Prediction of chaotic time-series based on online wavelet support vector regression

Support vector regression (SVR) is an effective method for the predication of chaotic time-series, which is a fundamental topic of nonlinear dynamics. Through analyzing the possible variation of support vector sets after new-samples are inserted to the training set, a novel SVR algorithm is proposed; thus an online learning algorithm is set up. In connection with the specific characteristics of chaotic signals, a wavelet kernel satisfying wavelet frames is also presented. The wavelet kernel can approximate arbitrary functions, and is especially suitable for local processing; hence the generalization ability of SVR is improved. To illustrate the good performance of the online wavelet SVR, a benchmark problem, i.e. the online prediction of chaotic Mackey-Glass time-series, is considered. The simulation results indicate that the online wavelet SVR algorithm outperforms the existing algorithms in higher efficiency of learning as well as better accuracy of prediction.