Batch Sequential Designs In Computer Experiments

With the development of science and technology,computer experiments be-comes more and more popular for the study of complex physical process.Computer experiments model physical phenomena through coded mathematical models.How-ever,many physical phenomena are extremely complex,as a result,the correspond-ing mathematical models are not easy to simulate.Thus,it is necessary to obtain as valuable experimental results as possible under the limited experimental resources.Sequential adaptive test designs can constantly bring new and useful experimen-tal information to computer simulations.This method can continuously establish an agent model based on the previous experimental information,and self-adjust the experimental points according to the established agent model and corresponding filling criteria to obtain more effective experimental information.?Compared with the fixed-point experimental design in which all experimental points are determined in advance at one time,the sequential adaptive design is more flexible and more efficient in obtaining information.However,the sequential adaptive designs is usually the single-point updating mechanism.In each iteration,only one run is selected,and the surrogate model is refitted,while it greatly wastes a,great deal of experimental resources,it is unable to perform efficient parallel calculations,which greatly wastes computing resources and increases the time cost of the experiment.In order to improve the operation efficiency of computer simulations,it is very important to update the experimental points in the way of batch sequence to adapt to the parallel technology.Hence,a barch sequential adaptive design framework based on translational quasi-Monte Carlo point set technology is programmed in this paper.This method can be applied to most of the traditional single-point filling criteria and can improve the simulated efficiency of computer experiments.In addition the points obtained based on this method can effectively avoid the clustering phenoinenon,improve the ability of space-filling,and ensure the fitting performance of the surface.Through munerical simulation experiments,the surface fitting ability of single point and batch sequential experinmental designs is compared under different filling criteria,and the peperform ance of batch sequential upating method is validated.At the same time.a batch sequential maximum entropy design based on Bayesian Gaussian process model is proposed.This design is programmed based on the flexible Bayesian framework and the multi-point maximum entropy criterion.The numerical simulation results show that the proposed method is superior to the batch sequential maximum entropy design based on the translation point set technology in terms of space filling ability of updating points and the surface fitting ability of the surrogate model.