Mercurial > hg > camir-aes2014

annotate toolboxes/FullBNT-1.0.7/docs/changelog.html @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 <title>History of changes to BNT</title>
wolffd@0 2 <h1>History of changes to BNT</h1>
wolffd@0 3
wolffd@0 4
wolffd@0 5 <h2>Changes since 4 Oct 2007</h2>
wolffd@0 6
wolffd@0 7 <pre>
wolffd@0 8 - 19 Oct 07 murphyk
wolffd@0 9
wolffd@0 10 * BNT\CPDs\@noisyor_CPD\CPD_to_CPT.m: 2nd half of the file is a repeat
wolffd@0 11 of the first half and was deleted (thanks to Karl Kuschner)
wolffd@0 12
wolffd@0 13 * KPMtools\myismember.m should return logical for use in "assert" so add line at end
wolffd@0 14 p=logical(p); this prevents "assert" from failing on an integer input.
wolffd@0 15 (thanks to Karl Kuschner)
wolffd@0 16
wolffd@0 17
wolffd@0 18
wolffd@0 19 - 17 Oct 07 murphyk
wolffd@0 20
wolffd@0 21 * Updated subv2ind and ind2subv in KPMtools to Tom Minka's implementation.
wolffd@0 22 His ind2subv is faster (vectorized), but I had to modify it so it
wolffd@0 23 matched the behavior of my version when called with siz=[].
wolffd@0 24 His subv2inv is slightly simpler than mine because he does not treat
wolffd@0 25 the siz=[2 2 ... 2] case separately.
wolffd@0 26 Note: there is now no need to ever use the C versions of these
wolffd@0 27 functions (or any others, for that matter).
wolffd@0 28
wolffd@0 29 * removed BNT/add_BNT_to_path since no longer needed.
wolffd@0 30
wolffd@0 31
wolffd@0 32
wolffd@0 33 - 4 Oct 07 murphyk
wolffd@0 34
wolffd@0 35 * moved code from sourceforge to UBC website, made version 1.0.4
wolffd@0 36
wolffd@0 37 * @pearl_inf_engine/pearl_inf_engine line 24, default
wolffd@0 38 argument for protocol changed from [] to 'parallel'.
wolffd@0 39 Also, changed private/parallel_protocol so it doesn't write to an
wolffd@0 40 empty file id (Matlab 7 issue)
wolffd@0 41
wolffd@0 42 * added foptions (Matlab 7 issue)
wolffd@0 43
wolffd@0 44 * changed genpathKPM to exclude svn. Put it in toplevel directory to
wolffd@0 45 massively simplify the installation process.
wolffd@0 46
wolffd@0 47 </pre>
wolffd@0 48
wolffd@0 49
wolffd@0 50 <h2>Sourceforge changelog</h2>
wolffd@0 51
wolffd@0 52 BNT was first ported to sourceforge on 28 July 2001 by yozhik.
wolffd@0 53 BNT was removed from sourceforge on 4 October 2007 by Kevin Murphy;
wolffd@0 54 that version is cached as <a
wolffd@0 55 href="FullBNT-1.0.3.zip">FullBNT-1.0.3.zip</a>.
wolffd@0 56 See <a href="ChangeLog.Sourceforge.txt">Changelog from
wolffd@0 57 sourceforge</a> for a history of that version of the code,
wolffd@0 58 which formed the basis of the branch currently on Murphy's web page.
wolffd@0 59
wolffd@0 60
wolffd@0 61 <h2> Changes from August 1998 -- July 2004</h2>
wolffd@0 62
wolffd@0 63 Kevin Murphy made the following changes to his own private copy.
wolffd@0 64 (Other small changes were made between July 2004 and October 2007, but were
wolffd@0 65 not documented.)
wolffd@0 66 These may or may not be reflected in the sourceforge version of the
wolffd@0 67 code (which was independently maintained).
wolffd@0 68
wolffd@0 69
wolffd@0 70 <ul>
wolffd@0 71 <li> 9 June 2004
wolffd@0 72 <ul>
wolffd@0 73 <li> Changed tabular_CPD/learn_params back to old syntax, to make it
wolffd@0 74 compatible with gaussian_CPD/learn_params (and re-enabled
wolffd@0 75 generic_CPD/learn_params).
wolffd@0 76 Modified learning/learn_params.m and learning/score_family
wolffd@0 77 appropriately.
wolffd@0 78 (In particular, I undid the change Sonia Leach had to make to
wolffd@0 79 score_family to handle this asymmetry.)
wolffd@0 80 Added examples/static/gaussian2 to test this new functionality.
wolffd@0 81
wolffd@0 82 <li> Added bp_mrf2 (for generic pairwise MRFs) to
wolffd@0 83 inference/static/@bp_belprop_mrf2_inf_engine. [MRFs are not
wolffd@0 84 "officially" supported in BNT, so this code is just for expert
wolffd@0 85 hackers.]
wolffd@0 86
wolffd@0 87 <li> Added examples/static/nodeorderExample.m to illustrate importance
wolffd@0 88 of using topological ordering.
wolffd@0 89
wolffd@0 90 <li> Ran dos2unix on all *.c files within BNT to eliminate compiler
wolffd@0 91 warnings.
wolffd@0 92
wolffd@0 93 </ul>
wolffd@0 94
wolffd@0 95 <li> 7 June 2004
wolffd@0 96 <ul>
wolffd@0 97 <li> Replaced normaliseC with normalise in HMM/fwdback, for maximum
wolffd@0 98 portability (and negligible loss in speed).
wolffd@0 99 <li> Ensured FullBNT versions of HMM, KPMstats etc were as up-to-date
wolffd@0 100 as stand-alone versions.
wolffd@0 101 <li> Changed add_BNT_to_path so it no longer uses addpath(genpath()),
wolffd@0 102 which caused Old versions of files to mask new ones.
wolffd@0 103 </ul>
wolffd@0 104
wolffd@0 105 <li> 18 February 2004
wolffd@0 106 <ul>
wolffd@0 107 <li> A few small bug fixes to BNT, as posted to the Yahoo group.
wolffd@0 108 <li> Several new functions added to KPMtools, KPMstats and Graphviz
wolffd@0 109 (none needed by BNT).
wolffd@0 110 <li> Added CVS to some of my toolboxes.
wolffd@0 111 </ul>
wolffd@0 112
wolffd@0 113 <li> 30 July 2003
wolffd@0 114 <ul>
wolffd@0 115 <li> qian.diao fixed @mpot/set_domain_pot and @cgpot/set_domain_pot
wolffd@0 116 <li> Marco Grzegorczyk found, and Sonia Leach fixed, a bug in
wolffd@0 117 do_removal inside learn_struct_mcmc
wolffd@0 118 </ul>
wolffd@0 119
wolffd@0 120
wolffd@0 121 <li> 28 July 2003
wolffd@0 122 <ul>
wolffd@0 123 <li> Sebastian Luehr provided 2 minor bug fixes, to HMM/fwdback (if any(scale==0))
wolffd@0 124 and FullBNT\HMM\CPDs\@hhmmQ_CPD\update_ess.m (wrong transpose).
wolffd@0 125 </ul>
wolffd@0 126
wolffd@0 127 <li> 8 July 2003
wolffd@0 128 <ul>
wolffd@0 129 <li> Removed buggy BNT/examples/static/MRF2/Old/mk_2D_lattice.m which was
wolffd@0 130 masking correct graph/mk_2D_lattice.
wolffd@0 131 <li> Fixed bug in graph/mk_2D_lattice_slow in the non-wrap-around case
wolffd@0 132 (line 78)
wolffd@0 133 </ul>
wolffd@0 134
wolffd@0 135
wolffd@0 136 <li> 2 July 2003
wolffd@0 137 <ul>
wolffd@0 138 <li> Sped up normalize(., 1) in KPMtools by avoiding general repmat
wolffd@0 139 <li> Added assign_cols and marginalize_table to KPMtools
wolffd@0 140 </ul>
wolffd@0 141
wolffd@0 142
wolffd@0 143 <li> 29 May 2003
wolffd@0 144 <ul>
wolffd@0 145 <li> Modified KPMstats/mixgauss_Mstep so it repmats Sigma in the tied
wolffd@0 146 covariance case (bug found by galt@media.mit.edu).
wolffd@0 147
wolffd@0 148 <li> Bob Welch found bug in gaussian_CPDs/maximize_params in the way
wolffd@0 149 cpsz was computed.
wolffd@0 150
wolffd@0 151 <li> Added KPMstats/mixgauss_em, because my code is easier to
wolffd@0 152 understand/modify than netlab's (at least for me!).
wolffd@0 153
wolffd@0 154 <li> Modified BNT/examples/dynamic/viterbi1 to call multinomial_prob
wolffd@0 155 instead of mk_dhmm_obs_lik.
wolffd@0 156
wolffd@0 157 <li> Moved parzen window and partitioned models code to KPMstats.
wolffd@0 158
wolffd@0 159 <li> Rainer Deventer fixed some bugs in his scgpot code, as follows:
wolffd@0 160 1. complement_pot.m
wolffd@0 161 Problems occured for probabilities equal to zero. The result is an
wolffd@0 162 division by zero error.
wolffd@0 163 <br>
wolffd@0 164 2. normalize_pot.m
wolffd@0 165 This function is used during the calculation of the log-likelihood.
wolffd@0 166 For a probability of zero a warning "log of zero" occurs. I have not
wolffd@0 167 realy fixed the bug. As a workaround I suggest to calculate the
wolffd@0 168 likelihhod based on realmin (the smallest real number) instead of
wolffd@0 169 zero.
wolffd@0 170 <br>
wolffd@0 171 3. recursive_combine_pots
wolffd@0 172 At the beginning of the function there was no test for the trivial case,
wolffd@0 173 which defines the combination of two potentials as equal to the direct
wolffd@0 174 combination. The result might be an infinite recursion which leads to
wolffd@0 175 a stack overflow in matlab.
wolffd@0 176 </ul>
wolffd@0 177
wolffd@0 178
wolffd@0 179
wolffd@0 180 <li> 11 May 2003
wolffd@0 181 <ul>
wolffd@0 182 <li> Fixed bug in gaussian_CPD/maximize_params so it is compatible
wolffd@0 183 with the new clg_Mstep routine
wolffd@0 184 <li> Modified KPMstats/cwr_em to handle single cluster case
wolffd@0 185 separately.
wolffd@0 186 <li> Fixed bug in netlab/gmminit.
wolffd@0 187 <li> Added hash tables to KPMtools.
wolffd@0 188 </ul>
wolffd@0 189
wolffd@0 190
wolffd@0 191 <li> 4 May 2003
wolffd@0 192 <ul>
wolffd@0 193 <li>
wolffd@0 194 Renamed many functions in KPMstats so the name of the
wolffd@0 195 distribution/model type comes first,
wolffd@0 196 Mstep_clg -> clg_Mstep,
wolffd@0 197 Mstep_cond_gauss -> mixgauss_Mstep.
wolffd@0 198 Also, renamed eval_pdf_xxx functions to xxx_prob, e.g.
wolffd@0 199 eval_pdf_cond_mixgauss -> mixgauss_prob.
wolffd@0 200 This is simpler and shorter.
wolffd@0 201
wolffd@0 202 <li>
wolffd@0 203 Renamed many functions in HMM toolbox so the name of the
wolffd@0 204 distribution/model type comes first,
wolffd@0 205 log_lik_mhmm -> mhmm_logprob, etc.
wolffd@0 206 mk_arhmm_obs_lik has finally been re-implemented in terms of clg_prob
wolffd@0 207 and mixgauss_prob (for slice 1).
wolffd@0 208 Removed the Demos directory, and put them in the main directory.
wolffd@0 209 This code is not backwards compatible.
wolffd@0 210
wolffd@0 211 <li> Removed some of the my_xxx functions from KPMstats (these were
wolffd@0 212 mostly copies of functions from the Mathworks stats toolbox).
wolffd@0 213
wolffd@0 214
wolffd@0 215 <li> Modified BNT to take into account changes to KPMstats and
wolffd@0 216 HMM toolboxes.
wolffd@0 217
wolffd@0 218 <li> Fixed KPMstats/Mstep_clg (now called clg_Mstep) for spherical Gaussian case.
wolffd@0 219 (Trace was wrongly parenthesised, and I used YY instead of YTY.
wolffd@0 220 The spherical case now gives the same result as the full case
wolffd@0 221 for cwr_demo.)
wolffd@0 222 Also, mixgauss_Mstep now adds 0.01 to the ML estimate of Sigma,
wolffd@0 223 to act as a regularizer (it used to add 0.01 to E[YY'], but this was
wolffd@0 224 ignored in the spherical case).
wolffd@0 225
wolffd@0 226 <li> Added cluster weighted regression to KPMstats.
wolffd@0 227
wolffd@0 228 <li> Added KPMtools/strmatch_substr.
wolffd@0 229 </ul>
wolffd@0 230
wolffd@0 231
wolffd@0 232
wolffd@0 233 <li> 28 Mar 03
wolffd@0 234 <ul>
wolffd@0 235 <li> Added mc_stat_distrib and eval_pdf_cond_prod_parzen to KPMstats
wolffd@0 236 <li> Fixed GraphViz/arrow.m incompatibility with matlab 6.5
wolffd@0 237 (replace all NaN's with 0).
wolffd@0 238 Modified GraphViz/graph_to_dot so it also works on windows.
wolffd@0 239 <li> I removed dag_to_jtree and added graph_to_jtree to the graph
wolffd@0 240 toolbox; the latter expects an undirected graph as input.
wolffd@0 241 <li> I added triangulate_2Dlattice_demo.m to graph.
wolffd@0 242 <li> Rainer Deventer fixed the stable conditional Gaussian potential
wolffd@0 243 classes (scgpot and scgcpot) and inference engine
wolffd@0 244 (stab_cond_gauss_inf_engine).
wolffd@0 245 <li> Rainer Deventer added (stable) higher-order Markov models (see
wolffd@0 246 inference/dynamic/@stable_ho_inf_engine).
wolffd@0 247 </ul>
wolffd@0 248
wolffd@0 249
wolffd@0 250 <li> 14 Feb 03
wolffd@0 251 <ul>
wolffd@0 252 <li> Simplified learning/learn_params so it no longer returns BIC
wolffd@0 253 score. Also, simplified @tabular_CPD/learn_params so it only takes
wolffd@0 254 local evidence.
wolffd@0 255 Added learn_params_dbn, which does ML estimation of fully observed
wolffd@0 256 DBNs.
wolffd@0 257 <li> Vectorized KPMstats/eval_pdf_cond_mixgauss for tied Sigma
wolffd@0 258 case (much faster!).
wolffd@0 259 Also, now works in log-domain to prevent underflow.
wolffd@0 260 eval_pdf_mixgauss now calls eval_pdf_cond_mixgauss and inherits these benefits.
wolffd@0 261 <li> add_BNT_to_path now calls genpath with 2 arguments if using
wolffd@0 262 matlab version 5.
wolffd@0 263 </ul>
wolffd@0 264
wolffd@0 265
wolffd@0 266 <li> 30 Jan 03
wolffd@0 267 <ul>
wolffd@0 268 <li> Vectorized KPMstats/eval_pdf_cond_mixgauss for scalar Sigma
wolffd@0 269 case (much faster!)
wolffd@0 270 <li> Renamed mk_dotfile_from_hmm to draw_hmm and moved it to the
wolffd@0 271 GraphViz library.
wolffd@0 272 <li> Rewrote @gaussian_CPD/maximize_params.m so it calls
wolffd@0 273 KPMstats/Mstep_clg.
wolffd@0 274 This fixes bug when using clamped means (found by Rainer Deventer
wolffd@0 275 and Victor Eruhimov)
wolffd@0 276 and a bug when using a Wishart prior (no gamma term in the denominator).
wolffd@0 277 It is also easier to read.
wolffd@0 278 I rewrote the technical report re-deriving all the equations in a
wolffd@0 279 clearer notation, making the solution to the bugs more obvious.
wolffd@0 280 (See www.ai.mit.edu/~murphyk/Papers/learncg.pdf)
wolffd@0 281 Modified Mstep_cond_gauss to handle priors.
wolffd@0 282 <li> Fixed bug reported by Ramgopal Mettu in which add_BNT_to_path
wolffd@0 283 calls genpath with only 1 argument, whereas version 5 requires 2.
wolffd@0 284 <li> Fixed installC and uninstallC to search in FullBNT/BNT.
wolffd@0 285 </ul>
wolffd@0 286
wolffd@0 287
wolffd@0 288 <li> 24 Jan 03
wolffd@0 289 <ul>
wolffd@0 290 <li> Major simplification of HMM code.
wolffd@0 291 The API is not backwards compatible.
wolffd@0 292 No new functionality has been added, however.
wolffd@0 293 There is now only one fwdback function, instead of 7;
wolffd@0 294 different behaviors are controlled through optional arguments.
wolffd@0 295 I renamed 'evaluate observation likelihood' (local evidence)
wolffd@0 296 to 'evaluate conditional pdf', since this is more general.
wolffd@0 297 i.e., renamed
wolffd@0 298 mk_dhmm_obs_lik to eval_pdf_cond_multinomial,
wolffd@0 299 mk_ghmm_obs_lik to eval_pdf_cond_gauss,
wolffd@0 300 mk_mhmm_obs_lik to eval_pdf_cond_mog.
wolffd@0 301 These functions have been moved to KPMstats,
wolffd@0 302 so they can be used by other toolboxes.
wolffd@0 303 ghmm's have been eliminated, since they are just a special case of
wolffd@0 304 mhmm's with M=1 mixture component.
wolffd@0 305 mixgauss HMMs can now handle a different number of
wolffd@0 306 mixture components per state.
wolffd@0 307 init_mhmm has been eliminated, and replaced with init_cond_mixgauss
wolffd@0 308 (in KPMstats) and mk_leftright/rightleft_transmat.
wolffd@0 309 learn_dhmm can no longer handle inputs (although this is easy to add back).
wolffd@0 310 </ul>
wolffd@0 311
wolffd@0 312
wolffd@0 313
wolffd@0 314
wolffd@0 315
wolffd@0 316 <li> 20 Jan 03
wolffd@0 317 <ul>
wolffd@0 318 <li> Added arrow.m to GraphViz directory, and commented out line 922,
wolffd@0 319 in response to a bug report.
wolffd@0 320 </ul>
wolffd@0 321
wolffd@0 322 <li> 18 Jan 03
wolffd@0 323 <ul>
wolffd@0 324 <li> Major restructuring of BNT file structure:
wolffd@0 325 all code that is not specific to Bayes nets has been removed;
wolffd@0 326 these packages must be downloaded separately. (Or just download FullBNT.)
wolffd@0 327 This makes it easier to ensure different toolboxes are consistent.
wolffd@0 328 misc has been slimmed down and renamed KPMtools, so it can be shared by other toolboxes,
wolffd@0 329 such as HMM and Kalman; some of the code has been moved to BNT/general.
wolffd@0 330 The Graphics directory has been slimmed down and renamed GraphViz.
wolffd@0 331 The graph directory now has no dependence on BNT (dag_to_jtree has
wolffd@0 332 been renamed graph_to_jtree and has a new API).
wolffd@0 333 netlab2 no longer contains any netlab files, only netlab extensions.
wolffd@0 334 None of the functionality has changed.
wolffd@0 335 </ul>
wolffd@0 336
wolffd@0 337
wolffd@0 338
wolffd@0 339 <li> 11 Jan 03
wolffd@0 340 <ul>
wolffd@0 341 <li> jtree_dbn_inf_engine can now support soft evidence.
wolffd@0 342
wolffd@0 343 <li> Rewrote graph/dfs to make it clearer.
wolffd@0 344 Return arguments have changed, as has mk_rooted_tree.
wolffd@0 345 The acyclicity check for large undirected graphs can cause a stack overflow.
wolffd@0 346 It turns out that this was not a bug, but is because Matlab's stack depth
wolffd@0 347 bound is very low by default.
wolffd@0 348
wolffd@0 349 <li> Renamed examples/dynamic/filter2 to filter_test1, so it does not
wolffd@0 350 conflict with the filter2 function in the image processing toolbox.
wolffd@0 351
wolffd@0 352 <li> Ran test_BNT on various versions of matlab to check compatibility.
wolffd@0 353 On matlab 6.5 (r13), elapsed time = 211s, cpu time = 204s.
wolffd@0 354 On matlab 6.1 (r12), elapsed time = 173s, cpu time = 164s.
wolffd@0 355 On matlab 5.3 (r11), elapsed time = 116s, cpu time = 114s.
wolffd@0 356 So matlab is apparently getting slower with time!!
wolffd@0 357 (All results were with a linux PIII machine.)
wolffd@0 358 </ul>
wolffd@0 359
wolffd@0 360
wolffd@0 361 <li> 14 Nov 02
wolffd@0 362 <ul>
wolffd@0 363 <li> Removed all ndx inference routines, since they are only
wolffd@0 364 marginally faster on toy problems,
wolffd@0 365 and are slower on large problems due to having to store and lookup
wolffd@0 366 the indices (causes cache misses).
wolffd@0 367 In particular, I removed jtree_ndx_inf_eng and jtree_ndx_dbn_inf_eng, all the *ndx*
wolffd@0 368 routines from potentials/Tables, and all the UID stuff from
wolffd@0 369 add_BNT_to_path,
wolffd@0 370 thus simplifying the code.
wolffd@0 371 This required fixing hmm_(2TBN)_inf_engine/marginal_nodes\family,
wolffd@0 372 and updating installC.
wolffd@0 373
wolffd@0 374
wolffd@0 375 <li> Removed jtree_C_inf_engine and jtree_C_dbn_inf_engine.
wolffd@0 376 The former is basically the same as using jtree_inf_engine with
wolffd@0 377 mutliply_by_table.c and marginalize_table.c.
wolffd@0 378 The latter benefited slightly by assuming potentials were tables
wolffd@0 379 (arrays not objects), but these negligible savings don't justify the
wolffd@0 380 complexity and code duplication.
wolffd@0 381
wolffd@0 382 <li> Removed stab_cond_gauss_inf_engine and
wolffd@0 383 scg_unrolled_dbn_inf_engine,
wolffd@0 384 written by shan.huang@intel.com, since the code was buggy.
wolffd@0 385
wolffd@0 386 <li> Removed potential_engine, which was only experimental anyway.
wolffd@0 387
wolffd@0 388 </ul>
wolffd@0 389
wolffd@0 390
wolffd@0 391
wolffd@0 392 <li> 13 Nov 02
wolffd@0 393 <ul>
wolffd@0 394 <li> <b>Released version 5</b>.
wolffd@0 395 The previous version, released on 7/28/02, is available
wolffd@0 396 <a href="BNT4.zip">here</a>.
wolffd@0 397
wolffd@0 398 <li> Moved code and documentation to MIT.
wolffd@0 399
wolffd@0 400 <li> Added repmat.c from Thomas Minka's lightspeed library.
wolffd@0 401 Modified it so it can return an empty matrix.
wolffd@0 402
wolffd@0 403 <li> Tomas Kocka fixed bug in the BDeu option for tabular_CPD,
wolffd@0 404 and contributed graph/dag_to_eg, to convert to essential graphs.
wolffd@0 405
wolffd@0 406 <!--<li> Wrote a <a href="../Papers/fastmult.pdf">paper</a> which explains
wolffd@0 407 the ndx methods and the ndx cache BNT uses for fast
wolffd@0 408 multiplication/ marginalization of multi-dimensional arrays.
wolffd@0 409 -->
wolffd@0 410
wolffd@0 411 <li> Modified definition of hhmmQ_CPD, so that Qps can now accept
wolffd@0 412 parents in either the current or previous slice.
wolffd@0 413
wolffd@0 414 <li> Added hhmm2Q_CPD class, which is simpler than hhmmQ (no embedded
wolffd@0 415 sub CPDs, etc), and which allows the conditioning parents, Qps, to
wolffd@0 416 be before (in the topological ordering) the F or Q(t-1) nodes.
wolffd@0 417 See BNT/examples/dynamic/HHMM/Map/mk_map_hhmm for an example.
wolffd@0 418 </ul>
wolffd@0 419
wolffd@0 420
wolffd@0 421 <li> 7/28/02
wolffd@0 422 <ul>
wolffd@0 423 <li> Changed graph/best_first_elim_order from min-fill to min-weight.
wolffd@0 424 <li> Ernest Chan fixed bug in Kalman/sample_lds (G{i} becomes G{m} in
wolffd@0 425 line 61).
wolffd@0 426 <li> Tal Blum <bloom@cs.huji.ac.il> fixed bug in HMM/init_ghmm (Q
wolffd@0 427 becomes K, the number of states).
wolffd@0 428 <li> Fixed jtree_2tbn_inf_engine/set_fields so it correctly sets the
wolffd@0 429 maximize flag to 1 even in subengines.
wolffd@0 430 <li> Gary Bradksi did a simple mod to the PC struct learn alg so you can pass it an
wolffd@0 431 adjacency matrix as a constraint. Also, CovMat.m reads a file and
wolffd@0 432 produces a covariance matrix.
wolffd@0 433 <li> KNOWN BUG in CPDs/@hhmmQ_CPD/update_ess.m at line 72 caused by
wolffd@0 434 examples/dynamic/HHMM/Square/learn_square_hhmm_cts.m at line 57.
wolffd@0 435 <li>
wolffd@0 436 The old version is available from www.cs.berkeley.edu/~murphyk/BNT.24june02.zip
wolffd@0 437 </ul>
wolffd@0 438
wolffd@0 439
wolffd@0 440 <li> 6/24/02
wolffd@0 441 <ul>
wolffd@0 442 <li> Renamed dag_to_dot as graph_to_dot and added support for
wolffd@0 443 undirected graphs.
wolffd@0 444 <li> Changed syntax for HHMM CPD constructors: no need to specify d/D
wolffd@0 445 anymore,so they can be used for more complex models.
wolffd@0 446 <li> Removed redundant first argument to mk_isolated_tabular_CPD.
wolffd@0 447 </ul>
wolffd@0 448
wolffd@0 449
wolffd@0 450 <li> 6/19/02
wolffd@0 451 <ul>
wolffd@0 452 <li>
wolffd@0 453 Fixed most probable explanation code.
wolffd@0 454 Replaced calc_mpe with find_mpe, which is now a method of certain
wolffd@0 455 inference engines, e.g., jtree, belprop.
wolffd@0 456 calc_mpe_global has become the find_mpe method of global_joint.
wolffd@0 457 calc_mpe_bucket has become the find_mpe method of var_elim.
wolffd@0 458 calc_mpe_dbn has become the find_mpe method of smoother.
wolffd@0 459 These routines now correctly find the jointly most probable
wolffd@0 460 explanation, instead of the marginally most probable assignments.
wolffd@0 461 See examples/static/mpe1\mpe2 and examples/dynamic/viterbi1
wolffd@0 462 for examples.
wolffd@0 463 Removed maximize flag from constructor and enter_evidence
wolffd@0 464 methods, since this no longer needs to be specified by the user.
wolffd@0 465
wolffd@0 466 <li> Rainer Deventer fixed in a bug in
wolffd@0 467 CPDs/@gaussian_CPD/udpate_ess.m:
wolffd@0 468 now, hidden_cps = any(hidden_bitv(cps)), whereas it used to be
wolffd@0 469 hidden_cps = all(hidden_bitv(cps)).
wolffd@0 470
wolffd@0 471 </ul>
wolffd@0 472
wolffd@0 473
wolffd@0 474 <li> 5/29/02
wolffd@0 475 <ul>
wolffd@0 476 <li> CPDs/@gaussian_CPD/udpate_ess.m fixed WX,WXX,WXY (thanks to Rainer Deventer and
wolffd@0 477 Yohsuke Minowa for spotting the bug). Does the C version work??
wolffd@0 478 <li> potentials/@cpot/mpot_to_cpot fixed K==0 case (thanks to Rainer Deventer).
wolffd@0 479 <li> CPDs/@gaussian_CPD/log_prob_node now accepts non-cell array data
wolffd@0 480 on self (thanks to rishi <rishi@capsl.udel.edu> for catching this).
wolffd@0 481 </ul>
wolffd@0 482
wolffd@0 483
wolffd@0 484 <li> 5/19/02
wolffd@0 485 <ul>
wolffd@0 486
wolffd@0 487 <!--
wolffd@0 488 <li> Finally added <a href="../Papers/wei_ndx.ps.gz">paper</a> by Wei Hu (written
wolffd@0 489 November 2001)
wolffd@0 490 describing ndxB, ndxD, and ndxSD.
wolffd@0 491 -->
wolffd@0 492
wolffd@0 493 <li> Wei Hu made the following changes.
wolffd@0 494 <ul>
wolffd@0 495 <li> Memory leak repair:
wolffd@0 496 a. distribute_evidence.c in static/@jtree_C directory
wolffd@0 497 b. distribute_evidence.c in static/@jtree_ndx directory
wolffd@0 498 c. marg_tablec. in Tables dir
wolffd@0 499
wolffd@0 500 <li> Add "@jtree_ndx_2TBN_inf_engine" in inference/online dir
wolffd@0 501
wolffd@0 502 <li> Add "@jtree_sparse_inf_engine" in inference/static dir
wolffd@0 503
wolffd@0 504 <li> Add "@jtree_sparse_2TBN_inf_engine" in inference/online dir
wolffd@0 505
wolffd@0 506 <li> Modify "tabular_CPD.m" in CPDs/@tabular_CPD dir , used for sparse
wolffd@0 507
wolffd@0 508 <li> In "@discrete_CPD" dir:
wolffd@0 509 a. modify "convert_to_pot.m", used for sparse
wolffd@0 510 b. add "convert_to_sparse_table.c"
wolffd@0 511
wolffd@0 512 <li> In "potentials/@dpot" dir:
wolffd@0 513 a. remove "divide_by_pot.c" and "multiply_by_pot.c"
wolffd@0 514 b. add "divide_by_pot.m" and "multiply_by_pot.m"
wolffd@0 515 c. modify "dpot.m", "marginalize_pot.m" and "normalize_pot.m"
wolffd@0 516
wolffd@0 517 <li> In "potentials/Tables" dir:
wolffd@0 518 a. modify mk_ndxB.c;(for speedup)
wolffd@0 519 b. add "mult_by_table.m",
wolffd@0 520 "divide_by_table.m",
wolffd@0 521 "divide_by_table.c",
wolffd@0 522 "marg_sparse_table.c",
wolffd@0 523 "mult_by_sparse_table.c",
wolffd@0 524 "divide_by_sparse_table.c".
wolffd@0 525
wolffd@0 526 <li> Modify "normalise.c" in misc dir, used for sparse.
wolffd@0 527
wolffd@0 528 <li>And, add discrete2, discrete3, filter2 and filter3 as test applications in test_BNT.m
wolffd@0 529 Modify installC.m
wolffd@0 530 </ul>
wolffd@0 531
wolffd@0 532 <li> Kevin made the following changes related to strong junction
wolffd@0 533 trees:
wolffd@0 534 <ul>
wolffd@0 535 <li> jtree_inf_engin line 75:
wolffd@0 536 engine.root_clq = length(engine.cliques);
wolffd@0 537 the last clq is guaranteed to be a strong root
wolffd@0 538
wolffd@0 539 <li> dag_to_jtree line 38: [jtree, root, B, w] =
wolffd@0 540 cliques_to_jtree(cliques, ns);
wolffd@0 541 never call cliques_to_strong_jtree
wolffd@0 542
wolffd@0 543 <li> strong_elim_order: use Ilya's code instead of topological sorting.
wolffd@0 544 </ul>
wolffd@0 545
wolffd@0 546 <li> Kevin fixed CPDs/@generic_CPD/learn_params, so it always passes
wolffd@0 547 in the correct hidden_bitv field to update_params.
wolffd@0 548
wolffd@0 549 </ul>.
wolffd@0 550
wolffd@0 551
wolffd@0 552 <li> 5/8/02
wolffd@0 553 <ul>
wolffd@0 554
wolffd@0 555 <li> Jerod Weinman helped fix some bugs in HHMMQ_CPD/maximize_params.
wolffd@0 556
wolffd@0 557 <li> Removed broken online inference from hmm_inf_engine.
wolffd@0 558 It has been replaced filter_inf_engine, which can take hmm_inf_engine
wolffd@0 559 as an argument.
wolffd@0 560
wolffd@0 561 <li> Changed graph visualization function names.
wolffd@0 562 'draw_layout' is now 'draw_graph',
wolffd@0 563 'draw_layout_dbn' is now 'draw_dbn',
wolffd@0 564 'plotgraph' is now 'dag_to_dot',
wolffd@0 565 'plothmm' is now 'hmm_to_dot',
wolffd@0 566 added 'dbn_to_dot',
wolffd@0 567 'mkdot' no longer exists': its functioality has been subsumed by dag_to_dot.
wolffd@0 568 The dot functions now all take optional args in string/value format.
wolffd@0 569 </ul>
wolffd@0 570
wolffd@0 571
wolffd@0 572 <li> 4/1/02
wolffd@0 573 <ul>
wolffd@0 574 <li> Added online inference classes.
wolffd@0 575 See BNT/inference/online and BNT/examples/dynamic/filter1.
wolffd@0 576 This is work in progress.
wolffd@0 577 <li> Renamed cmp_inference to cmp_inference_dbn, and made its
wolffd@0 578 interface and behavior more similar to cmp_inference_static.
wolffd@0 579 <li> Added field rep_of_eclass to bnet and dbn, to simplify
wolffd@0 580 parameter tying (see ~murphyk/Bayes/param_tieing.html).
wolffd@0 581 <li> Added gmux_CPD (Gaussian mulitplexers).
wolffd@0 582 See BNT/examples/dynamic/SLAM/skf_data_assoc_gmux for an example.
wolffd@0 583 <li> Modified the forwards sampling routines.
wolffd@0 584 general/sample_dbn and sample_bnet now take optional arguments as
wolffd@0 585 strings, and can sample with pre-specified evidence.
wolffd@0 586 sample_bnet can only generate a single sample, and it is always a cell
wolffd@0 587 array.
wolffd@0 588 sample_node can only generate a single sample, and it is always a
wolffd@0 589 scalar or vector.
wolffd@0 590 This eliminates the false impression that the function was
wolffd@0 591 ever vectorized (which was only true for tabular_CPDs).
wolffd@0 592 (Calling sample_bnet inside a for-loop is unlikely to be a bottleneck.)
wolffd@0 593 <li> Updated usage.html's description of CPDs (gmux) and inference
wolffd@0 594 (added gibbs_sampling and modified the description of pearl).
wolffd@0 595 <li> Modified BNT/Kalman/kalman_filter\smoother so they now optionally
wolffd@0 596 take an observed input (control) sequence.
wolffd@0 597 Also, optional arguments are now passed as strings.
wolffd@0 598 <li> Removed BNT/examples/static/uci_data to save space.
wolffd@0 599 </ul>
wolffd@0 600
wolffd@0 601 <li> 3/14/02
wolffd@0 602 <ul>
wolffd@0 603 <li> pearl_inf_engine now works for (vector) Gaussian nodes, as well
wolffd@0 604 as discrete. compute_pi has been renamed CPD_to_pi. compute_lambda_msg
wolffd@0 605 has been renamed CPD_to_lambda_msg. These are now implemented for
wolffd@0 606 the discrete_CPD class instead of tabular_CPD. noisyor and
wolffd@0 607 Gaussian have their own private implemenations.
wolffd@0 608 Created examples/static/Belprop subdirectory.
wolffd@0 609 <li> Added examples/dynamic/HHMM/Motif.
wolffd@0 610 <li> Added Matt Brand's entropic prior code.
wolffd@0 611 <li> cmp_inference_static has changed. It no longer returns err. It
wolffd@0 612 can check for convergence. It can accept 'observed'.
wolffd@0 613 </ul>
wolffd@0 614
wolffd@0 615
wolffd@0 616 <li> 3/4/02
wolffd@0 617 <ul>
wolffd@0 618 <li> Fixed HHMM code. Now BNT/examples/dynamic/HHMM/mk_abcd_hhmm
wolffd@0 619 implements the example in the NIPS paper. See also
wolffd@0 620 Square/sample_square_hhmm_discrete and other files.
wolffd@0 621
wolffd@0 622 <li> Included Bhaskara Marthi's gibbs_sampling_inf_engine. Currently
wolffd@0 623 this only works if all CPDs are tabular and if you call installC.
wolffd@0 624
wolffd@0 625 <li> Modified Kalman/tracking_demo so it calls plotgauss2d instead of
wolffd@0 626 gaussplot.
wolffd@0 627
wolffd@0 628 <li> Included Sonia Leach's speedup of mk_rnd_dag.
wolffd@0 629 My version created all NchooseK subsets, and then picked among them. Sonia
wolffd@0 630 reorders the possible parents randomly and choose
wolffd@0 631 the first k. This saves on having to enumerate the large number of
wolffd@0 632 possible subsets before picking from one.
wolffd@0 633
wolffd@0 634 <li> Eliminated BNT/inference/static/Old, which contained some old
wolffd@0 635 .mexglx files which wasted space.
wolffd@0 636 </ul>
wolffd@0 637
wolffd@0 638
wolffd@0 639
wolffd@0 640 <li> 2/15/02
wolffd@0 641 <ul>
wolffd@0 642 <li> Removed the netlab directory, since most of it was not being
wolffd@0 643 used, and it took up too much space (the goal is to have BNT.zip be
wolffd@0 644 less than 1.4MB, so if fits on a floppy).
wolffd@0 645 The required files have been copied into netlab2.
wolffd@0 646 </ul>
wolffd@0 647
wolffd@0 648 <li> 2/14/02
wolffd@0 649 <ul>
wolffd@0 650 <li> Shan Huang fixed most (all?) of the bugs in his stable CG code.
wolffd@0 651 scg1-3 now work, but scg_3node and scg_unstable give different
wolffd@0 652 behavior than that reported in the Cowell book.
wolffd@0 653
wolffd@0 654 <li> I changed gaussplot so it plots an ellipse representing the
wolffd@0 655 eigenvectors of the covariance matrix, rather than numerically
wolffd@0 656 evaluating the density and using a contour plot; this
wolffd@0 657 is much faster and gives better pictures. The new function is
wolffd@0 658 called plotgauss2d in BNT/Graphics.
wolffd@0 659
wolffd@0 660 <li> Joni Alon <jalon@cs.bu.edu> fixed some small bugs:
wolffd@0 661 mk_dhmm_obs_lik called forwards with the wrong args, and
wolffd@0 662 add_BNT_to_path should quote filenames with spaces.
wolffd@0 663
wolffd@0 664 <li> I added BNT/stats2/myunidrnd which is called by learn_struct_mcmc.
wolffd@0 665
wolffd@0 666 <li> I changed BNT/potentials/@dpot/multiply_by_dpot so it now says
wolffd@0 667 Tbig.T(:) = Tbig.T(:) .* Ts(:);
wolffd@0 668 </ul>
wolffd@0 669
wolffd@0 670
wolffd@0 671 <li> 2/6/02
wolffd@0 672 <ul>
wolffd@0 673 <li> Added hierarchical HMMs. See BNT/examples/dynamic/HHMM and
wolffd@0 674 CPDs/@hhmmQ_CPD and @hhmmF_CPD.
wolffd@0 675 <li> sample_dbn can now sample until a certain condition is true.
wolffd@0 676 <li> Sonia Leach fixed learn_struct_mcmc and changed mk_nbrs_of_digraph
wolffd@0 677 so it only returns DAGs.
wolffd@0 678 Click <a href="sonia_mcmc.txt">here</a> for details of her changes.
wolffd@0 679 </ul>
wolffd@0 680
wolffd@0 681
wolffd@0 682 <li> 2/4/02
wolffd@0 683 <ul>
wolffd@0 684 <li> Wei Hu fixed a bug in
wolffd@0 685 jtree_ndx_inf_engine/collect\distribute_evidence.c which failed when
wolffd@0 686 maximize=1.
wolffd@0 687 <li>
wolffd@0 688 I fixed various bugs to do with conditional Gaussians,
wolffd@0 689 so mixexp3 now works (thansk to Gerry Fung <gerry.fung@utoronto.ca>
wolffd@0 690 for spotting the error). Specifically:
wolffd@0 691 Changed softmax_CPD/convert_to_pot so it now puts cts nodes in cdom, and no longer inherits
wolffd@0 692 this function from discrete_CPD.
wolffd@0 693 Changed root_CPD/convert_to_put so it puts self in cdom.
wolffd@0 694 </ul>
wolffd@0 695
wolffd@0 696
wolffd@0 697 <li> 1/31/02
wolffd@0 698 <ul>
wolffd@0 699 <li> Fixed log_lik_mhmm (thanks to ling chen <real_lingchen@yahoo.com>
wolffd@0 700 for spotting the typo)
wolffd@0 701 <li> Now many scripts in examples/static call cmp_inference_static.
wolffd@0 702 Also, SCG scripts have been simplified (but still don't work!).
wolffd@0 703 <li> belprop and belprop_fg enter_evidence now returns [engine, ll,
wolffd@0 704 niter], with ll=0, so the order of the arguments is compatible with other engines.
wolffd@0 705 <li> Ensured that all enter_evidence methods support optional
wolffd@0 706 arguments such as 'maximize', even if they ignore them.
wolffd@0 707 <li> Added Wei Hu's potentials/Tables/rep_mult.c, which is used to
wolffd@0 708 totally eliminate all repmats from gaussian_CPD/update_ess.
wolffd@0 709 </ul>
wolffd@0 710
wolffd@0 711
wolffd@0 712 <li> 1/30/02
wolffd@0 713 <ul>
wolffd@0 714 <li> update_ess now takes hidden_bitv instead of hidden_self and
wolffd@0 715 hidden_ps. This allows gaussian_CPD to distinguish hidden discrete and
wolffd@0 716 cts parents. Now learn_params_em, as well as learn_params_dbn_em,
wolffd@0 717 passes in this info, for speed.
wolffd@0 718
wolffd@0 719 <li> gaussian_CPD update_ess is now vectorized for any case where all
wolffd@0 720 the continuous nodes are observed (eg., Gaussian HMMs, AR-HMMs).
wolffd@0 721
wolffd@0 722 <li> mk_dbn now automatically detects autoregressive nodes.
wolffd@0 723
wolffd@0 724 <li> hmm_inf_engine now uses indexes in marginal_nodes/family for
wolffd@0 725 speed. Marginal_ndoes can now only handle single nodes.
wolffd@0 726 (SDndx is hard-coded, to avoid the overhead of using marg_ndx,
wolffd@0 727 which is slow because of the case and global statements.)
wolffd@0 728
wolffd@0 729 <li> add_ev_to_dmarginal now retains the domain field.
wolffd@0 730
wolffd@0 731 <li> Wei Hu wrote potentials/Tables/repmat_and_mult.c, which is used to
wolffd@0 732 avoid some of the repmat's in gaussian_CPD/update_ess.
wolffd@0 733
wolffd@0 734 <li> installC now longer sets the global USEC, since USEC is set to 0
wolffd@0 735 by add_BNT_to_path, even if the C files have already been compiled
wolffd@0 736 in a previous session. Instead, gaussian_CPD checks to
wolffd@0 737 see if repmat_and_mult exists, and (bat1, chmm1, water1, water2)
wolffd@0 738 check to see if jtree_C_inf_engine/collect_evidence exists.
wolffd@0 739 Note that checking if a file exists is slow, so we do the check
wolffd@0 740 inside the gaussian_CPD constructor, not inside update_ess.
wolffd@0 741
wolffd@0 742 <li> uninstallC now deletes both .mex and .dll files, just in case I
wolffd@0 743 accidently ship a .zip file with binaries. It also deletes mex
wolffd@0 744 files from jtree_C_inf_engine.
wolffd@0 745
wolffd@0 746 <li> Now marginal_family for both jtree_limid_inf_engine and
wolffd@0 747 global_joint_inf_engine returns a marginal structure and
wolffd@0 748 potential, as required by solve_limid.
wolffd@0 749 Other engines (eg. jtree_ndx, hmm) are not required to return a potential.
wolffd@0 750 </ul>
wolffd@0 751
wolffd@0 752
wolffd@0 753
wolffd@0 754 <li> 1/22/02
wolffd@0 755 <ul>
wolffd@0 756 <li> Added an optional argument to mk_bnet and mk_dbn which lets you
wolffd@0 757 add names to nodes. This uses the new assoc_array class.
wolffd@0 758
wolffd@0 759 <li> Added Yimin Zhang's (unfinished) classification/regression tree
wolffd@0 760 code to CPDs/tree_CPD.
wolffd@0 761
wolffd@0 762 </ul>
wolffd@0 763
wolffd@0 764
wolffd@0 765
wolffd@0 766 <li> 1/14/02
wolffd@0 767 <ul>
wolffd@0 768 <li> Incorporated some of Shan Huang's (still broken) stable CG code.
wolffd@0 769 </ul>
wolffd@0 770
wolffd@0 771
wolffd@0 772 <li> 1/9/02
wolffd@0 773 <ul>
wolffd@0 774 <li> Yimin Zhang vectorized @discrete_CPD/prob_node, which speeds up
wolffd@0 775 structure learning considerably. I fixed this to handle softmax CPDs.
wolffd@0 776
wolffd@0 777 <li> Shan Huang changed the stable conditional Gaussian code to handle
wolffd@0 778 vector-valued nodes, but it is buggy.
wolffd@0 779
wolffd@0 780 <li> I vectorized @gaussian_CPD/update_ess for a special case.
wolffd@0 781
wolffd@0 782 <li> Removed denom=min(1, ... Z) from gaussian_CPD/maximize_params
wolffd@0 783 (added to cope with negative temperature for entropic prior), which
wolffd@0 784 gives wrong results on mhmm1.
wolffd@0 785 </ul>
wolffd@0 786
wolffd@0 787
wolffd@0 788 <li> 1/7/02
wolffd@0 789
wolffd@0 790 <ul>
wolffd@0 791 <li> Removed the 'xo' typo from mk_qmr_bnet.
wolffd@0 792
wolffd@0 793 <li> convert_dbn_CPDs_to_tables has been vectorized; it is now
wolffd@0 794 substantially faster to compute the conditional likelihood for long sequences.
wolffd@0 795
wolffd@0 796 <li> Simplified constructors for tabular_CPD and gaussian_CPD, so they
wolffd@0 797 now both only take the form CPD(bnet, i, ...) for named arguments -
wolffd@0 798 the CPD('self', i, ...) format is gone. Modified mk_fgraph_given_ev
wolffd@0 799 to use mk_isolated_tabular_CPD instead.
wolffd@0 800
wolffd@0 801 <li> Added entropic prior to tabular and Gaussian nodes.
wolffd@0 802 For tabular_CPD, changed name of arguments to the constructor to
wolffd@0 803 distinguish Dirichlet and entropic priors. In particular,
wolffd@0 804 tabular_CPD(bnet, i, 'prior', 2) is now
wolffd@0 805 tabular_CPD(bnet, i, 'prior_type', 'dirichlet', 'dirichlet_weight', 2).
wolffd@0 806
wolffd@0 807 <li> Added deterministic annealing to learn_params_dbn_em for use with
wolffd@0 808 entropic priors. The old format learn(engine, cases, max_iter) has
wolffd@0 809 been replaced by learn(engine, cases, 'max_iter', max_iter).
wolffd@0 810
wolffd@0 811 <li> Changed examples/dynamic/bat1 and kjaerulff1, since default
wolffd@0 812 equivalence classes have changed from untied to tied.
wolffd@0 813 </ul>
wolffd@0 814
wolffd@0 815 <li> 12/30/01
wolffd@0 816 <ul>
wolffd@0 817 <li> DBN default equivalence classes for slice 2 has changed, so that
wolffd@0 818 now parameters are tied for nodes with 'equivalent' parents in slices
wolffd@0 819 1 and 2 (e.g., observed leaf nodes). This essentially makes passing in
wolffd@0 820 the eclass arguments redundant (hooray!).
wolffd@0 821 </ul>
wolffd@0 822
wolffd@0 823
wolffd@0 824 <li> 12/20/01
wolffd@0 825 <ul>
wolffd@0 826 <li> <b>Released version 4</b>.
wolffd@0 827 Version 4 is considered a major new release
wolffd@0 828 since it is not completely backwards compatible with V3.
wolffd@0 829 Observed nodes are now specified when the bnet/dbn is created,
wolffd@0 830 not when the engine is created. This changes the interface to many of
wolffd@0 831 the engines, making the code no longer backwards compatible.
wolffd@0 832 Hence support for non-named optional arguments (BNT2 style) has also
wolffd@0 833 been removed; hence mk_dbn etc. requires arguments to be passed by name.
wolffd@0 834
wolffd@0 835 <li> Ilya Shpitser's C code for triangulation now compiles under
wolffd@0 836 Windows as well as Unix, thanks to Wei Hu.
wolffd@0 837
wolffd@0 838 <li> All the ndx engines have been combined, and now take an optional
wolffd@0 839 argument specifying what kind of index to use.
wolffd@0 840
wolffd@0 841 <li> learn_params_dbn_em is now more efficient:
wolffd@0 842 @tabular_CPD/update_ess for nodes whose families
wolffd@0 843 are hidden does not need need to call add_evidence_to_dmarginal, which
wolffd@0 844 is slow.
wolffd@0 845
wolffd@0 846 <li> Wei Hu fixed bug in jtree_ndxD, so now the matlab and C versions
wolffd@0 847 both work.
wolffd@0 848
wolffd@0 849 <li> dhmm_inf_engine replaces hmm_inf_engine, since the former can
wolffd@0 850 handle any kind of topology and is slightly more efficient. dhmm is
wolffd@0 851 extended to handle Gaussian, as well as discrete,
wolffd@0 852 observed nodes. The new hmm_inf_engine no longer supports online
wolffd@0 853 inference (which was broken anyway).
wolffd@0 854
wolffd@0 855 <li> Added autoregressive HMM special case to hmm_inf_engine for
wolffd@0 856 speed.
wolffd@0 857
wolffd@0 858 <li> jtree_ndxSD_dbn_inf_engine now computes likelihood of the
wolffd@0 859 evidence in a vectorized manner, where possible, just like
wolffd@0 860 hmm_inf_engine.
wolffd@0 861
wolffd@0 862 <li> Added mk_limid, and hence simplified mk_bnet and mk_dbn.
wolffd@0 863
wolffd@0 864
wolffd@0 865 <li> Gaussian_CPD now uses 0.01*I prior on covariance matrix by
wolffd@0 866 default. To do ML estimation, set 'cov_prior_weight' to 0.
wolffd@0 867
wolffd@0 868 <li> Gaussian_CPD and tabular_CPD
wolffd@0 869 optional binary arguments are now set using 0/1 rather no 'no'/'yes'.
wolffd@0 870
wolffd@0 871 <li> Removed Shan Huang's PDAG and decomposable graph code, which will
wolffd@0 872 be put in a separate structure learning library.
wolffd@0 873 </ul>
wolffd@0 874
wolffd@0 875
wolffd@0 876 <li> 12/11/01
wolffd@0 877 <ul>
wolffd@0 878 <li> Wei Hu fixed jtree_ndx*_dbn_inf_engine and marg_table.c.
wolffd@0 879
wolffd@0 880 <li> Shan Huang contributed his implementation of stable conditional
wolffd@0 881 Gaussian code (Lauritzen 1999), and methods to search through the
wolffd@0 882 space of PDAGs (Markov equivalent DAGs) and undirected decomposable
wolffd@0 883 graphs. The latter is still under development.
wolffd@0 884 </ul>
wolffd@0 885
wolffd@0 886
wolffd@0 887 <li> 12/10/01
wolffd@0 888 <ul>
wolffd@0 889 <li> Included Wei Hu's new versions of the ndx* routines, which use
wolffd@0 890 integers instead of doubles. The new versions are about 5 times faster
wolffd@0 891 in C. In general, ndxSD is the best choice.
wolffd@0 892
wolffd@0 893 <li> Fixed misc/add_ev_to_dmarginal so it works with the ndx routines
wolffd@0 894 in bat1.
wolffd@0 895
wolffd@0 896 <li> Added calc_mpe_dbn to do Viterbi parsing.
wolffd@0 897
wolffd@0 898 <li> Updated dhmm_inf_engine so it computes marginals.
wolffd@0 899 </ul>
wolffd@0 900
wolffd@0 901
wolffd@0 902
wolffd@0 903 <li> 11/23/01
wolffd@0 904 <ul>
wolffd@0 905 <li> learn_params now does MAP estimation (i.e., uses Dirichlet prior,
wolffd@0 906 if define). Thanks to Simon Keizer skeizer@cs.utwente.nl for spotting
wolffd@0 907 this.
wolffd@0 908 <li> Changed plotgraph so it calls ghostview with the output of dotty,
wolffd@0 909 instead of converting from .ps to .tif. The resulting image is much
wolffd@0 910 easier to read.
wolffd@0 911 <li> Fixed cgpot/multiply_by_pots.m.
wolffd@0 912 <li> Wei Hu fixed ind2subv.c.
wolffd@0 913 <li> Changed arguments to compute_joint_pot.
wolffd@0 914 </ul>
wolffd@0 915
wolffd@0 916
wolffd@0 917 <li> 11/1/01
wolffd@0 918 <ul>
wolffd@0 919 <li> Changed sparse to dense in @dpot/multiply_pots, because sparse
wolffd@0 920 arrays apparently cause a bug in the NT version of Matlab.
wolffd@0 921
wolffd@0 922 <li> Fixed the bug in gaussian_CPD/log_prob_node.m which
wolffd@0 923 incorrectly called the vectorized gaussian_prob with different means
wolffd@0 924 when there were continuous parents and more than one case.
wolffd@0 925 (Thanks to Dave Andre for finding this.)
wolffd@0 926
wolffd@0 927 <li> Fixed the bug in root_CPD/convert_to_pot which did not check for
wolffd@0 928 pot_type='g'.
wolffd@0 929 (Thanks to Dave Andre for finding this.)
wolffd@0 930
wolffd@0 931 <li> Changed calc_mpe and calc_mpe_global so they now return a cell array.
wolffd@0 932
wolffd@0 933 <li> Combine pearl and loopy_pearl into a single inference engine
wolffd@0 934 called 'pearl_inf_engine', which now takes optional arguments passed
wolffd@0 935 in using the name/value pair syntax.
wolffd@0 936 marginal_nodes/family now takes the optional add_ev argument (same as
wolffd@0 937 jtree), which is the opposite of the previous shrink argument.
wolffd@0 938
wolffd@0 939 <li> Created pearl_unrolled_dbn_inf_engine and "resurrected"
wolffd@0 940 pearl_dbn_inf_engine in a simplified (but still broken!) form.
wolffd@0 941
wolffd@0 942 <li> Wei Hi fixed the bug in ind2subv.c, so now ndxSD works.
wolffd@0 943 He also made C versions of ndxSD and ndxB, and added (the unfinished) ndxD.
wolffd@0 944
wolffd@0 945 </ul>
wolffd@0 946
wolffd@0 947
wolffd@0 948 <li> 10/20/01
wolffd@0 949
wolffd@0 950 <ul>
wolffd@0 951 <li> Removed the use_ndx option from jtree_inf,
wolffd@0 952 and created 2 new inference engines: jtree_ndxSD_inf_engine and
wolffd@0 953 jtree_ndxB_inf_engine.
wolffd@0 954 The former stores 2 sets of indices for the small and difference
wolffd@0 955 domains; the latter stores 1 set of indices for the big domain.
wolffd@0 956 In Matlab, the ndxB version is often significantly faster than ndxSD
wolffd@0 957 and regular jree, except when the clique size is large.
wolffd@0 958 When compiled to C, the difference between ndxB and ndxSD (in terms of
wolffd@0 959 speed) vanishes; again, both are faster than compiled jtree, except
wolffd@0 960 when the clique size is large.
wolffd@0 961 Note: ndxSD currently has a bug in it, so it gives the wrong results!
wolffd@0 962 (The DBN analogs are jtree_dbn_ndxSD_inf_engine and
wolffd@0 963 jtree_dbn_ndxB_inf_engine.)
wolffd@0 964
wolffd@0 965 <li> Removed duplicate files from the HMM and Kalman subdirectories.
wolffd@0 966 e.g., normalise is now only in BNT/misc, so when compiled to C, it
wolffd@0 967 masks the unique copy of the Matlab version.
wolffd@0 968 </ul>
wolffd@0 969
wolffd@0 970
wolffd@0 971
wolffd@0 972 <li> 10/17/01
wolffd@0 973 <ul>
wolffd@0 974 <li> Fixed bugs introduced on 10/15:
wolffd@0 975 Renamed extract_gaussian_CPD_params_given_ev_on_dps.m to
wolffd@0 976 gaussian_CPD_params_given_dps.m since Matlab can't cope with such long
wolffd@0 977 names (this caused cg1 to fail). Fixed bug in
wolffd@0 978 gaussian_CPD/convert_to_pot, which now calls convert_to_table in the
wolffd@0 979 discrete case.
wolffd@0 980
wolffd@0 981 <li> Fixed bug in bk_inf_engine/marginal_nodes.
wolffd@0 982 The test 'if nodes < ss' is now
wolffd@0 983 'if nodes <= ss' (bug fix due to Stephen seg_ma@hotmail.com)
wolffd@0 984
wolffd@0 985 <li> Simplified uninstallC.
wolffd@0 986 </ul>
wolffd@0 987
wolffd@0 988
wolffd@0 989 <li> 10/15/01
wolffd@0 990 <ul>
wolffd@0 991
wolffd@0 992 <li> Added use_ndx option to jtree_inf and jtree_dbn_inf.
wolffd@0 993 This pre-computes indices for multiplying, dividing and marginalizing
wolffd@0 994 discrete potentials.
wolffd@0 995 This is like the old jtree_fast_inf_engine, but we use an extra level
wolffd@0 996 of indirection to reduce the number of indices needed (see
wolffd@0 997 uid_generator object).
wolffd@0 998 Sometimes this is faster than the original way...
wolffd@0 999 This is work in progress.
wolffd@0 1000
wolffd@0 1001 <li> The constructor for dpot no longer calls myreshape, which is very
wolffd@0 1002 slow.
wolffd@0 1003 But new dpots still must call myones.
wolffd@0 1004 Hence discrete potentials are only sometimes 1D vectors (but should
wolffd@0 1005 always be thought of as multi-D arrays). This is work in progress.
wolffd@0 1006 </ul>
wolffd@0 1007
wolffd@0 1008
wolffd@0 1009 <li> 10/6/01
wolffd@0 1010 <ul>
wolffd@0 1011 <li> Fixed jtree_dbn_inf_engine, and added kjaerulff1 to test this.
wolffd@0 1012 <li> Added option to jtree_inf_engine/marginal_nodes to return "full
wolffd@0 1013 sized" marginals, even on observed nodes.
wolffd@0 1014 <li> Clustered BK in examples/dynamic/bat1 seems to be broken,
wolffd@0 1015 so it has been commented out.
wolffd@0 1016 BK will be re-implemented on top of jtree_dbn, which should much more
wolffd@0 1017 efficient.
wolffd@0 1018 </ul>
wolffd@0 1019
wolffd@0 1020 <li> 9/25/01
wolffd@0 1021 <ul>
wolffd@0 1022 <li> jtree_dbn_inf_engine is now more efficient than calling BK with
wolffd@0 1023 clusters = exact, since it only uses the interface nodes, instead of
wolffd@0 1024 all of them, to maintain the belief state.
wolffd@0 1025 <li> Uninstalled the broken C version of strong_elim_order.
wolffd@0 1026 <li> Changed order of arguments to unroll_dbn_topology, so that intra1
wolffd@0 1027 is no longer required.
wolffd@0 1028 <li> Eliminated jtree_onepass, which can be simulated by calling
wolffd@0 1029 collect_evidence on jtree.
wolffd@0 1030 <li> online1 is no longer in the test_BNT suite, since there is some
wolffd@0 1031 problem with online prediction with mixtures of Gaussians using BK.
wolffd@0 1032 This functionality is no longer supported, since doing it properly is
wolffd@0 1033 too much work.
wolffd@0 1034 </ul>
wolffd@0 1035 </li>
wolffd@0 1036
wolffd@0 1037 <li> 9/7/01
wolffd@0 1038 <ul>
wolffd@0 1039 <li> Added Ilya Shpitser's C triangulation code (43x faster!).
wolffd@0 1040 Currently this only compiles under linux; windows support is being added.
wolffd@0 1041 </ul>
wolffd@0 1042
wolffd@0 1043
wolffd@0 1044 <li> 9/5/01
wolffd@0 1045 <ul>
wolffd@0 1046 <li> Fixed typo in CPDs/@tabular_kernel/convert_to_table (thanks,
wolffd@0 1047 Philippe!)
wolffd@0 1048 <li> Fixed problems with clamping nodes in tabular_CPD, learn_params,
wolffd@0 1049 learn_params_tabular, and bayes_update_params. See
wolffd@0 1050 BNT/examples/static/learn1 for a demo.
wolffd@0 1051 </ul>
wolffd@0 1052
wolffd@0 1053
wolffd@0 1054 <li> 9/3/01
wolffd@0 1055 <ul>
wolffd@0 1056 <li> Fixed typo on line 87 of gaussian_CPD which caused error in cg1.m
wolffd@0 1057 <li> Installed Wei Hu's latest version of jtree_C_inf_engine, which
wolffd@0 1058 can now compute marginals on any clique/cluster.
wolffd@0 1059 <li> Added Yair Weiss's code to compute the Bethe free energy
wolffd@0 1060 approximation to the log likelihood in loopy_pearl (still need to add
wolffd@0 1061 this to belprop). The return arguments are now: engine, loglik and
wolffd@0 1062 niter, which is different than before.
wolffd@0 1063 </ul>
wolffd@0 1064
wolffd@0 1065
wolffd@0 1066
wolffd@0 1067 <li> 8/30/01
wolffd@0 1068 <ul>
wolffd@0 1069 <li> Fixed bug in BNT/examples/static/id1 which passed hard-coded
wolffd@0 1070 directory name to belprop_inf_engine.
wolffd@0 1071
wolffd@0 1072 <li> Changed tabular_CPD and gaussian_CPD so they can now be created
wolffd@0 1073 without having to pass in a bnet.
wolffd@0 1074
wolffd@0 1075 <li> Finished mk_fgraph_given_ev. See the fg* files in examples/static
wolffd@0 1076 for demos of factor graphs (work in progress).
wolffd@0 1077 </ul>
wolffd@0 1078
wolffd@0 1079
wolffd@0 1080
wolffd@0 1081 <li> 8/22/01
wolffd@0 1082 <ul>
wolffd@0 1083
wolffd@0 1084 <li> Removed jtree_compiled_inf_engine,
wolffd@0 1085 since the C code it generated was so big that it would barf on large
wolffd@0 1086 models.
wolffd@0 1087
wolffd@0 1088 <li> Tidied up the potentials/Tables directory.
wolffd@0 1089 Removed mk_marg/mult_ndx.c,
wolffd@0 1090 which have been superceded by the much faster mk_marg/mult_index.c
wolffd@0 1091 (written by Wei Hu).
wolffd@0 1092 Renamed the Matlab versions mk_marginalise/multiply_table_ndx.m
wolffd@0 1093 to be mk_marg/mult_index.m to be compatible with the C versions.
wolffd@0 1094 Note: nobody calls these routines anymore!
wolffd@0 1095 (jtree_C_inf_engine/enter_softev.c has them built-in.)
wolffd@0 1096 Removed mk_ndx.c, which was only used by jtree_compiled.
wolffd@0 1097 Removed mk_cluster_clq_ndx.m, mk_CPD_clq_ndx, and marginalise_table.m
wolffd@0 1098 which were not used.
wolffd@0 1099 Moved shrink_obs_dims_in_table.m to misc.
wolffd@0 1100
wolffd@0 1101 <li> In potentials/@dpot directory: removed multiply_by_pot_C_old.c.
wolffd@0 1102 Now marginalize_pot.c can handle maximization,
wolffd@0 1103 and divide_by_pot.c has been implmented.
wolffd@0 1104 marginalize/multiply/divide_by_pot.m no longer have useC or genops options.
wolffd@0 1105 (To get the C versions, use installC.m)
wolffd@0 1106
wolffd@0 1107 <li> Removed useC and genops options from jtree_inf_engine.m
wolffd@0 1108 To use the C versions, install the C code.
wolffd@0 1109
wolffd@0 1110 <li> Updated BNT/installC.m.
wolffd@0 1111
wolffd@0 1112 <li> Added fclose to @loopy_pearl_inf/enter_evidence.
wolffd@0 1113
wolffd@0 1114 <li> Changes to MPE routines in BNT/general.
wolffd@0 1115 The maximize parameter is now specified inside enter_evidence
wolffd@0 1116 instead of when the engine is created.
wolffd@0 1117 Renamed calc_mpe_given_inf_engine to just calc_mpe.
wolffd@0 1118 Added Ron Zohar's optional fix to handle the case of ties.
wolffd@0 1119 Now returns log-likelihood instead of likelihood.
wolffd@0 1120 Added calc_mpe_global.
wolffd@0 1121 Removed references to genops in calc_mpe_bucket.m
wolffd@0 1122 Test file is now called mpe1.m
wolffd@0 1123
wolffd@0 1124 <li> For DBN inference, filter argument is now passed by name,
wolffd@0 1125 as is maximize. This is NOT BACKWARDS COMPATIBLE.
wolffd@0 1126
wolffd@0 1127 <li> Removed @loopy_dbn_inf_engine, which will was too complicated.
wolffd@0 1128 In the future, a new version, which applies static loopy to the
wolffd@0 1129 unrolled DBN, will be provided.
wolffd@0 1130
wolffd@0 1131 <li> discrete_CPD class now contains the family sizes and supports the
wolffd@0 1132 method dom_sizes. This is because it could not access the child field
wolffd@0 1133 CPD.sizes, and mysize(CPT) may give the wrong answer.
wolffd@0 1134
wolffd@0 1135 <li> Removed all functions of the form CPD_to_xxx, where xxx = dpot, cpot,
wolffd@0 1136 cgpot, table, tables. These have been replaced by convert_to_pot,
wolffd@0 1137 which takes a pot_type argument.
wolffd@0 1138 @discrete_CPD calls convert_to_table to implement a default
wolffd@0 1139 convert_to_pot.
wolffd@0 1140 @discrete_CPD calls CPD_to_CPT to implement a default
wolffd@0 1141 convert_to_table.
wolffd@0 1142 The convert_to_xxx routines take fewer arguments (no need to pass in
wolffd@0 1143 the globals node_sizes and cnodes!).
wolffd@0 1144 Eventually, convert_to_xxx will be vectorized, so it will operate on
wolffd@0 1145 all nodes in the same equivalence class "simultaneously", which should
wolffd@0 1146 be significantly quicker, at least for Gaussians.
wolffd@0 1147
wolffd@0 1148 <li> Changed discrete_CPD/sample_node and prob_node to use
wolffd@0 1149 convert_to_table, instead of CPD_to_CPT, so mlp/softmax nodes can
wolffd@0 1150 benefit.
wolffd@0 1151
wolffd@0 1152 <li> Removed @tabular_CPD/compute_lambda_msg_fast and
wolffd@0 1153 private/prod_CPD_and_pi_msgs_fast, since no one called them.
wolffd@0 1154
wolffd@0 1155 <li> Renamed compute_MLE to learn_params,
wolffd@0 1156 by analogy with bayes_update_params (also because it may compute an
wolffd@0 1157 MAP estimate).
wolffd@0 1158
wolffd@0 1159 <li> Renamed set_params to set_fields
wolffd@0 1160 and get_params to get_field for CPD and dpot objects, to
wolffd@0 1161 avoid confusion with the parameters of the CPD.
wolffd@0 1162
wolffd@0 1163 <li> Removed inference/doc, which has been superceded
wolffd@0 1164 by the web page.
wolffd@0 1165
wolffd@0 1166 <li> Removed inference/static/@stab_cond_gauss_inf_engine, which is
wolffd@0 1167 broken, and all references to stable CG.
wolffd@0 1168
wolffd@0 1169 </ul>
wolffd@0 1170
wolffd@0 1171
wolffd@0 1172
wolffd@0 1173
wolffd@0 1174
wolffd@0 1175 <li> 8/12/01
wolffd@0 1176 <ul>
wolffd@0 1177 <li> I removed potentials/@dpot/marginalize_pot_max.
wolffd@0 1178 Now marginalize_pot for all potential classes take an optional third
wolffd@0 1179 argument, specifying whether to sum out or max out.
wolffd@0 1180 The dpot class also takes in optional arguments specifying whether to
wolffd@0 1181 use C or genops (the global variable USE_GENOPS has been eliminated).
wolffd@0 1182
wolffd@0 1183 <li> potentials/@dpot/marginalize_pot has been simplified by assuming
wolffd@0 1184 that 'onto' is always in ascending order (i.e., we remove
wolffd@0 1185 Maynard-Reid's patch). This is to keep the code identical to the C
wolffd@0 1186 version and the other class implementations.
wolffd@0 1187
wolffd@0 1188 <li> Added Ron Zohar's general/calc_mpe_bucket function,
wolffd@0 1189 and my general/calc_mpe_given_inf_engine, for calculating the most
wolffd@0 1190 probable explanation.
wolffd@0 1191
wolffd@0 1192
wolffd@0 1193 <li> Added Wei Hu's jtree_C_inf_engine.
wolffd@0 1194 enter_softev.c is about 2 times faster than enter_soft_evidence.m.
wolffd@0 1195
wolffd@0 1196 <li> Added the latest version of jtree_compiled_inf_engine by Wei Hu.
wolffd@0 1197 The 'C' ndx_method now calls potentials/Tables/mk_marg/mult_index,
wolffd@0 1198 and the 'oldC' ndx_method calls potentials/Tables/mk_marg/mult_ndx.
wolffd@0 1199
wolffd@0 1200 <li> Added potentials/@dpot/marginalize_pot_C.c and
wolffd@0 1201 multiply_by_pot_C.c by Wei Hu.
wolffd@0 1202 These can be called by setting the 'useC' argument in
wolffd@0 1203 jtree_inf_engine.
wolffd@0 1204
wolffd@0 1205 <li> Added BNT/installC.m to compile all the mex files.
wolffd@0 1206
wolffd@0 1207 <li> Renamed prob_fully_instantiated_bnet to log_lik_complete.
wolffd@0 1208
wolffd@0 1209 <li> Added Shan Huang's unfinished stable conditional Gaussian
wolffd@0 1210 inference routines.
wolffd@0 1211 </ul>
wolffd@0 1212
wolffd@0 1213
wolffd@0 1214
wolffd@0 1215 <li> 7/13/01
wolffd@0 1216 <ul>
wolffd@0 1217 <li> Added the latest version of jtree_compiled_inf_engine by Wei Hu.
wolffd@0 1218 <li> Added the genops class by Doug Schwarz (see
wolffd@0 1219 BNT/genopsfun/README). This provides a 1-2x speed-up of
wolffd@0 1220 potentials/@dpot/multiply_by_pot and divide_by_pot.
wolffd@0 1221 <li> The function BNT/examples/static/qmr_compiled compares the
wolffd@0 1222 performance gains of these new functions.
wolffd@0 1223 </ul>
wolffd@0 1224
wolffd@0 1225 <li> 7/6/01
wolffd@0 1226 <ul>
wolffd@0 1227 <li> Made bk_inf_engine use the name/value argument syntax. This can
wolffd@0 1228 now do max-product (Viterbi) as well as sum-product
wolffd@0 1229 (forward-backward).
wolffd@0 1230 <li> Changed examples/static/mfa1 to use the new name/value argument
wolffd@0 1231 syntax.
wolffd@0 1232 </ul>
wolffd@0 1233
wolffd@0 1234
wolffd@0 1235 <li> 6/28/01
wolffd@0 1236
wolffd@0 1237 <ul>
wolffd@0 1238
wolffd@0 1239 <li> <b>Released version 3</b>.
wolffd@0 1240 Version 3 is considered a major new release
wolffd@0 1241 since it is not completely backwards compatible with V2.
wolffd@0 1242 V3 supports decision and utility nodes, loopy belief propagation on
wolffd@0 1243 general graphs (including undirected), structure learning for non-tabular nodes,
wolffd@0 1244 a simplified way of handling optional
wolffd@0 1245 arguments to functions,
wolffd@0 1246 and many other features which are described below.
wolffd@0 1247 In addition, the documentation has been substantially rewritten.
wolffd@0 1248
wolffd@0 1249 <li> The following functions can now take optional arguments specified
wolffd@0 1250 as name/value pairs, instead of passing arguments in a fixed order:
wolffd@0 1251 mk_bnet, jtree_inf_engine, tabular_CPD, gaussian_CPD, softmax_CPD, mlp_CPD,
wolffd@0 1252 enter_evidence.
wolffd@0 1253 This is very helpful if you want to use default values for most parameters.
wolffd@0 1254 The functions remain backwards compatible with BNT2.
wolffd@0 1255
wolffd@0 1256 <li> dsoftmax_CPD has been renamed softmax_CPD, replacing the older
wolffd@0 1257 version of softmax. The directory netlab2 has been updated, and
wolffd@0 1258 contains weighted versions of some of the learning routines in netlab.
wolffd@0 1259 (This code is still being developed by P. Brutti.)
wolffd@0 1260
wolffd@0 1261 <li> The "fast" versions of the inference engines, which generated
wolffd@0 1262 matlab code, have been removed.
wolffd@0 1263 @jtree_compiled_inf_engine now generates C code.
wolffd@0 1264 (This feature is currently being developed by Wei Hu of Intel (China),
wolffd@0 1265 and is not yet ready for public use.)
wolffd@0 1266
wolffd@0 1267 <li> CPD_to_dpot, CPD_to_cpot, CPD_to_cgpot and CPD_to_upot
wolffd@0 1268 are in the process of being replaced by convert_to_pot.
wolffd@0 1269
wolffd@0 1270 <li> determine_pot_type now takes as arguments (bnet, onodes)
wolffd@0 1271 instead of (onodes, cnodes, dag),
wolffd@0 1272 so it can detect the presence of utility nodes as well as continuous
wolffd@0 1273 nodes.
wolffd@0 1274 Hence this function is not backwards compatible with BNT2.
wolffd@0 1275
wolffd@0 1276 <li> The structure learning code (K2, mcmc) now works with any node
wolffd@0 1277 type, not just tabular.
wolffd@0 1278 mk_bnets_tabular has been eliminated.
wolffd@0 1279 bic_score_family and dirichlet_score_family will be replaced by score_family.
wolffd@0 1280 Note: learn_struct_mcmc has a new interface that is not backwards
wolffd@0 1281 compatible with BNT2.
wolffd@0 1282
wolffd@0 1283 <li> update_params_complete has been renamed bayes_update_params.
wolffd@0 1284 Also, learn_params_tabular has been replaced by learn_params, which
wolffd@0 1285 works for any CPD type.
wolffd@0 1286
wolffd@0 1287 <li> Added decision/utility nodes.
wolffd@0 1288 </ul>
wolffd@0 1289
wolffd@0 1290
wolffd@0 1291 <li> 6/6/01
wolffd@0 1292 <ul>
wolffd@0 1293 <li> Added soft evidence to jtree_inf_engine.
wolffd@0 1294 <li> Changed the documentation slightly (added soft evidence and
wolffd@0 1295 parameter tying, and separated parameter and structure learning).
wolffd@0 1296 <li> Changed the parameters of determine_pot_type, so it no longer
wolffd@0 1297 needs to be passed a DAG argument.
wolffd@0 1298 <li> Fixed parameter tying in mk_bnet (num. CPDs now equals num. equiv
wolffd@0 1299 classes).
wolffd@0 1300 <li> Made learn_struct_mcmc work in matlab version 5.2 (thanks to
wolffd@0 1301 Nimrod Megiddo for finding this bug).
wolffd@0 1302 <li> Made 'acyclic.m' work for undirected graphs.
wolffd@0 1303 </ul>
wolffd@0 1304
wolffd@0 1305
wolffd@0 1306 <li> 5/23/01
wolffd@0 1307 <ul>
wolffd@0 1308 <li> Added Tamar Kushnir's code for the IC* algorithm
wolffd@0 1309 (learn_struct_pdag_ic_star). This learns the
wolffd@0 1310 structure of a PDAG, and can identify the presence of latent
wolffd@0 1311 variables.
wolffd@0 1312
wolffd@0 1313 <li> Added Yair Weiss's code for computing the MAP assignment using
wolffd@0 1314 junction tree (i.e., a new method called @dpot/marginalize_pot_max
wolffd@0 1315 instead of marginalize_pot.)
wolffd@0 1316
wolffd@0 1317 <li> Added @discrete_CPD/prob_node in addition to log_prob_node to handle
wolffd@0 1318 deterministic CPDs.
wolffd@0 1319 </ul>
wolffd@0 1320
wolffd@0 1321
wolffd@0 1322 <li> 5/12/01
wolffd@0 1323 <ul>
wolffd@0 1324 <li> Pierpaolo Brutti updated his mlp and dsoftmax CPD classes,
wolffd@0 1325 and improved the HME code.
wolffd@0 1326
wolffd@0 1327 <li> HME example now added to web page. (The previous example was non-hierarchical.)
wolffd@0 1328
wolffd@0 1329 <li> Philippe Leray (author of the French documentation for BNT)
wolffd@0 1330 pointed out that I was including netlab.tar unnecessarily.
wolffd@0 1331 </ul>
wolffd@0 1332
wolffd@0 1333
wolffd@0 1334 <li> 5/4/01
wolffd@0 1335 <ul>
wolffd@0 1336 <li> Added mlp_CPD which defines a CPD as a (conditional) multi-layer perceptron.
wolffd@0 1337 This class was written by Pierpaolo Brutti.
wolffd@0 1338
wolffd@0 1339 <li> Added hierarchical mixtures of experts demo (due to Pierpaolo Brutti).
wolffd@0 1340
wolffd@0 1341 <li> Fixed some bugs in dsoftmax_CPD.
wolffd@0 1342
wolffd@0 1343 <li> Now the BNT distribution includes the whole
wolffd@0 1344 <a href="http://www.ncrg.aston.ac.uk/netlab/">Netlab</a> library in a
wolffd@0 1345 subdirectory.
wolffd@0 1346 It also includes my HMM and Kalman filter toolboxes, instead of just
wolffd@0 1347 fragments of them.
wolffd@0 1348 </ul>
wolffd@0 1349
wolffd@0 1350
wolffd@0 1351 <li> 5/2/01
wolffd@0 1352 <ul>
wolffd@0 1353 <li> gaussian_inf_engine/enter_evidence now correctly returns the
wolffd@0 1354 loglik, even if all nodes are instantiated (bug fix due to
wolffd@0 1355 Michael Robert James).
wolffd@0 1356
wolffd@0 1357 <li> Added dsoftmax_CPD which allows softmax nodes to have discrete
wolffd@0 1358 and continuous parents; the discrete parents act as indices into the
wolffd@0 1359 parameters for the continuous node, by analogy with conditional
wolffd@0 1360 Gaussian nodes. This class was written by Pierpaolo Brutti.
wolffd@0 1361 </ul>
wolffd@0 1362
wolffd@0 1363
wolffd@0 1364 <li> 3/27/01
wolffd@0 1365 <ul>
wolffd@0 1366 <li> learn_struct_mcmc no longer returns sampled_bitv.
wolffd@0 1367 <li> Added mcmc_sample_to_hist to post-process the set of samples.
wolffd@0 1368 </ul>
wolffd@0 1369
wolffd@0 1370 <li> 3/21/01
wolffd@0 1371 <ul>
wolffd@0 1372 <li> Changed license from UC to GNU Library GPL.
wolffd@0 1373
wolffd@0 1374 <li> Made all CPD constructors accept 0 arguments, so now bnets can be
wolffd@0 1375 saved to and loaded from files.
wolffd@0 1376
wolffd@0 1377 <li> Improved the implementation of sequential and batch Bayesian
wolffd@0 1378 parameter learning for tabular CPDs with completely observed data (see
wolffd@0 1379 log_marg_lik_complete and update_params_complete). This code also
wolffd@0 1380 handles interventional data.
wolffd@0 1381
wolffd@0 1382 <li> Added MCMC structure learning for completely observed, discrete,
wolffd@0 1383 static BNs.
wolffd@0 1384
wolffd@0 1385 <li> Started implementing Bayesian estimation of linear Gaussian
wolffd@0 1386 nodes. See root_gaussian_CPD and
wolffd@0 1387 linear_gaussian_CPD. The old gaussian_CPD class has not been changed.
wolffd@0 1388
wolffd@0 1389 <li> Renamed evaluate_CPD to log_prob_node, and simplified its
wolffd@0 1390 arguments.
wolffd@0 1391
wolffd@0 1392 <li> Renamed sample_CPD to sample_node, simplified its
wolffd@0 1393 arguments, and vectorized it.
wolffd@0 1394
wolffd@0 1395 <li> Renamed "learn_params_tabular" to "update_params_complete".
wolffd@0 1396 This does Bayesian updating, but no longer computes the BIC score.
wolffd@0 1397
wolffd@0 1398 <li> Made routines for completely observed networks (sampling,
wolffd@0 1399 complete data likelihood, etc.) handle cell arrays or regular arrays,
wolffd@0 1400 which are faster.
wolffd@0 1401 If some nodes are not scalars, or are hidden, you must use cell arrays.
wolffd@0 1402 You must convert to a cell array before passing to an inference routine.
wolffd@0 1403
wolffd@0 1404 <li> Fixed bug in gaussian_CPD constructor. When creating CPD with
wolffd@0 1405 more than 1 discrete parent with random parameters, the matrices were
wolffd@0 1406 the wrong shape (Bug fix due to Xuejing Sun).
wolffd@0 1407 </ul>
wolffd@0 1408
wolffd@0 1409
wolffd@0 1410
wolffd@0 1411 <li> 11/24/00
wolffd@0 1412 <ul>
wolffd@0 1413 <li> Renamed learn_params and learn_params_dbn to learn_params_em/
wolffd@0 1414 learn_params_dbn_em. The return arguments are now [bnet, LLtrace,
wolffd@0 1415 engine] instead of [engine, LLtrace].
wolffd@0 1416 <li> Added structure learning code for static nets (K2, PC).
wolffd@0 1417 <li> Renamed learn_struct_inter_full_obs as learn_struct_dbn_reveal,
wolffd@0 1418 and reimplemented it to make it simpler and faster.
wolffd@0 1419 <li> Added sequential Bayesian parameter learning (learn_params_tabular).
wolffd@0 1420 <li> Major rewrite of the documentation.
wolffd@0 1421 </ul>
wolffd@0 1422
wolffd@0 1423 <!--
wolffd@0 1424 <li> 6/1/00
wolffd@0 1425 <ul>
wolffd@0 1426 <li> Subtracted 1911 off the counter, so now it counts hits from
wolffd@0 1427 5/22/00. (The initial value of 1911 was a conservative lower bound on the number of
wolffd@0 1428 hits from the time the page was created.)
wolffd@0 1429 </ul>
wolffd@0 1430 -->
wolffd@0 1431
wolffd@0 1432 <li> 5/22/00
wolffd@0 1433 <ul>
wolffd@0 1434 <li> Added online filtering and prediction.
wolffd@0 1435 <li> Added the factored frontier and loopy_dbn algorithms.
wolffd@0 1436 <li> Separated the online user manual into two, for static and dynamic
wolffd@0 1437 networks.
wolffd@0 1438 <!--
wolffd@0 1439 <li> Added a counter to the BNT web page, and initialized it to 1911,
wolffd@0 1440 which is the number of people who have downloaded my software (BNT and
wolffd@0 1441 other toolboxes) since 8/24/98.
wolffd@0 1442 -->
wolffd@0 1443 <li> Added a counter to the BNT web page.
wolffd@0 1444 <!--
wolffd@0 1445 Up to this point, 1911 people had downloaded my software (BNT and
wolffd@0 1446 other toolboxes) since 8/24/98.
wolffd@0 1447 -->
wolffd@0 1448 </ul>
wolffd@0 1449
wolffd@0 1450
wolffd@0 1451 <li> 4/27/00
wolffd@0 1452 <ul>
wolffd@0 1453 <li> Fixed the typo in bat1.m
wolffd@0 1454 <li> Added preliminary code for online inference in DBNs
wolffd@0 1455 <li> Added coupled HMM example
wolffd@0 1456 </ul>
wolffd@0 1457
wolffd@0 1458 <li> 4/23/00
wolffd@0 1459 <ul>
wolffd@0 1460 <li> Fixed the bug in the fast inference routines where the indices
wolffd@0 1461 are empty (arises in bat1.m).
wolffd@0 1462 <li> Sped up marginal_family for the fast engines by precomputing indices.
wolffd@0 1463 </ul>
wolffd@0 1464
wolffd@0 1465 <li> 4/17/00
wolffd@0 1466 <ul>
wolffd@0 1467 <li> Simplified implementation of BK_inf_engine by using soft evidence.
wolffd@0 1468 <li> Added jtree_onepass_inf_engine (which computes a single marginal)
wolffd@0 1469 and modified jtree_dbn_fast to use it.
wolffd@0 1470 </ul>
wolffd@0 1471
wolffd@0 1472 <li> 4/14/00
wolffd@0 1473 <ul>
wolffd@0 1474 <li> Added fast versions of jtree and BK, which are
wolffd@0 1475 designed for models where the division into hidden/observed is fixed,
wolffd@0 1476 and all hidden variables are discrete. These routines are 2-3 times
wolffd@0 1477 faster than their non-fast counterparts.
wolffd@0 1478
wolffd@0 1479 <li> Added graph drawing code
wolffd@0 1480 contributed by Ali Taylan Cemgil from the University of Nijmegen.
wolffd@0 1481 </ul>
wolffd@0 1482
wolffd@0 1483 <li> 4/10/00
wolffd@0 1484 <ul>
wolffd@0 1485 <li> Distinguished cnodes and cnodes_slice in DBNs so that kalman1
wolffd@0 1486 works with BK.
wolffd@0 1487 <li> Removed dependence on cellfun (which only exists in matlab 5.3)
wolffd@0 1488 by adding isemptycell. Now the code works in 5.2.
wolffd@0 1489 <li> Changed the UC copyright notice.
wolffd@0 1490 </ul>
wolffd@0 1491
wolffd@0 1492
wolffd@0 1493
wolffd@0 1494 <li> 3/29/00
wolffd@0 1495 <ul>
wolffd@0 1496 <li><b>Released BNT 2.0</b>, now with objects!
wolffd@0 1497 Here are the major changes.
wolffd@0 1498
wolffd@0 1499 <li> There are now 3 classes of objects in BNT:
wolffd@0 1500 Conditional Probability Distributions, potentials (for junction tree),
wolffd@0 1501 and inference engines.
wolffd@0 1502 Making an inference algorithm (junction tree, sampling, loopy belief
wolffd@0 1503 propagation, etc.) an object might seem counter-intuitive, but in
wolffd@0 1504 fact turns out to be a good idea, since the code and documentation
wolffd@0 1505 can be made modular.
wolffd@0 1506 (In Java, each algorithm would be a class that implements the
wolffd@0 1507 inferenceEngine interface. Since Matlab doesn't support interfaces,
wolffd@0 1508 inferenceEngine is an abstract (virtual) base class.)
wolffd@0 1509
wolffd@0 1510 <p>
wolffd@0 1511 <li>
wolffd@0 1512 In version 1, instead of Matlab's built-in objects,
wolffd@0 1513 I used structs and a
wolffd@0 1514 simulated dispatch mechanism based on the type-tag system in the
wolffd@0 1515 classic textbook by Abelson
wolffd@0 1516 and Sussman ("Structure and Interpretation of Computer Programs",
wolffd@0 1517 MIT Press, 1985). This required editing the dispatcher every time a
wolffd@0 1518 new object type was added. It also required unique (and hence long)
wolffd@0 1519 names for each method, and allowed the user unrestricted access to
wolffd@0 1520 the internal state of objects.
wolffd@0 1521
wolffd@0 1522 <p>
wolffd@0 1523 <li> The Bayes net itself is now a lightweight struct, and can be
wolffd@0 1524 used to specify a model independently of the inference algorithm used
wolffd@0 1525 to process it.
wolffd@0 1526 In version 1, the inference engine was stored inside the Bayes net.
wolffd@0 1527
wolffd@0 1528 <!--
wolffd@0 1529 See the list of <a href="differences2.html">changes from version
wolffd@0 1530 1</a>.
wolffd@0 1531 -->
wolffd@0 1532 </ul>
wolffd@0 1533
wolffd@0 1534
wolffd@0 1535
wolffd@0 1536 <li> 11/24/99
wolffd@0 1537 <ul>
wolffd@0 1538 <li> Added fixed lag smoothing, online EM and the ability to learn
wolffd@0 1539 switching HMMs (POMDPs) to the HMM toolbox.
wolffd@0 1540 <li> Renamed the HMM toolbox function 'mk_dhmm_obs_mat' to
wolffd@0 1541 'mk_dhmm_obs_lik', and similarly for ghmm and mhmm. Updated references
wolffd@0 1542 to these functions in BNT.
wolffd@0 1543 <li> Changed the order of return params from kalman_filter to make it
wolffd@0 1544 more natural. Updated references to this function in BNT.
wolffd@0 1545 </ul>
wolffd@0 1546
wolffd@0 1547
wolffd@0 1548
wolffd@0 1549 <li>10/27/99
wolffd@0 1550 <ul>
wolffd@0 1551 <li>Fixed line 42 of potential/cg/marginalize_cgpot and lines 32-39 of bnet/add_evidence_to_marginal
wolffd@0 1552 (thanks to Rainer Deventer for spotting these bugs!)
wolffd@0 1553 </ul>
wolffd@0 1554
wolffd@0 1555
wolffd@0 1556 <li>10/21/99
wolffd@0 1557 <ul>
wolffd@0 1558 <li>Completely changed the blockmatrix class to make its semantics
wolffd@0 1559 more sensible. The constructor is not backwards compatible!
wolffd@0 1560 </ul>
wolffd@0 1561
wolffd@0 1562 <li>10/6/99
wolffd@0 1563 <ul>
wolffd@0 1564 <li>Fixed all_vals = cat(1, vals{:}) in user/enter_evidence
wolffd@0 1565 <li>Vectorized ind2subv and sub2indv and removed the C versions.
wolffd@0 1566 <li>Made mk_CPT_from_mux_node much faster by having it call vectorized
wolffd@0 1567 ind2subv
wolffd@0 1568 <li>Added Sondhauss's bug fix to line 68 of bnet/add_evidence_to_marginal
wolffd@0 1569 <li>In dbn/update_belief_state, instead of adding eps to likelihood if 0,
wolffd@0 1570 we leave it at 0, and set the scale factor to 0 instead of dividing.
wolffd@0 1571 </ul>
wolffd@0 1572
wolffd@0 1573 <li>8/19/99
wolffd@0 1574 <ul>
wolffd@0 1575 <li>Added Ghahramani's mfa code to examples directory to compare with
wolffd@0 1576 fa1, which uses BNT
wolffd@0 1577 <li>Changed all references of assoc to stringmatch (e.g., in
wolffd@0 1578 examples/mk_bat_topology)
wolffd@0 1579 </ul>
wolffd@0 1580
wolffd@0 1581 <li>June 1999
wolffd@0 1582 <ul>
wolffd@0 1583 <li><b>Released BNT 1.0</b> on the web.
wolffd@0 1584 </ul>
wolffd@0 1585
wolffd@0 1586
wolffd@0 1587 <li>August 1998
wolffd@0 1588 <ul>
wolffd@0 1589 <li><b>Released BNT 0.0</b> via email.
wolffd@0 1590 </ul>
wolffd@0 1591
wolffd@0 1592
wolffd@0 1593 <li>October 1997
wolffd@0 1594 <ul>
wolffd@0 1595 <li>First started working on Matlab version of BNT.
wolffd@0 1596 </ul>
wolffd@0 1597
wolffd@0 1598 <li>Summer 1997
wolffd@0 1599 <ul>
wolffd@0 1600 <li> First started working on C++ version of BNT while working at DEC (now Compaq) CRL.
wolffd@0 1601 </ul>
wolffd@0 1602
wolffd@0 1603 <!--
wolffd@0 1604 <li>Fall 1996
wolffd@0 1605 <ul>
wolffd@0 1606 <li>Made a C++ program that generates DBN-specific C++ code
wolffd@0 1607 for inference using the frontier algorithm.
wolffd@0 1608 </ul>
wolffd@0 1609
wolffd@0 1610 <li>Fall 1995
wolffd@0 1611 <ul>
wolffd@0 1612 <li>Arrive in Berkeley, and first learn about Bayes Nets. Start using
wolffd@0 1613 Geoff Zweig's C++ code.
wolffd@0 1614 </ul>
wolffd@0 1615 -->
wolffd@0 1616
wolffd@0 1617 </ul>