Factor Screening Based On Simulation Experiments And Application Research

Experimental design on the basis of statistics is an important aspect of designing for quality and quality improvement.Classical analysis methods used for physical experiments have been widely developed and gained much attention in the fields of agriculture,industry,and medical science.However,sometimes moral and ethical reasons such as damage to environment make physical experiment infeasible.Since the 20^th century,with the rapid development of computer techniques,simulation experiments have gradually become popular.The real system is abstracted and realized through computer programming which is called simulator.Meanwhile,studies on design and analysis of simulation experiments represents a growing research field,and some relative problems are emerged.On the one hand,simulation experiments involve much more variables(or factors)than physical experiments,and the causing computational and time costs highlight the importance of efficiency of designs.On the other hand,simulation experiments are easily to conduct sequentially.Those features of simulator bring both opportunities and challenges for the design and analysis of simulation experiments.And it inspires the topic of factor screening for simulation expeirments which can identify important factors with as few runs as possible.In this dissertation,we comprehensively analyzed the features of simulation system,and consider possible response distributions and data types.Sequential bifurcation(SB)screening menthod with high efficiency was further improved and modified under those situations of response by developing response models and constructing statistics for significance testing.The main contents are as follows:(1)Under the case of normally distributed response,we emphasized the importance of dispersion effect which plays an important role in performance evaluation of system.In addition,different classes of factors can be implemented in different stages of process improvemnent.Hence,we constructed both location and dispersion models and proposed the screening and classification of factors under homogeneous and heterogeneous response data.Firstly,for homogeneous response data,the location and dispersion models were assumed to contain only fixed factor effects,and sequential probability ratio test and fixed-interval width test were developed for significance testing in every stage of SB.According to the reults of significance test,those identified factors are classified in three classes,including adjust factors,robust factors and important factors.Secondly,for the heterogeneous response data caused by heterogeneous system structures,random effect was used to describe effect level of structures.The mean and variance of response were constructed through mixed effect models.Based on fixed interval method,the hypothesis testing procedure was modified for screening and classifying significant factors.(2)Under the case of non-normally distributed response,we proposed the screening methods under data contamination and asymmetric response data respectively.Firstly,data contanmination is hard to avoid especially when data volume is large and veracity is doubt.Hence,the influence of some common types of contamination on classical SB was analyzed.A robust statistic was incorporated in hypothesis testing procedure of SB,so that the improved SB can obtain robust screening results with high efficiency.Secondly,in the case of asymmetric response,the importance of skewness feature was emphasized.In order to describe the asymmetric response,a skew-normal response model was constructed.After that,the SB was modified by incorperating a parametric bootstrap procedure for significance testing of factors.Hence,the modified SB can identify those factors which have impacts on the location,dispersion,or skewness of response distribution.Through concluding above-mentioned screening methods,strategies of choosing screening methods were suggested.In addition,a practical simulation system was introduced to guide the application of proposed screening methods in real life.At last,some open issues were discussed.