Evaluating new functionality for inclusion in bnlearn requires many small (and big) decisions for which the optimal choice, if any, is not obvious nor available in the literature. Below are a number of small simulation studies which were used to choose default argument values and to compare the trade-offs alternative implementations of specific algorithms.
- What is a good default value for the gamma parameter of the extended BIC
score for discrete networks (
"ebic")? - Does permuting the columns of the data while bootstrapping improve the
quality of the averaged network structures produced by
boot.strength()andaveraged.network()? - How to scale SHD and SID to a common scale, for instance, for plotting?
