基于自适应在线极限学习机模型的预测方法

Study on Prediction Method Based on Adaptive Ensemble Online Sequential Extreme learning machine

本文针对单个在线极限学习机输出不稳定的情况,提出一种自适应集成在线极限学习机算法(ASEOSELM)。算法首先初始化多个在线极限学习机模型,然后根据到达的每一批次数据的训练误差及其方差自适应地调整各个在线极限学习机的集成权重,并动态删除那些小于设定阈值的模型以提高算法的训练速度,最后选择准确度高、泛化能力好的模型用于集成预测。通过函数拟合、UCI数据集以及真实股价预测实验表明,文中提出的ASE-OSELM算法相比传统的OSELM、LS-SVM和BPNN算法具有更高的预测准确度和抗干扰能力。

Since the single online sequential extreme learning machine＇ s performance is unstable,it propose an adaptive and selective OSELM. Firstly,it initializes the multiple online sequential extreme learning machine model,then adjustes adaptively the integrated weight of every online sequential extreme learning machine according to their training error and variance for each batch of data,and deletes the model that its integrated weight is smaller than the threshold to improve the training speed dynamically. Finally,the high accuracy and good generalization＇ s model will be selected for integrated prediction. Experimental results show that the ASE-OSELM has higher forecast accuracy and generalization ability than BPNN、LS-SVM and OSELM.