摘要
针对NPSVR训练速度和预测精度问题,提出一种基于L1范数损失的非平行支持向量回归机L1NPSVR模型,用于预测数值输出。L1NPSVR通过求解两个较小规模的凸规划问题,建立一个ε_(1)-不敏感的下界函数和一个ε_(2)-不敏感的上界函数。在L1NPSVR模型中,每个优化问题同时最小化训练样本的L1范数损失和铰链损失,以保证模型的稳定性,减轻噪声和异常值的影响。L1NPSVR通过求解一对更小的优化问题来提高模型的运行效率。仿真结果验证了所提出方法的可行性及有效性。
Aiming at the problem of training speed and prediction accuracy of NPSVR,a novel nonparallel support vector regression(L1NPSVR)based on norm-L1 loss is proposed for predicting numerical outputs.L1NPSVR constructs aε_(1)-insensitive lower bound function and aε_(2)-insensitive upper bound function by solving two smaller quadratic programming problems.In the L1NPSVR,each optimization problem simultaneously minimizes thenorm-L1 based losses and hinge losses,which can ensure the stability of the model and reduce the adverse of noise and outliers.L1NPSVR improves the efficiency of the model by solving a pair of smaller optimization problems.The simulation results demonstrate the feasibility and effectiveness of the proposed method.
作者
刘历铭
巩荣芬
储茂祥
LIU Liming;GONG Rongfen;CHU Maoxiang(School of Electronic and Information Engineering,University of Science and Technology Liaoning,Anshan 114051,China)
出处
《辽宁科技大学学报》
CAS
2023年第2期101-110,共10页
Journal of University of Science and Technology Liaoning
基金
辽宁省自然科学基金(2022-MS-353)
辽宁省教育厅项目(2020LNZD06,LJKMZ20220640)。
