Support Vector Machine Parameters Optimization by Enhanced Fireworks Algorithm

Support vector machines are widely used as superior classifiers for many different applications. Accuracy of the constructed support vector machine classifier depends on the proper parameter tuning. One of the most common used techniques for parameter determination is grid search. This optimization can be done more precisely and computationally more efficiently by using stochastic search metaheuristics. In this paper we propose using enhanced fireworks algorithm for support vector machine parameter optimization. We tested our approach on standard benchmark datasets from the UCI Machine Learning Repository and compared the results with grid search and with results obtained by other swarm intelligence approaches from the literature. Enhanced fireworks algorithm proved to be very successful, but most importantly it significantly outperformed other algorithms for more realistic cases for which there were separate test sets, rather than doing only cross validation.