1 libDAI-0.2.2 (2008-??-??)

2 -------------------------

4 * Now compiles also under Visual C++ and cygwin.

5 * Added Exceptions framework.

6 * Approximate inference methods now report the number of iterations needed.

7 * Added damping to various algorithms to improve convergence properties.

8 * Replaced ENUM2,ENUM3,ENUM4,ENUM5,ENUM6 by single DAI_ENUM macro.

9 * Removed utils/remove_short_loops and matlab/remove_short_loops.

10 * Added more features to utils/createfg for creating factor graphs.

11 * Replaced sub_nb class in mr.h by boost::dynamic_bitset.

12 * Pervasive change of BipartiteGraph implementation (based on an idea by

13 Giuseppe Passino). BipartiteGraph no longer stores the node properties

14 (former _V1 and _V2), nor does it store a dense adjacency matrix anymore,

15 nor an edge list. Instead, it stores the graph structure as lists of

16 neighboring nodes. This yields a significant memory/speed improvement for

17 large factor graphs, and is more elegant as well. Iterating over neighbors is

18 made easy by using boost::foreach.

19 * Improved index.h (merged from SVN head):

20 - Renamed Index -> IndexFor

21 - Added some .reserve()'s to IndexFor methods which yields a

22 25% speedup of testregression

23 - Replaced multind by Permute

24 - Added MultiFor

25 - Added State

26 * Improved factor.h (merged from SVN head): mainly new functionality.

27 * Moved everything into namespace "dai".

28 * Added ExactInf class for brute force exact inference.

29 * Renamed DEBUG to DAI_DEBUG to avoid conflicts.

30 * Added conditional compilation of inference methods.

31 * Contributions by Peter Gober:

32 - Renamed variable _N in mr.* for compatibility with g++ under cygwin.

33 * Contributions by Giuseppe Passino:

34 - removed "using namespace std;" from header files - bad practice;

35 - moved header files in include/dai and sources in src;

36 - changed #ifndefs to GNU style;

37 - added extra warning checks (-W -Wextra) and fixed resulting warnings;

38 - dai::TProb:

39 o removed copy constructor and assignment operators (redundant);

40 o implementation of some methods via STL algorithms;

41 o added methods takeExp, takeLog, takeLog0 for transformation in-place;

42 o explicit constructor (prevents implicit conversion from size_t to TProb);

43 o added operator+,+=,-,-=, with argument T (for calculations in log-scale);

44 - Added "logdomain" property to BP, a boolean that controls whether

45 calculations are done in the log-domain or in the linear domain;

46 doing calculations in the log-domain may help if the numerical range

47 of a double is too small.

48 * Contributions by Christian Wojek:

49 - New FactorGraph constructor that constructs from given ranges of factors

50 and variables;

51 - Optimization of FactorGraph constructors using tr1::unordered_map.

52 * Contributions by Claudio Lima:

53 - Added Max-Product functinoality to BP.

54 * FactorGraph constructors no longer check for short loops (huge speed

55 increase for large factor graphs), nor for negative entries. Also, the

56 normtype is now Prob::NORMPROB by default.

57 * VarSet is now implemented using a std::vector<Var> instead of a

58 std::set<Var>, which yields a significant speed improvement. Furthermore,

59 the implementation has been generalized, resulting in the small_set<T> class

60 which can be used to represent sets of small cardinality; VarSet is the

61 specialization with T = Var.

62 * Improved MaxSpanningTreePrims algorithm (uses boost::graph).

63 * Small optimization in Diffs.

64 * Replaced Complex with real numbers (negative potentials is just not

65 used enough to warrant the additional "complexity" :)).

66 * Moved Properties and MaxDiff frameworks from InfAlg to each individual

67 inference algorithm, because the Properties framework was not as

68 convenient as I hoped, and not every inference algorithm needs a maxdiff

69 variable. Also, replaced some FactorGraph functionality in InfAlg by a

70 function that returns the FactorGraph. The result is cleaner (less

71 entangled) code.

72 * Improved ClusterGraph implementation, yielding significant speedups

73 for the JunctionTree algorithm on large factorgraphs.

74 * Improved documetation.

75 * Removed x2x.

76 * Interface changes:

77 - VarSet::

78 VarSet::stateSpace() -> nrStates(const VarSet &)

79 VarSet( const std::set<Var> ) -> VarSet( begin, end, sizeHint=0 )

80 VarSet( const std::vector<Var> ) -> VarSet( begin, end, sizeHint=0 )

81 removed bool operator||

82 operator&&(const VarSet&) -> intersects(const VarSet&)

83 operator&&(const Var&) -> contains(const Var&)

84 - FactorGraph::

85 delta(const Var &) -> delta(size_t)

86 Delta(const Var &) -> Delta(size_t)

87 makeCavity(const Var &) -> makeCavity(size_t)

88 vars() -> vars

89 factors() -> factors

90 removed MakeFactorCavity(size_t)

91 removed ExactMarginal(const VarSet &)

92 removed ExactlogZ()

93 removed updatedFactor(size_t)

94 removed _normtype and NormType()

95 removed hasShortLoops(...) and removeShortLoops(...)

96 WriteToDotFile(const char *filename) -> printDot( std::ostream& os )

97 undoProb(size_t) -> restoreFactor(size_t)

98 saveProb(size_t) -> backupFactor(size_t)

99 undoProbs(const VarSet &) -> restoreFactors(const VarSet &)

100 saveProbs(const VarSet &) -> backupFactors(const VarSet &)

101 ReadFromFile(const char*) returns void (throws on error)

102 WriteToFile(const char*) returns void (throws on error)

103 removed hasNegatives()

104 - RegionGraph::

105 nr_ORs() -> nrORs()

106 nr_IRs() -> nrIRs()

107 ORs() -> ORs

108 IRs() -> IRs

109 - *::Regenerate() -> *::construct()

110 - Renamed Index -> IndexFor

111 - Diffs:

112 max() -> maxDiff()

113 max_size() -> maxSize()

114 - Prob::max() -> Prob::maxVal()

115 - Factor::

116 max() -> maxVal()

117 part_sum() -> partSum()

118 - toc() in util.h now returns seconds as a double

119 - VarSet::operator&&

120 - Properties -> PropertySet

123 libDAI-0.2.1 (2008-05-26)

124 -------------------------

126 Bugfix release.

127 * added missing cstdio header in util.h

128 * fixed Properties in MR_CLAMPING_* and MR_EXACT_*

129 * added description of the factor graph fileformat

130 * improved Makefile

133 libDAI-0.2.0 (2006-11-30)

134 -------------------------

136 First public release.

139 0.1.5 (2006-11-30)

140 ------------------

142 Regressions

144 - tests/testlcbp and tests/testlcbp are broken.

145 - EXACT method does not work anymore.

146 - The Properties framework gives a speed penalty because of the lookup

147 costs involved; inner loops should be optimized.

149 General framework

151 - DAIAlg is now a template class; typedefs for DAIAlg<FactorGraph> and for

152 DAIAlg<RegionGraph> are provided. In this way, we do not have to write "wrapper"

153 functions to forward functionality from either FactorGraph or RegionGraph

154 to DAIAlg. Functionality like clamping can be implemented in FactorGraph

155 and in RegionGraph and no explicit interface is needed in descendants.

156 - New abstract base class InfAlg added, representing an inference algorithm,

157 from which DAIAlg<T> inherits. This solves the incompatibility problems of

158 DAIAlg<T> for different T (e.g. DAIAlg<FactorGraph> was incompatible with

159 DAIAlg<RegionGraph>). More work is required to reduce code duplication

160 (make FactorGraph part of InfAlg).

161 - Added generic interface (nrVars(), Vars(), nrFactors(), factor(size_t),

162 beliefs(), belief(VarSet &), ...) to InfAlg and descendants.

163 - Added a saveProbs/undoProbs interface to InfAlg and descendants that enables

164 one to save a few factors, modify them (e.g. clamp them), and then restore them

165 to their old values. Undo should also init the corresponding messages / beliefs.

166 This can be used if a given factor graph repeatedly needs to be clamped in

167 different ways and an approximation method is run for each clamping; using the

168 saveProbs/undoProbs can give a significant speed increase.

169 - Switched to a general Properties framework that handles the parameters of

170 all inference methods in a uniform manner. The Properties class is a map of

171 several properties in boost::any objects, indexed by their names (strings).

172 It can read from a stream and write to a stream. It is recursive, in the sense

173 that a Properties object can hold a variable of type Properties as well.

174 - Added a generic way of constructing inference algorithms from a factor graph,

175 name and properties object. Added the newInfAlg function which constructs

176 the requested object. This is used by LCBP, the Matlab interface and the

177 command line (test) interface.

178 - Added a generic enum framework for enum parameters. Although implemented as a

179 hack, it is quite useful in that it drastically simplifies and reduces the

180 amount of code for handling enum parameters.

181 - Provided generic functions for calculating marginals in different ways that

182 work for all approximate inference methods.

184 Bugfixes

186 - Fixed GBP free energy.

187 - Fixed bug in junctiontree (it didn't init the _vars variable).

188 - Corrected two bugs in operator&& and operator|| in VarSet (they returned

189 the logical NOT of what they should return).

190 - Fixed bug in RegionGraph::RecomputeOR(s).

191 - Fixed bug in utils/create_dreg_fg:

192 graph structure was not random for given parameters (forgot to call srand()).

193 - TreeEP bug workaround: use the complete junction tree instead of a subtree.

194 - Fixed bug in JTree::HUGIN() and JTree:ShaferShenoy() in case of junction tree

195 that consists of one outer region only.

196 - Fixed INIT bug in LCBP2::UpdatePancake().

197 - Fixed MaxDiffs flow (except for MR).

199 New functionality

201 - HAK supports several default cluster choices:

202 minimal (only factors)

203 delta (Markov blankets)

204 loop (all loops consisting of loops consisting of <loopdepth> or less variables)

205 Only the maximal clusters are used as outer clusters.

206 - Implemented TreeEP. It generalizes the heuristic method described in the

207 Minka & Qi paper for obtaining a tree with the most relevant interactions to

208 higher order interactions. Almost all optimizations described in the Minka & Qi

209 paper are used, except that evidence is passed over the whole tree instead of

210 relevant subsets (the latter is almost implemented but buggy). Also added

211 alternative (worst-case) algorithm that uses a maximum spanning tree on the

212 weighted graph where the weight between neighbours i and j is given by

213 N(psi,i,j), where psi is the product of all the factors involving both i and j

214 (which is an upper bound on the effective interaction between i and j).

215 - Implemented MR (MontanariRizzo) based on Bastian's code, but extended it

216 to be able to handle connectivities larger than 3 (in principle, up to 32).

217 It supports different initialization methods (the original RESPPROP,

218 the CLAMPING method and EXACT which uses JTree) and different update methods

219 (FULL and LINEAR).

220 - Implemented LCBP2, an analogon of LCBP which represents pancakes as little

221 networks and uses some approximate inference method on them for calculating

222 marginals.

223 - Now there are several LCBP variants (LCBP, LCBPI, LCBPJ, LCBPK, LCBPL);

224 LCBPJ works only for pairwise, LCBPK is LCBP improved for higher order

225 interactions and LCBPL is LCBPI improved for higher-order interactions.

226 - Wrote one single program utils/createfg for creating various types of

227 random factor graphs.

228 - Wrote utility to visualize factor graphs using graphviz.

229 (it uses the BOOST Program Options library)

230 - Added fginfo utility that displays some info about a .fg file.

231 - Implemented Factor::strength function that calculates the potential strength

232 N(psi,i,j) between variables i and j as described in cs.IT:0504030

233 - Wrote a general MatLab interface matlab/dai (similar to tests/test);

234 this unified the matlab functions dai, dai_bp, dai_mf, dai_jt, dai_tep, dai_cvm.

235 - Added MATLAB routine that returns contraction matrix A for BP convergence analysis.

236 - Implemented a MATLAB interface ai_potstrength for Factor::strength

237 - Added Martijn's x2x

239 Improvements of existing code

241 - Reimplemented RegionGraph and descendants: a RegionGraph ISA FactorGraph

242 and also a BipartiteGraph<FRegion,Region>. It now also keeps a map that

243 associates outer region indices to factor indices (no powers yet, this

244 is deemed superfluous) and provides functions to recompute (some of) the

245 outer regions from the factors.

246 - InfAlg descendants run() methods now stop immediately and return NAN in case

247 they detect NANs. Only BP does not do NAN checking for performance reasons.

248 - LCBP now works with factors containing zeroes (by defining x/0 = 0).

249 - HAK, GBP and DoubleLoop now conform to the standards for verbose reporting,

250 timing and convergence criteria.

251 - Implemented logZ() for JTree. It does the calculation during message-passing.

252 - Marginal2ndO now optionally divides by the single node beliefs (to the power n-2);

253 hopefully this will give more accurate approximations.

254 - Marginal and Marginal2ndO (optionally) use the new saveProbs/undoProbs functionality

255 for a faster way of calculating marginals, which does not require a call to init()

256 nor cloning the whole object for each clamping. This leads to a significant speedup.

257 - LCBP (and LCBP2) now have complete flexibility in the specification of the

258 inner method, i.e. the method used to generate the initial cavity approximations.

259 One can pass two strings, a name and a properties string, and LCBP simply invokes

260 newInfAlg to construct the corresponding inference algorithm and uses the generic

261 marginal functions to approximate cavity marginals.

262 - Replaced the global "method" variable by local properties and removed ai.h

263 - Added some methods to Factor (operators *, *=, /, /= with doubles as

264 second argument, operators -, +=, -= with other Factors as second

265 arguments, randomize(), RemoveFirstOrderInteractions) and similar

266 operations to Prob

267 - Moving towards boost::program_options for handling command line arguments

268 (tests/test is done).

269 - Renamed some FactorGraph methods:

270 nr_vars -> nrVars

271 nr_factors -> nrFactors

272 varind -> findVar

273 factorind -> findFactor

274 makeFacCavity -> makeFactorCavity

275 - LCBP_SEQMAXRES has been removed because it did strange things.

276 - Implemented RandomDRegularGraph

277 - Implemented JTree::calcMarginal for marginals that are not confined

278 within one cluster (using cut-set conditioning).

279 - Added isConnected() method to FactorGraph (some methods do not work with

280 disconnected factor graphs).

281 - Pair beliefs are now calculated in a symmetrical way by calcPairBeliefs

282 - Removed single node interaction "correction" code from clamping methods

283 - Removed calcCavityDist and calcCavityDist2ndO

284 - No longer depends on GSL.

285 - Increased portability by combining platform dependant utility functions

286 in util.{h,cpp}.

287 - Wrote *.m files providing help

289 Testing framework

291 - Made a new and significantly improved testing framework that provides most

292 functionality from the command line.

293 - The basis is provided by tests/test, which uses the newInfAlg functionality

294 and enables the user to easily compare from the command line different

295 inference methods on a given factorgraph. All parameters can be specified.

296 Output consists of CPU time, average and maximum single variable marginal

297 errors, relative logZ error and MaxDiff().

298 - tests/aliases.conf contains aliases for standard combinations of methods

299 and their options (which can be used in tests/test).

300 - tests/large contains several bash/python scripts that create random factor

301 graphs, compare several approximate inference algorithms (using tests/test) and

302 allow for easy visualization of the results using PyX.

303 - Added several .fg files for test purposes to /tests (e.g. two ALARM versions

304 alarm.fg and alarm_bnt.fg; testfast.fg, a 4x4 periodic Ising grid for

305 regression testing).

306 - Added a regression test to the Makefile which is included in the standard

307 target. It compares all inference methods on tests/testfast.fg with the

308 results stored in tests/testfast.out

310 Miscellaneous

312 - Expanded all tabs to spaces (":set tabstop 4\n:set expandtab\n:retab" in vim)

313 - Experimental MATLAB code added for StarEP approximation on cavity

314 - Renamed project to libDAI and changed directory name accordingly.

315 - Renamed JunctionTree to JTree.

316 - Fixed licensing (now it's officially GPL).

317 - Improved README

320 revision 252

321 ------------

323 Functionality

324 - Added RegionGraph, GBP, CVM and HAK (double-loop).

325 - Added JunctionTree (with two update algorithms, HUGIN and Shafer-Shenoy), which is a

326 RegionGraph.

327 - NormType is now chosen automatically (in case of negative factors, Prob::NORMLINF is used,

328 otherwise the default Prob::NORMPROB is used). Also, in case of negative factors, the

329 RegionGraph constructors assign each Factor to a unique outer region instead of dividing

330 it over all subsuming outer regions. See README for when negative factors are known to work

331 and when not.

332 - FactorGraph::FactorGraph(const vector<Factor>) only gives a warning in case of short loops,

333 it does not automatically merge factors anymore.

334 - Removed BP_SEQMAXRESNOCLEAR (all cavity initialization methods now are implicitly NOCLEAR)

335 - Added MATLAB interface functions ai_readfg, ai_removeshortloops and ai_bp

336 - Added LCBP-III type that should be equivalent to LCBP-II, but can handle zeroes

337 in potentials. Note that it is significantly slower than LCBP-II (and has to be reimplemented

338 such that it does not store the complete pancakes, but represents them as little factor graphs).

340 Implementation / code

341 - Made a seperate type WeightedGraph, which until now only implements Prim's

342 maximal spanning tree algorithm and is only used by the junction tree code. It might

343 be good to make it a class somewhere in the future and extend it's interface.

344 - Made a seperate class ClusterGraph, which is only used by the junction tree

345 code. It's main purpose is a graph-theoretical variable elimination algorithm.

346 - Implemented the heuristic "minimum-new-edges-in-clique-graph" for variable elimination.

347 - Massive code cleanup, moving towards "generic" programming style, using

348 multiple inheritance and polymorphism.

349 o BP, LCBP, MF, HAK and JunctionTree now inherit from a common DAIAlg class

350 o Made generic functions Marginal, Marginal2ndO, calcCavityDist, calcCavityDist2ndO, clamp

351 that can be used by FactorGraph-based DAIAlgs.

352 o Created TProb<T> class, which stores a probability vector (without the accompanying indexing

353 and VarSet) and provides functionality for it (which is used by TFactor<T>).

354 o Rewrote the VarSet class. It now caches its statespace(). It now privately inherits from set<Var>.

355 I had to overload the insert methods of set<Var> so that they calculate the new statespace.

356 o Rewrote the TFactor class. The TFactor class now HAS a TProb and HAS a VarSet.

357 - Rewrote BP to use the new TProb<T> interface. Performance of BP improved up to a factor 6 by:

358 o using Prob's instead of Factor's;

359 o splitting the multiplication of the messages into two parts (thanks to Vicenc!);

360 o optimizing the calculation of the beliefs (only the message calculations were optimized till now).

361 o replacing FactorGraph::_nb1 and _nb2 (which were set<size_t>) by vector<size_t>

362 - LCBP now seperately stores cavitydists and pancakes. Added InitPancakes() method

363 that takes the cavitydists and multiplies them with the relevant factors. This

364 resulted in an API change in AI which now accepts and returns initial cavitydists

365 instead of initial pancakes.

367 Minor changes

368 - Started writing DoxyGen documentation

369 - Renamed lcptab2fg.m matlab/ai_writefg.m

370 - Moved all matlab stuff to matlab/

371 - More detailed reporting (also reports clocks used).

372 - Marginal and Marginal2ndO now use *differences* in logZ to avoid NaNs.

373 - Improved testing suite.

374 - Removed logreal support.

375 - FactorGraph now also supports input streams and ignores comment lines in .fg files.

376 - Added tests/create_full_fg.cpp and tests/create_ising_fg.cpp which create

377 full and periodic 2D Ising networks according to some command line parameters.

378 - Now logZ really returns logZ instead of -logZ.

379 - Added FactorGraph::WriteToDotFile

382 0.1.4 (2006-04-13)

383 ------------------

384 - Added file IO routines to read and write factorgraphs.

385 - Added L-infinity normalization for use with (partially) negative factors.

386 - Renamed BetheF, MFF to logZ, which now use complex numbers to be able to

387 handle (partially) negative factors.

388 - Added test suite.

389 - All probabilities are now represented using double instead of LogReal<double>.

390 - Combined Alg and Alg3 into LCBP. Several update schemes possible.

391 - Combined several variants of BP into doBP. Several update schemes possible.

392 Now uses precalculated indices for optimization.

393 - Renamed Node -> Var and Potential -> Factor.

394 - Extensive code cleanup. More use of OO techniques. Changed API.

395 - MaxIter now means maximum number of passes (corresponding to number of

396 _parallel_ updates).

397 - MaxDiff now means the maximum L-infinity distance between the updated and

398 original single variable beliefs, for all AI methods.

399 - Implemented RBP which converges faster than BP for difficult problems.

400 - Now uses method parameter which is a bitmask combining outmethod and inmethod

401 (see ai.h).

404 0.1.3 (2006-03-22)

405 --------------------

406 - All AI methods now return maxdiff

407 - ai.cpp:

408 o Now returns maxdiffinner and maxdiffouter

409 o New BP2ndO innermethod (estimate only 2nd order cavity interactions)

410 o New InitCav outermethod (only create initial cavity distributions)

411 - bp.cpp:

412 o New CavityDist2ndO which estimates 2nd order cavity interactions

413 - Makefile:

414 o Bugfix: removed dependencies on algwim.*

417 0.1.2 (2006-02-28)

418 --------------------

419 - Cleaned up alg.cpp (removed Alg2 and its corresponding data structures).

420 - Added the possibility to provide initial cavity distributions as an input

421 argument to ai (not much error checking is done, so be careful).

422 - Potentials2mx now correctly sets the dimensions of the P field (i.e. for

423 the output arguments Q, Q0 of ai).

424 - Removed algwim.* since it does not work.

427 0.1.1 (2006-02-28)

428 --------------------

429 - The constructors of (Log)FactorGraph and LogFactorGraph from a

430 vector<(Log)Potential> now merge potentials to prevent short loops (of length

431 4) in the factor graph. These are used in ai to construct the factor graphs

432 from the psi argument. If compiled with DEBUG defined, the method calc_nb()

433 of BipGraph checks for the existence of short loops.

434 - Changed calling syntax of ai (now the actual syntax *does* correspond to its

435 description in the help).

436 - ai does not hook cout anymore (which caused weird segfaults).

437 - Fixed a bug in an assert statement in the matlab interface code in ai.cpp.

438 - Removed network.* since it is not useful.

441 0.1.0 (2006-02-28)

442 --------------------

443 First version worthy a version number.