All notable changes to policytree will be documented in this file.
The format is based on Keep a Changelog and this project adheres to Semantic Versioning.
hybrid_policy_tree
consistent with policy_tree
(level first). #156
hybrid_policy_tree
for building deeper trees by
using policy_tree
to look ahead search.depth
(typically equal to 2) splits. #118
policy_tree
in case the input type (Gamma, X
) is beyond what is
computationally feasible with exact tree search. #116, #129
double_robust_scores
for GRF’s
causal_survival_forest
. #126
double_robust_scores
. #125
IMPORTANT Some of these changes might cause small differences in results compared to previous releases, even if the same random seed is used.
policytree
to use GRF version 2.0.0. The “one vs
all” multi_causal_forest
is deprecated and we instead
utilize the new GRF estimator multi_arm_causal_forest
which
supports multiple treatment arms natively.
multi_causal_forest
will continue to work until the next
release, but dispatches to multi_arm_causal_forest
and
emits a warning. Note that this allows for a drop-in replacement in
workflows that rely on calls to double_robust_scores
for
policy learning, but not for workflows involving point predictions
(predict(forest)
) as the new GRF estimator will for K
treatment arms predict
a K-1 contrast matrix. #67
leaf.label
argument to
plot.policy_tree
allowing custom treatment names. #60
"type = node.id"
argument to
predict.policy_tree
returning the leaf node the test sample
falls into. #78.min.node.size
to policy_tree
specifying the smallest permissible node size. #77
split.step
approximation for discrete covariates by
redefining to skip observations instead of unique values. Note: this may
cause small differences in results compared to previous releases when
using the split.step
approximation parameter. #73
pow
call to retain CRAN
compatibility. #54
help(<package-name>)
in the R
package. #41