Support vector machine for nonlinear system on-line identification

Neural networks is a very popular black-box identification tool. But it suffers some weaknesses for nonlinear on-line identification. For example, the learning process can only arrive local minima. The training algorithms are slow. Support vector machine (SVM) can overcome these problems. But the SVM needs all data to find optimal solution, it is not suitable for online identification. In this paper, we propose a new method to use SVM for on-fine identification. We call it as Recursive Support Vector Machine (RSVM), where the kernel is not depended on all data, it is calculated by a recursive method, the SVM is also recursive. So we can realize on-line identification via SVM. Two examples are proposed to compare our RSVM with normal SVM.