Adaptive Genetic Algorithm to Select Training Data for Support Vector Machines

This paper presents a new adaptive genetic algorithm (AGA) to select training data for support vector machines (SVMs). SVM training data selection strongly influences the classification accuracy and time, especially in the case of large and noisy data sets. In the proposed AGA, a population of solutions evolves with time. The AGA parameters, including the chromosome length, are adapted according to the current state of exploring the solution space. We propose a new multi-parent crossover operator for an efficient search. A new metric of distance between individuals is introduced and applied in the AGA. It is based on the fast analysis of the vectors distribution in the feature space obtained using principal component analysis. An extensive experimental study performed on the well-known benchmark sets along with the real-world and artificial data sets, confirms that the AGA outperforms a standard GA in terms of the convergence capabilities. Also, it reduces the number of support vectors and allows for faster SVM classification.