Abstract: A method to restructure the performance function is proposed. Based on the minimum misclassified sampling rule in pattern recognition it can achieve good effects to classify samples, then a good approximation to the performance function can be obtained. An iterative algorithm is used to approximate the performance function accurately around the true design point through modification to the solution of design points in each step. It overcomes the disadvantage that the general response surface method needs the performance function values. Numerical examples show that the method is of high precision with fewer samples.