compile {jti}

Compile information

Description

Compiled objects are used as building blocks for junction tree inference

Usage

compile(
  x,
  evidence = NULL,
  root_node = "",
  joint_vars = NULL,
  tri = "min_fill",
  pmf_evidence = NULL,
  alpha = NULL,
  initialize_cpts = TRUE
)

## S3 method for class 'cpt_list'
compile(
  x,
  evidence = NULL,
  root_node = "",
  joint_vars = NULL,
  tri = "min_fill",
  pmf_evidence = NULL,
  alpha = NULL,
  initialize_cpts = TRUE
)

Arguments

x

An object returned from cpt_list (baeysian network) or pot_list (decomposable markov random field)

evidence

A named vector. The names are the variabes and the elements are the evidence.

root_node

A node for which we require it to live in the root clique (the first clique).

joint_vars

A vector of variables for which we require them to be in the same clique. Edges between all these variables are added to the moralized graph.

tri

The optimization strategy used for triangulation if x originates from a Baeysian network. One of

• 'min_fill'

• 'min_rfill'

• 'min_sp'

• 'min_ssp'

• 'min_lsp'

• 'min_lssp'

• 'min_elsp'

• 'min_elssp'

• 'min_nei'

• 'minimal'

• 'alpha'

pmf_evidence

A named vector of frequencies of the expected missingness of a variable. Variables with frequencies of 1 can be neglected; these are inferrred. A value of 0.25 means, that the given variable is expected to be missing (it is not a evidence node) in one fourth of the future cases. Relevant for tri methods 'min_elsp' and 'min_elssp'.

alpha

Character vector. A permutation of the nodes in the graph. It specifies a user-supplied eliminination ordering for triangulation of the moral graph.

initialize_cpts

TRUE if the CPTs should be initialized, i.e. multiplied together to form the clique potentials. If FALSE, the compiled object will save the triangulation and other information that needs only bee computed once. Herafter, it is possible to enter evidence into the CPTs, using set_evidence, saving a lot of computations.

Details

The Junction Tree Algorithm performs both a forward and inward message pass (collect and distribute). However, when the forward phase is finished, the root clique potential is guaranteed to be the joint pmf over the variables involved in the root clique. Thus, if it is known in advance that a specific variable is of interest, the algortihm can be terminated after the forward phase. Use the root_node to specify such a variable and specify propagate = "collect" in the juntion tree algortihm function jt.

Moreover, if interest is in some joint pmf for variables that end up being in different cliques these variables must be specified in advance using the joint_vars argument. The compilation step then adds edges between all of these variables to ensure that at least one clique contains all of them.

Evidence can be entered either at compile stage or after compilation. Hence, one can also combine evidence from before compilation with evidence after compilation. Before refers to entering evidence in the 'compile' function and after refers to entering evidence in the 'jt' function.

Finally, one can either use a Bayesian network or a decomposable Markov random field (use the ess package to fit these). Bayesian networks must be constructed with cpt_list and decomposable MRFs can be constructed with both pot_list and cpt_list. However, pot_list is just an alias for cpt_list which handles both cases internally.

Examples

cptl <- cpt_list(asia2)
cp1  <- compile(cptl, evidence = c(bronc = "yes"), joint_vars = c("bronc", "tub"))
print(cp1)
names(cp1)
dim_names(cp1)
plot(get_graph(cp1))