A method that accelerates Mixed-Integer Linear Program solving by learning a reduced, equivalent model (active constraints plus integer-variable values) rather than the solution directly. Its novelty is treating the relative performance (objective cost and feasibility) of all candidate reduced models on an instance as preferences and learning from them with an attention-based encoder over ranked preferences, plus a SETCOVER-based pruning to control the number of reduced-model labels. It yields large accuracy gains over prior model-reduction ML and orders-of-magnitude speedups over commercial solvers.
A machine-learning method that solves large Mixed-Integer Linear Programs (MILPs) fast by learning a reduced, equivalent model (the active constraints and integer-variable values needed to recover th…