Research On Kriging Optimization Algorithm Based On The Gradient Method

Global optimization algorithm which is based on response surface can reduce the estimated degree of source function by the way of approximating source function.Hence, it lowers the consumption of computation. With the characteristic of the fast computation speed and high precision, Global Optimized method based on Kriging model has been widely researched and applied. In which the representative algorithm is the Efficient Global Optimization algorithm, i.e. the EGO algorithm. The modification of EGO algorithm is meaningfully for theory research and practice. Based on all of these, we mainly studied the following:Firstly, this paper introduced the study process of Kriging model based on response surface and EGO algorithm. And this paper also gives the illustration of the characteristic of the algorithm.I introduce the method of Design Of Experiment, usually write DOE for short, which is frequently-used in the construction of response surface and analysis the advantage and disadvantage of the methods. Using Latin Hypercube Sampling, i.e. LHS, for example, I gain the sampling graph for some methods.Secondly, combing the character of Kriging model, I give detailed explain about the process of combination. For the low accuracy of mode search method in traditional parameter optimization, I put forward the optimization algorithm based on gradient algorithm of space relevant function. The core of the algorithm is to compute the gradient of the target function about q. And we can simplify the solving process by holding the value of b in step-by-step solving, in case of the solution is more complicated. In this way, we can reduce the computation consumption.Numerical-value experiment indicates that we can gain more accuracy solution by this algorithm.Thirdly, this paper also illustrate the process of the realizing of EGO algorithm.Adopting projected gradient method to the core of EGO algorithm, that is the EI function of sample filling criteria, and combining the parameter optimization of the corresponding space function, I also put forward the EGO algorithm based on the projected gradient method. Large number of experiment tests show that comparedwith the traditional EGO method, this algorithm makes an apparent progress in the stability of the mode and the predictive results.