New identification approach for nonlinear systems based on the combination network model of least squares and support vector machines

A novel combination network model of least squares and support vector machines(MLS-SVMs) and the associate learning algorithm for identifying nonlinear systems based on the input-output data are proposed. In the model, the identification task is dynamically decomposed into several subtasks according to the physical or statistical natures of the problem. The SVMs are applied as learning machines to every subtask. After analyzing the statistical characteristics of the model in the formal characterization, we give an algorithm for training the MLS-SVMs, based on the frame optimizing principle. The expectation conditional maximization(ECM) algorithm is applied to solve the dependence problem of parameters. Regularization theory and least squares method assure the identification principle of minimal construction risk for expert modules. Experiment illustrates good performance of the proposed method by high approximation accuracy and generalization levels.