Module rules 

Rules that decide, at each node, whether the DL8.5 search goes on.

The search consults two sets of rules. Node rules (common) encode the problem: maximum depth, minimum support, pure nodes and bounds. Search rules restrict the search to make it anytime: they cut branches such as the ones beyond a discrepancy budget (DiscrepancyRule, LDS-DL8.5) or beyond the k best-ranked features (TopkRule, Top-k-DL8.5).

When a relaxable rule has cut part of the search, the search is restarted with every rule relaxed (its budget widened), until no rule cuts anything and the tree is optimal. This is the CA-DL8.5 framework of Kiossou and Schaus, A Generic Complete Anytime Beam Search for Optimal Decision Trees (IDA 2026).

Modules

common

The rules that define the DL8.5 problem, plus the time limit. None of them is relaxable except, optionally, the time limit.

Structs

DecreasingTopkRule

Like TopkRule, but the number of features allowed halves at each level: k / 2^depth, and at least one.

DiscrepancyRule

Limited discrepancy search (LDS-DL8.5).

Exponential

1, b, b², b³, …

GainRule

Restricts a pass to the paths whose accumulated heuristic loss stays within a gap.

Luby

Running sums of the Luby sequence 1, 1, 2, 1, 1, 2, 4, …, scaled by a multiplier, as used for restarts in SAT solvers.

Monotonic

0, n, 2n, 3n, …

PurityRule

Stops at nodes that are already pure enough.

RuleContext

What the rules know about the node being evaluated.

RuleManager

An ordered set of rules, evaluated by decreasing priority.

RuleResult

What a rule decided for a node.

TopkRule

Top-k search: at each node, a pass only branches on the k + 1 best-ranked features. Relaxing the rule raises k, following its StepStrategy, up to limit.

Enums

Reason

Why the search stopped at a node.

RuleState

Whether a rule is applied.

Traits

Rule

A condition checked at every node of the search.

StepStrategy

A sequence of budgets used to relax a rule between passes.