Approximate Optimal Active Learning of Decision Trees

Abstract

We consider the problem of actively learning an unknown binary decision tree using only membership queries, a setting in which the learner must reason about a large hypothesis space while maintaining formal guarantees. Rather than enumerating candidate trees or relying on heuristic impurity or entropy measures, we encode the entire space of bounded-depth decision trees symbolically in SAT formulas. We propose a symbolic method for active learning of decision trees, in which approximate model counting is used to estimate the reduction of the hypothesis space caused by each potential query, enabling near-optimal query selection without full model enumeration. The resulting learner incrementally strengthens a CNF representation based on observed query outcomes, and approximate model counter ApproxMC is invoked to quantify the remaining version space in a sound and scalable manner. Additionally, when ApproxMC stagnates, a functional equivalence check is performed to verify that all remaining hypotheses are functionally identical. Experiments on decision trees show that the method reliably converges to the correct model using only a handful of queries, while retaining a rigorous SAT-based foundation suitable for formal analysis and verification.