Chaotic Time Series Prediction Based on Optimal Training Subset Online Fuzzy LSSVM

An optimal training subset online fuzzy least squares support vector machine(OTSOF-LSSVM) is proposed for chaotic time series prediction.Samples nearest to the prediction sample in both time and space are chosen to form the optimal training subset.An e-insensitive function is introduced to formulate the fuzzy membership.Thus a prediction model is established by fuzzy LSSVM.The subset and model are updated with the moving time window.Computational complexity is reduced by matrix partitioning.Experiment of predicting the time-variant chaotic time series Ikeda shows that the proposed method has better accuracy and high training speed as compared to offline and online LSSVM.