【作者】 杨兴；

【导师】 朱劼；

【作者基本信息】 武汉大学 ， 信息与通信工程， 2019， 硕士

【摘要】 仿真可信度的评估对仿真的应用以及基于仿真的决策具有重要意义,可信度评估是指决策者/仿真用户对应用仿真模型解决所定义的目标问题的可信程度的量化。网络仿真是建模仿真研究的一个具体的领域,基于已有成熟的网络仿真建模工具进行网络建模是其目前的主流形式。因此,目前网络仿真的应用主要是在用户的角度,仿真程序对于用户是一个灰盒。因此在网络仿真可信度评估时,在用户的视角下需要考虑哪些方面的内容,以及在各个方面如何度量仿真结果的表现,是提高网络仿真用户对其建模结果应用信心程度的关键。所以,本文对基于用户视角的网络仿真可信度评估的关键技术进行了研究,具体包括:1.介绍了基于用户视角的网络仿真可信度评估的分析框架。根据目前应用网络建模仿真的特点以及可信度评估中需要考虑的关键问题,阐述了在用户视角下进行网络仿真可信度评估的具体考察的分析指标,以及其中的关键技术。2.研究了网络仿真数据有效性分析的方法。仿真的有效性是仿真可信的基础,其主要内容是灵敏度分析,由于网络仿真的输出多个指标之间具有相互影响,为了综合地分析仿真的输入参数对仿真输出响应的影响程度,提出了自编码-随机森林回归的组合模型。自编码网络提取网络性能指标的综合特征,并将其和网络仿真输入参数构建随机森林的回归模型,通过其变量重要性评分度量网络仿真的输入对输出响应的影响程度。3.研究了网络仿真数据准确性量化的方法。准确性的量化作为网络仿真可信度评估的目标,主要是度量网络仿真的数据和参考数据的近似程度。基于拓扑数据分析进行了数据集拓扑性质量化的正确估计,并通过仿真数据集和参考数据集的“环”个数的概率分布的JS散度量化仿真数据和参考数据的拓扑性质的相似程度,即仿真数据和参考数据在全局结构和行为上的相似程度。本文提出的方法适用于仿真用户对网络仿真输出的结果进行验证分析的情况,是一种结合定性验证和客观量化分析的全局综合分析方法,并在最后通过一个数据中心网络性能仿真分析案例,验证了本文方法的可行、有效、可操作以及可解释的特点。更多还原

【Abstract】 Simulation credibility evaluation refers to the quantification of the credibility,when the simulation user or the decision maker apply the simulation model to solve the problem defined by themselves.The evaluation of credibility is of great significance to the application of simulation and the decision-making based on simulation.Network modeling and simulation is a specific field of modeling and simulation research,of which mainstream is based on some mature network simulation modeling tools.Therefore,the current network simulation is mainly from the perspective of the user,and the simulation program is a gray box for the user.Therefore,on the network simulation credibility evaluation,from the perspective of user,which aspects should be used to examine the credibility of the network simulation,and how to measure the performance of the simulation results in various aspects,is the key technologies for advance the level of confidence of application,as the network simulation user apply its modeling results.Therefore,this paper studies the key technologies of credibility evaluation of network simulation from the perspective of users,including:1.The method framework of network simulation credibility evaluation from the perspective of users is introduced.According to the characteristics of current network modeling and simulation and the key issues to be considered in the evaluation of simulation credibility,the framework of credibility evaluation of network simulation based on user’s perspective is presented,then detailed process of the credibility evaluation of network simulation as well as the key technologies therein is introduced.2.The approach of analysis of effectiveness of network simulation data is studied.Effective simulation of simulation is credible premise.Sensitivity analysis is the main content of the effectiveness analysis of network simulation.Since the output of network simulation has certain mutual influences,in order to comprehensively analyze the influence degree of the input parameters of the simulation on the simulation output response,Auto-Encoder-RFR,the combined model of Auto-Encoder and random forest regression,is proposed.Auto-Encoder is used to extract the comprehensive characteristics of network performance indicators,and the extracted features and network simulation input parameters are used to construct a random forest regression model.Degree of influence by network simulation input parameters of the output response is measured via the variable importance score of random forest.3.The method of accuracy measurement of network simulation data is studied.The accuracy quantification is the goal of network simulation credibility evaluation.It mainly measures the similarity between network simulation data and reference data.Based on the topological data analysis,the correct estimation of the data set topology property is performed,then,the similarity of the topological properties of the simulation data and the reference data is quantified by the JS divergence of the probability distribution of the “holes” number of the simulation data set and the reference data set.That is,the similarity between the simulation data and the reference data in terms of global structure and behavior.The method proposed in this paper is suitable for users to verify and analyze the results of network simulation output.It is a global comprehensive analysis method combining qualitative verification and objective quantitative analysis.Which is feasible,effective,operational,and interpretable,and finally,it is verified by a data center network performance simulation analysis case.更多还原