无偏置v-SVM分类优化问题研究

在高维空间中,分类超平面倾向于通过原点,即不需要偏置(b).为了研究在v- SVM分类问题中是否需要b,该文提出了无(b)的v-SVM的对偶优化问题并给出了其优化问题求解方法.该方法通过有效集策略将对偶优化问题转化为等式约束子优化问题,然后通过拉格朗日乘子法将子优化问题转化为线程方程组来求解.实验表明偏置(b)的存在会...     

二阶传输条件撕裂互连法在电大辐射中的应用

针对于大尺寸电磁辐射问题,将撕裂互连法应用于三维电大尺寸辐射问题的计算仿真。该算法是区域分解方法中的一种,求解区域划分成互不重叠的子区域,各子区域之间通过拉格朗日乘子将交界面的连续边界条件或者传输条件耦合。鉴于原来的传输条件存在不收敛或者收敛较慢情况,提出同时能在TE/TM凋落模式快速收敛的二阶传输边界条件作为子区域之间交界面的边界条件,并且引入虚拟激励流为交换信息。数值结果表明,该改进算法有效地提高了撕裂互连法在迭代求解中的收敛性。在求解三维电大辐射问题时,该算法仿真结果与有限元算法结果一致,表明撕裂互连算法是计算大尺寸电磁辐射问题的一种有效方法。     

一种基于神经网络的广义熵模糊聚类算法

以模糊聚类为基础,将广义熵引入到模糊聚类的目标函数中,提出一种基于模糊熵的模糊聚类的统一形式,即广义熵模糊聚类模型;利用增广拉格朗日求解方法,以及Hopfield神经网络和复突触神经网络解决了基于广义熵的目标函数的优化问题,提出了基于神经网络的广义熵模糊聚类算法,表明了使用神经网络求解的收敛性;同时,给出一种用于确定增广拉格朗日乘子的迭代方法.实验中选取人工生成数据集和UCI标准数据集对提出的算法进行了实验研究,并与常用的聚类算法进行了性能比较.     

基于并行分离增广拉格朗日乘子法的字符矫正

字符矫正是光学字符识别(OCR)系统预处理过程中 的重要步骤,针对传统的增广拉格朗日乘子法(ALM)求解字符矫正问题时收敛性和计算速度 的不足,本文研究了并行分离的增广拉格朗日乘子法,综合考虑字符矫正模型的建立过程, 提出并行分离方法与ALM相结合的思想解决字符 矫正问题。用并行方式将迭代问题分解成3个子问题,计算时能够同时求解分解后的这3个 子问题,然后进行凸组合,最 后收敛到问题的最优解。实验结果表明,本文算法能够快速准确地对变形的字符图像进 行矫正,并且具有良好的实时性和适 应性,可用于OCR系统的矫正预处理中,提高OCR系统的识别率。     

求解任意凸规划问题的神经网络模型

本文提出求解任意凸规划问题的神经网络模型，证明了采用松弛变量的拉格朗日乘子法用凸规划求解时所构造的函数，在一定条件下，其平衡点即为问题的最优解，通过附加惩罚函数项，克服了单纯使用拉格朗日乘子法或惩罚函数法时存在的问题，计算机模拟结果证明了本文的有效性。     

无偏置ν-SVM分类优化问题研究

在高维空间中,分类超平面倾向于通过原点,即不需要偏置(b)。为了研究在ν-SVM分类问题中是否需要b,该文提出了无(b)的ν-SVM的对偶优化问题并给出了其优化问题求解方法。该方法通过有效集策略将对偶优化问题转化为等式约束子优化问题,然后通过拉格朗日乘子法将子优化问题转化为线程方程组来求解。实验表明偏置(b)的存在会降低ν-SVM的泛化性能,ν-SVM只能得到无(b)ν-SVM的次优解。     

基于目标容量的网络化雷达功率分配方案

针对现有网络化雷达功率资源利用率低的问题,该文提出一种基于目标容量的功率分配(TC-PA)方案以提升保精度跟踪目标个数.TC-PA方案首先将网络化雷达功率分配模型制定为非光滑非凸优化问题;而后引入Sigmoid函数将原问题松弛为光滑非凸优化问题;最后运用近端非精确增广拉格朗日乘子法(PI-ALMM)对松弛后的非凸问题进行求解.仿真结果表明,PI-ALMM对于求解线性约束非凸优化问题可以较快地收敛到一个稳态点.另外,相比传统功率均分方法和遗传算法,所提TC-PA方案可以最大限度地提升目标容量.     

MIMO-MC-CDMA中基于拉氏算子的资源分配算法

提出一种应用于多输入多输出MC-CDMA系统的功率、码道自适应分配算法。基于接收端的信道反馈信息,发送端在总发射功率受限的条件下通过多用户间功率和码道的分配最大化系统吞吐量。算法通过限制用户的最大和最小码道数来实现用户之间的带宽公平性。算法最终归结为一个约束优化问题并利用拉格朗日乘子法进行求解。提出一种用于加快拉格朗日乘子收敛速度的搜索算法,计算机仿真验证了算法的有效性。     

基于ADMM的低副瓣阵列区域聚焦照射

区域聚焦照射(regional focusing irradiation)能将信号能量准确投送至指定区域,但阵列的超稀疏性使得副瓣区域能量较高,在实施精确电子战时极大增加了干扰系统遭受打击的风险,因此副瓣区域能量抑制问题必须加以解决.本文提出一种基于L∞范数评估栅瓣区域能量的方法,将其作为正则项引入区域聚焦照射模型以抑制副瓣区域最大能量,并建立了多目标优化模型.采用交替方向乘子(alternating direction method of multipliers)框架将原问题分解为两个子问题:对于恒模约束下二次规划子问题I,采用贪婪算法给出其闭式解;对于无约束L2?L∞范数的子问题II,将L∞范数做近似光滑化处理,并通过梯度下降法求解.交替求解两个子问题至收敛,以求解发射信号.仿真实验表明,本文模型在副瓣区域能量抑制上较基于L1范数的区域聚焦照射模型具有更优性能,且本文算法实用性更强.     

视频编码率失真优化技术研究综述

率失真优化技术能有效地提升编码器的压缩性能,是视频编码领域重要的研究内容.本文综合论述了基于拉格朗日乘子的率失真优化方法,从独立率失真优化、依赖率失真优化和码率控制中的比特分配等几方面归纳和分析最新研究进展.首先,通过形象的举例解释了约束优化问题通过拉格朗日乘子法转换为非约束优化问题,并根据相关文献分析拉格朗日乘子与量化参数的关系.其次,分析了编码过程中存在的率失真依赖性,并分类讨论近年提出的依赖率失真优化方法.然后,简要介绍了率失真依赖关系在码率控制比特分配中的影响.最后通过分析比较不同率失真优化技术的特征和性能,提出当前面临的挑战和值得进一步探索的工作.     

Domain-decomposed MEI method for field computation in single andmultiple regions

The domain-decomposed measured equation of invariance (DDMEI) method is proposed for field computation in single and multiple regions. The whole computing domain is partitioned into a cluster of subdomains. For single region problems, this partition splits the computing domain into many subdomains artificially. For multiple regions problems, these subdomains can be taken as those regions separated geometrically. The contribution of sources residing in a subdomain is approximated by a set of sources selected out of these original sources with greatly reduced amounts. The approximation is implemented numerically by the MEI method. The resultant MEI matrices are blocked matrices and each submatrix is highly sparse. Approaches and numerical results are given respectively for the applications of the DDMEI to the scattering of single conducting cylinders, radiation of wire arrays, and capacitance matrix computation for multiconductor transmission lines. The DDMEI proposed in this paper is an improved version of the surface current MEI method (SCMEI). Compared with the SCMEI, the DDMEI improves the sparsity of the MEI matrices and the feasibility of measuring out the MEI coefficients. Furthermore, the DDMEI makes it possible to apply the kind of on-surface MEI methods (OSMEI) to multiple region problems for the first time     

棱边元分区的区域分解算法及其在电磁问题中的应用

针对三维电磁问题,该文提出了采用非结构化网格剖分计算区域,并按单元进行区域划分的区域分解算法。将原求解区域划分为若干个不重叠的子区域,先通过求解容量矩阵获得子区域之间连接边界上的场值,再利用矢量有限元快速计算出每个子区域内部的场值,显著地降低了计算复杂度和存储量。通过引入预条件的Krylov子空间法求解容量矩阵方程,加速了收敛,进一步提高了效率。数值算例验证了该方法的准确性和有效性。     

矩形波导E面不连续性的DDM/FEM分析

应用区域分解法(DDM)结合有限元法(FEM)分析矩形波导E面不连续性问题,将原求解区域分解为若干个非重叠的子域,在子域的虚拟边界上采用吸收虚拟边界条件,以保证相邻子域间的波传播。分别计算了矩形波导中加载双金属膜片和双介质柱时的散射系数,结果与有关文献一致。采用这种技术,大大地减少了对计算机内存的需求。     

A generalized algorithm for the capacitance extraction of 3D VLSIinterconnects

In this paper, a generalized algorithm based upon the nonoverlapping domain decomposition method (NDDM) is presented for the capacitance extraction of three-dimensional (3-D) VLSI interconnects. The subdomains with conductors are analyzed by the finite-difference method, while the subdomains with pure dielectric layers are analyzed with the eigenmode expansion technique. The central processing unit time and memory size used by the NDDM are unrelated to the thickness of pure dielectric layers. NDDM's computing time grows as O(n) if the number of domain iterations is bounded. Also, benchmarks show that it is approximately 15 times less than those used by Ansoft's Maxwell SpiceLink. In addition, only a two-dimensional mesh is needed to analyze 3-D structures. This greatly reduces the algorithm complexity and makes it easy and straightforward to interface with layout automation software     

三维非均匀介质成像问题的区域分解方法

在利用数值方法分析非均匀介质问题时,容易生成大型系数矩阵,从而在求解时常常造成计算机内存不足或者计算时间过长。该文利用区域分解方法对三维非均匀介质成像问题进行分析,通过将求解区域划分为几个子区域,在子区域上以迭代求解子问题的方式解决以上问题。文中给出的迭代收敛速度曲线证明区域分解算法的收敛速度很快。该文对一些复杂的非均匀介质问题给出了模拟测量成像的结果。     

Scattering from 3-D cavities with a plug and play numerical schemecombining IE, PDE, and modal techniques

In this paper, we present a multidomain and multi-method coupling scheme called FACTOPO, based on generalized scattering matrix computations on three-dimensional (3-D) subdomains. The global target Ω is split in N_V subdomains (V_i)(i=1, N_V), separated by N_I fictitious surfaces (Γ _j)(j=1,N_I). We use a modal representation of the tangent fields on the interfaces. In each domain, the generalized scattering matrix S_i is computed with different methods such as the 3-D finite-element method (FEM) or the electric field integral equation (EFIE). This coupling scheme leads to an important reduction in computational resources, especially for cavities with one dimension much larger than the other two. The advantages of this formulation for parametric studies is illustrated by two cases: computing the RCS of an air-intake terminated with a flat PEC or a fan (CHANNEL) and of an antenna structure coupled to an electronic feed with a varying parameter (DENEB). Numerical as well as experimental results are presented     

分解区域方法在涡流检测仿真中的应用

随着现代制造业的发展,涡流检测仿真问题日渐复杂。传统的单一求解区域方法的运用遇到瓶颈,而分解区域方法将复杂的单一区域上的求解转化为较简单的若干个子区域上的求解,克服了传统方法遇到的困难。该文主要介绍分解区域方法在涡流检测仿真中的应用。     

电大尺寸开口腔体电磁散射的子结构法研究

将子结构法与矢量有限元法相结合对无限大接地的三维开口腔体的电磁散射特性进行分析.将原尺寸较大的腔体分解成若干个不重叠的子腔体,在各子腔体内应用矢量有限元法进行分析,在原腔体开口面应用边界积分方程描述.通过求解容量矩阵获得子腔体之间连接边界上的场值,可以快速获得腔体开口面上的场值,极大地减少了存储量和计算量,易于对电大尺寸腔体的电磁散射问题进行分析.数值算例验证了该方法的准确性和高效性.     

A domain decomposition method for the vector wave equation

A nonoverlapping domain decomposition method (DDM) is presented for the finite-element (FE) solution of electromagnetic scattering problems by inhomogeneous three-dimensional (3-D) bodies. The computational domain is partitioned into concentric subdomains on the interfaces of which conformal vector transmission conditions are prescribed and that can be implemented in the inhomogeneous part. The DDM is numerically implemented when a conformal vector absorbing boundary condition (ABC) is utilized on the outer boundary terminating the FE mesh, while employing the standard edge-based FE formulation. Then, numerical experiments are performed on a sphere and a cone sphere that emphasize the advantages of this technique in terms of memory storage and computing times, especially when the total number of unknowns is very large. Also, these numerical experiments serve as a severe test for the performances of the ABC     

The finite-element method with domain decomposition for electromagnetic bistatic scattering from the comprehensive model of a ship on and a target above a large-scale rough sea surface

The domain decomposition method (DDM) and finite-element method (FEM) are developed for numerical solution of bistatic scattering from the composite model of a ship on and a target above a two-dimensional (2-D) randomly rough sea surface under an electromagnetic (EM) wave incidence at low grazing angle. The coupling boundary conditions on the interface between two adjacent subdomains are derived when the conformal perfectly matched layer is used as the truncation boundary of the FEM, and the final coupling matrices are obtained by using the inward-looking approach. Because the computational domain with several millions of unknowns can be solved on a personal computer, our FEM-DDM method is powerful for scattering simulation of a very large-scale rough surface with targets presence. In addition to reduction of the memory storage, the superiority of this method in computing time over the conventional FEM is also demonstrated. Our codes are examined by the FEM without DDM, the forward-backward method (FBM), and generalized FBM for some simple cases. Numerical simulations of bistatic scattering from a comprehensive model of a ship on and a target above the 2-D randomly rough perfectly conducting sea surface in large electric scale are obtained, and its functional dependence on many physical parameters of the targets and oceanic status are discussed.