K插值单纯形法核极限学习机的研究  

Kernel Extreme Learning Machine Based on K Interpolation Simplex Method

作　　者：苏一丹[1] 李若愚 覃华[1] 陈琴[1] SU Yidan;LI Ruoyu;QIN Hua;CHEN Qin(College of Computer and Electronic Information,Guangxi University,Nanning 530004,Chin)

机构地区：[1]广西大学计算机与电子信息学院,南宁530004

出　　处：《电子与信息学报》2018年第8期1860-1866,共7页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61762009)~~

摘　　要：针对核极限学习机高斯核函数参数选优难,影响学习机训练收敛速度和分类精度的问题,该文提出一种K插值单纯形法的核极限学习机算法。把核极限学习机的训练看作一个无约束优化问题,在训练迭代过程中,用Nelder-Mead单纯形法搜索高斯核函数的最优核参数,提高所提算法的分类精度。引入K插值为Nelder-Mead单纯形法提供合适的初值,减少单纯形法的迭代次数,提高了新算法的训练收敛效率。通过在UCI数据集上的仿真实验并与其它算法比较,新算法具有更快的收敛速度和更高的分类精度。The kernel Extreme Learning Machine （ELM） has a problem that the kernel parameter of the Gauss kernel function is hard to be optimized. As a result, training speed and classification accuracy of kernel ELM are negatively affected. To deal with that problem, a novel kernel ELM based on K interpolation simplex method is proposed. The training process of kernel ELM is considered as an unconstrained optimal problem. Then, the Nelder-Mead Simplex Method （NMSM） is used as an optimal method to search the optimized kernel parameter, which improves the classification accuracy of kernel ELM. Furthermore, the K interpolation method is used to provide appropriate initial values for the Nelder-Mead simplex to reduce the number of iterations, and as a result, the training speed of ELM is improved. Comparative results on UCI dataset demonstrate that the novel ELM algorithm has better training speed and higher classification accuracy.

关 键 词：核极限学习机 核参数 Nelder—Mead单纯形法 旆值法 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]