【作者】 张思源； 谢林柏；

【Author】 Siyuan Zhang;Linbo Xie;Jiangnan University,School of Internet of Things Engineering;

【机构】 江南大学物联网工程学院；

【摘要】 增量型随机权神经网络是一类前馈型神经网络,其模型中的神经元在构建中采用依次增加的方式生成,且输入权值和隐含层神经元的偏置都是随机产生,其中每一次新增节点的过程分为两个步骤:首先是通过各种优化算法设定隐含层节点参数,然后在输出层通过最小二乘法更新所有的输出权值.上述过程存在的问题是两个计算步骤中没有考虑隐含层和输出层权值的同步优化问题,其结果是产生了大量冗余无效的节点,降低了算法的收敛速度和模型的泛化能力.针对这一重要问题,本文提出了一种新的参数融合优化方法,主要创新性是:(1)针对新增节点提出新的融合优化指标,确保隐含层与输出层权值的同步优化,给出了对应的参数融合优化的随机权神经网络学习算法(IOI-RVFL).(2)提出了一种新的模型学习算法收敛性分析方法,得到了随机赋权搜索应满足的约束条件.实验结果表明,相比已有的增量型随机权神经网络,本文的模型具有更紧凑的网络结构和更优的泛化能力.更多还原

【Abstract】 Incremental random vector functional-link network(I-RVFL) is a kind of feed-forward neural network,whose model is constructed by adding the neurons gradually and the input weights and bias of the hidden layer neurons are randomly generated.There are two phases in constructing the newly added nodes.One is to obtain the hidden weights with different optimization algorithms.The other is to adjust all the output weights by the least squares method.Nevertheless,it has a basic deficiency in the aforementioned construction scheme that there is no guarantee on the simultaneous optimization of the weights in the hidden and output layers,which may produce a lot of redundant nodes in the final model and affect the convergence speed and generalization ability of the obtained model.In this paper,a new integrated optimization method is proposed and the main innovations are:(1) a new integrated optimization index is developed to construct the simultaneously optimized weights in the first and second phases and the corresponding integrated optimization incremental random vector functional-link network(termed as IOI-RVFL) is also established;(2) a novel convergence analysis method is developed to provide a constraint searching condition for the stochastic selection of the parameters.Experimental results show that our proposed algorithm has a more compact network structure and better generalization than other algorithms.更多还原