jt_nbinary_ops and new triangulation
methods.jt_binary_ops breaking api slightly. The algorithm is
now much faster though.compile gets a new argumen initialize_cpts
in order to speed up computations when insertion of evidence at CPT
level are of interest.initialize to initialize an object from
compile if initialize_cpts was set to
FALSEpot_list deprecated. Use cpt_list for both
BNs and MRFsjt_nbinary_ops more than twice as fast.pot_list for markov random fields which
is more efficient and idiomaticjoint_vars is specified, the root node is
automatically set to the clique where these variables are located.
Hence, one only needs to collect to query probabilies about
these variablesjt
algorithm now proceeds assuming a uniform prior distribution on the
affected tables. In this regard, one can not query the evidence since it
has no meaning. The print method flags if there are inconsistencies and
it can be obtained by has_inconsistencies. This new feature
means, that jti can now also be seen as a
“machine-learning” algorithm that can be very useful in connection with
class-prediction e.g.new_mpd renamed to mpd. Works on
cpt_list objects now.triangulate method for cpt_list
objects.evidence, for triangulation.jt_binary_ops for calculating the number
of binary operations to perform a full message passingmpd for finding maximal prime
decompositions is now include includedtriangulate a graph before
compilation in order to investigate the size of the cliques etc.There was a bug in the creation of the junction tree when calling Kruskals algorithm.
It is now possible to specify variables of interest in advance, such that we are guaranteed to be able to query the joint pmf of these variables.
Some refactoring making compilation much faster. When potentials is assigned to a clique we no longer start by creating a unity table and then multiply. This was killing the advantage of the sparsity.