Mixed integer programming (MIP) solvers employ many heuristics to solve large-scale instances that are encountered in practice. However, these heuristics are general-purpose and are designed to have reasonable performance across a broad distribution of MIP applications. In this talk we show that deep neural networks trained on data from a single problem family can learn application-specific heuristics that significantly outperform generic, application-agnostic policies on a holdout set. This goes some way towards eliminating the need for human experts to spend time and effort handcrafting application-specific policies, which up to now has been required for some time-sensitive applications. We demonstrate the improvement on several large, real-world application datasets, covering a range of problem sizes and application areas.
Copyright. All Rights Reserved.