基于证据框架的模糊超球面支持向量机超参数优化

Hyperparameter Optimization of Fuzzy Hypersphere Support Vector Machine Based on Evidence Framework

模糊超球面支持向量机(FHS-SVM)在处理一类分类问题时比超平面支持向量机泛化能力更强,特别是在雷达目标检测中得到了成功应用.FHS-SVM训练时需要预设一些超参数,不同的超参数得到的FHS-SVM性能差异很大.文中首先证明了FHS-SVM训练过程与证据框架第一层贝叶斯推理的等价性,然后在证据框架下提出了FHS-SVM超参数优化迭代方法.基于超宽带合成孔径雷达探雷数据,通过与穷举方法结果的对比检验了迭代优化方法的有效性.

Fuzzy hypersphere support vector machine （FHS-SVM） has stronger generalization capability than hyperplane support vector machine in the one-class classification problem, being successful in radar target detection. Some hyperparameters have to be predefined before the FHS-SVM training, with different hyperparameters leading to significant difference in the FHS-SVM performance. In this paper, equivalence between FHS-SVM training and the level 1 Bayesian inference of the evidence framework is proved. Then, an FHS-SVM hyperparameter optimization iteration method is proposed based on the evidence framework. Using landmine detection data obtained with uhra-wide band synthetic aperture radar, the proposed iteration method is verified by comparing it with an exhaustive search method.