Bayes Nets: 'RHugin' Emulation with 'gRain'

Description

Wrappers for functions in the 'gRain' package to emulate some 'RHugin' functionality, allowing the building of Bayesian networks consisting on discrete chance nodes incrementally, through adding nodes, edges and conditional probability tables, the setting of evidence, both 'hard' (boolean) or 'soft' (likelihoods), querying marginal probabilities and normalizing constants, and generating sets of high-probability configurations. Computations will typically not be so fast as they are with 'RHugin', but this package should assist users without access to 'Hugin' to use code written to use 'RHugin'.

Author(s)

Maintainer: Peter Green <P.J.Green@bristol.ac.uk>

Examples

library(gRaven)
rm(list=ls(all=TRUE))

yn <- c("yes","no")
chest<-hugin.domain()

add.node(chest,"asia",states=yn)
add.node(chest,"tub",states=yn)
add.node(chest,"smoke",states=yn)
add.node(chest,"lung",states=yn)
add.node(chest,"bronc",states=yn)
add.node(chest,"either",states=yn)
add.node(chest,"xray",states=yn)
add.node(chest,"dysp",states=yn)

add.edge(chest,"tub","asia")
add.edge(chest,"lung","smoke")
add.edge(chest,"bronc","smoke")
add.edge(chest,"either","lung")
add.edge(chest,"either","tub")
add.edge(chest,"xray","either")
add.edge(chest,"dysp","bronc")
add.edge(chest,"dysp","either")

set.table(chest,"asia",c(0.01,0.99))
set.table(chest,"tub",c(0.05,0.95,0.01,0.99))
set.table(chest,"smoke",c(0.5,0.5))
set.table(chest,"lung",c(0.1,0.9,0.01,0.99))
set.table(chest,"bronc",c(0.6,0.4,0.3,0.7))
set.table(chest,"either",c(1,0,1,0,1,0,0,1))
set.table(chest,"xray",c(0.98,0.02,0.05,0.95))
set.table(chest,"dysp",c(0.9,0.1,0.7,0.3,0.8,0.2,0.1,0.9))

chest
get.nodes(chest)
chest$states
sapply(get.nodes(chest),function(x) get.parents(chest,x))
compile(chest)
chest
get.belief(chest,"asia")
get.belief(chest,"tub")
get.belief(chest,"lung")
get.belief(chest,"bronc")
sapply(get.nodes(chest),function(x) get.belief(chest,x)) 

c2<-clone.domain(chest)

set.finding(chest,"asia","yes")
set.finding(chest,"dysp","yes")
propagate(chest)

get.belief(chest,"asia")
get.belief(chest,"tub")
get.belief(chest,"lung")
get.belief(chest,"bronc")

pEvidence(chest$net)
get.normalization.constant(chest)

pEvidence(c2$net,evidence=list(asia="yes",dysp="yes"))
get.normalization.constant(c2)

Add or Delete Edge(s) in a gRaven domain

Description

Add or delete a directed edge from one or more parents.

Usage

add.edge(domain, child, parent)
delete.edge(domain, child, parent)

Arguments

Details

Emulates function of the same name in the RHugin package by calls to gRain functions

Value

a NULL value is invisibly returned.

Differences from RHugin

Deleting an edge nullifies the CPT for node n, if any.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo(chest.setup,package="gRaven")
chest
delete.edge(chest,"tub","asia")
add.edge(chest,"bronc","asia")
chest

Add or Delete Node in a gRaven domain

Description

Add or delete a node in a gRaven domain.

Usage

add.node(domain, name, category = c("chance", 
    "decision", "utility", "function"), kind = c("discrete", "continuous", "other"), 
    subtype, states)
delete.node(domain, name)

Arguments

Details

Emulates functions of the same name in the RHugin package by calls to gRain functions, delete.node deletes the specified node and all incident edges; nullifies all corresponding CPTs.

Value

a NULL value is invisibly returned.

Differences from RHugin

Only discrete chance nodes are currently handled in gRaven.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo(chest.setup,package="gRaven")
chest
add.node(chest,'dummy',states=1:3)
delete.node(chest,'either')
chest

Compile a gRaven domain

Description

Compile a gRaven domain

Usage

## S3 method for class 'gRaven'
compile(object, ...)
check.compiled(object)

Arguments

Details

check.compiled compiles the domain if necessary, having first ensured that necessary information such as cptables is in place. If the domain is already compiled, compile triggers a warning, but proceeds to re-initialise the grain object domain$net, in particular deleting any evidence set.

Value

a NULL value is invisibly returned.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

chest<-hugin.domain()
add.node(chest,"asia",states=c("yes","no"))
add.node(chest,"tub",states=c("yes","no"))
add.edge(chest,"tub","asia")
chest
compile(chest)
chest
set.finding(chest,"asia","yes")
set.finding(chest,"dysp","no")
propagate(chest)
chest$nodes
chest$states
chest$parents

Dummy network compression routine

Description

Dummy routine.

Usage

compress(domain)

Arguments

Details

compression in the sense used in Hugin is not available in gRain, so this function has no effect.

Value

1

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
compile(chest)
compress(chest)

Get beliefs in a gRaven domain

Description

Get beliefs (marginal joint probabilities) in a gRaven domain

Usage

get.belief(domain, nodes)
get.marginal(domain, nodes, class = c("data.frame", "table", "ftable", "numeric"))

Arguments

Details

Emulates functions of the same name in the RHugin package by calls to gRain functions. Unlike with RHugin, gRaven conditions on all entered evidence in reporting probabilities, not only propagated evidence.

Value

For get.marginal, a list with one component "table" which is a data frame, table, flat table or numeric vector of marginal joint probabilities, as specified by class. For get.belief, a vector in the case of a single node (if there is no evidence on that node. a vector of ones), otherwise as get.marginal.

Differences from RHugin

get.belief handles more than one node at a time.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
compile(chest)
chest
set.finding(chest,"asia","yes")
set.finding(chest,"dysp","no")
propagate(chest)
get.belief(chest,"asia")
get.belief(chest,"tub")
get.marginal(chest,c('asia','xray','tub'))

Get the nodes of a gRaven domain

Description

Get the nodes of a gRaven domain

Usage

get.nodes(domain)

Arguments

Details

Emulates function of the same name in the RHugin package by calls to gRain functions

Value

Character vector of node names.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
get.nodes(chest)
chest$states
sapply(get.nodes(chest),function(x) get.parents(chest,x))

Get the normalisation constant of a gRaven domain

Description

Get the normalisation constant of a gRaven domain

Usage

get.normalization.constant(domain, log = FALSE)

Arguments

Value

numeric, the value of the normalisation constant (or its logarithm). If no evidence has been entered, the constant is defined as 1.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

yn <- c("yes","no")
chest<-hugin.domain()

add.node(chest,"asia",states=yn)
add.node(chest,"tub",states=yn)
add.node(chest,"smoke",states=yn)
add.node(chest,"lung",states=yn)
add.node(chest,"bronc",states=yn)
add.node(chest,"either",states=yn)
add.node(chest,"xray",states=yn)
add.node(chest,"dysp",states=yn)

add.edge(chest,"tub","asia")
add.edge(chest,"lung","smoke")
add.edge(chest,"bronc","smoke")
add.edge(chest,"either","lung")
add.edge(chest,"either","tub")
add.edge(chest,"xray","either")
add.edge(chest,"dysp","bronc")
add.edge(chest,"dysp","either")

set.table(chest,"asia",c(0.01,0.99))
set.table(chest,"tub",c(0.05,0.95,0.01,0.99))
set.table(chest,"smoke",c(0.5,0.5))
set.table(chest,"lung",c(0.1,0.9,0.01,0.99))
set.table(chest,"bronc",c(0.6,0.4,0.3,0.7))
set.table(chest,"either",c(1,0,1,0,1,0,0,1))
set.table(chest,"xray",c(0.98,0.02,0.05,0.95))
set.table(chest,"dysp",c(0.9,0.1,0.7,0.3,0.8,0.2,0.1,0.9))

compile(chest)
get.normalization.constant(chest)

set.finding(chest,"asia","yes")
set.finding(chest,"dysp","yes")
get.normalization.constant(chest)

Get the parents of a node in a gRaven domain

Description

Get the parents of a node in a gRaven domain

Usage

get.parents(domain, n, type = "parents")

Arguments

Value

A named list of character vectors of names of parents.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
get.nodes(chest)
chest$states
sapply(get.nodes(chest),function(x) get.parents(chest,x))

Get states, edges and children in a gRaven domain

Description

Get states, edges and children in a gRaven domain

Usage

get.states(domain, nodes=domain$nodes)
get.children(domain, nodes)
get.edges(domain, nodes=domain$nodes)

Arguments

Details

Emulates functions of the same name in the RHugin package by calls to gRain functions

Value

For get.states, vector of state values. For get.edges, a list with one element for each node in domain. Each element is in turn a list with a single element edges which is a character vector of names of the node's children. An empty vector indicates that the node has no children. For get.children, character vector of children's node names in the case of a single node, otherwise list of children of each node, each element being a character vector of names of the node's children.

Differences from RHugin

get.states and get.children allow more than one node; get.edges allows selecting nodes.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
get.states(chest,"asia")
get.children(chest,"asia")
get.edges(chest)

Create or Clone a gRaven Domain

Description

Create or Clone a gRaven domain object.

Usage

hugin.domain()
clone.domain(domain)

Arguments

Details

Emulates functions of the same name in the RHugin package. A gRaven domain is an environment, with additional class attribute 'gRaven'. The environment holds structures such as nodes, states, parents and cptables, populated incrementally by functions in the package, preparing the information needed by the gRain package to create and manipulate the structure net in the domain, which is a grain object.

Value

character string naming a gRaven domain

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

chest<-hugin.domain()

yn <- c("yes","no")
add.node(chest,"asia",states=yn)
add.node(chest,"smoke",states=yn)
add.node(chest,"tub",states=yn)
add.node(chest,"lung",states=yn)
add.node(chest,"either",states=yn)

add.edge(chest,"tub","asia")
add.edge(chest,"lung","smoke")
add.edge(chest,"either",c("lung","tub"))

set.table(chest,"asia",c(0.01,0.99))
set.table(chest,"tub",c(0.05,0.95,0.01,0.99))
set.table(chest,"smoke",c(0.5,0.5))
set.table(chest,"lung",c(0.1,0.9,0.01,0.99))
set.table(chest,"either",c(1,0,1,0,1,0,0,1))

chest

chest2<-clone.domain(chest)

add.node(chest2,"bronc",states=yn)
add.node(chest2,"dysp",states=yn)
add.node(chest2,"xray",states=yn)

add.edge(chest2,"bronc","smoke")
add.edge(chest2,"dysp",c("bronc","either"))
add.edge(chest2,"xray","either")

set.table(chest2,"bronc",c(0.6,0.4,0.3,0.7))
set.table(chest2,"dysp",c(0.9,0.1,0.7,0.3,0.8,0.2,0.1,0.9))
set.table(chest2,"xray",c(0.98,0.02,0.05,0.95))

chest2

Re-initialise a gRaven domain

Description

Re-initialise a gRaven domain

Usage

initialize.domain(domain)

Arguments

Details

Emulates function of the same name in the RHugin package. Restores the domain to the state it was in after the call to compile, in particular erasing all entered evidence.

Value

a NULL value is invisibly returned.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
compile(chest)
chest
set.finding(chest,"asia","yes")
set.finding(chest,"dysp","no")
propagate(chest)
sapply(get.nodes(chest),function(x) get.belief(chest,x)) 
get.normalization.constant(chest)

initialize.domain(chest)
sapply(get.nodes(chest),function(x) get.belief(chest,x)) 
get.normalization.constant(chest)

List gRaven domain Objects

Description

List gRaven domain Objects

Usage

list.domains(print=TRUE)

Arguments

Details

gRaven domains are R environment objects, with named slots typically including nodes, parents, cptables and net, where the net component is a gRain grain object. The domains that this function finds and lists are all the objects in .GlobalEnv of class "gRaven", and all those objects in this class that are components of a component named 'domains' of a list object in .GlobalEnv.

Value

The function invisibly returns a character vector of the domain names found.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo(chest,package="gRaven")
chest
chest2<-clone.domain(chest)
chest2

set.finding(chest,"asia","yes")
set.finding(chest,"dysp","no")
propagate(chest)
get.belief(chest,"asia")
get.belief(chest,"tub")

propagate(chest2)
get.belief(chest2,"asia")
get.belief(chest2,"tub")

list.domains()

Get belief in a gRaven domain

Description

Find the configurations of the specified nodes that occur with probability pmin or greater. These configurations are known as most probable configurations or maximum a posteriori (MAP) configurations.

Usage

map.configurations(domain, nodes, pmin)

Arguments

Details

Emulates function of the same name in the RHugin package by calls to gRain functions

Value

Data.frame with one column for each node in nodes. Each row contains a most probable configuration. The final column of the data.frame (Prob) gives the probability of the configuration.

Author(s)

Therese Graversen, theg@itu.dk

Examples

demo("chest",package="gRaven",echo=FALSE)
set.finding(chest,"asia","yes")
set.finding(chest,"dysp","no")
map.configurations(chest,c("lung","bronc"),0.0001)

Print a gRaven domain

Description

Print method for a gRaven domain

Usage

## S3 method for class 'gRaven'
print(x, ...)

Arguments

Details

Prints summary information describing the domain.

Value

a NULL value is invisibly returned.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo(chest,package="gRaven")
print(chest)

Propagate method for a gRaven domain

Description

Propagate method for a gRaven domain

Usage

## S3 method for class 'gRaven'
propagate(object, ...)

Arguments

Details

Propagates all previously entered evidence through the network. This call is required by functions such as get.belief and map.configurations, but not before get.normalization.constant. In the implementation, findings will have been accumulated in the variable net$cache in the domain, by calls to set.finding, and are moved to net$evid by the propagate function.

Value

a NULL value is invisibly returned.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
compile(chest)
chest
set.finding(chest,"asia","yes")
set.finding(chest,"dysp","no")
propagate(chest)
get.belief(chest,"asia")
get.belief(chest,"tub")

Set, retract and inspect findings

Description

Set, retract and inspect findings

Usage

set.finding(domain, node, finding)
retract(domain, nodes=domain$nodes)
get.finding(domain, nodes=domain$nodes, type = c("entered", "propagated"), 
	namestates=FALSE)

Arguments

Details

For type, partial matching is used.

Value

For set.finding and retract, a NULL value is invisibly returned. For get.finding, the likelihood vector is invisibly returned for each specified node, in a format governed by namestates. This is a named vector in the case of a single node, otherwise a list of such vectors. For a single node, in the absence of any evidence, a named vectors of 1's is returned.

Differences from RHugin

Findings (or evidence) are handled differently in gRaven (and gRain) than in RHugin, and that is reflected in differences in results from these functions, especially get.finding. For set.finding in gRaven, the case argument is not supported. For get.finding in both packages, previously-set evidence is displayed as a non-negative vector indexed by the states, typically a probability distribution; two enhancements in gRaven are that findings on more than one node can be displayed, and that the format of the output can be controlled by namestates. In gRaven when new evidence is set by set.finding, it replaces any existing evidence on the same node, and all existing evidence on all nodes is "unpropagated". Evidence is held in a data structure cache until propagation, when it is moved to evid; these names are used in labelling the output from get.finding. Finally if no evidence has been entered on a node, RHugin reports a vector of all ones, while gRaven returns such a vector invisibly.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo(chest,package="gRaven",echo=FALSE)
chest
set.finding(chest,"asia","yes")
set.finding(chest,"dysp","no")
propagate(chest)

get.belief(chest,"asia")
get.belief(chest,"tub")
get.finding(chest)

retract(chest,"dysp")
get.belief(chest,"tub")
get.finding(chest)

Set or get CPT in a gRaven domain

Description

Set or get CPT in a gRaven domain

Usage

set.table(domain, n, tab = 1, type = c("cpt", "experience", 
    "fading"))
get.table(domain, n, type = c("cpt", "experience", 
    "fading"), class = c("data.frame", "table", 
    "ftable", "numeric"))

Arguments

Value

For set.table, a NULL value is invisibly returned; for get.table, an object of required class.

Differences from RHugin

Only type="cpt" and class="data.frame" are currently implemented in gRaven.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

chest<-hugin.domain()
add.node(chest,"asia",states=c("yes","no"))
add.node(chest,"tub",states=c("yes","no"))
add.edge(chest,"tub","asia")
compile(chest)
chest
get.table(chest,"asia")
set.table(chest,"asia",c(0.01,0.99))
get.table(chest,"asia")

method for a gRaven domain

Description

Simulate method for a gRaven domain

Usage

## S3 method for class 'gRaven'
simulate(object, nsim = 1, seed = NULL, ...)

Arguments

Value

a data frame.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
compile(chest)
chest
set.finding(chest,"asia","yes")
set.finding(chest,"dysp","no")
newdata<-simulate(chest,100)
head(newdata)

Summary method for a gRaven domain

Description

Summary method for a gRaven domain

Usage

## S3 method for class 'gRaven'
summary(object, domain = TRUE, nodes = FALSE, jt = FALSE, print.cliques = FALSE, 
    ...)

Arguments

Value

a data frame.

Differences from RHugin

In gRaven, not all of the options are yet implemented, and generally the information delivered is less detailed than in the RHugin version.

Author(s)

Peter J. Green, P.J.Green@bristol.ac.uk

Examples

demo("chest",package="gRaven",echo=FALSE)
chest
summary(chest,jt=TRUE)