Ensemble classification for constraint solver configuration

Abstract

The automatic tuning of the parameters of algorithms and automatic selection of algorithms has received a lot of attention recently. One possible approach is the use of machine learning techniques to learn classifiers which, given the characteristics of a particular problem, make a decision as to which algorithm or what parameters to use. Little research has been done into which machine learning algorithms are suitable and the impact of picking the "right" over the "wrong" technique. This paper investigates the differences in performance of several techniques on different data sets. It furthermore provides evidence that by using a meta-technique which combines several machine learning algorithms, we can avoid the problem of having to pick the "best" one and still achieve good performance. © 2010 Springer-Verlag.