On-line Adaptation Algorithm for RBF Kernel Based FS-SVM

The performance of Fixed-Size least squares support vector machines (FS-SVM) has been illustrates on the large-scale modeling problem. This paper presents an adaptive RBF kernel based FS-SVM and an on-line adaptation algorithm for time-varying nonlinear systems. The key feature of this algorithm method is the direct approach used for formulating the training target. Based on the feature of RBF kernel, the error (objective) function between actual active model and target model is formulated and can be minimized by Gradient descent algorithm. The proposed algorithm is capable of maintaining the accuracy of learned patterns even when a large number of aged patterns are replaced by new ones through the adaptation process. The simulation results show the effectiveness of this architecture for adaptive modeling.