Year: 2024
Type: conference-paper
Abstract: Presolving has become an important component of modern Mixed-integer linear programming (MILP) solvers. Empirically, it has been observed that the performance of the solver is significantly influenced by the presolving algorithm (presolver), and selecting an appropriate combination of presolvers can improve solving efficiency. In industry, it is common to manually control the switch of presolvers to find a more efficient combination. However, with the emergence of new presolvers, it has become increasingly challenging to manually implement an optimal selection strategy in the vast combination space. Therefore, this paper proposes presolver selection, which needs to consider two key issues: (P1) How many presolvers should be selected? (P2) Which presolvers should be preferred among them? To address this challenge, this paper uses a hierarchical sequence model (HEM) to learn the presolver selection strategy through reinforcement learning. Specifically, the high-level model learns how many presolvers should be selected, and the low-level model learns to select a subset of presolvers within the determined size. The experimental results show that the method used can solve (P1) and (P2) better scompared to the designed baseline. It effectively improves the performance of the solver on real-world and synthetic MILPs.
Authors: Wentao Song, Naijie Gu
Institution: University of Science and Technology of China
Language: English
Subfield: Numerical Analysis
Field: Mathematics
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4399418743