【作者】 褚菲； 卢新宇； 张倩； 陈俊龙； 王雪松； 马小平；

【Author】 Chu Fei;Lu Xin-yu;Zhang Qian;Chen Jun-long;Wang Xue-song;Ma Xiao-ping;Research Center of Underground Space Intelligent Control Engineering of the Ministry of Education,China University of Mining and Technology;School of Information and Control Engineering,China University of Mining and Technology;State Key Laboratory of Automatic Control Technology for Mining and Metallurgy Process,Beijing General Research Institute of Mining & Metallurgy;School of Computer Science and Engineering,South China University of Technology;Faculty of Science and Technology,University of Macau;

【机构】 中国矿业大学地下空间智能控制教育部工程研究中心； 中国矿业大学信息与控制工程学院； 北京矿冶科技集团有限公司矿冶过程自动控制技术国家重点实验室； 华南理工大学计算机科学与工程学院； 澳门大学科技学院；

【摘要】 该文提出了一种基于自适应lasso正则化技术的宽度学习系统(Broad learning system,BLS),将原始BLS网络的L2范数惩罚项替换为自适应lasso,选取该正则化技术应用于BLS网络的原因有:优秀的正则化方法在变量选择上应具有连续性,无偏性,稀疏性以及Oracle等性质,而L1范数对权重的参数估计是有偏的,且对所有权重施加一致的惩罚,导致增加网络的预测误差。当BLS网络节点过多时,L2范数则保留了BLS网络中所有节点的信息,无法降低数据维度,不满足变量选择的稀疏性。而自适应lasso则同时满足以上性质,改善了以上正则化技术的缺陷。由于自适应lasso引入自适应权重,相当于对输出权重进行二次惩罚,对于重要的权重施加较小的惩罚,对于作用较小的权重施加较大的惩罚。当BLS网络节点过多时,这种正则化技术不仅在一定程度上精简了BLS的网络结构,而且提高了BLS网络的预测精度,改善了BLS泛化性。本文通过对一些回归数据集进行实验,与原有的几种正则化技术比较,实验表明该方法既能稀疏BLS网络结构,也能提高BLS网络的泛化性能。更多还原

【Abstract】 This paper proposes a broad learning system based on adaptive lasso regularization.（Broad learning system,BLS）,This method replaces the L2-norm penalty term of the original BLS with adaptive lasso.The reasons for choosing this regularization method to apply the BLS are followed:An excellent regularization method should have the properties of continuity,unbiasedness,sparsity and Oracle in variable selection.The L1-norm is biased in the parameter estimation of the weights and imposes a consistent penalty on all the weights,That increase the prediction error of the network.When there are too many nodes in the BLS,the L₂-norm retains the information of all the nodes in the BLS,which unable to reduce the data dimensionality and does not satisfy the sparsity of variable selection,The adaptive lasso satisfies the above properties and improves the drawback of the above regularization method.Since adaptive lasso have adaptive weights,it is equivalent to penalty the output weights again.A smaller penalty is imposed on important weights,and large penalty is imposed on unimportant weights.When there are too many nodes in the BLS,this regularization method not only simplifies the BLS to a certain extent,but also improves the prediction accuracy and generalization performance of the BLS.Through the experiments on some regression datasets,Compared with the original regularization methods,Experiments show that this method can not only sparse the BLS,but also improve the generalization performance of the BLS.更多还原