融合概率分布和单调性的支持向量回归算法

发布时间：2018-06-03 15:20

本文选题：支持向量回归 + 概率分布　；参考：《控制理论与应用》2017年05期

【摘要】：传统支持向量回归是单纯基于样本数据的输入输出值建模,仅使用样本数据信息,未充分利用其他已知信息,模型泛化能力不强.为了进一步提高其性能,提出一种融合概率分布和单调性先验知识的支持向量回归算法.首先将对偶二次规划问题简化为线性规划问题,在求解时,加入与拉格朗日乘子相关的单调性约束条件;通过粒子群算法优化惩罚参数和核参数,优化目标包括四阶矩估计表示的输出样本概率分布特性.实验结果表明,融合这两部分信息的模型,能使预测值较好地满足训练样本隐含的概率分布特性及已知的单调性,既提高了预测精度,又增加了模型的可解释性.
[Abstract]:The traditional support vector regression is an input and output Zhi Jianmo only based on sample data, only using sample data information and not fully utilizing other known information, the model generalization ability is not strong. In order to further improve its performance, a support vector regression algorithm which combines probability distribution and monotonicity prior knowledge is proposed. First, the dual two times are combined. The programming problem is simplified as a linear programming problem. In the solution, the monotonicity constraint conditions related to the Lagrange multiplier are added. The particle swarm optimization is used to optimize the penalty parameters and the kernel parameters. The optimization target includes the probability distribution characteristics of the output samples represented by the four order moment estimation. The experimental results show that the model which combines the two parts of information can make the preview of the model. The measured values satisfactorily satisfy the implicit probability distribution characteristics and known monotonicity of training samples, which not only improve the prediction accuracy, but also increase the interpretability of the models.
【作者单位】：华东理工大学化工过程先进控制和优化技术教育部重点实验室;
【基金】：国家自然科学基金项目(21176073) 国家“973”计划(2013CB733605)资助~~
【分类号】：TP18

【相似文献】