An Efficient Framework for Automatic Algorithm Selection using Meta-Learning
Mohammed Elmahgiubi · The Atrium (University of Guelph) · 2016
With the unprecedented growth of Information and communication technology (ICT) industry, core networking devices become highly stringent elements of the network due to the increase of packet classification (PC) requirements. Although many PC algorithms with variable performances and capabilities are available, no single algorithm is guaranteed to outperform every other one in every case. This research provides a generic and efficient framework for algorithm selection using Meta-Learning and Artificial Neural Networks (ANN). The developed framework was tested in different scenarios comprising different PC algorithms with different performance measures. Using ANN as the learning model and 10-fold cross validation as the evaluation criteria, the framework was able to achieve an average accuracy of 92.5% on predicting the most suitable algorithm that maximizes classification speed for an unseen ruleset, and 88% when minimizing memory footprint on a larger set of algorithms using the same evaluation criteria.