Algorithmic Configuration by Learning and Optimization

EasyChair Preprint 2634

Abstract

We propose a methodology, based on machine learning and optimization, for selecting a solver configuration for a given instance. First, we employ a set of solved instances and configurations in order to learn a performance function of the solver. Secondly, we formulate the learning outcome as components of a mixed-integer nonlinear program, which we solve, upon the arrival of an unknown instance, in order to find the best solver configuration for that instance, based on the performance function. The main novelties of our approach are: the search problem is formulated as a mathematical program, which allows us to enforce hard dependence and compatibility constraints on the configurations; furthermore, since it contains an explicit formulation of the mathematical properties of the learning methodology, it can be solved efficiently with known optimization techniques.

Keyphrases: Algorithm Configuration, MIP solver, machine learning, mathematical programming

@booklet{EasyChair:2634,
  author    = {Claudia D'Ambrosio and Antonio Frangioni and Gabriele Iommazzo and Leo Liberti},
  title     = {Algorithmic Configuration by Learning and Optimization},
  howpublished = {EasyChair Preprint 2634},
  year      = {EasyChair, 2020}}