2005-173,

803 * http://research.microsoft.com/en-us/um/people/minka/papers/message-passing/minka-divergence.pdf

804 *

805 * \anchor MiQ04 \ref MiQ04

806 * T. Minka and Y. Qi (2004):

807 * "Tree-structured Approximations by Expectation Propagation",

808 * <em>Advances in Neural Information Processing Systems</em> (NIPS) 16,

809 * http://books.nips.cc/papers/files/nips16/NIPS2003_AA25.pdf

810 *

811 * \anchor MoK07 \ref MoK07

812 * J. M. Mooij and H. J. Kappen (2007):

813 * "Loop Corrections for Approximate Inference on Factor Graphs",

814 * <em>Journal of Machine Learning Research</em> 8:1113-1143,

815 * http://www.jmlr.org/papers/volume8/mooij07a/mooij07a.pdf

816 *

817 * \anchor MoK07b \ref MoK07b

818 * J. M. Mooij and H. J. Kappen (2007):

819 * "Sufficient Conditions for Convergence of the Sum-Product Algorithm",

820 * <em>IEEE Transactions on Information Theory</em> 53(12):4422-4437,

821 * http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=4385778

822 *

823 * \anchor Moo08 \ref Moo08

824 * J. M. Mooij (2008):

825 * "Understanding and Improving Belief Propagation",

826 * <em>Ph.D. Thesis</em> Radboud University Nijmegen

827 * http://webdoc.ubn.ru.nl/mono/m/mooij_j/undeanimb.pdf

828 *

829 * \anchor MoR05 \ref MoR05

830 * A. Montanari and T. Rizzo (2005):

831 * "How to Compute Loop Corrections to the Bethe Approximation",

832 * <em>Journal of Statistical Mechanics: Theory and Experiment</em> 2005(10)-P10011,

833 * http://stacks.iop.org/1742-5468/2005/P10011

834 *

835 * \anchor StW99 \ref StW99

836 * A. Steger and N. C. Wormald (1999):

837 * "Generating Random Regular Graphs Quickly",

838 * <em>Combinatorics, Probability and Computing</em> Vol 8, Issue 4, pp. 377-396,

839 * http://www.math.uwaterloo.ca/~nwormald/papers/randgen.pdf

840 *

841 * \anchor WiH03 \ref WiH03

842 * W. Wiegerinck and T. Heskes (2003):

843 * "Fractional Belief Propagation",

844 * <em>Advances in Neural Information Processing Systems</em> (NIPS) 15, pp. 438-445,

845 * http://books.nips.cc/papers/files/nips15/LT16.pdf

846 *

847 * \anchor WJW03 \ref WJW03

848 * M. J. Wainwright, T. S. Jaakkola and A. S. Willsky (2003):

849 * "Tree-reweighted belief propagation algorithms and approximate ML estimation by pseudo-moment matching",

850 * <em>9th Workshop on Artificial Intelligence and Statistics</em>,

851 * http://www.eecs.berkeley.edu/~wainwrig/Papers/WJW_AIStat03.pdf

852 *

853 * \anchor YFW05 \ref YFW05

854 * J. S. Yedidia and W. T. Freeman and Y. Weiss (2005):

855 * "Constructing Free-Energy Approximations and Generalized Belief Propagation Algorithms",

856 * <em>IEEE Transactions on Information Theory</em> 51(7):2282-2312,

857 * http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1459044

858 */

861 /** \page discussion Ideas not worth exploring

862 * \section discuss_extendedgraphs Extended factorgraphs/regiongraphs

863 *

864 * A FactorGraph and a RegionGraph are often equipped with

865 * additional properties for nodes and edges. The code to initialize those

866 * is often quite similar. Maybe one could abstract this, e.g.:

867 * \code

868 * template <typename Node1Properties, typename Node2Properties, typename EdgeProperties>

869 * class ExtFactorGraph : public FactorGraph {

870 * public:

871 * std::vector<Node1Properties> node1Props;

872 * std::vector<Node2Properties> node2Props;

873 * std::vector<std::vector<EdgeProperties> > edgeProps;

874 * // ...

875 * }

876 * \endcode

877 *

878 * Advantages:

879 * - Less code duplication.

880 * - Easier maintainability.

881 * - Easier to write new inference algorithms.

882 *

883 * Disadvantages:

884 * - Cachability may be worse.

885 * - A problem is the case where there are no properties for either type of nodes or for edges.

886 * Maybe this can be solved using specializations, or using variadac template arguments?

887 * Another possible solution would be to define a "class Empty {}", and add some code

888 * that checks for the typeid, comparing it with Empty, and doing something special in that case

889 * (e.g., not allocating memory).

890 * - The main disadvantage of this approach seems to be that it leads to even more entanglement.

891 * Therefore this is probably a bad idea.

892 *

893 * \section discuss_templates Polymorphism by template parameterization

894 *

895 * Instead of polymorphism by inheritance, use polymorphism by template parameterization.

896 * For example, the real reason for introducing the complicated inheritance scheme of dai::InfAlg

897 * was for functions like dai::calcMarginal. Instead, one could use a template function:

898 * \code

899 * template<typename InfAlg>

900 * Factor calcMarginal( const InfAlg &obj, const VarSet &ns, bool reInit );

901 * \endcode

902 * This would assume that the type InfAlg supports certain methods. Ideally, one would use

903 * concepts to define different classes of inference algorithms with different capabilities,

904 * for example the ability to calculate logZ, the ability to calculate marginals, the ability to

905 * calculate bounds, the ability to calculate MAP states, etc. Then, one would use traits

906 * classes in order to be able to query the capabilities of the model. For example, one would be

907 * able to query whether the inference algorithm supports calculation of logZ. Unfortunately,

908 * this is compile-time polymorphism, whereas tests/testdai needs runtime polymorphism.

909 * Therefore this is probably a bad idea.

910 */