Performance prediction using support vector machine for the configuration of optimization algorithms

The aim of this paper is to propose a machine learning approach for predicting the performance of each configuration of optimization algorithms. Our approach consists of advocating making the decision of finding the most suitable configuration on a per-instance analysis based on a supervised machine learning model. That is, it consists of building a support vector machine (SVM) model to predict the performance of each configuration on each instance and then to select the adapted setting depending on the instance. Furthermore, feature selection has been used as a pre-processing step to select the relevant features in order to enhance the predictive capacity of SVM. The experiment consists of predicting algorithm performance metrics for two well known optimization problems using SVM in its continuous and binary form depending on the metric of each problem.