Interpretability of tree-aware branching policies

Ascertain how the neural branching policy of Zarpellon et al. (2021), which parameterizes branch-and-bound search trees using solver statistics and search-tree descriptors, leverages these features to compute branching variable scores; identify the contributions of specific features and characterize decision behavior across different parts of the tree.

Background

To improve generalization across heterogeneous MILP classes, Zarpellon et al. (2021) propose incorporating search-tree context and solver statistics into the variable scoring process. While this approach shows promise, the survey notes that it is unclear how the proposed features are used by the neural scoring function, reflecting a broader challenge of interpretability in learned branching strategies.

Clarifying the mapping from features to scores would help understand when and why tree-aware policies succeed or fail, guide feature design, and potentially inform adaptive strategies that switch behaviors depending on tree state.