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

first hg version after svn

<title>History of changes to BNT</title>
<h1>History of changes to BNT</h1>


<h2>Changes since 4 Oct 2007</h2>

<pre>
- 19 Oct 07 murphyk

* BNT\CPDs\@noisyor_CPD\CPD_to_CPT.m: 2nd half of the file is a repeat
of the first half and was deleted (thanks to Karl Kuschner)

* KPMtools\myismember.m should return logical for use in "assert" so add line at end
                p=logical(p);  this prevents "assert" from failing on an integer input.
(thanks to Karl Kuschner)



- 17 Oct 07 murphyk

* Updated subv2ind and ind2subv in KPMtools to Tom Minka's implementation.
His ind2subv is faster (vectorized), but I had to modify it so it
matched the behavior of my version when called with siz=[].
His subv2inv is slightly simpler than mine because he does not treat
the siz=[2 2 ... 2] case separately.
Note: there is now no need to ever use the C versions of these
functions (or any others, for that matter).

* removed BNT/add_BNT_to_path since no longer needed.



- 4 Oct 07 murphyk

*  moved code from sourceforge to UBC website, made version 1.0.4

* @pearl_inf_engine/pearl_inf_engine line 24, default
argument for protocol changed from [] to 'parallel'.
Also, changed private/parallel_protocol so it doesn't write to an
empty file id (Matlab 7 issue)

* added foptions (Matlab 7 issue)

* changed genpathKPM to exclude svn. Put it in toplevel directory to
massively simplify the installation process.

</pre>


<h2>Sourceforge changelog</h2>

BNT was first ported to sourceforge on 28 July 2001 by yozhik.
BNT was removed from sourceforge on 4 October 2007 by Kevin Murphy;
that version is cached as <a
href="FullBNT-1.0.3.zip">FullBNT-1.0.3.zip</a>. 
See  <a href="ChangeLog.Sourceforge.txt">Changelog from
sourceforge</a> for a history of that version of the code,
which formed the basis of the branch currently on Murphy's web page.


<h2> Changes from August 1998 -- July 2004</h2>

Kevin Murphy made the following changes to his own private copy.
(Other small changes were made between July 2004 and October 2007, but were
not documented.)
These may or may not be reflected in the sourceforge version of the
code (which was independently maintained).


<ul>
<li> 9 June 2004
<ul>
<li> Changed tabular_CPD/learn_params back to old syntax, to make it
compatible with gaussian_CPD/learn_params (and re-enabled
generic_CPD/learn_params).
Modified learning/learn_params.m and learning/score_family
appropriately.
(In particular, I undid the change Sonia Leach had to make to
score_family to handle this asymmetry.)
Added examples/static/gaussian2 to test this new functionality.

<li> Added bp_mrf2 (for generic pairwise MRFs) to
inference/static/@bp_belprop_mrf2_inf_engine. [MRFs are not
"officially" supported in BNT, so this code is just for expert
hackers.]

<li> Added examples/static/nodeorderExample.m to illustrate importance
of using topological ordering.

<li> Ran dos2unix on all *.c files within BNT to eliminate compiler
warnings.

</ul>

<li> 7 June 2004
<ul>
<li> Replaced normaliseC with normalise in HMM/fwdback, for maximum
portability (and negligible loss in speed).
<li> Ensured FullBNT versions of HMM, KPMstats etc were as up-to-date
as stand-alone versions.
<li> Changed add_BNT_to_path so it no longer uses addpath(genpath()),
which caused Old versions of files to mask new ones.
</ul>

<li> 18 February 2004
<ul>
<li> A few small bug fixes to BNT, as posted to the Yahoo group.
<li> Several new functions added to KPMtools, KPMstats and Graphviz
(none needed by BNT).
<li> Added CVS to some of my toolboxes.
</ul>

<li> 30 July 2003
<ul>
<li> qian.diao fixed @mpot/set_domain_pot and @cgpot/set_domain_pot
<li> Marco Grzegorczyk found, and Sonia Leach fixed, a bug in
do_removal inside learn_struct_mcmc
</ul>


<li> 28 July 2003
<ul>
<li> Sebastian Luehr provided 2 minor bug fixes, to HMM/fwdback (if any(scale==0))
and FullBNT\HMM\CPDs\@hhmmQ_CPD\update_ess.m (wrong transpose).
</ul>

<li> 8 July 2003
<ul>
<li> Removed buggy  BNT/examples/static/MRF2/Old/mk_2D_lattice.m which was
masking correct graph/mk_2D_lattice.
<li> Fixed bug in graph/mk_2D_lattice_slow in the non-wrap-around case
(line 78)
</ul>


<li> 2 July 2003
<ul>
<li> Sped up normalize(., 1) in KPMtools by avoiding general repmat
<li> Added assign_cols and marginalize_table to KPMtools
</ul>


<li> 29 May 2003
<ul>
<li> Modified KPMstats/mixgauss_Mstep so it repmats Sigma in the tied
covariance case (bug found by [e-mail]).

<li> Bob Welch found bug in gaussian_CPDs/maximize_params in the way
cpsz was computed.

<li> Added KPMstats/mixgauss_em, because my code is easier to
understand/modify than netlab's (at least for me!).

<li> Modified BNT/examples/dynamic/viterbi1 to call multinomial_prob
instead of mk_dhmm_obs_lik.

<li> Moved parzen window and partitioned models code to KPMstats.

<li> Rainer Deventer fixed some bugs in his scgpot code, as follows:
1. complement_pot.m
Problems occured for probabilities equal to zero. The result is an
division by zero error.
<br>
2. normalize_pot.m
This function is used during the calculation of the log-likelihood.
For a probability of zero a warning "log of zero" occurs. I have not
realy fixed the bug. As a workaround I suggest to calculate the 
likelihhod based on realmin (the smallest real number) instead of
zero.
<br>
3. recursive_combine_pots
At the beginning of the function there was no test for the trivial case,
which defines the combination of two potentials as equal to the direct
combination. The result might be an infinite recursion which leads to
a stack overflow in matlab.
</ul>



<li> 11 May 2003
<ul> 
<li> Fixed bug in gaussian_CPD/maximize_params so it is compatible
with the new clg_Mstep routine
<li> Modified KPMstats/cwr_em to handle single cluster case 
separately.
<li> Fixed bug in netlab/gmminit.
<li> Added hash tables to KPMtools.
</ul>


<li> 4 May 2003
<ul>
<li>
Renamed many functions in KPMstats so the name of the
distribution/model type comes first, 
Mstep_clg -> clg_Mstep,
Mstep_cond_gauss -> mixgauss_Mstep.
Also, renamed eval_pdf_xxx functions to xxx_prob, e.g.
eval_pdf_cond_mixgauss -> mixgauss_prob.
This is simpler and shorter.

<li> 
Renamed many functions in HMM toolbox so the name of the
distribution/model type comes first, 
log_lik_mhmm -> mhmm_logprob, etc.
mk_arhmm_obs_lik has finally been re-implemented in terms of clg_prob
and mixgauss_prob (for slice 1).
Removed the Demos directory, and put them in the main directory.
This code is not backwards compatible.

<li> Removed some of the my_xxx functions from KPMstats (these were
mostly copies of functions from the Mathworks stats toolbox).


<li> Modified BNT to take into account changes to KPMstats and
HMM toolboxes.

<li> Fixed KPMstats/Mstep_clg (now called clg_Mstep) for spherical Gaussian case.
(Trace was wrongly parenthesised, and I used YY instead of YTY.
The spherical case now gives the same result as the full case
for cwr_demo.)
Also, mixgauss_Mstep now adds 0.01 to the ML estimate of Sigma,
to act as a regularizer (it used to add 0.01 to  E[YY'], but this was
ignored in the spherical case).

<li> Added cluster weighted regression to KPMstats.

<li> Added KPMtools/strmatch_substr.
</ul>



<li>  28 Mar 03 
<ul>
<li> Added mc_stat_distrib and eval_pdf_cond_prod_parzen to KPMstats
<li> Fixed GraphViz/arrow.m incompatibility with matlab 6.5
(replace all NaN's with 0).
Modified GraphViz/graph_to_dot so it also works on windows.
<li> I removed dag_to_jtree and added graph_to_jtree to the graph
toolbox; the latter expects an undirected graph as input.
<li> I added triangulate_2Dlattice_demo.m to graph.
<li> Rainer Deventer fixed the stable conditional Gaussian potential
classes (scgpot and scgcpot) and inference engine
(stab_cond_gauss_inf_engine).
<li> Rainer Deventer added (stable) higher-order Markov models (see
inference/dynamic/@stable_ho_inf_engine).
</ul>


<li> 14 Feb 03
<ul>
<li> Simplified learning/learn_params so it no longer returns BIC
score. Also, simplified @tabular_CPD/learn_params so it only takes
local evidence.
Added learn_params_dbn, which does ML estimation of fully observed
DBNs.
<li> Vectorized KPMstats/eval_pdf_cond_mixgauss for tied Sigma
case (much faster!).
Also, now works in log-domain to prevent underflow.
eval_pdf_mixgauss now calls eval_pdf_cond_mixgauss and inherits these benefits.
<li> add_BNT_to_path now calls genpath with 2 arguments if using
matlab version 5.
</ul>


<li>  30 Jan 03
<ul>
<li> Vectorized KPMstats/eval_pdf_cond_mixgauss for scalar Sigma
case (much faster!)
<li> Renamed mk_dotfile_from_hmm to draw_hmm and moved it to the
GraphViz library.
<li> Rewrote @gaussian_CPD/maximize_params.m so it calls
KPMstats/Mstep_clg.
This fixes bug when using clamped means (found by Rainer Deventer
and Victor Eruhimov)
and a bug when using a Wishart prior (no gamma term in the denominator).
It is also easier to read.
I rewrote the technical report re-deriving all the equations in a
clearer notation, making the solution to the bugs more obvious.
(See www.ai.mit.edu/~murphyk/Papers/learncg.pdf)
Modified Mstep_cond_gauss to handle priors.
<li> Fixed bug reported by Ramgopal Mettu in which add_BNT_to_path
calls genpath with only 1 argument, whereas version 5 requires 2.
<li> Fixed installC and uninstallC to search in FullBNT/BNT.
</ul>


<li> 24 Jan 03
<ul>
<li> Major simplification of HMM code.
The API is not backwards compatible.
No new functionality has been added, however.
There is now only one fwdback function, instead of 7;
different behaviors are controlled through optional arguments. 
I renamed 'evaluate observation likelihood' (local evidence)
to 'evaluate conditional pdf', since this is more general.
i.e., renamed
mk_dhmm_obs_lik to eval_pdf_cond_multinomial,
mk_ghmm_obs_lik to eval_pdf_cond_gauss,
mk_mhmm_obs_lik to eval_pdf_cond_mog.
These functions have been moved to KPMstats,
so they can be used by other toolboxes.
ghmm's have been eliminated, since they are just a special case of
mhmm's with M=1 mixture component.
mixgauss HMMs can now handle a different number of
mixture components per state.
init_mhmm has been eliminated, and replaced with init_cond_mixgauss
(in KPMstats) and mk_leftright/rightleft_transmat.
learn_dhmm can no longer handle inputs (although this is easy to add back).
</ul>





<li> 20 Jan 03
<ul>
<li> Added arrow.m to GraphViz directory, and commented out line 922,
in response to a bug report.
</ul>

<li> 18 Jan 03
<ul>
<li> Major restructuring of BNT file structure:
all code that is not specific to Bayes nets has been removed;
these packages must be downloaded separately. (Or just download FullBNT.)
This makes it easier to ensure different toolboxes are consistent.
misc has been slimmed down and renamed KPMtools, so it can be shared by other toolboxes,
such as HMM and Kalman; some of the code has been moved to BNT/general.
The Graphics directory has been slimmed down and renamed GraphViz.
The graph directory now has no dependence on BNT (dag_to_jtree has
been renamed graph_to_jtree and has a new API).
netlab2 no longer contains any netlab files, only netlab extensions.
None of the functionality has changed.
</ul>



<li> 11 Jan 03
<ul>
<li> jtree_dbn_inf_engine can now support soft evidence.

<li> Rewrote graph/dfs to make it clearer.
Return arguments have changed, as has mk_rooted_tree.
The acyclicity check for large undirected graphs can cause a stack overflow.
It turns out that this was not a bug, but is because Matlab's stack depth
bound is very low by default.

<li> Renamed examples/dynamic/filter2 to filter_test1, so it does not
conflict with the filter2 function in the image processing toolbox.

<li> Ran test_BNT on various versions of matlab to check compatibility.
On matlab 6.5 (r13), elapsed time = 211s, cpu time = 204s.
On matlab 6.1 (r12), elapsed time = 173s, cpu time = 164s.
On matlab 5.3 (r11), elapsed time = 116s, cpu time = 114s.
So matlab is apparently getting slower with time!!
(All results were with a linux PIII machine.)
</ul>


<li> 14 Nov 02
<ul>
<li> Removed all ndx inference routines, since they are only
marginally faster on toy problems,
and are slower on large problems due to having to store and lookup 
the indices (causes cache misses).
In particular, I removed jtree_ndx_inf_eng and jtree_ndx_dbn_inf_eng, all the *ndx*
routines from potentials/Tables, and all the UID stuff from
add_BNT_to_path,
thus simplifying the code.
This required fixing hmm_(2TBN)_inf_engine/marginal_nodes\family,
and updating installC.


<li> Removed jtree_C_inf_engine and jtree_C_dbn_inf_engine.
The former is basically the same as using jtree_inf_engine with
mutliply_by_table.c and marginalize_table.c.
The latter benefited slightly by assuming potentials were tables
(arrays not objects), but these negligible savings don't justify the
complexity and code duplication.

<li> Removed stab_cond_gauss_inf_engine and
scg_unrolled_dbn_inf_engine,
written by [e-mail], since the code was buggy.

<li> Removed potential_engine, which was only experimental anyway.

</ul>



<li> 13 Nov 02
<ul>
<li> <b>Released version 5</b>.
The previous version, released on 7/28/02, is available
<a href="BNT4.zip">here</a>.

<li> Moved code and documentation to MIT.

<li> Added repmat.c from Thomas Minka's lightspeed library.
Modified it so it can return an empty matrix.

<li> Tomas Kocka fixed bug in the BDeu option for tabular_CPD,
and contributed graph/dag_to_eg, to convert to essential graphs.

<!--<li> Wrote a <a href="../Papers/fastmult.pdf">paper</a> which explains
the ndx methods and the ndx cache BNT uses for fast
multiplication/ marginalization of multi-dimensional arrays.
-->

<li> Modified definition of hhmmQ_CPD, so that Qps can now accept
parents in either the current or previous slice.

<li> Added hhmm2Q_CPD class, which is simpler than hhmmQ (no embedded
sub CPDs, etc), and which allows the conditioning parents, Qps, to
be before (in the topological ordering) the F or Q(t-1) nodes.
See BNT/examples/dynamic/HHMM/Map/mk_map_hhmm for an example.
</ul>


<li> 7/28/02
<ul>
<li> Changed graph/best_first_elim_order from min-fill to min-weight.
<li> Ernest Chan fixed bug in Kalman/sample_lds (G{i} becomes G{m} in
line 61).
<li> Tal Blum <[e-mail]> fixed bug in HMM/init_ghmm (Q
becomes K, the number of states).
<li> Fixed jtree_2tbn_inf_engine/set_fields so it correctly sets the
maximize flag to 1 even in subengines.
<li> Gary Bradksi did a simple mod to the PC struct learn alg so you can pass it an
adjacency matrix as a constraint. Also, CovMat.m reads a file and
produces a covariance matrix.
<li> KNOWN BUG in CPDs/@hhmmQ_CPD/update_ess.m at line 72 caused by 
examples/dynamic/HHMM/Square/learn_square_hhmm_cts.m at line 57.
<li>
The old version is available from www.cs.berkeley.edu/~murphyk/BNT.24june02.zip
</ul>


<li> 6/24/02
<ul>
<li> Renamed dag_to_dot as graph_to_dot and added support for
undirected graphs.
<li> Changed syntax for HHMM CPD constructors: no need to specify d/D
anymore,so they can be used for more complex models.
<li> Removed redundant first argument to mk_isolated_tabular_CPD.
</ul>


<li> 6/19/02
<ul>
<li>
Fixed most probable explanation code.
Replaced calc_mpe with find_mpe, which is now a method of certain
inference engines, e.g., jtree, belprop.
calc_mpe_global has become the find_mpe method of global_joint.
calc_mpe_bucket has become the find_mpe method of var_elim.
calc_mpe_dbn has become the find_mpe method of smoother.
These routines now correctly find the jointly most probable
explanation, instead of the marginally most probable assignments.
See examples/static/mpe1\mpe2 and examples/dynamic/viterbi1
for examples.
Removed maximize flag from constructor and enter_evidence
methods, since this no longer needs to be specified by the user.

<li> Rainer Deventer fixed in a bug in 
CPDs/@gaussian_CPD/udpate_ess.m: 
now, hidden_cps = any(hidden_bitv(cps)), whereas it used to be
hidden_cps = all(hidden_bitv(cps)).

</ul>


<li> 5/29/02
<ul>
<li> CPDs/@gaussian_CPD/udpate_ess.m fixed WX,WXX,WXY (thanks to Rainer Deventer and
Yohsuke Minowa for spotting the bug). Does the C version work??
<li> potentials/@cpot/mpot_to_cpot fixed K==0 case (thanks to Rainer Deventer).
<li> CPDs/@gaussian_CPD/log_prob_node now accepts non-cell array data
on self (thanks to rishi <[e-mail]> for catching this).
</ul>


<li> 5/19/02
<ul>

<!--
<li> Finally added <a href="../Papers/wei_ndx.ps.gz">paper</a> by Wei Hu (written
November 2001)
describing ndxB, ndxD, and ndxSD.
-->

<li> Wei Hu made the following changes.
<ul>
<li>   Memory leak repair:
     a.  distribute_evidence.c   in  static/@jtree_C directory
     b.  distribute_evidence.c   in  static/@jtree_ndx directory
     c.  marg_tablec.            in  Tables dir

<li>   Add "@jtree_ndx_2TBN_inf_engine"  in inference/online dir

<li>   Add "@jtree_sparse_inf_engine"    in inference/static dir

<li>   Add "@jtree_sparse_2TBN_inf_engine"   in inference/online  dir

<li>   Modify "tabular_CPD.m" in CPDs/@tabular_CPD dir , used for sparse

<li>   In "@discrete_CPD" dir:
     a.  modify "convert_to_pot.m", used for sparse
     b.  add "convert_to_sparse_table.c"

<li>   In "potentials/@dpot" dir:
     a.  remove "divide_by_pot.c" and "multiply_by_pot.c"
     b.  add "divide_by_pot.m" and "multiply_by_pot.m"
     c.  modify "dpot.m", "marginalize_pot.m" and "normalize_pot.m"

<li>   In "potentials/Tables" dir:
     a.  modify mk_ndxB.c;(for speedup)
     b.  add "mult_by_table.m", 
             "divide_by_table.m",
             "divide_by_table.c",
             "marg_sparse_table.c",
             "mult_by_sparse_table.c",
             "divide_by_sparse_table.c".

<li>   Modify "normalise.c" in misc dir, used for sparse.

<li>And, add discrete2, discrete3, filter2 and filter3 as test applications in test_BNT.m
Modify installC.m
</ul>

<li> Kevin made the following changes related to strong junction
trees:
<ul>
<li> jtree_inf_engin line 75:
engine.root_clq = length(engine.cliques);
the last clq is guaranteed to be a strong root

<li> dag_to_jtree line 38: [jtree, root, B, w] =
cliques_to_jtree(cliques, ns);
never call cliques_to_strong_jtree

<li> strong_elim_order: use Ilya's code instead of topological sorting.
</ul>

<li> Kevin fixed CPDs/@generic_CPD/learn_params, so it always passes
in the correct hidden_bitv field to update_params.

</ul>.


<li> 5/8/02
<ul>

<li> Jerod Weinman helped fix some bugs in HHMMQ_CPD/maximize_params.

<li> Removed broken online inference from hmm_inf_engine.
It has been replaced filter_inf_engine, which can take hmm_inf_engine
as an argument.

<li> Changed graph visualization function names.
'draw_layout' is now 'draw_graph',
'draw_layout_dbn' is now 'draw_dbn',
'plotgraph' is now 'dag_to_dot',
'plothmm' is now 'hmm_to_dot',
added 'dbn_to_dot',
'mkdot' no longer exists': its functioality has been subsumed by dag_to_dot.
The dot functions now all take optional args in string/value format.
</ul>


<li> 4/1/02
<ul>
<li> Added online inference classes.
See BNT/inference/online and BNT/examples/dynamic/filter1.
This is work in progress.
<li> Renamed cmp_inference to cmp_inference_dbn, and made its
    interface and behavior more similar to cmp_inference_static.
<li> Added field rep_of_eclass to bnet and dbn, to simplify
parameter tying (see ~murphyk/Bayes/param_tieing.html).
<li> Added gmux_CPD (Gaussian mulitplexers).
See BNT/examples/dynamic/SLAM/skf_data_assoc_gmux for an example.
<li> Modified the forwards sampling routines.
general/sample_dbn and sample_bnet now take optional arguments as
strings, and can sample with pre-specified evidence.
sample_bnet can only generate a single sample, and it is always a cell
array. 
sample_node can only generate a single sample, and it is always a
scalar or vector.
This eliminates the false impression that the function was
ever vectorized (which was only true for tabular_CPDs).
(Calling sample_bnet inside a for-loop is unlikely to be a bottleneck.)
<li> Updated usage.html's description of CPDs (gmux) and inference
(added gibbs_sampling and modified the description of pearl).
<li> Modified BNT/Kalman/kalman_filter\smoother so they now optionally
take an observed input (control) sequence.
Also, optional arguments are now passed as strings.
<li> Removed BNT/examples/static/uci_data to save space.
</ul>

<li> 3/14/02
<ul>
<li> pearl_inf_engine now works for (vector) Gaussian nodes, as well
as discrete. compute_pi has been renamed CPD_to_pi. compute_lambda_msg
    has been renamed CPD_to_lambda_msg. These are now implemented for
    the discrete_CPD class instead of tabular_CPD. noisyor and
    Gaussian have their own private implemenations.
Created examples/static/Belprop subdirectory.
<li> Added examples/dynamic/HHMM/Motif.
<li> Added Matt Brand's entropic prior code.
<li> cmp_inference_static has changed. It no longer returns err. It
    can check for convergence. It can accept 'observed'.
</ul>


<li> 3/4/02
<ul>
<li> Fixed HHMM code. Now BNT/examples/dynamic/HHMM/mk_abcd_hhmm
implements the example in the NIPS paper. See also
Square/sample_square_hhmm_discrete and other files.

<li> Included Bhaskara Marthi's gibbs_sampling_inf_engine. Currently
this only works if all CPDs are tabular and if you call installC.

<li> Modified Kalman/tracking_demo  so it calls plotgauss2d instead of
    gaussplot.

<li> Included Sonia Leach's speedup of mk_rnd_dag.
My version created all NchooseK subsets, and then picked among them. Sonia
reorders the possible parents randomly and choose
the first k. This saves on having to enumerate the large number of
possible subsets before picking from one. 

<li> Eliminated BNT/inference/static/Old, which contained some old
.mexglx files which wasted space.
</ul>



<li> 2/15/02
<ul>
<li> Removed the netlab directory, since most of it was not being
used, and it took up too much space (the goal is to have BNT.zip be
less than 1.4MB, so if fits on a floppy).
The required files have been copied into netlab2.
</ul>

<li> 2/14/02
<ul>
<li> Shan Huang fixed most (all?) of the bugs in his stable CG code.
scg1-3 now work, but scg_3node and scg_unstable give different
    behavior than that reported in the Cowell book.

<li> I changed gaussplot so it plots an ellipse representing the
    eigenvectors of the covariance matrix, rather than numerically
    evaluating the density and using a contour plot; this
    is much faster and gives better pictures. The new function is
    called plotgauss2d in BNT/Graphics.

<li> Joni Alon <[e-mail]> fixed some small bugs:
    mk_dhmm_obs_lik called forwards with the wrong args, and
    add_BNT_to_path should quote filenames with spaces.

<li> I added BNT/stats2/myunidrnd which is called by learn_struct_mcmc.

<li> I changed BNT/potentials/@dpot/multiply_by_dpot so it now says
Tbig.T(:) = Tbig.T(:) .* Ts(:);
</ul>


<li> 2/6/02
<ul>
<li> Added hierarchical HMMs. See BNT/examples/dynamic/HHMM and
CPDs/@hhmmQ_CPD and @hhmmF_CPD.
<li> sample_dbn can now sample until a certain condition is true.
<li> Sonia Leach fixed learn_struct_mcmc and changed mk_nbrs_of_digraph
so it only returns DAGs.
Click <a href="sonia_mcmc.txt">here</a> for details of her changes.
</ul>


<li> 2/4/02
<ul>
<li> Wei Hu fixed a bug in
jtree_ndx_inf_engine/collect\distribute_evidence.c which failed when
maximize=1.
<li>
I fixed various bugs to do with conditional Gaussians,
so mixexp3 now works (thansk to Gerry Fung <[e-mail]>
      for spotting the error). Specifically:
Changed softmax_CPD/convert_to_pot so it now puts cts nodes in cdom, and no longer inherits 
      this function from discrete_CPD.
     Changed root_CPD/convert_to_put so it puts self in cdom.
</ul>


<li> 1/31/02
<ul>
<li> Fixed log_lik_mhmm (thanks to ling chen <[e-mail]>
for spotting the typo)
<li> Now many scripts in examples/static  call cmp_inference_static.
Also, SCG scripts have been simplified (but still don't work!).
<li> belprop and belprop_fg enter_evidence now returns [engine, ll,
      niter], with ll=0, so the order of the arguments is compatible with other engines.
<li> Ensured that all enter_evidence methods support optional
      arguments such as 'maximize', even if they ignore them.
<li> Added Wei Hu's potentials/Tables/rep_mult.c, which is used to
      totally eliminate all repmats from gaussian_CPD/update_ess.
</ul>


<li> 1/30/02
<ul>
<li> update_ess now takes hidden_bitv instead of hidden_self and
hidden_ps. This allows gaussian_CPD to distinguish hidden discrete and 
cts parents. Now learn_params_em, as well as learn_params_dbn_em,
    passes in this info, for speed.

<li> gaussian_CPD update_ess is now vectorized for any case where all
    the continuous nodes are observed (eg., Gaussian HMMs,  AR-HMMs).

<li> mk_dbn now automatically detects autoregressive nodes.

<li> hmm_inf_engine now uses indexes in marginal_nodes/family for
    speed. Marginal_ndoes can now only handle single nodes.
   (SDndx is hard-coded, to avoid the overhead of using marg_ndx,
    which is slow because of the case and global statements.)

<li> add_ev_to_dmarginal now retains the domain field.

<li> Wei Hu wrote potentials/Tables/repmat_and_mult.c, which is used to
    avoid some of the repmat's in gaussian_CPD/update_ess.

<li> installC now longer sets the global USEC, since USEC is set to 0
    by add_BNT_to_path, even if the C files have already been compiled 
    in a previous session. Instead, gaussian_CPD checks to
    see if repmat_and_mult exists, and (bat1, chmm1, water1, water2)
    check to see if jtree_C_inf_engine/collect_evidence exists.
    Note that checking if a file exists is slow, so we do the check
    inside the gaussian_CPD constructor, not inside update_ess.

<li> uninstallC now deletes both .mex and .dll files, just in case I
    accidently ship a .zip file with binaries.  It also deletes mex
   files from jtree_C_inf_engine.
   
<li> Now marginal_family for both jtree_limid_inf_engine and
	    global_joint_inf_engine returns a marginal structure and
	    potential, as required by solve_limid.
    Other engines (eg. jtree_ndx, hmm) are not required to return a potential.
</ul>



<li> 1/22/02
<ul>
<li> Added an optional argument to mk_bnet and mk_dbn which lets you
add names to nodes. This uses the new assoc_array class.

<li> Added Yimin Zhang's (unfinished) classification/regression tree
code to CPDs/tree_CPD.

</ul>



<li> 1/14/02
<ul> 
<li> Incorporated some of Shan Huang's (still broken) stable CG code.
</ul>


<li> 1/9/02
<ul>
<li> Yimin Zhang vectorized @discrete_CPD/prob_node, which speeds up
structure learning considerably. I fixed this to handle softmax CPDs.

<li> Shan Huang changed the stable conditional Gaussian code to handle
vector-valued nodes, but it is buggy.

<li> I vectorized @gaussian_CPD/update_ess for a special case.

<li> Removed denom=min(1, ... Z) from gaussian_CPD/maximize_params
(added to cope with negative temperature for entropic prior), which
gives wrong results on mhmm1.
</ul>


<li> 1/7/02

<ul>
<li> Removed the 'xo' typo from mk_qmr_bnet.

<li> convert_dbn_CPDs_to_tables has been vectorized; it is now
substantially faster to compute the conditional likelihood for long sequences.

<li> Simplified constructors for tabular_CPD and gaussian_CPD, so they
now both only take the form CPD(bnet, i, ...) for named arguments -
the CPD('self', i, ...) format is gone. Modified mk_fgraph_given_ev
to use mk_isolated_tabular_CPD instead.

<li> Added entropic prior to tabular and Gaussian nodes.
For tabular_CPD, changed name of arguments to the constructor to
distinguish Dirichlet and entropic priors. In particular,
tabular_CPD(bnet, i, 'prior', 2) is now
tabular_CPD(bnet, i, 'prior_type', 'dirichlet', 'dirichlet_weight', 2).

<li> Added deterministic annealing to learn_params_dbn_em for use with
entropic priors. The old format learn(engine, cases, max_iter) has
been replaced by learn(engine, cases, 'max_iter', max_iter).

<li> Changed examples/dynamic/bat1 and kjaerulff1, since default
equivalence classes have changed from untied to tied.
</ul>

<li> 12/30/01
<ul>
<li> DBN default equivalence classes for slice 2 has changed, so that
now parameters are tied for nodes with 'equivalent' parents in slices
1 and 2 (e.g., observed leaf nodes). This essentially makes passing in
the eclass arguments redundant (hooray!).
</ul>


<li> 12/20/01
<ul>
<li> <b>Released version 4</b>.
Version 4 is considered a major new release
since it is not completely backwards compatible with V3.
Observed nodes are now specified when the bnet/dbn is created,
not when the engine is created. This changes the interface to many of
the engines, making the code no longer backwards compatible.
Hence support for non-named optional arguments (BNT2 style) has also
been removed; hence mk_dbn etc. requires arguments to be passed by name.

<li> Ilya Shpitser's C code for triangulation now compiles under
Windows as well as Unix, thanks to Wei Hu.

<li> All the ndx engines have been combined, and now take an optional
argument specifying what kind of index to use.

<li> learn_params_dbn_em is now more efficient:
@tabular_CPD/update_ess for nodes whose families
are hidden does not need need to call add_evidence_to_dmarginal, which
is slow.

<li> Wei Hu fixed bug in jtree_ndxD, so now the matlab and C versions
both work.

<li> dhmm_inf_engine replaces hmm_inf_engine, since the former can
handle any kind of topology and is slightly more efficient. dhmm is
extended to handle Gaussian, as well as discrete,
observed nodes. The new hmm_inf_engine no longer supports online
inference (which was broken anyway).

<li> Added autoregressive HMM special case to hmm_inf_engine for
speed.

<li> jtree_ndxSD_dbn_inf_engine now computes likelihood of the
evidence in a vectorized manner, where possible, just like
hmm_inf_engine.

<li> Added mk_limid, and hence simplified mk_bnet and mk_dbn.


<li> Gaussian_CPD now uses 0.01*I prior on covariance matrix by
default. To do ML estimation, set 'cov_prior_weight' to 0.

<li> Gaussian_CPD and tabular_CPD
optional binary arguments are now set using 0/1 rather no 'no'/'yes'.

<li> Removed Shan Huang's PDAG and decomposable graph code, which will
be put in a separate structure learning library.
</ul>


<li> 12/11/01
<ul>
<li> Wei Hu fixed jtree_ndx*_dbn_inf_engine and marg_table.c.

<li> Shan Huang contributed his implementation of stable conditional
Gaussian code (Lauritzen 1999), and methods to search through the
space of PDAGs (Markov equivalent DAGs) and undirected decomposable
graphs. The latter is still under development.
</ul>


<li> 12/10/01
<ul>
<li> Included Wei Hu's new versions of the ndx* routines, which use
integers instead of doubles. The new versions are about 5 times faster
in C. In general, ndxSD is the best choice.

<li> Fixed misc/add_ev_to_dmarginal so it works with the ndx routines
in bat1.

<li> Added calc_mpe_dbn to do Viterbi parsing.

<li> Updated dhmm_inf_engine so it computes marginals.
</ul>



<li> 11/23/01
<ul>
<li> learn_params now does MAP estimation (i.e., uses Dirichlet prior,
if define). Thanks to Simon Keizer [e-mail] for spotting
this.
<li> Changed plotgraph so it calls ghostview with the output of dotty,
instead of converting from .ps to .tif. The resulting image is much
easier to read.
<li> Fixed cgpot/multiply_by_pots.m.
<li> Wei Hu fixed ind2subv.c.
<li> Changed arguments to compute_joint_pot.
</ul>


<li> 11/1/01
<ul>
<li> Changed sparse to dense in @dpot/multiply_pots, because sparse
arrays apparently cause a bug in the NT version of Matlab.

<li> Fixed the bug in gaussian_CPD/log_prob_node.m which
incorrectly called the vectorized gaussian_prob with different means
when there were continuous parents and more than one case.
(Thanks to Dave Andre for finding this.)

<li> Fixed the bug in root_CPD/convert_to_pot which did not check for
pot_type='g'. 
(Thanks to Dave Andre for finding this.)

<li> Changed calc_mpe and calc_mpe_global so they now return a cell array.

<li> Combine pearl and loopy_pearl into a single inference engine
called 'pearl_inf_engine', which now takes optional arguments passed
in using the name/value pair syntax.
marginal_nodes/family now takes the optional add_ev argument (same as
jtree), which is the opposite of the previous shrink argument.

<li> Created pearl_unrolled_dbn_inf_engine and "resurrected"
pearl_dbn_inf_engine in a simplified (but still broken!) form.

<li> Wei Hi fixed the bug in ind2subv.c, so now ndxSD works.
He also made C versions of ndxSD and ndxB, and added (the unfinished) ndxD.

</ul>


<li> 10/20/01

<ul>
<li>  Removed the use_ndx option from jtree_inf,
and created 2 new inference engines: jtree_ndxSD_inf_engine and
jtree_ndxB_inf_engine.
The former stores 2 sets of indices for the small and difference
domains; the latter stores 1 set of indices for the big domain.
In Matlab, the ndxB version is often significantly faster than ndxSD
and regular jree, except when the clique size is large.
When compiled to C, the difference between ndxB and ndxSD (in terms of
speed) vanishes; again, both are faster than compiled jtree, except
when the clique size is large.
Note: ndxSD currently has a bug in it, so it gives the wrong results!
(The DBN analogs are jtree_dbn_ndxSD_inf_engine and 
jtree_dbn_ndxB_inf_engine.)

<li> Removed duplicate files from the HMM and Kalman subdirectories.
e.g., normalise is now only in BNT/misc, so when compiled to C, it
masks the unique copy of the Matlab version.
</ul>



<li> 10/17/01
<ul>
<li> Fixed bugs introduced on 10/15:
Renamed extract_gaussian_CPD_params_given_ev_on_dps.m to
gaussian_CPD_params_given_dps.m since Matlab can't cope with such long
names (this caused cg1 to fail). Fixed bug in
gaussian_CPD/convert_to_pot, which now calls convert_to_table in the
discrete case.

<li> Fixed bug in bk_inf_engine/marginal_nodes.
The test 'if nodes < ss' is now
'if nodes <= ss' (bug fix due to Stephen [e-mail])

<li> Simplified uninstallC.
</ul>


<li> 10/15/01
<ul>

<li> Added use_ndx option to jtree_inf and jtree_dbn_inf.
This pre-computes indices for multiplying, dividing and marginalizing
discrete potentials.
This is like the old jtree_fast_inf_engine, but we use an extra level
of indirection to reduce the number of indices needed (see
uid_generator object).
Sometimes this is faster than the original way...
This is work in progress.

<li> The constructor for dpot no longer calls myreshape, which is very
slow.
But new dpots still must call myones.
Hence discrete potentials are only sometimes 1D vectors (but should
always be thought of as multi-D arrays). This is work in progress.
</ul>


<li> 10/6/01
<ul>
<li> Fixed jtree_dbn_inf_engine, and added kjaerulff1 to test this.
<li> Added option to jtree_inf_engine/marginal_nodes to return "full
sized" marginals, even on observed nodes.
<li> Clustered BK in examples/dynamic/bat1 seems to be broken,
so it has been commented out.
BK will be re-implemented on top of jtree_dbn, which should much more
efficient.
</ul>

<li> 9/25/01
<ul>
<li> jtree_dbn_inf_engine is now more efficient than calling BK with
clusters = exact, since it only uses the interface nodes, instead of
all of them, to maintain the belief state.
<li> Uninstalled the broken C version of strong_elim_order.
<li> Changed order of arguments to unroll_dbn_topology, so that intra1
is no longer required.
<li> Eliminated jtree_onepass, which can be simulated by calling
collect_evidence on jtree.
<li> online1 is no longer in the test_BNT suite, since there is some
problem with online prediction with mixtures of Gaussians using BK.
This functionality is no longer supported, since doing it properly is
too much work.
</ul>
</li>

<li> 9/7/01
<ul>
<li> Added Ilya Shpitser's C triangulation code (43x faster!).
Currently this only compiles under linux; windows support is being added.
</ul>


<li> 9/5/01
<ul>
<li> Fixed typo in CPDs/@tabular_kernel/convert_to_table (thanks,
Philippe!)
<li> Fixed problems with clamping nodes in tabular_CPD, learn_params,
learn_params_tabular, and bayes_update_params. See
BNT/examples/static/learn1 for a demo.
</ul>


<li> 9/3/01
<ul>
<li> Fixed typo on line 87 of gaussian_CPD which caused error in cg1.m
<li> Installed Wei Hu's latest version of jtree_C_inf_engine, which
can now compute marginals on any clique/cluster.
<li> Added Yair Weiss's code to compute the Bethe free energy
approximation to the log likelihood in loopy_pearl (still need to add
this to belprop). The return arguments are now: engine, loglik and
niter, which is different than before.
</ul>



<li> 8/30/01
<ul>
<li> Fixed bug in BNT/examples/static/id1 which passed hard-coded
directory name to belprop_inf_engine.

<li> Changed tabular_CPD and gaussian_CPD so they can now be created
without having to pass in a bnet.

<li> Finished mk_fgraph_given_ev. See the fg* files in examples/static
for demos of factor graphs (work in progress).
</ul>



<li> 8/22/01
<ul>

<li> Removed jtree_compiled_inf_engine,
since the C code it generated was so big that it would barf on large
models.

<li> Tidied up the potentials/Tables directory.
Removed mk_marg/mult_ndx.c,
which have been superceded by the much faster mk_marg/mult_index.c
(written by Wei Hu).
Renamed the Matlab versions mk_marginalise/multiply_table_ndx.m
to be mk_marg/mult_index.m to be compatible with the C versions.
Note: nobody calls these routines anymore!
(jtree_C_inf_engine/enter_softev.c has them built-in.)
Removed mk_ndx.c, which was only used by jtree_compiled.
Removed mk_cluster_clq_ndx.m, mk_CPD_clq_ndx, and marginalise_table.m
which were not used.
Moved shrink_obs_dims_in_table.m to misc.

<li> In potentials/@dpot directory: removed multiply_by_pot_C_old.c.
Now marginalize_pot.c can handle maximization,
and divide_by_pot.c has been implmented.
marginalize/multiply/divide_by_pot.m no longer have useC or genops options.
(To get the C versions, use installC.m)

<li> Removed useC and genops options from jtree_inf_engine.m
To use the C versions, install the C code.

<li> Updated BNT/installC.m.

<li> Added fclose to @loopy_pearl_inf/enter_evidence.

<li> Changes to MPE routines in BNT/general.
The maximize parameter is now specified inside enter_evidence
instead of when the engine is created.
Renamed calc_mpe_given_inf_engine to just calc_mpe.
Added Ron Zohar's optional fix to handle the case of ties.
Now returns log-likelihood instead of likelihood.
Added calc_mpe_global.
Removed references to genops in calc_mpe_bucket.m
Test file is now called mpe1.m

<li> For DBN inference, filter argument is now passed by name,
as is maximize. This is NOT BACKWARDS COMPATIBLE.

<li> Removed @loopy_dbn_inf_engine, which will was too complicated.
In the future, a new version, which applies static loopy to the
unrolled DBN, will be provided.

<li> discrete_CPD class now contains the family sizes and supports the
method dom_sizes. This is because it could not access the child field
CPD.sizes, and mysize(CPT) may give the wrong answer.

<li> Removed all functions of the form CPD_to_xxx, where xxx = dpot, cpot,
cgpot, table, tables.  These have been replaced by convert_to_pot,
which takes a pot_type argument.
@discrete_CPD calls convert_to_table to implement a default
convert_to_pot.
@discrete_CPD calls CPD_to_CPT to implement a default
convert_to_table.
The convert_to_xxx routines take fewer arguments (no need to pass in
the globals node_sizes and cnodes!).
Eventually, convert_to_xxx will be vectorized, so it will operate on
all nodes in the same equivalence class "simultaneously", which should
be significantly quicker, at least for Gaussians.

<li> Changed discrete_CPD/sample_node and prob_node to use
convert_to_table, instead of CPD_to_CPT, so mlp/softmax nodes can
benefit.

<li> Removed @tabular_CPD/compute_lambda_msg_fast and
private/prod_CPD_and_pi_msgs_fast, since no one called them.

<li> Renamed compute_MLE to learn_params,
by analogy with bayes_update_params (also because it may compute an
MAP estimate).

<li> Renamed set_params to set_fields
and get_params to get_field for CPD and dpot objects, to
avoid confusion with the parameters of the CPD.

<li> Removed inference/doc, which has been superceded
by the web page.

<li> Removed inference/static/@stab_cond_gauss_inf_engine, which is
broken, and all references to stable CG.

</ul>





<li> 8/12/01
<ul>
<li> I removed potentials/@dpot/marginalize_pot_max.
Now marginalize_pot for all potential classes take an optional third
argument, specifying whether to sum out or max out.
The dpot class also takes in optional arguments specifying whether to
use C or genops (the global variable USE_GENOPS has been eliminated).

<li> potentials/@dpot/marginalize_pot has been simplified by assuming
that 'onto' is always in ascending order (i.e., we remove
Maynard-Reid's patch). This is to keep the code identical to the C
version and the other class implementations.

<li> Added Ron Zohar's general/calc_mpe_bucket function,
and my general/calc_mpe_given_inf_engine, for calculating the most
probable explanation.


<li> Added Wei Hu's jtree_C_inf_engine.
enter_softev.c is about 2 times faster than enter_soft_evidence.m.

<li> Added the latest version of jtree_compiled_inf_engine by Wei Hu.
The 'C' ndx_method now calls potentials/Tables/mk_marg/mult_index,
and the 'oldC' ndx_method calls potentials/Tables/mk_marg/mult_ndx.

<li> Added potentials/@dpot/marginalize_pot_C.c and
multiply_by_pot_C.c by Wei Hu.
These can be called by setting the 'useC' argument in
jtree_inf_engine. 

<li> Added BNT/installC.m to compile all the mex files.

<li> Renamed prob_fully_instantiated_bnet to log_lik_complete.

<li> Added Shan Huang's unfinished stable conditional Gaussian
inference routines.
</ul>



<li> 7/13/01
<ul>
<li> Added the latest version of jtree_compiled_inf_engine by Wei Hu.
<li> Added the genops class by Doug Schwarz (see
BNT/genopsfun/README). This provides a 1-2x speed-up of
potentials/@dpot/multiply_by_pot and divide_by_pot.
<li> The function BNT/examples/static/qmr_compiled compares the
performance gains of these new functions.
</ul>

<li> 7/6/01
<ul>
<li> Made bk_inf_engine use the name/value argument syntax. This can
now do max-product (Viterbi) as well as sum-product
(forward-backward).
<li> Changed examples/static/mfa1 to use the new name/value argument
syntax.
</ul>


<li> 6/28/01

<ul>

<li> <b>Released version 3</b>.
Version 3 is considered a major new release
since it is not completely backwards compatible with V2.
V3 supports decision and utility nodes, loopy belief propagation on
general graphs (including undirected), structure learning for non-tabular nodes,
a simplified way of handling optional
arguments to functions,
and many other features which are described below.
In addition, the documentation has been substantially rewritten.

<li> The following functions can now take optional arguments specified
as name/value pairs, instead of passing arguments in a fixed order:
mk_bnet, jtree_inf_engine, tabular_CPD, gaussian_CPD, softmax_CPD, mlp_CPD,
enter_evidence.
This is very helpful if you want to use default values for most parameters.
The functions remain backwards compatible with BNT2.

<li> dsoftmax_CPD has been renamed softmax_CPD, replacing the older
version of softmax. The directory netlab2 has been updated, and
contains weighted versions of some of the learning routines in netlab.
(This code is still being developed by P. Brutti.)

<li> The "fast" versions of the inference engines, which generated
matlab code, have been removed.
@jtree_compiled_inf_engine now generates C code.
(This feature is currently being developed by Wei Hu of Intel (China), 
and is not yet ready for public use.)

<li> CPD_to_dpot, CPD_to_cpot, CPD_to_cgpot and CPD_to_upot 
are in the process of being replaced by convert_to_pot.

<li> determine_pot_type now takes as arguments (bnet, onodes)
instead of (onodes, cnodes, dag),
so it can detect the presence of utility nodes as well as continuous
nodes.
Hence this function is not backwards compatible with BNT2.

<li> The structure learning code (K2, mcmc) now works with any node
type, not just tabular.
mk_bnets_tabular has been eliminated.
bic_score_family and dirichlet_score_family will be replaced by score_family.
Note: learn_struct_mcmc has a new interface that is not backwards
compatible with BNT2.

<li> update_params_complete has been renamed bayes_update_params.
Also, learn_params_tabular has been replaced by learn_params, which
works for any CPD type.

<li> Added decision/utility nodes.
</ul>


<li> 6/6/01
<ul>
<li> Added soft evidence to jtree_inf_engine.
<li> Changed the documentation slightly (added soft evidence and
parameter tying, and separated parameter and structure learning).
<li> Changed the parameters of determine_pot_type, so it no longer
needs to be passed a DAG argument.
<li> Fixed parameter tying in mk_bnet (num. CPDs now equals num. equiv
classes).
<li> Made learn_struct_mcmc work in matlab version 5.2 (thanks to
Nimrod Megiddo for finding this bug).
<li> Made 'acyclic.m' work for undirected graphs.
</ul>


<li> 5/23/01
<ul>
<li> Added Tamar Kushnir's code for the IC* algorithm
(learn_struct_pdag_ic_star). This learns the
structure of a PDAG, and can identify the presence of latent
variables.

<li> Added Yair Weiss's code for computing the MAP assignment using
junction tree (i.e., a new method called @dpot/marginalize_pot_max
instead of marginalize_pot.)

<li> Added @discrete_CPD/prob_node in addition to log_prob_node to handle
deterministic CPDs.
</ul>


<li> 5/12/01
<ul>
<li> Pierpaolo Brutti updated his mlp and dsoftmax CPD classes,
and improved the HME code.

<li> HME example now added to web page. (The previous example was non-hierarchical.)

<li> Philippe Leray (author of the French documentation for BNT)
pointed out that I was including netlab.tar unnecessarily.
</ul>


<li> 5/4/01
<ul>
<li> Added mlp_CPD which defines a CPD as a (conditional) multi-layer perceptron.
This class was written by Pierpaolo Brutti.

<li> Added hierarchical mixtures of experts demo (due to Pierpaolo Brutti).

<li> Fixed some bugs in dsoftmax_CPD.

<li> Now the BNT distribution includes the whole
<a href="http://www.ncrg.aston.ac.uk/netlab/">Netlab</a> library in a
subdirectory.
It also includes my HMM and Kalman filter toolboxes, instead of just
fragments of them.
</ul>


<li> 5/2/01
<ul>
<li> gaussian_inf_engine/enter_evidence now correctly returns the
loglik, even if all nodes are instantiated (bug fix due to
Michael Robert James).

<li> Added dsoftmax_CPD which allows softmax nodes to have discrete
and continuous parents; the discrete parents act as indices into the
parameters for the continuous node, by analogy with conditional
Gaussian nodes. This class was written by Pierpaolo Brutti.
</ul>


<li> 3/27/01
<ul>
<li> learn_struct_mcmc  no longer returns sampled_bitv.
<li> Added mcmc_sample_to_hist to post-process the set of samples.
</ul>

<li> 3/21/01
<ul>
<li> Changed license from UC to GNU Library GPL.

<li> Made all CPD constructors accept 0 arguments, so now bnets can be
saved to and loaded from files.

<li> Improved the implementation of sequential and batch Bayesian
parameter learning for tabular CPDs with completely observed data (see
log_marg_lik_complete and update_params_complete). This code also
handles interventional data.

<li> Added MCMC structure learning for completely observed, discrete,
static BNs.

<li> Started implementing Bayesian estimation of linear Gaussian
nodes. See root_gaussian_CPD and
linear_gaussian_CPD. The old gaussian_CPD class has not been changed.

<li> Renamed evaluate_CPD to log_prob_node, and simplified its
arguments.

<li> Renamed sample_CPD to sample_node, simplified its
arguments, and vectorized it.

<li> Renamed "learn_params_tabular" to "update_params_complete".
This does Bayesian updating, but no longer computes the BIC score.

<li> Made routines for completely observed networks (sampling,
complete data likelihood, etc.) handle cell arrays or regular arrays,
which are faster.
If some nodes are not scalars, or are hidden, you must use cell arrays.
You must convert to a cell array before passing to an inference routine.

<li> Fixed bug in gaussian_CPD constructor. When creating CPD with
more than 1 discrete parent with random parameters, the matrices were
the wrong shape (Bug fix due to Xuejing Sun).
</ul>



<li> 11/24/00
<ul>
<li> Renamed learn_params and learn_params_dbn to learn_params_em/
learn_params_dbn_em. The return arguments are now [bnet, LLtrace,
engine] instead of [engine, LLtrace].
<li> Added structure learning code for static nets (K2, PC).
<li> Renamed learn_struct_inter_full_obs as learn_struct_dbn_reveal,
and reimplemented it to make it simpler and faster.
<li> Added sequential Bayesian parameter learning (learn_params_tabular).
<li> Major rewrite of the documentation.
</ul>

<!--
<li> 6/1/00
<ul>
<li> Subtracted 1911 off the counter, so now it counts hits from
5/22/00. (The initial value of 1911 was a conservative lower bound on the number of
hits from the time the page was created.)
</ul>
-->

<li> 5/22/00
<ul>
<li> Added online filtering and prediction.
<li> Added the factored frontier and loopy_dbn algorithms.
<li> Separated the online user manual into two, for static and dynamic
networks.
<!--
<li> Added a counter to the BNT web page, and initialized it to 1911,
which is the number of people who have downloaded my software (BNT and
other toolboxes) since 8/24/98.
-->
<li> Added a counter to the BNT web page.
<!--
Up to this point, 1911 people had downloaded my software (BNT and
other toolboxes) since 8/24/98.
-->
</ul>


<li> 4/27/00
<ul>
<li> Fixed the typo in bat1.m
<li> Added preliminary code for online inference in DBNs
<li> Added coupled HMM example
</ul>

<li> 4/23/00
<ul>
<li> Fixed the bug in the fast inference routines where the indices
are empty (arises in bat1.m).
<li> Sped up marginal_family for the fast engines by precomputing indices.
</ul>

<li> 4/17/00
<ul>
<li> Simplified implementation of BK_inf_engine by using soft evidence.
<li> Added jtree_onepass_inf_engine (which computes a single marginal)
and modified jtree_dbn_fast to use it.
</ul>

<li> 4/14/00
<ul>
<li> Added fast versions of jtree and BK, which are
designed for models where the division into hidden/observed is fixed,
and all hidden variables are discrete. These routines are 2-3 times
faster than their non-fast counterparts.

<li> Added graph drawing code
contributed by Ali Taylan Cemgil from the University of Nijmegen.
</ul>

<li> 4/10/00
<ul>
<li> Distinguished cnodes and cnodes_slice in DBNs so that kalman1
works with BK.
<li> Removed dependence on cellfun (which only exists in matlab 5.3)
by adding isemptycell. Now the code works in 5.2.
<li> Changed the UC copyright notice.
</ul>


 
<li> 3/29/00
<ul>
<li><b>Released BNT 2.0</b>, now with objects!
Here are the major changes.

<li> There are now 3 classes of objects in BNT:
Conditional Probability Distributions, potentials (for junction tree),
and inference engines. 
Making an inference algorithm (junction tree, sampling, loopy belief
propagation, etc.) an object might seem counter-intuitive, but in
fact turns out to be a good idea, since the code and documentation
can be made modular.
(In Java, each algorithm would be a class that implements the
inferenceEngine interface. Since Matlab doesn't support interfaces,
inferenceEngine is an abstract (virtual) base class.)

<p>
<li>
In version 1, instead of Matlab's built-in objects,
I used structs and a
  simulated dispatch mechanism based on the type-tag system in the
  classic textbook by Abelson
  and Sussman ("Structure and Interpretation of Computer Programs",
  MIT Press, 1985). This required editing the dispatcher every time a
  new object type was added. It also required unique (and hence long)
  names for each method, and allowed the user unrestricted access to
  the internal state of objects.

<p>
<li> The Bayes net itself is now a lightweight struct, and can be
used to specify a model independently of the inference algorithm used
to process it.
In version 1, the inference engine was stored inside the Bayes net.
              
<!--
See the list of <a href="differences2.html">changes from version
1</a>.
-->
</ul>



<li> 11/24/99
<ul>
<li> Added fixed lag smoothing, online EM and the ability to learn
switching HMMs (POMDPs) to the HMM toolbox.
<li> Renamed the HMM toolbox function 'mk_dhmm_obs_mat' to
'mk_dhmm_obs_lik', and similarly for ghmm and mhmm. Updated references
to these functions in BNT.
<li> Changed the order of return params from kalman_filter to make it
more natural. Updated references to this function in BNT.
</ul>



<li>10/27/99
<ul>
<li>Fixed line 42 of potential/cg/marginalize_cgpot and lines 32-39 of bnet/add_evidence_to_marginal
(thanks to Rainer Deventer for spotting these bugs!)
</ul>


<li>10/21/99
<ul>
<li>Completely changed the blockmatrix class to make its semantics
more sensible. The constructor is not backwards compatible!
</ul>

<li>10/6/99
<ul>
<li>Fixed all_vals = cat(1, vals{:}) in user/enter_evidence
<li>Vectorized ind2subv and sub2indv and removed the C versions.
<li>Made mk_CPT_from_mux_node much faster by having it call vectorized
ind2subv
<li>Added Sondhauss's bug fix to line 68 of bnet/add_evidence_to_marginal
<li>In dbn/update_belief_state, instead of adding eps to likelihood if 0,
we leave it at 0, and set the scale factor to 0 instead of dividing.
</ul>

<li>8/19/99
<ul>
<li>Added Ghahramani's mfa code to examples directory to compare with
fa1, which uses BNT
<li>Changed all references of assoc to stringmatch (e.g., in
examples/mk_bat_topology)
</ul>

<li>June 1999
<ul>
<li><b>Released BNT 1.0</b> on the web.
</ul>


<li>August 1998
<ul>
<li><b>Released BNT 0.0</b> via email.
</ul>


<li>October 1997
<ul>
<li>First started working on Matlab version of BNT.
</ul>

<li>Summer 1997
<ul>
<li> First started working on C++ version of BNT while working at DEC (now Compaq) CRL.
</ul>

<!--
<li>Fall 1996
<ul>
<li>Made a C++ program that generates DBN-specific C++ code
for inference using the frontier algorithm.
</ul>

<li>Fall 1995
<ul>
<li>Arrive in Berkeley, and first learn about Bayes Nets. Start using
Geoff Zweig's C++ code.
</ul>
-->

</ul>