合作交流 / 学术活动

【06-25】知行融创论坛:Parallel MIP Solving with Dynamic Task Decomposition

知行融创论坛-实验室月度学术交流
Speaker: 林鹏
Time: 2025年6月25日
Venue: 中国科学院软件园区5号楼三层 334报告厅
Abstract: Mixed Integer Programming (MIP) is a foundational model in operations research.
Although significant progress has been made in enhancing sequential MIP solvers through sophisticated techniques and heuristics, remarkable developments in computing resources have made parallel solving a promising direction for performance improvement.
In this work, we propose a novel parallel MIP solving framework that employs dynamic task decomposition in a divide-and-conquer paradigm.
Our framework incorporates a hardness estimate heuristic to identify challenging solving tasks and a reward decaying mechanism to reinforce the task decomposition decision.
We apply our framework to two state-of-the-art open-source MIP solvers, SCIP and HiGHS, yielding efficient parallel solvers.
Extensive experiments on the full MIPLIB benchmark, using up to 128 cores, demonstrate that our framework yields substantial performance improvements over modern divide-and-conquer parallel solvers.
Moreover, our parallel solvers have established new best known solutions for 16 open MIPLIB instances.
Download Parallel MIP Solving with Dynamic Task Decomposition