A new approach is proposed to model nonlinear dynamic systems by combining SOM(self-organizing feature map) with support vector regression (SVR) based on expert system. Thewhole system has a two-stage neural network a...A new approach is proposed to model nonlinear dynamic systems by combining SOM(self-organizing feature map) with support vector regression (SVR) based on expert system. Thewhole system has a two-stage neural network architecture. In the first stage SOM is used as a clus-tering algorithm to partition the whole input space into several disjointed regions. A hierarchicalarchitecture is adopted in the partition to avoid the problem of predetermining the number of parti-tioned regions. Then, in the second stage, multiple SVR, also called SVR experts, that best fit eachpartitioned region by the combination of di?erent kernel function of SVR and promote the configura-tion and tuning of SVR. Finally, to apply this new approach to time-series prediction problems basedon the Mackey-Glass di?erential equation and Santa Fe data, the results show that SVR experts hase?ective improvement in the generalization performance in comparison with the single SVR model.展开更多
基金Supported by the National High Technology Research and Development Program of P.R.China(2002AA412010)the Technology Development Program of the Ministry of Science and Technology of P.R.China(2003EG113016)
文摘A new approach is proposed to model nonlinear dynamic systems by combining SOM(self-organizing feature map) with support vector regression (SVR) based on expert system. Thewhole system has a two-stage neural network architecture. In the first stage SOM is used as a clus-tering algorithm to partition the whole input space into several disjointed regions. A hierarchicalarchitecture is adopted in the partition to avoid the problem of predetermining the number of parti-tioned regions. Then, in the second stage, multiple SVR, also called SVR experts, that best fit eachpartitioned region by the combination of di?erent kernel function of SVR and promote the configura-tion and tuning of SVR. Finally, to apply this new approach to time-series prediction problems basedon the Mackey-Glass di?erential equation and Santa Fe data, the results show that SVR experts hase?ective improvement in the generalization performance in comparison with the single SVR model.
