A novel multi-innovation gradient support vector machine regression method.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Traditional support vector machine (SVM) regression methods rely on offline datasets, limiting their application in online scenarios.
• Accurate online parameter identification is crucial for dynamic systems and real-time data analysis.

Purpose of the Study:

• To develop a novel online approach for identifying unknown parameters in support vector machine regression.
• To enhance the efficiency and applicability of SVM regression in dynamic environments.

Main Methods:

• A modified cost function and gradient descent approach are used to propose a stochastic gradient support vector machine (SGSVM) algorithm.
• Multi-innovation identification theory with a moving data window is employed to increase information used in parameter estimation.
• Forgetting factor recursive algorithms are derived for improved online performance.

Main Results:

• The proposed methods enable online parameter identification for SVM regression, overcoming limitations of offline approaches.
• The integration of multi-innovation theory and forgetting factors enhances the algorithm's efficiency and information utilization.
• Numerical simulations on the MatLab platform validate the effectiveness of the developed methodologies.

Conclusions:

• The developed online recursive identification methods offer an efficient and computationally advantageous solution for SVM regression parameter estimation.
• The novel stochastic gradient support vector machine algorithm with enhancements is suitable for real-time applications and dynamic data analysis.