abstract bibtex

Perhaps surprisingly, it is possible to predict how long an algorithm will take to run on previously unseen input data, using machine learning techniques to build a model of the algorithm's runtime as a function of domain-specific problem features. Such models have important applications to algorithm analysis, portfolio-based algorithm selection, and the automatic configuration of parameterized algorithms. Over the past decade, a wide variety of techniques have been studied for building such models. Here, we describe extensions and improvements of previous models, new families of models, and –- perhaps most importantly –- a much more thorough treatment of algorithm parameters as model inputs. We also describe novel features for predicting algorithm runtime for the propositional satisfiability (SAT), mixed integer programming (MIP), and travelling salesperson (TSP) problems. We evaluate these innovations through the largest empirical analysis of its kind, comparing to all previously proposed modeling techniques of which we are aware. Our experiments consider 11 algorithms and 35 instance distributions; they also span a very wide range of SAT, MIP, and TSP instances, with the least structured having been generated uniformly at random and the most structured having emerged from real industrial applications. Overall, we demonstrate that our new models yield substantially better runtime predictions than previous approaches in terms of their generalization to new problem instances, to new algorithms from a parameterized space, and to both simultaneously.

@Article{Hutter2012, author = {Hutter, Frank and Xu, Lin and Hoos, Holger H and Leyton-Brown, Kevin}, title = {Algorithm Runtime Prediction: The State of the Art}, journal = {arXiv.org}, volume = {cs.AI}, number = {}, pages = {}, year = {2012}, abstract = {Perhaps surprisingly, it is possible to predict how long an algorithm will take to run on previously unseen input data, using machine learning techniques to build a model of the algorithm\'s runtime as a function of domain-specific problem features. Such models have important applications to algorithm analysis, portfolio-based algorithm selection, and the automatic configuration of parameterized algorithms. Over the past decade, a wide variety of techniques have been studied for building such models. Here, we describe extensions and improvements of previous models, new families of models, and --- perhaps most importantly --- a much more thorough treatment of algorithm parameters as model inputs. We also describe novel features for predicting algorithm runtime for the propositional satisfiability (SAT), mixed integer programming (MIP), and travelling salesperson (TSP) problems. We evaluate these innovations through the largest empirical analysis of its kind, comparing to all previously proposed modeling techniques of which we are aware. Our experiments consider 11 algorithms and 35 instance distributions; they also span a very wide range of SAT, MIP, and TSP instances, with the least structured having been generated uniformly at random and the most structured having emerged from real industrial applications. Overall, we demonstrate that our new models yield substantially better runtime predictions than previous approaches in terms of their generalization to new problem instances, to new algorithms from a parameterized space, and to both simultaneously.}, location = {}, keywords = {}}

Downloads: 0