混沌时间序列的高斯过程混合模型预测

Chaotic Time Series Prediction Based on Gaussian Processes Mixture

高斯过程混合(Gaussian Processes Mixture,GPM)模型现有的学习算法如马尔科夫链蒙特卡洛法、变分法或留一法等,计算复杂度偏高,提出一种隐变量后验硬划分迭代学习算法,简化模型的学习过程,基于该算法将GPM模型用于混沌时间序列预测,并讨论嵌入维、时间延迟、学习样本和测试样本数目等参数对预测性能的影响。实验结果表明,GPM模型预测精度高于支持向量机(Support Vector Machine,SVM)、高斯过程(Gaussian Process,GP)和径向基(Radical Basis Function,RBF)网络,学习速度介于RBF网络、GP和SVM之间。

Aiming at the problem that the existing learning algorithms of Gaussian processes mixture(GPM) model, such as Markov Chain Monte Carlo(MCMC), variation or leave one out, have high computational complexity, a hidden variables posterior hard-cut iterative training algorithm is proposed,which simplifies the training process of the model. The GPM model based on the proposed algorithm is applied to chaotic time series prediction. The effects of embedding dimension, time delay, learning sample number, and testing sample numbers on predictive ability are discussed. It is demonstrated by the experimental results that the prediction of the GPM model is more accurate than SVM, GP and RBF network, and the training speed of GPM model falls in between RBF network, GP model, and SVM.