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

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

4 * Now compiles also under Visual C++.

5 * Added Exceptions framework.

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

7 * Removed utils/remove_short_loops and matlab/remove_short_loops.

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

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

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

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

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

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

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

15 made easy by using boost::foreach.

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

17 - Renamed Index -> IndexFor

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

19 25% speedup of testregression

20 - Replaced multind by Permute

21 - Added MultiFor

22 - Added State

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

24 * Moved everything into namespace "dai"

25 * Added ExactInf class for brute force exact inference

26 * Renamed DEBUG to DAI_DEBUG to avoid conflicts

27 * Added conditional compilation of inference methods

28 * Contributions by Peter Gober:

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

30 * Contributions by Giuseppe Passino:

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

32 - moved header files in include/dai and sources in src

33 - changed #ifndefs to GNU style

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

35 - dai::TProb:

36 o removed copy constructor and assignment operators (redundant)

37 o implementation of some methods via STL algorithms

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

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

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

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

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

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

44 of a double is too small.

45 * Contributions by Christian Wojek:

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

47 and variables

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

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

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

51 normtype is now Prob::NORMPROB by default.

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

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

54 * Improved MaxSpanningTreePrims algorithm (uses boost::graph)

55 * Small optimization in Diffs

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

57 used enough to warrant the additional "complexity" :))

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

59 inference algorithm, because the Properties framework was not as

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

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

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

63 entangled) code.

64 * Improved ClusterGraph implementation, yielding significant speedups

65 for the JunctionTree algorithm on large factorgraphs.

66 * Improved documetation

67 * Interface changes:

68 - VarSet::

69 stateSpace() -> states()

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

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

72 removed bool operator||

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

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

75 - FactorGraph::

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

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

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

79 vars() -> vars

80 factors() -> factors

81 removed MakeFactorCavity(size_t)

82 removed ExactMarginal(const VarSet &)

83 removed ExactlogZ()

84 removed updatedFactor(size_t)

85 removed _normtype and NormType()

86 removed hasShortLoops(...) and removeShortLoops(...)

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

88 undoProb(size_t) -> restoreFactor(size_t)

89 saveProb(size_t) -> backupFactor(size_t)

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

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

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

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

94 removed hasNegatives()

95 - RegionGraph::

96 nr_ORs() -> nrORs()

97 nr_IRs() -> nrIRs()

98 ORs() -> ORs

99 IRs() -> IRs

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

101 - Renamed Index -> IndexFor

102 - Diffs:

103 max() -> maxDiff()

104 max_size() -> maxSize()

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

106 - Factor::

107 max() -> maxVal()

108 part_sum() -> partSum()

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

110 - VarSet::operator&&

111 - Properties -> PropertySet

114 libDAI-0.2.1 (2008-05-26)

115 -------------------------

117 Bugfix release.

118 * added missing cstdio header in util.h

119 * fixed Properties in MR_CLAMPING_* and MR_EXACT_*

120 * added description of the factor graph fileformat

121 * improved Makefile

124 libDAI-0.2.0 (2006-11-30)

125 -------------------------

127 First public release.

130 0.1.5 (2006-11-30)

131 ------------------

133 Regressions

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

136 - EXACT method does not work anymore.

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

138 costs involved; inner loops should be optimized.

140 General framework

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

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

144 functions to forward functionality from either FactorGraph or RegionGraph

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

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

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

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

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

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

151 (make FactorGraph part of InfAlg).

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

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

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

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

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

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

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

159 saveProbs/undoProbs can give a significant speed increase.

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

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

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

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

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

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

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

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

168 command line (test) interface.

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

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

171 amount of code for handling enum parameters.

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

173 work for all approximate inference methods.

175 Bugfixes

177 - Fixed GBP free energy.

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

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

180 the logical NOT of what they should return).

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

182 - Fixed bug in utils/create_dreg_fg:

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

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

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

186 that consists of one outer region only.

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

188 - Fixed MaxDiffs flow (except for MR).

190 New functionality

192 - HAK supports several default cluster choices:

193 minimal (only factors)

194 delta (Markov blankets)

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

196 Only the maximal clusters are used as outer clusters.

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

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

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

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

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

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

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

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

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

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

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

208 It supports different initialization methods (the original RESPPROP,

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

210 (FULL and LINEAR).

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

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

213 marginals.

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

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

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

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

218 random factor graphs.

219 - Wrote utility to visualize factor graphs using graphviz.

220 (it uses the BOOST Program Options library)

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

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

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

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

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

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

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

228 - Added Martijn's x2x

230 Improvements of existing code

232 - Reimplemented RegionGraph and descendants: a RegionGraph ISA FactorGraph

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

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

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

236 outer regions from the factors.

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

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

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

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

241 timing and convergence criteria.

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

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

244 hopefully this will give more accurate approximations.

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

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

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

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

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

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

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

252 marginal functions to approximate cavity marginals.

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

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

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

256 arguments, randomize(), RemoveFirstOrderInteractions) and similar

257 operations to Prob

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

259 (tests/test is done).

260 - Renamed some FactorGraph methods:

261 nr_vars -> nrVars

262 nr_factors -> nrFactors

263 varind -> findVar

264 factorind -> findFactor

265 makeFacCavity -> makeFactorCavity

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

267 - Implemented RandomDRegularGraph

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

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

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

271 disconnected factor graphs).

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

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

274 - Removed calcCavityDist and calcCavityDist2ndO

275 - No longer depends on GSL.

276 - Increased portability by combining platform dependant utility functions

277 in util.{h,cpp}.

278 - Wrote *.m files providing help

280 Testing framework

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

283 functionality from the command line.

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

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

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

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

288 errors, relative logZ error and MaxDiff().

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

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

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

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

293 allow for easy visualization of the results using PyX.

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

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

296 regression testing).

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

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

299 results stored in tests/testfast.out

301 Miscellaneous

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

304 - Experimental MATLAB code added for StarEP approximation on cavity

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

306 - Renamed JunctionTree to JTree.

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

308 - Improved README

311 revision 252

312 ------------

314 Functionality

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

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

317 RegionGraph.

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

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

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

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

322 and when not.

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

324 it does not automatically merge factors anymore.

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

326 - Added MATLAB interface functions ai_readfg, ai_removeshortloops and ai_bp

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

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

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

331 Implementation / code

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

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

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

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

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

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

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

339 multiple inheritance and polymorphism.

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

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

342 that can be used by FactorGraph-based DAIAlgs.

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

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

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

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

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

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

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

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

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

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

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

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

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

356 instead of initial pancakes.

358 Minor changes

359 - Started writing DoxyGen documentation

360 - Renamed lcptab2fg.m matlab/ai_writefg.m

361 - Moved all matlab stuff to matlab/

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

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

364 - Improved testing suite.

365 - Removed logreal support.

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

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

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

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

370 - Added FactorGraph::WriteToDotFile

373 0.1.4 (2006-04-13)

374 ------------------

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

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

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

378 handle (partially) negative factors.

379 - Added test suite.

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

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

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

383 Now uses precalculated indices for optimization.

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

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

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

387 _parallel_ updates).

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

389 original single variable beliefs, for all AI methods.

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

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

392 (see ai.h).

395 0.1.3 (2006-03-22)

396 --------------------

397 - All AI methods now return maxdiff

398 - ai.cpp:

399 o Now returns maxdiffinner and maxdiffouter

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

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

402 - bp.cpp:

403 o New CavityDist2ndO which estimates 2nd order cavity interactions

404 - Makefile:

405 o Bugfix: removed dependencies on algwim.*

408 0.1.2 (2006-02-28)

409 --------------------

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

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

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

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

414 the output arguments Q, Q0 of ai).

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

418 0.1.1 (2006-02-28)

419 --------------------

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

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

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

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

424 of BipGraph checks for the existence of short loops.

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

426 description in the help).

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

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

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

432 0.1.0 (2006-02-28)

433 --------------------

434 First version worthy a version number.