AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
Interpretable Cutting Plane Selection for Mixed-Integer Linear Programming Using Genetic Programming
Cutting planes are a key ingredient in mixed-integer linear programming (MILP) solvers, but current handcrafted cut strategies are static, while neural-network-based learned strategies offer limited interpretability and no convergence guarantees. The authors propose GPSCS, the first learning-based framework that generalizes handcrafted heuristics into a fully interpretable policy: genetic programming evolves a symbolic scoring function that ranks candidate cuts, and a lightweight statistics-based mechanism adaptively decides how many top-ranked cuts to add each separation round. Evaluated in SCIP on seven classes of synthetic and real-world benchmarks whose evaluation instances are on average 41 times harder than the training ones, GPSCS improves average solving time by 6% over SCIP's default and achieves a two-fold speedup over both a recent neural approach and a no-cut baseline.