Wikipedia
defines a graphical model as follows: A graphical model is a
probabilistic model for which a graph denotes the conditional
independence structure between random variables. They are commonly
used in probability theory, statistics  particularly Bayesian
statistics and machine learning.
A supplementary view is that graphical models are based on
exploiting conditional independencies for constructing complex
stochastic models with a modular structure. That is, a complex
stochastic model is built up by simpler building blocks.
This task view is a collection of packages intended to supply R code
to deal with graphical models.
The packages can be roughly structured into the following topics
(although several of them have functionalities which go across these categories):
Representation, manipulation and display of graphs

diagram
Visualises simple graphs (networks) based on a transition matrix,
utilities to plot flow diagrams, visualising webs, electrical
networks, ...

DiagrammeR
Create Graph Diagrams and Flowcharts Using R

dynamicGraph
Interactive graphical tool for manipulating graphs

graph
A package that implements some simple graph handling capabilities.

gRbase
The gRbase package provides certain general constructs which are used
by other graphical modelling packages. This includes 1) the
concept of gmData (graphical meta data), 2) several graph
algorithms 3) facilities for table operations, 4) functions for
testing for conditional independence. gRbase also illustrates
how hierarchical loglinear models (hllm) may be implemented.

igraph
Routines for simple graphs, network analysis.

network
Tools to create and modify network objects. The network class can
represent a range of relational data types, and supports arbitrary
vertex/edge/graph attributes.

Rgraphviz
Provides plotting capabilities for R graph objects.

RBGL
A fairly extensive and comprehensive interface to the graph algorithms contained in the BOOST library.
(based on graph
objects from the graph package).
Classical models  General purpose packages

ggm
Fitting graphical Gaussian models.

gRbase
The gRbase package provides certain general constructs
which are used by other graphical modelling packages (in
particular by gRain).
This includes 1) the concept of gmData (graphical meta
data), 2) several graph algorithms 3) facilities for
table operations, 4) functions for testing for
conditional independence. gRbase also illustrates how
hierarchical loglinear models (hllm) may be
implemented.
Link: doi:10.18637/jss.v014.i17

gRim
Implements graphical interaction models for contingency tables
(i.e. loglinear models) and graphical Gaussian models for the
multivariate normal data (i.e. covariance selection models) and
mixed interaction models.
Miscellaneous: Model search, structure learning, specialized types of models etc.

BDgraph Bayesian Graph Selection Based on BirthDeath
MCMC Approach. Bayesian inference for structure learning in
undirected graphical models. The main target is to uncover
complicated patterns in multivariate data wherein either
continuous or discrete variables.

bnstruct Bayesian Network Structure Learning from
Data with Missing Values.
Bayesian Network Structure Learning from Data with Missing
Values. The package implements the SilanderMyllymaki
complete search, the MaxMin ParentsandChildren, the
HillClimbing, the MaxMin Hillclimbing heuristic searches,
and the Structural ExpectationMaximization
algorithm. Available scoring functions are BDeu, AIC,
BIC. The package also implements methods for generating and
using bootstrap samples, imputed data, inference.

catnet
A package that handles discrete Bayesian network models and provides
inference using the frequentist approach

FBFsearch
Algorithm for searching the space of Gaussian directed acyclic
graphical models through moment fractional Bayes factors

GeneNet
Modeling and Inferring Gene Networks. GeneNet
is a package for analyzing gene expression (time series) data with
focus on the inference of gene networks.

huge Highdimensional Undirected Graph Estimation.

lvnetlvnet: Latent Variable Network Modeling.
Estimate, fit and compare Structural Equation Models (SEM) and
network models (Gaussian Graphical Models; GGM) using
OpenMx. Allows for two possible generalizations to include GGMs
in SEM: GGMs can be used between latent variables (latent
network modeling; LNM) or between residuals (residual network
modeling; RNM).

MXM Feature Selection (Including Multiple Solutions)
and Bayesian Networks. Feature selection methods for
identifying minimal, statisticallyequivalent and
equallypredictive feature subsets. Bayesian network algorithms
and related functions are also included. The package name 'MXM'
stands for "Mens eX Machina", meaning "Mind from the Machine" in
Latin. Link: doi:10.18637/jss.v080.i07

ndtv Network Dynamic Temporal Visualizations. Renders
dynamic network data from "networkDynamic" objects as animated
movies or other representations of relational structure and node
attributes that change over time.

networkDynamic
Dynamic Extensions for Network Objects. Simple interface routines to facilitate the
handling of network objects with complex intertemporal
data. "networkDynamic" is a part of the "statnet" suite of
packages for network analysis.

parcor
The package estimates the matrix of partial correlations based on different regularized
regression methods: lasso, adaptive lasso, PLS, and Ridge Regression. In addition, the package
provides model selection for lasso, adaptive lasso and Ridge
regression based on
crossvalidation.

pcalg
Standard and robust estimation of the skeleton (ugraph) and the equivalence class
of a Directed Acyclic Graph (DAG) via the PCAlgorithm. The equivalence class is
represented by its (unique) Completed Partially Directed Acyclic Graph (CPDAG).

qp
This package is deprecated and it is now only a stub for the newer
version called qpgraph available through the Bioconductor
project. The qorder partial correlation graph search algorithm,
qpartial, or qp, algorithm for short, is a robust procedure for
structure learning of undirected Gaussian graphical Markov
models from "small n, large p" data, that is, multivariate
normal data coming from a number of random variables p larger
than the number of multidimensional data points n as in the case
of, e.g., microarray data.

qpgraph
qorder partial correlation graphs, or qpgraphs for short, are
undirected Gaussian graphical Markov models that represent
qorder partial correlations. They are useful for learning
undirected graphical Gaussian Markov models from data sets where
the number of random variables p exceeds the available sample
size n as, for instance, in the case of microarray data where
they can be employed to reverse engineer a molecular regulatory
network.

QUIC
Regularized sparse inverse covariance matrix estimation. Use
Newton's method and coordinate descent to solve the regularized
inverse covariance matrix estimation problem.

SIN
This package provides routines to perform SIN model selection as
described in Drton and Perlman (2004). The selected models are
represented in the format of the 'ggm' package, which allows in
particular parameter estimation in the selected model.
Bayesian Networks/Probabilistic expert systems

abn
A graphical modelling formulation is used to construct Bayesian
regression models for analyses of multivariate data.

bnlearn
Bayesian network structure learning via constraintbased (also known as
'conditional independence') and scorebased algorithms. This package implements the
GrowShrink (GS) algorithm, the Incremental Association (IAMB) algorithm, the
InterleavedIAMB (InterIAMB) algorithm, the FastIAMB (FastIAMB) algorithm, the
MaxMin Parents and Children (MMPC) algorithm and the HillClimbing (HC) greedy
search algorithm for both discrete and Gaussian networks, along with many score
functions and conditional independence tests. Some utility functions (model
comparison and manipulation, random data generation, arc orientation testing) are
also included.

gRain
A package for probability propagation in graphical
independence networks, also known as probabilistic
expert systems (which includes Bayesian networks as
a special case).
Link: doi:10.18637/jss.v046.i10

RHugin
The Hugin Decision Engine (HDE) is commercial software produced by
HUGIN EXPERT A/S for building and making inference from Bayesian
belief networks. The RHugin package provides a suite of
functions allowing the HDE to be controlled from within the R
environment for statistical computing. The RHugin package can
thus be used to build Bayesian belief networks, enter and
propagate evidence, and to retrieve beliefs. Additionally, the
RHugin package can read and write hkb and NET files, making it
easy to work simultaneously with both the RHugin package and the
Hugin GUI. A licensed copy of the HDE (or the trial version) is
required for the RHugin package to function, hence the target
audience for the package is Hugin users who would like to take
advantage of the statistical and programmatic capabilities of R.
Notice: RHugin is NOT on CRAN.
Link: http://rhugin.rforge.rproject.org/

sparsebn Fast methods for learning sparse Bayesian
networks from highdimensional data using coordinate descent and
sparse regularization. Designed to handle mixed experimental and
observational data with thousands of variables with either
continuous or discrete observations.
BUGS models

bayesmix
Bayesian mixture models of univariate Gaussian distributions using
JAGS.

dclone
Data Cloning and MCMC Tools for Maximum Likelihood Methods. Low
level functions for implementing maximum likelihood estimating
procedures for complex models using data cloning and Bayesian
Markov chain Monte Carlo methods with support for JAGS, WinBUGS
and OpenBUGS. Parallel MCMC computation is supported and can
result in considerable speedup.

boa
boa: Bayesian Output Analysis Program (BOA) for MCMC. A
menudriven program and library of functions for carrying out
convergence diagnostics and statistical and graphical analysis of
Markov chain Monte Carlo sampling output.

BRugs
BRugs: R interface to the OpenBUGS MCMC software.
Fullyinteractive R interface to the OpenBUGS software for Bayesian
analysis using MCMC sampling. Runs natively and stably in 32bit
R under Windows. Versions running on Linux and on 64bit R under
Windows are in "beta" status and less efficient.

coda
coda: Output analysis and diagnostics for MCMC.
Output analysis and diagnostics for Markov Chain Monte Carlo simulations.

ergm
ergm: Fit, Simulate and Diagnose ExponentialFamily Models for
Networks.
An integrated set of tools to analyze and simulate networks based on
exponentialfamily random graph models (ERGM). "ergm" is a part of
the statnet suite of packages for network analysis.

R2OpenBUGS
R2OpenBUGS: Running OpenBUGS from R.
Using this package, it is possible to call a BUGS model, summarize
inferences and convergence in a table and graph, and save the
simulations in arrays for easy access in R.

R2WinBUGS
Running WinBUGS and OpenBUGS from R / SPLUS.
Using this package, it is possible to call a BUGS model, summarize
inferences and convergence in a table and graph, and save the
simulations in arrays for easy access in R / SPLUS. In SPLUS,
the openbugs functionality and the windows emulation functionality
is not yet available.

rbugs
Fusing R and OpenBugs.
Functions to prepare files needed for running BUGS in batchmode, and
running BUGS from R. Support for Linux and Windows systems with
OpenBugs is emphasized.

rjags
rjags: Bayesian graphical models using MCMC.
Interface to the JAGS MCMC library.