Learning to Accelerate the Global Optimization of Quadratically-Constrained Quadratic Programs

We learn optimal instance-specific heuristics for the global minimization of nonconvex quadratically-constrained quadratic programs (QCQPs). Specifically, we consider partitioning-based mixed-integer programming relaxations for nonconvex QCQPs and propose the novel problem of strong partitioning to optimally partition variable domains without sacrificing global optimality. We design a local optimization method for solving this challenging max-min strong partitioning problem and replace this expensive benchmark strategy with a machine learning (ML) approximation for homogeneous families of QCQPs. We present a detailed computational study on randomly generated families of QCQPs, including instances of the pooling problem, using the open-source global solver Alpine. Our numerical experiments demonstrate that strong partitioning and its ML approximation significantly reduce Alpine's solution time by factors of 3.5 - 16.5 and 2 - 4.5 on average and by maximum factors of 15 - 700 and 10 - 200, respectively, over the different QCQP families.

Article

Download

View Learning to Accelerate the Global Optimization of Quadratically-Constrained Quadratic Programs