Added max_lag() / set_max_lag() to
knowledge(), which forbids edges between variables that are
more than a given number of tiers apart. max_lag() is set
inside tier(), alongside its formulas:
tier(1 ~ V1, 2 ~ V2, max_lag(1)).
Tetrad algorithms now expose more of Tetrad’s parameters, all defaulting to Tetrad’s own defaults:
grasp() gains use_score,
allow_internal_randomness, output_cpdag and
seed. Set allow_internal_randomness = FALSE
for reproducible results.grasp_fci() gains use_score,
use_heuristic, exclude_selection_bias and
seed.pc() gains collider_orientation_style,
allow_bidirected and fdr_q.fci() gains collider_orientation_style,
do_possible_dsep, exclude_selection_bias and
fdr_q.rfci() gains exclude_selection_bias and
fdr_q.ges() gains num_threads;
boss() gains num_threads and
seed.boss_fci(), gfci() and
sp_fci() gain exclude_selection_bias.plot() gains equal_node_size. By
default, all nodes are now drawn the same size, fitted to the longest
label. Use equal_node_size = FALSE for the previous
behaviour, where each node is sized to its own label.
tpc_run() now supports
orientation_method = "standard" and
"maj_rule".
make_tikz() gains:
style = "classic": plain-text nodes, gray tier bands,
and a time axis.node_labels and tier_labels for LaTeX
labels, e.g. c(X1 = "$X_1$").plot(), and bent edges leaving a node together are spread
apart. An edge whose bend would pass through a node is bent the other
way, or by the nearest angle to bend that clears it.
edge_style = list(route = FALSE) turns both adjustments
off.bend sets the bend of all edges in degrees
(0 = straight, negative = the other way), or of individual
edges with a list, e.g.
list(default = 20, A = list(B = 0)).var,
link, tier, tier label) instead
of every node/every path, which also styled
the tiers.orientation_method = "maj.rule" in
tpc_run() and tfci_run() is deprecated in
favour of "maj_rule".
make_tikz(bend_edges, bend_angle) are deprecated in
favour of bend:
bend_edges = TRUE, bend_angle = 15 is now
bend = 15, and bend_edges = FALSE is
bend = 0.
verify_tetrad() no longer reports Tetrad as not
installed after a successful install_tetrad() when the
TETRAD_DIR environment variable is set. The (undocumented)
TETRAD_DIR is no longer used; set
options(causalDisco.tetrad_cache = ) to use a custom
directory instead.
make_tikz() now draws PAG edge marks
(o->, o-o, --o) instead of
warning.
make_tikz() with bent edges now bends edges in both
directions (A --> B, B --> A) to
opposite sides instead of drawing them on top of each other.
install_tetrad(version = ) now saves a non-default
Tetrad version in its own cache directory instead of the default
version’s, and Java loads only the JAR for the configured
version.
tpc_run() now orients conservatively by default, as
documented; it previously used standard orientation.
Removed Tetrad algorithm arguments that Tetrad ignored, so setting them had no effect:
pc(): conflict_rule,
guarantee_cpdagges(): parallelized (use
num_threads)rfci(): stable_fas,
complete_rule_set_usedgrasp_fci(): stable_fas,
covered_depthsp_fci(): depthTetradSearch$set_alg(): cpc and
pc_max (conflict_rule,
guarantee_cpdag, use_heuristic,
max_disc_path_length), fcit
(guarantee_pag, max_disc_path_length),
restricted_boss (allow_internal_randomness)
and ccd (depth).use_heuristic in gfci() and
sp_fci() had no effect because it set the wrong Tetrad
parameter. It now controls the max-p heuristic as documented.
boss_fci()’s num_threads was never
passed to Tetrad. It is now, and its default changes from 0 to Tetrad’s
default of 1.
pc(engine = "pcalg") no longer errors when the
estimated graph contains a cycle. Now emits a message and returns the
graph with its class downgraded to "UNKNOWN".
Added support for the RFCI algorithm via rfci(),
interfacing with implementations from pcalg and Tetrad.
The infix edge operators %-->% and
%!-->% in knowledge() now accept
+ on both sides to specify multiple variables,
e.g. A + B %-->% C + D is equivalent to
c(A, B) %-->% c(C, D).
The infix edge operators %-->% and
%!-->% in knowledge() now support
tidyselect set operations such as !, &,
and | on either side, e.g.
child_x1 %!-->% !starts_with("child").
You can now plug a fully custom, user-written search algorithm
into the causalDisco engine and run it through disco() (via
CausalDiscoSearch$set_alg() and
make_method()). See the “Extending causalDisco with new
algorithms” article for a worked example.
Added register_engine(), which lets you plug in an
entirely new engine backend (i.e. one not built on bnlearn, causalDisco,
pcalg, or Tetrad) so. See the “Extending causalDisco with new
algorithms” article for a worked example.
The summary() methods for Knowledge and
Disco objects are deprecated; use print()
instead.
The directed_as_undirected argument of
tpc_run(), tfci_run(), and
CausalDiscoSearch$set_knowledge() is deprecated and
ignored; see the related bug-fix entry below for the new handling of
directed forbidden edges.
The internal-only directed_as_undirected_knowledge
argument of make_runner() has been removed, and the
directed_as_undirected argument of
PcalgSearch$set_knowledge() and
as_pcalg_constraints() is deprecated and ignored.
as_pcalg_constraints() no longer returns a
fixed_edges matrix. pcalg’s fixedEdges
argument only protects an undirected adjacency in the skeleton, so a
directed required edge (A %-->% B) cannot be honored
through it.
Fixed tpc() and tfci() silently
treating directed forbidden edges as undirected: forbidding
i %!-->% j also forbidden j !--> i.
Forbidden edges now follow the same rule as the pcalg engine: both
directions must be supplied explicitly, otherwise an error is raised.
The exception is if the Knowledge forbidden knowledge is
equivalent to tiered knowledge.
Fixed knowledge() dropping variables from
tier() formulas when infix edge operators
(%-->%, %!-->%) appeared before the
tier() call and had already registered some (but not all)
of the tier’s variables. This only affected the no-data-frame form of
knowledge().
The print() methods for Knowledge and
Disco objects now give a more concise and readable
summary.
Disco objects now store an attribute
graph_class of the actual learned graph class.
Knowledge objects generated from
knowledge() now verifies the knowledge for requires edges
doesn’t contain a directed cycle.
Improved the documentation.
Reduced the number of package dependencies.
Exports list_registered_tetrad_algorithms() to view
the custom registered Tetrad algorithms.
Allows the user to specify custom conditional independence tests
for use with the engines bnlearn, causalDisco, and pcalg. The custom
test should have the signature
function(x, y, conditioning_set, suff_stat) (optionally
with an additional args argument for additional
parameters), and return a p-value. See the documentation and the
associated vignette for details.
Added support for micd tests such as
micd::gaussCItestMI() for pcalg and causalDisco engine. See
all the new available tests in ?PcalgSearch and
?CausalDiscoSearch under the test field.
Simplified the interface for extending causalDisco with new
algorithms. Exports make_method() and
make_runner() to do this. Please see the documentation and
the associated vignette for details.
Improved speed of tfci(), tges(), and
tpc() algorithms by at least 3x, and often much
more.
Updated knowledge_to_caugi() to include information
about required and forbidden edges in the output.
Improve documentation for many functions.
This release should be considered essentially a new package with many new features, and some breaking changes (mainly renaming functions, arguments, and object classes) to be consistent.
Please see the vignettes and documentation for details on using the new features.