An Online Parameters Optimization Based On Gamma Test For LSSVM

In recent years, time series regression model based on machine learning have been used to solve the forecasting problems encountered in practical industrial production. Among them, the least square support vector machine (LSSVM) has been widely used.In this paper, we study parameters optimization methods for LSSVM based on effective noise estimation of sample, named as Gamma Test, and applied to the prediction for time series prediction problem. For the optimal embedding dimensionality of the studied time series, this paper employs Gamma Test to determine it. On the basis of full study for LSSVM parameter, a gradient descent based hyperparameters optimization based on efficient noise estimation is proposed, and two aspects of the algorithm including hyperparameters selection and validation are further analyzed. Firstly, the effective noise of the observed sample is estimated, and then a conjugate gradient algorithm is developed to optimize the width of Gaussian kernels and the value of the regularization factor. The proposed method can not only overcome the slow computational efficiency and the blindness of the grid search method in parameter selection, but also avoid the tedious calculations of cross-validation method in validation process.To verify the effectiveness of the proposed method, a series comparative experiments using the proposed method and the other parameter optimization method for LSSVM are carried out, in which the classical Sinc function and the practical data of byproduct gas system in steel industry are employed. The experimental results demonstrate that compared to the common parameters optimizations, the proposed method has the advantages of high prediction accuracy and computational efficiency. These two merits make this optimization algorithm based on effective noise estimation applicable to real-time prediction for some industrial application problems.