Collaborative Proposal: Machine Learning Aided Global Optimization of MINLP
Abstract
Mixed-integer nonlinear programming (MINLP) has widespread applications in chemical engineering, power system, machine learning, and cybersecurity. Most of these applications involve mixed-integer variables and generic nonconvex nonlinear constraints. Compared with the state-of-the-art mixed-integer linear programming (MILP) solvers that are scalable to millions of variables and constraints, global optimization solvers for MINLP are less mature and typically scalable to at most thousands of variables and constraints. However, MINLP applications arising from modern energy, manufacturing, and cybersecurity problems can involve hundreds of thousands of variables. For these high-stakes decision-making problems, solving MINLP to global optimality in a reasonable time is strongly preferred to guarantee safe operations and bring economic savings. In this document, we propose to use machine learning techniques to accelerate the solution of MINLP. We focus on four important tasks in global optimization algorithms: (1) learning to select and predictconvex relaxations# (2) learning to tighten variable bounds# (3) learning to branch# and (4) learning to decompose. We will developnovel machine learning architectures such as graph neural networks and novel reinforcement learning algorithms. We will implement these tasks as part of the Ecole package.Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 09, 2024
- Source ID
- N000142412645
Entities
People
- Andrea Lodi
Organizations
- Cornell University
- Office of Naval Research
- United States Navy