MipConfigBench: A dataset for learning in the space of Mixed-Integer Programming algorithms

Aditya Paliwal, Nick Doudchenko, Miles Lubin, Ross Anderson

Abstract: Algorithm configuration (AC), algorithm selection (AS), and parallel algorithm portfolio (AP) are techniques for systematically choosing algorithm parameters to improve performance. Respectively, AC finds a single configuration with good performance over a family of instances, AS predicts good configurations on a per-instance basis by observing instance features, and parallel AP identifies combinations of multiple parameter settings that achieve better performance than any individual parameter setting. See [8] for a recent survey of these general techniques. In the context of mixed-integer programming (MIP), commercial solver developers have very recently begun to recognize and exploit the promise of choosing algorithm parameters based on instance features [1, 2]