Research On Global Optimization Algorithm Based On Kriging Model

Traditional global optimization algorithm can not meet the requirements of Multi-domain modeling and simulation tools that less evaluation times are wanted. Global optimization based on meta-model evaluates the agent model of source function so that the total number of evaluations is greatly reduced. With the support of National Natural Science Foundation, research on global optimization based on Kriging model is done in this paper. The main work is as follows:First of all, several design of experimental methods commonly used in construction of response surface are introduced, basic principles of these methods are described, respective advantages and disadvantages are analyzed.Then, the characteristics of Kriging model and the process of construction are anglicized. This paper proposes the increment Kriging method (IKM) in which the inversion of the correlation matrix and the new data points are manipulated to get the coefficients of the Kriging model, while coefficients of correlation function are optimized and the inversion of new correlation matrix is directly calculated in traditional Kriging method. Lots of experiments are taken and the results demonstrate that IKM greatly reduces the time of modeling with little loss of accuracy.EGO algorithm is researched and analyzed in this paper. The time of rebuilding the Kriging model increases rapidly with the increment of samples'size, and premature convergence may exist when the range of the object function is too large. To conquer these problems, this paper proposes the improved EGO (IMEGO) algorithm which the IKM is applied to. And stopping criterions on expected improvement, response value and argument are used in the IMEGO algorithm. Finally the experimental results based on standard global optimization benchmark functions demonstrate that the improved EGO method has higher efficiency and better stability.