Research On Efficient Experimental Design Methods For Complex Simulation System

The design of simulation experiments is a critical method to realize efficient simulation,which ensures the research work of simulation evaluation,analysis,and optimization to be processed smoothly.With the complexity of simulation objects and users' requirements increasing,simulation systems tend to be more complicated.The simulation system has been brought with some new characteristics,including high dimensional,large scale and mixed experiment space,which usually leads to time consuming.However,it is usually necessary to obtain results of simulation experiment as quickly as possible in practice.Therefore,the hot spot and difficulty in the current study lie in how to improve the efficiency of complex simulation experiment and its design.Aiming at these problems,the main contents of this dissertation are as follows.First,faced with a wide range of candidate experiment design methods,a great challenge is to pick an appropriate method in practice.The problem formulation of simulation experiment design is given,followed by establishing experiment design method system of complex simulation.An intelligent selection method,which combines case-based reasoning and rule-based flexible layer-by-layer reasoning,is proposed.By analyzing the characteristics of the alternative methods for simulation experiment design,the purposes and process model of simulation experiment are given.The simulation experiment design methods are classified from four perspectives,including experiment execution phase,experiment point distribution,model independence,and sequential design.Furthermore,considering the characteristic of simulation experiment design methods and user requirements,the framework of intelligent design for simulation experiment is established.Finally,the intelligent selection method is studied from two aspects.On the one hand,the attributes of cases are divided into three types and the case retrieval strategy is given.On the other hand,the method of rule-based flexible layer-by-layer reasoning(RBFLR)is proposed.On the premise of guaranteeing the reasoning results reliability,RBFLR can reduce the number of rules to a large extent.Meanwhile,the condition attributes of rules can be managed flexibly.Second,to improve the optimization efficiency of optimal Latin hypercube sampling(OLHS)method and alleviate the oversampling problem of sequential design for OLHS,two kinds of methods are proposed respectively.An OLHS method based on an improved differential evolution(DE)algorithm is proposed according to the characteristics of optimization problem for OLHS.The strategies of local optimization and parameters self-adaptive adjustment are introduced to improve the local searching ability and avoid premature convergence of the basic DE algorithm,respectively.The proposed method appears to perform well in optimization efficiency of OLHS in terms of space-filling performance.Furthermore,subject to the strict LHS structure and preservation of existing sampling points,a sequential design method of OLHS is proposed,which can extend sampling points to k times of initial sample size.Compared with the alternative sequential design methods,the proposed method can avoid oversampling effectively and improve the efficiency of simulation experiments.Third,the alternative methods of screening experiment design perform poorly in dealing with factor screening problem for complex simulation experiments.To solve this issue,two kinds of factor screening methods are proposed according to the different situations.Considering the large number of factors,an Apriori algorithm coupled with Gaussian process(GP)model is proposed.If the sample data is sparsity,the GP model is used to generate sufficient sample data.The relationship between input and output is mined by Apriori algorithm.Furthermore,the sensitivity index of each factor is calculated and the important factors are identified.When the number of factors is small,a sensitivity analysis method based on sequential design is proposed.The proposed method is improved from two aspects.On the one hand,a generalized estimator of first order sensitivity indices is proposed,and its asymptotic unbiasedness and efficiency are verified.On the other hand,based on the strategy of sequential design,a new sequential design method is proposed to evaluate the sensitivity indices.Finally,confronted with the fact that the alternative simulation platforms and the general experiment design and analysis software cannot meet the application requirements of complex simulation experiment design and analysis,a complex simulation experiment design and analysis platform is designed and implemented.The overall structure of the platform is designed after analyzing the application requirements of the platform.The class diagram and sequence diagram of the subsystems are given,including simulation experiment design,data management,metamodelling and data analysis.The platform is implemented,which appears to perform well in interface,reusability and extensibility.The experiment design and analysis of a system-of-systems operational simulation system is applied to validate the effectiveness and practicability of the platform.