First-Order Primal-Dual Repairs for Scalable Branch and Bound

Abstract

Mixed-integer programs have found applications in many real-world problems but remain difficult to solve at scale. To improve scalability, there has been increasing interest in incorporating first-order methods as subroutines for solving node relaxations, with success primarily on problems involving specific types of constraints. However, when more general constraints are considered, first-order methods cannot be directly applied because, at finite iterations, they typically produce infeasible primal and dual solutions, preventing their direct integration into branch-and-bound. To overcome this limitation, we introduce repair procedures that extract valid primal and dual bounds from infeasible first-order iterates. These bounds enable node pruning, incumbent updates, and early stopping at any iteration. Empirically, our approach achieves one to two orders of magnitude speedups when solving real-world instances (with another order of magnitude speedup when running on GPUs) and certifies optimality on instances that are difficult for commercial solvers.