| With the popularity of high-performance GPU(graphics processing unit)devices,the computing power of computers has been greatly improved,the scale of data sets available for training is becoming larger and larger,and the traditional support vector machine model has also ushered in a great challenge.Firstly,the massive types of data and noise require a higher discrimination ability of the support vector machine model.Most of the SVR(Support Vector Regression)models have limited optimization magnitude and only focus on the optimization of hyper-parameters.Secondly,the adjustable parameters of the traditional support vector machine model represented byε-SVR(ε-insensitive support Vector Regression)are limited and cannot assign different tolerance intervals according to data distribution and trend.Finally,the model with the fixed-radiusε-tube uses the same penalty function for the key data and noise data distributed in different locations,which does not exploit the full information of the dataset.Therefore,it is of certain significance to study the support vector machine model which has variable tolerance and can flexibly punish data points in different locations.In this paper,in view of the shortcomings of the traditionalε-SVR model,the following studies are carried out:(1)The concept of variable tolerance and a new support vector regression model_-SVR(Alterable_-Support Vector Regression)based on variable tolerance are proposed.This model can apply different penalty values according to the distribution position of the data points,and adjust the SVR model through multiple parameters.(2)A non-deterministic adaptive algorithm is put forward.The proposed_-SVR model contains a parameter_,=1,2,…,.It is difficult to find the optimal parameter combination by conventional methods,so a non-deterministic algorithm(non-deterministic adaptive algorithm)is proposed to solve the optimal combination of_values.By setting a reliability index as the evaluation standard and continuous iterative optimization,the algorithm finds out the globally optimal_combination and solves the corresponding_-SVR model.(3)According to the characteristics of_-SVR model,the Slime mould algorithm is optimized to find the optimal super-parameter combination and further improve the performance of the model.The experimental results show that,compared with the conventional SVR model,the_-SVR model can effectively improve the prediction accuracy and stability on both simulated and real datasets. |
