基于CMA—ES算法的支持向量机模型选择

研究模型选择对支持向量机(SVM)的泛化性能有着重要影响。针对传统梯度算法对初始值敏感及网格搜索法计算复杂的缺点,为了提高全面优化能力和分类精度,提出了一种基于协方差矩阵自适应进化策略(CMA-ES)的支持向量机(SVM)模型优化算法,通过对SVM泛化性能界(Bounds on Generalization Performance)的优化求解,实现了基于CMA-ES算法的SVM模型选择。在标准数据集上的实验结果表明:相比遗传算法和梯度算法,上述方法能够在较小计算代价下得到更优的超参数,提高支持向量机的预测精度稳定性,尤其适合大样本数据条件下的模型选择。

Model Selection for SVMs Based on CMA-ES Algorithm

Model selection plays a key role in SVM application.Traditional methods,such as the gradient based method and grid search method,respectively surfer from the sensitivity to the initial point and intensive computations.A new model selection method is proposed in this paper based on the Covariance Matrix Adaptation-Evolution Strategy(CMA-ES) using the bounds on generalization performance of SVM.Compared with the Genetic Algorithms(GA) and the Broyden-Fletcher-Goldfarb-Shanno(BFGS) method,the experimental resu...