面向网络化系统的分布式估计及优化算法研究

发布时间：2018-06-19 08:56

本文选题：网络化系统 + 分布式算法　；参考：《北京邮电大学》2017年博士论文

【摘要】：网络化系统作为计算机网络未来发展的趋势已经受到越来越广泛的关注。所谓网络化系统即指具备如下两点特征的系统。第一,大规模的网络化系统可以被划分为许多子模块,各模块间能够按照特定的网络通信模型进行互联。第二,个体间能够相互协作,整个网络拓扑呈现为分布式或者半分布式结构。我们日常生活中的大多数网络都可以看做为网络化系统中的一种,例如:车联网、无线传感网、移动数据网以及工业互联网等。当前,面向全分布式的网络化系统已有大量的研究,其中,大多数研究主要集中在相对应的分布式算法上。所谓分布式算法,即各个网络节点仅通过自身与周围节点的信息交互来共同解决网络中存在的问题。与传统的集中式算法不同,在分布式算法中,往往不需要控制中心节点,网络中每个节点的地位都是等价的。能够有效避免单点失效、关键节点负载过大等问题,鲁棒性,实时性高。但同时分布式算法会增加整个网络的通信量,带宽消耗巨大,且对网络中的拓扑有一定的要求。另外,现有面向网络化系统的分布式算法研究中,大多数都对网络环境进行了过于理想的假设,导致大多数研究在实际问题分析中不能得到有效的应用。本文重点围绕网络化系统下如何建立符合实际应用环境的模型、如何采用有效的分布式优化算法对模型进行理论分析以及针对网络化系统中具体应用问题进行分布式优化算法构建求解等内容展开深入研究。主要工作体现在对网络化系统下分布式算法的理论推导分析与具体应用建模求解两个方面,具体研究内容如下:(1)针对无线虚拟传感网络下链路状态不稳定特点,提出了基于概率稳定性的链路动态拓扑模型,并提出了动态扩散性卡尔曼滤波算法,该算法对于实时系统中的运动物体轨迹的估计、跟踪及预测都有着良好的性能表现。为了进一步证明算法的有效性,本研究点从理论层面对算法进行了深入分析,分别对其均值与均方误差值进行了严格的推导证明,以此证明了所提算法是无偏且收敛的,能够很好的用于上述网络化系统模型中。最后,通过仿真对算法进行了验证,并通过与相关实验对比分析,进一步证明了所提算法在所提动态拓扑网络模型下的良好性能表现。(2)考虑移动群智感知网络系统内大量分布的低功耗设备,由于计算、存储资源不足、带宽受限,设备间不能够进行大规模高精度数据传输。针对上述问题,首先,提出了基于抖动随机量的量化模型,其次,基于该模型提出了量化扩散卡尔曼滤波算法。在该算法中,将节点间交互的不同信息进行量化,包括观测值,噪声值等。由于量化后的数据会对原算法的稳定性和收敛性产生负面的影响,本研究点着重分析了量化后算法的各种理论性能表现,并通过仿真实验来对此进行验证。通过分析可知,该算法在在网络资源受限情况下仍具有较高的数值估计准确率。(3)针对5G内容中心网络下内容缓存能效最优化问题,采用基于投影的扩散分布式优化算法进行解决。为了能够有效运用该算法,首先针对内容缓存策略问题进行抽象建模,并提出连续性及凸性等假设条件来保障算法的有效实现。实际上,本研究点重点分析了扩散性分布式优化算法在具体应用问题进行建模求解的方法,具有普适性。最后,将所提方案在几种具有代表性的场景下进行了仿真实验,结果表明其能够达到预期的效果。由于网络化系统和分布式算法种类繁多,本文中提出的所有算法及应用都只是提供了一种研究该类问题的思想和方法。因此本文的研究具有很高的可扩展性。在具体研究过程中,本文通过对相关工作进行调研,并通过理论建模、算法设计、分析以及实验仿真验证等一系列方法对相关问题进行了深入研究,对所提算法的优越性以及实用性进行了验证。本文取得的研究成果对网络化系统下分布式算法的研究与发展具有很好的借鉴意义。
[Abstract]:The network system as the future development trend of the computer network has been paid more and more attention. The so-called network system refers to the system with the following two characteristics. First, the large-scale network system can be divided into many sub modules, each module can be interconnected according to the network communication model. Second, The whole network topology is distributed or semi distributed. Most of the networks in our daily life can be regarded as one of the network systems, such as car networking, wireless sensor networks, mobile data networks, and industrial Internet. Most of the studies mainly focus on the corresponding distributed algorithms. The so-called distributed algorithms, that is, each network node can solve the problems in the network only through the information interaction between themselves and the surrounding nodes. Unlike the traditional centralized algorithm, the central node is often not required in the distributed algorithm. The status of each node in the collaterals is equivalent. It can effectively avoid the single point failure, the key node is overloaded and so on, it is robust and real-time. At the same time, the distributed algorithm will increase the communication amount of the whole network, the bandwidth consumption is huge, and the topology of the network is required. In addition, the existing distributed system oriented distributed system is distributed. In the research of algorithm, most of the network environment is too ideal. Most of the research can not be applied effectively in the actual problem analysis. This paper focuses on how to establish a model which conforms to the actual application environment under the network system, and how to use an effective distributed optimization algorithm to analyze the model. The main work is two aspects: theoretical derivation analysis and specific application modeling solution of distributed algorithm under networked system. The specific research contents are as follows: (1) the chain under the wireless virtual sensor network The link dynamic topology model based on the probability stability is proposed, and the dynamic diffusion Calman filter algorithm is proposed. The algorithm has good performance performance for the estimation, tracking and prediction of the moving object trajectory in the real-time system. In order to further prove the effectiveness of the algorithm, this research point is based on the theory. In this paper, the algorithm is deeply analyzed, and the mean and mean square error value are strictly derived, which proves that the proposed algorithm is unbiased and convergent, and can be used in the networked system model well. Finally, the algorithm is verified by simulation, and the analysis is further compared with the related experiments. The performance performance of the proposed algorithm is proved under the proposed dynamic topology network model. (2) considering the large number of low-power devices distributed in the mobile swarm intelligence network system, because of the lack of storage resources and limited bandwidth, the high precision data transmission can not be carried out among the devices. Secondly, based on the model, a quantitative diffusion Calman filtering algorithm is proposed. In this algorithm, the different information between nodes is quantified, including observation value, noise value, etc. because the quantized data will have a negative impact on the stability and convergence of the original algorithm, this research point focuses on the analysis of quantization. All kinds of theoretical performance of the post algorithm are presented and verified by simulation experiments. Through analysis, it can be found that the algorithm still has a high accuracy rate in the case of network resource constraints. (3) aiming at the problem of content cache energy efficiency optimization under the 5G content center network, the distributed distributed optimization algorithm based on projection is adopted. In order to effectively use the algorithm, first of all, it aims at the abstract modeling of the problem of content caching strategy, and puts forward the hypothesis of continuity and convexity to ensure the effective implementation of the algorithm. In fact, this research point focuses on the analysis of the method of modeling and solving the diffusion distributed optimization algorithm in specific application problems. Finally, the proposed scheme is simulated in several representative scenes, and the results show that they can achieve the desired results. All the algorithms and applications proposed in this paper provide a kind of thought and method for studying the problem because of the variety of network and distributed algorithms. In the course of specific research, this paper makes a thorough study of related problems through a series of methods, such as theoretical modeling, algorithm design, analysis and experimental simulation verification, and proves the advantages and practicability of the proposed algorithm. The results have a good reference for the research and development of distributed algorithm under networked system.
【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP393.0;TN713

【参考文献】