利用神经元拓展正则极端学习机预测时间序列

为实现对于时间序列预测数据的准确预测,提出一种神经元拓展正则极端学习机(NERELM,Neuron-Expanding Regularized Extreme Learning Machine),并研究了其在时间序列预测中的应用.NERELM根据结构风险最小化原理权衡经验风险与结构风险,以逐次拓展隐层神经元的方式自动确定最佳的网络结构,以避免传统神经网络训练过程中需人为确定网络结构的弊端.应用于时间序列的仿真结果表明:NERELM可有效实现对于RELM最佳网络结构的自动确定,具有预测精度高与计算速度快的优点.

Time series prediction using neuron-expanding regularized extreme learning machine

A new algorithm called neuron-expanding regularized extreme learning machine (NERELM) was proposed and applied to time series prediction. In order to enhance the generalization performance of NERELM, the empirical risk and the structural risk were balanced on the basis of structural risk minimization theory. The output weights of NERELM were analytically determined at extremely fast learning speed instead of using gradient-based learning algorithm, and the optimal network structure of NERELM was automatically determined by expanding its hidden neuron nodes iteratively. Experiments on time series prediction indicate that NERELM has better performance in training computation cost and prediction accuracy in comparison with conventional gradient-based neural networks.