Abstract: For a class of discrete-time hybrid system in the piecewise a±ne form, its model identification problem is equivalent to the problems of classification of the cluster data of the system, the optimized classification of boundary and linear regression. Using improved G-K fuzzy cluster algorithm to solve the numerical problems in iterated processes, the optimal cluster data can be obtained. The number of sub-models can be estimated from multi-performance indexes. In each cluster, the parameters of submodel are obtained by the weighted least squares method. Two adjacent regions were achieved with the nearest distance among the cluster centers. The boundary hyper-plane can be estimated by using a soft margin support vector machine. Simulation results show good performances of this effective technique.