A Multiobjective Memetic Algorithm Based on Particle Swarm Optimization

合集下载

模糊逻辑中英文对照外文翻译文献

模糊逻辑中英文对照外文翻译文献

模糊逻辑中英文对照外文翻译文献(文档含英文原文和中文翻译)译文:模糊逻辑欢迎进入模糊逻辑的精彩世界,你可以用新科学有力地实现一些东西。

在你的技术与管理技能的领域中,增加了基于模糊逻辑分析和控制的能力,你就可以实现除此之外的其他人与物无法做到的事情。

以下就是模糊逻辑的基础知识:随着系统复杂性的增加,对系统精确的阐述变得越来越难,最终变得无法阐述。

于是,终于到达了一个只有靠人类发明的模糊逻辑才能解决的复杂程度。

模糊逻辑用于系统的分析和控制设计,因为它可以缩短工程发展的时间;有时,在一些高度复杂的系统中,这是唯一可以解决问题的方法。

虽然,我们经常认为控制是和控制一个物理系统有关系的,但是,扎德博士最初设计这个概念的时候本意并非如此。

实际上,模糊逻辑适用于生物,经济,市场营销和其他大而复杂的系统。

模糊这个词最早出现在扎德博士于1962年在一个工程学权威刊物上发表论文中。

1963年,扎德博士成为加州大学伯克利分校电气工程学院院长。

那就意味着达到了电气工程领域的顶尖。

扎德博士认为模糊控制是那时的热点,不是以后的热点,更不应该受到轻视。

目前已经有了成千上万基于模糊逻辑的产品,从聚焦照相机到可以根据衣服脏度自我控制洗涤方式的洗衣机等。

如果你在美国,你会很容易找到基于模糊的系统。

想一想,当通用汽车告诉大众,她生产的汽车其反刹车是根据模糊逻辑而造成的时候,那会对其销售造成多么大的影响。

以下的章节包括:1)介绍处于商业等各个领域的人们他们如果从模糊逻辑演变而来的利益中得到好处,以及帮助大家理解模糊逻辑是怎么工作的。

2)提供模糊逻辑是怎么工作的一种指导,只有人们知道了这一点,才能运用它用于做一些对自己有利的事情。

这本书就是一个指导,因此尽管你不是电气领域的专家,你也可以运用模糊逻辑。

需要指出的是有一些针对模糊逻辑的相反观点和批评。

一个人应该学会观察反面的各个观点,从而得出自己的观点。

我个人认为,身为被表扬以及因写关于模糊逻辑论文而受到赞赏的作者,他会认为,在这个领域中的这种批评有点过激。

人工智能基础(习题卷9)

人工智能基础(习题卷9)

人工智能基础(习题卷9)第1部分:单项选择题,共53题,每题只有一个正确答案,多选或少选均不得分。

1.[单选题]由心理学途径产生,认为人工智能起源于数理逻辑的研究学派是( )A)连接主义学派B)行为主义学派C)符号主义学派答案:C解析:2.[单选题]一条规则形如:,其中“←"右边的部分称为(___)A)规则长度B)规则头C)布尔表达式D)规则体答案:D解析:3.[单选题]下列对人工智能芯片的表述,不正确的是()。

A)一种专门用于处理人工智能应用中大量计算任务的芯片B)能够更好地适应人工智能中大量矩阵运算C)目前处于成熟高速发展阶段D)相对于传统的CPU处理器,智能芯片具有很好的并行计算性能答案:C解析:4.[单选题]以下图像分割方法中,不属于基于图像灰度分布的阈值方法的是( )。

A)类间最大距离法B)最大类间、内方差比法C)p-参数法D)区域生长法答案:B解析:5.[单选题]下列关于不精确推理过程的叙述错误的是( )。

A)不精确推理过程是从不确定的事实出发B)不精确推理过程最终能够推出确定的结论C)不精确推理过程是运用不确定的知识D)不精确推理过程最终推出不确定性的结论答案:B解析:6.[单选题]假定你现在训练了一个线性SVM并推断出这个模型出现了欠拟合现象,在下一次训练时,应该采取的措施是()0A)增加数据点D)减少特征答案:C解析:欠拟合是指模型拟合程度不高,数据距离拟合曲线较远,或指模型没有很好地捕 捉到数据特征,不能够很好地拟合数据。

可通过增加特征解决。

7.[单选题]以下哪一个概念是用来计算复合函数的导数?A)微积分中的链式结构B)硬双曲正切函数C)softplus函数D)劲向基函数答案:A解析:8.[单选题]相互关联的数据资产标准,应确保()。

数据资产标准存在冲突或衔接中断时,后序环节应遵循和适应前序环节的要求,变更相应数据资产标准。

A)连接B)配合C)衔接和匹配D)连接和配合答案:C解析:9.[单选题]固体半导体摄像机所使用的固体摄像元件为( )。

acm算法源代码

acm算法源代码
| DIJKSTRA O(E * LOG E)............................................. 4 | BELLMANFORD单源最短路O(VE)................................... 4
| SPFA(SHORTEST PATH FASTER ALGORITHM) .............. 4 | 第K短路(DIJKSTRA)................................................... 5 | 第K短路(A*) .............................................................. 5 | PRIM求MST ..................................................................... 6 | 次小生成树O(V^2)....................................................... 6 | 最小生成森林问题(K颗树)O(MLOGM). ....................... 6 | 有向图最小树形图 ......................................................... 6
(O(NLOGN + Q)).............................................................19 | RMQ离线算法 O(N*LOGN)+O(1)求解LCA...............19 | LCA离线算法 O(E)+O(1).........................................20 | 带权值的并查集 ...........................................................20 | 快速排序 .......................................................................20 | 2 台机器工作调度........................................................20 | 比较高效的大数 ...........................................................20 | 普通的大数运算 ...........................................................21 | 最长公共递增子序列 O(N^2)....................................22 | 0-1 分数规划...............................................................22 | 最长有序子序列(递增/递减/非递增/非递减) ....22 | 最长公共子序列 ...........................................................23 | 最少找硬币问题(贪心策略-深搜实现) .................23 | 棋盘分割 .......................................................................23 | 汉诺塔 ...........................................................................23

作为极限建筑空间设计依据的人体运动包络体研究

作为极限建筑空间设计依据的人体运动包络体研究

摘要城市化进程不断的发展导致了城市中心的地块不停的被分隔,因此出现了许多在空间极为局促、环境极为苛刻或使用者行为活动受到一定限制的条件下的极限建筑空间。

在此情况下,根据行为建筑学相关理论及设计方法,计算出满足使用者功能需求的最小建筑空间,显得十分重要。

然而现有的极限建筑空间的设计数据主要是根据人体百分位参数进行建筑空间以及空间中固定物的设计。

这样的设计方式,在很大程度上存在着缺少设计针对性、空间尺寸不合理、空间使用效率低、建筑能耗大等问题。

针对这一现象,本研究将首先详细阐述通过计算机编程方式模拟人体运动方式,并通过运动轨迹计算得出人体运动包络体。

人体运动包络体模拟是行为建筑学理论研究推理过程中所采用的一种模拟法。

从而克服了传统实验法存在的样本人体尺度从二维平面研究转化为三维立体空间研究。

在此基础之上,该论文将探讨现存极限建筑存在的问题以及如何在实际建筑设计中,通过计算空间使用者运动包络体得到他们的详细数据,并以此确定使用者在空间中的活动范围,作为极限建筑空间设计的重要参考依据。

这样的设计方式,可以计算出可以满足使用需求的极限建筑空间形态与体积,从而保证建筑空间可以满足使用者对使用功能的基本需求,提高建筑空间使用效率。

另一方面,人体运动包络体可以用于优化极限空间中固定物的位置与尺寸、形状,根据具体使用者的实际测量参数的进行个性化的私人定制,并保证了固定物的基本使用功能。

关键词:运动包络体;极限建筑空间;行为建筑学;模拟法;空间效率AbstractThe land in the center of the city is constantly divided for the sake of urbanization development. As a result, an increasing number of limited architectural space was designed and built. The environment of such kind of space is usually cramped. And the users’ behavior is also limited. In this case, it is of great importance to calculate the minimum size of space which can meet the basic functional needs of the users. However, the existing data for limited architectural extent, leads to an increasing number serious issues, such as lacking pertinence, unreasonable space size, low space efficiency and high energy consumption.In order to solve this issue, this essay will first simulate the movement of human body by computer programming. After that, enveloping solid will be calculated by the trail of human body. Enveloping solid simulation is a basic simulating method in the inference procedure of behavioral architecture. Compared with traditional experiments, there will be no sample quantity limitation and anthropogenic factor in simulating process. And the 2-dimensional human parameter comes to 3 dimensional.Based on which, this essay will explore the existing problems on limited architectural space design and how to use enveloping solid simulation in architecture design. In the first stage, the design data of users can be get from the process of enveloping solid simulation. And the users’ parameter shows the range of activity, which is important reference frame in design procedure. By this method, the functional needs of users can be meet. And space efficiency can also be improved. What’s more, enveloping solid can be used in optimizing the shape and location of fixtures in building as well.Keywords:enveloping solid, limited architectural space,behavioral architecture, simulation, space efficiency目录摘要 (1)Abstract (2)第1章绪论 (1)1.1课题背景及研究的目的和意义 (1)1.1.1 课题的研究背景 (1)1.1.2 课题的研究目的和意义 (2)1.2相关概念概述 (3)1.2.1 极限建筑空间的概念 (3)1.2.2 “包络体”的概念及构成概述 (3)1.3国内外研究现状及分析 (4)1.3.1 行为建筑学 (4)1.3.2 极限建筑空间 (4)1.3.3 包络体的应用及计算方式 (6)1.4研究内容、方法与框架 (11)1.4.1 课题的研究内容 (11)1.4.2 研究方法 (12)1.4.3 课题的研究框架 (14)第2章研究基础 (15)2.1人体运动学、运动解剖学 (15)2.1.1 人体运动形式 (15)2.1.2 人体运动的特性与坐标系建立 (15)2.2人体测量学与程序人体基本参数设定 (17)2.2.1 人体上肢静态尺寸测量 (17)2.2.2 程序人体基本参数设定 (18)2.3计算机编程 (19)2.3.1 模拟软件 (19)2.3.2 Toxiclibs类库引用与运动轨迹的向量表示 (19)2.3.3 HE_Mesh类库引用与包络曲面生成 (20)2.4本章小结 (20)第3章程序模拟 (21)3.1程序逻辑 (21)3.1.1 程序参数设定 (21)3.1.2 上肢运动轨迹模拟 (22)3.1.3 上肢运动包络体生成 (30)3.2不同人体参数对模拟结果的影响 (30)3.2.1 儿童(四肢长度对模拟结果的影响) (30)3.2.2 老年人(活动角度对模拟结果的影响) (33)3.2.3 残疾人(残肢对模拟结果的影响) (34)3.2.4 数据对比 (35)3.3“人体运动包络体”程序对行为建筑学研究方法的扩展 (36)3.3.1 行为建筑学研究的一般方法以及主要存在问题 (36)3.3.2 “人体运动包络体”模拟对行为建筑学研究方法的贡献 (37)3.4本章小结 (39)第4章 (40)4.1计算满足使用需求的极限建筑空间形态与体积 (40)4.1.1 满足功能需求,提高空间使用效率 (40)4.1.2 根据运动轨迹预测使用者所需的三维建筑空间 (45)4.1.3 节约能源 (49)4.2优化极限空间中固定物的位置与尺寸、形状 (50)4.2.1 包络体与极限空间中固定物的位置 (51)4.2.2 包络体与极限空间中固定物的尺寸 (55)4.2.3 包络体与固定物的三维空间组合 (57)4.3本章小结 (58)结论 (59)参考文献 (60)附录 (63) (74)致谢 (75)第1章绪论1.1 课题背景及研究的目的和意义1.1.1 课题的研究背景古代有蜗居的说法,用“蜗舍”比喻“圆舍”“蜗”字描述的是空间的形状,后来逐渐演变为居住空间狭小的意思。

纹理物体缺陷的视觉检测算法研究--优秀毕业论文

纹理物体缺陷的视觉检测算法研究--优秀毕业论文

摘 要
在竞争激烈的工业自动化生产过程中,机器视觉对产品质量的把关起着举足 轻重的作用,机器视觉在缺陷检测技术方面的应用也逐渐普遍起来。与常规的检 测技术相比,自动化的视觉检测系统更加经济、快捷、高效与 安全。纹理物体在 工业生产中广泛存在,像用于半导体装配和封装底板和发光二极管,现代 化电子 系统中的印制电路板,以及纺织行业中的布匹和织物等都可认为是含有纹理特征 的物体。本论文主要致力于纹理物体的缺陷检测技术研究,为纹理物体的自动化 检测提供高效而可靠的检测算法。 纹理是描述图像内容的重要特征,纹理分析也已经被成功的应用与纹理分割 和纹理分类当中。本研究提出了一种基于纹理分析技术和参考比较方式的缺陷检 测算法。这种算法能容忍物体变形引起的图像配准误差,对纹理的影响也具有鲁 棒性。本算法旨在为检测出的缺陷区域提供丰富而重要的物理意义,如缺陷区域 的大小、形状、亮度对比度及空间分布等。同时,在参考图像可行的情况下,本 算法可用于同质纹理物体和非同质纹理物体的检测,对非纹理物体 的检测也可取 得不错的效果。 在整个检测过程中,我们采用了可调控金字塔的纹理分析和重构技术。与传 统的小波纹理分析技术不同,我们在小波域中加入处理物体变形和纹理影响的容 忍度控制算法,来实现容忍物体变形和对纹理影响鲁棒的目的。最后可调控金字 塔的重构保证了缺陷区域物理意义恢复的准确性。实验阶段,我们检测了一系列 具有实际应用价值的图像。实验结果表明 本文提出的纹理物体缺陷检测算法具有 高效性和易于实现性。 关键字: 缺陷检测;纹理;物体变形;可调控金字塔;重构
Keywords: defect detection, texture, object distortion, steerable pyramid, reconstruction
II

Modified PSO algorithm for solving planar graph coloring problem

Modified PSO algorithm for solving planar graph coloring problem
4.3.1. Instance 1
4.3.2. Instance 2
4.3.3. Instance 3
5. Inter-cluster load balancing through self-organizing cluster approach
5.1. Performance evaluation
935
On Efficient Sparse Integer Matrix Smith Normal Form Computations Original Research Article
Journal of Symbolic Computation, Volume 32, Issues 1-2, July 2001, Pages 71-99
6.3.2. Liveliness property
6.3.3. Deadlock
7. Conclusion
References
Vitae Purchase
Research highlights
? A hybrid load balancing (HLB) approach in trusted clusters is proposed for HPC. ? HLB reduces network traffic by 80%–90% and increases CPU utilization by 40%–50%. ? The AWT and MRT of remote processes are reduced by 13%–26% using ReJAM. ? The stability analysis of JMM using PA ensures the finite sequences of transitions. ? On the basis of these properties, JM model has been proved safe and reliable.

Finding community structure in networks using the eigenvectors of matrices

Finding community structure in networks using the eigenvectors of matrices
Finding community structure in networks using the eigenvectors of matrices
M. E. J. Newman
Department of Physics and Center for the Study of Complex Systems, University of Michigan, Ann Arbor, MI 48109–1040
We consider the problem of detecting communities or modules in networks, groups of vertices with a higher-than-average density of edges connecting them. Previous work indicates that a robust approach to this problem is the maximization of the benefit function known as “modularity” over possible divisions of a network. Here we show that this maximization process can be written in terms of the eigenspectrum of a matrix we call the modularity matrix, which plays a role in community detection similar to that played by the graph Laplacian in graph partitioning calculations. This result leads us to a number of possible algorithms for detecting community structure, as well as several other results, including a spectral measure of bipartite structure in neteasure that identifies those vertices that occupy central positions within the communities to which they belong. The algorithms and measures proposed are illustrated with applications to a variety of real-world complex networks.

Multiobjective Optimization Problems With Complicated Pareto Sets, MOEA_D and NSGA-II

Multiobjective Optimization Problems With Complicated Pareto Sets, MOEA_D and NSGA-II
H. Li was with the Department of Computing and Electronic Systems, University of Essex, Colchester CO4 3SQ, U.K. He is now with the School of Computer Science, University of Nottingham, Nottingham NG8 1BB, U.K. (e-mail: hzl@).
I. INTRODUCTION
A multiobjective optimization problem (MOP) can be stated as follows:
minimize
subject to
(1)
where is the decision (variable) space, is the objective
MOP.
Let
,
be two
vectors, is said to dominate if
for all
,
and
.1 A point
is called (globally) Pareto optimal
if there is no
such that
Hale Waihona Puke dominates. The set
of all the Pareto optimal points, denoted by , is called the
Pareto set. The set of all the Pareto objective vectors,
, is called the Pareto front [1].

基于低秩约束的熵加权多视角模糊聚类算法

基于低秩约束的熵加权多视角模糊聚类算法

基于低秩约束的熵加权多视角模糊聚类算法张嘉旭 1王 骏 1, 2张春香 1林得富 1周 塔 3王士同1摘 要 如何有效挖掘多视角数据内部的一致性以及差异性是构建多视角模糊聚类算法的两个重要问题. 本文在Co-FKM 算法框架上, 提出了基于低秩约束的熵加权多视角模糊聚类算法(Entropy-weighting multi-view fuzzy C-means with low rank constraint, LR-MVEWFCM). 一方面, 从视角之间的一致性出发, 引入核范数对多个视角之间的模糊隶属度矩阵进行低秩约束; 另一方面, 基于香农熵理论引入视角权重自适应调整策略, 使算法根据各视角的重要程度来处理视角间的差异性. 本文使用交替方向乘子法(Alternating direction method of multipliers, ADMM)进行目标函数的优化. 最后, 人工模拟数据集和UCI (University of California Irvine)数据集上进行的实验结果验证了该方法的有效性.关键词 多视角模糊聚类, 香农熵, 低秩约束, 核范数, 交替方向乘子法引用格式 张嘉旭, 王骏, 张春香, 林得富, 周塔, 王士同. 基于低秩约束的熵加权多视角模糊聚类算法. 自动化学报, 2022,48(7): 1760−1770DOI 10.16383/j.aas.c190350Entropy-weighting Multi-view Fuzzy C-means With Low Rank ConstraintZHANG Jia-Xu 1 WANG Jun 1, 2 ZHANG Chun-Xiang 1 LIN De-Fu 1 ZHOU Ta 3 WANG Shi-Tong 1Abstract Effective mining both internal consistency and diversity of multi-view data is important to develop multi-view fuzzy clustering algorithms. In this paper, we propose a novel multi-view fuzzy clustering algorithm called en-tropy-weighting multi-view fuzzy c-means with low-rank constraint (LR-MVEWFCM). On the one hand, we intro-duce the nuclear norm as the low-rank constraint of the fuzzy membership matrix. On the other hand, the adaptive adjustment strategy of view weight is introduced to control the differences among views according to the import-ance of each view. The learning criterion can be optimized by the alternating direction method of multipliers (ADMM). Experimental results on both artificial and UCI (University of California Irvine) datasets show the effect-iveness of the proposed method.Key words Multi-view fuzzy clustering, Shannon entropy, low-rank constraint, nuclear norm, alternating direction method of multipliers (ADMM)Citation Zhang Jia-Xu, Wang Jun, Zhang Chun-Xiang, Lin De-Fu, Zhou Ta, Wang Shi-Tong. Entropy-weighting multi-view fuzzy C-means with low rank constraint. Acta Automatica Sinica , 2022, 48(7): 1760−1770随着多样化信息获取技术的发展, 人们可以从不同途径或不同角度来获取对象的特征数据, 即多视角数据. 多视角数据包含了同一对象不同角度的信息. 例如: 网页数据中既包含网页内容又包含网页链接信息; 视频内容中既包含视频信息又包含音频信息; 图像数据中既涉及颜色直方图特征、纹理特征等图像特征, 又涉及描述该图像内容的文本.多视角学习能有效地对多视角数据进行融合, 避免了单视角数据数据信息单一的问题[1−4].多视角模糊聚类是一种有效的无监督多视角学习方法[5−7]. 它通过在多视角聚类过程中引入各样本对不同类别的模糊隶属度来描述各视角下样本属于该类别的不确定性程度. 经典的工作有: 文献[8]以经典的单视角模糊C 均值(Fuzzy C-means, FCM)算法作为基础模型, 利用不同视角间的互补信息确定协同聚类的准则, 提出了Co-FC (Collaborative fuzzy clustering)算法; 文献[9]参考文献[8]的协同思想提出Co-FKM (Multiview fuzzy clustering algorithm collaborative fuzzy K-means)算法, 引入双视角隶属度惩罚项, 构造了一种新型的无监督多视角协同学习方法; 文献[10]借鉴了Co-FKM 和Co-FC 所使用的双视角约束思想, 通过引入视角权重, 并采用集成策略来融合多视角的模糊隶属收稿日期 2019-05-09 录用日期 2019-07-17Manuscript received May 9, 2019; accepted July 17, 2019国家自然科学基金(61772239), 江苏省自然科学基金(BK20181339)资助Supported by National Natural Science Foundation of China (61772239) and Natural Science Foundation of Jiangsu Province (BK20181339)本文责任编委 刘艳军Recommended by Associate Editor LIU Yan-Jun1. 江南大学数字媒体学院 无锡 2141222. 上海大学通信与信息工程学院 上海 2004443. 江苏科技大学电子信息学院 镇江2121001. School of Digital Media, Jiangnan University, Wuxi 2141222. School of Communication and Information Engineering,Shanghai University, Shanghai 2004443. School of Electronic Information, Jiangsu University of Science and Technology,Zhenjiang 212100第 48 卷 第 7 期自 动 化 学 报Vol. 48, No. 72022 年 7 月ACTA AUTOMATICA SINICAJuly, 2022度矩阵, 提出了WV-Co-FCM (Weighted view colla-borative fuzzy C-means) 算法; 文献[11]通过最小化双视角下样本与聚类中心的欧氏距离来减小不同视角间的差异性, 基于K-means 聚类框架提出了Co-K-means (Collaborative multi-view K-means clustering)算法; 在此基础上, 文献[12]提出了基于模糊划分的TW-Co-K-means (Two-level wei-ghted collaborative K-means for multi-view clus-tering)算法, 对Co-K-means 算法中的双视角欧氏距离加入一致性权重, 获得了比Co-K-means 更好的多视角聚类结果. 以上多视角聚类方法都基于成对视角来构造不同的正则化项来挖掘视角之间的一致性和差异性信息, 缺乏对多个视角的整体考虑.一致性和差异性是设计多视角聚类算法需要考虑的两个重要原则[10−14]. 一致性是指在多视角聚类过程中, 各视角的聚类结果应该尽可能保持一致.在设计多视角聚类算法时, 往往通过协同、集成等手段来构建全局划分矩阵, 从而得到最终的聚类结果[14−16]. 差异性是指多视角数据中的每个视角均反映了对象在不同方面的信息, 这些信息互为补充[10],在设计多视角聚类算法时需要对这些信息进行充分融合. 综合考虑这两方面的因素, 本文拟提出新型的低秩约束熵加权多视角模糊聚类算法(Entropy-weigh-ting multi-view fuzzy C-means with low rank con-straint, LR-MVEWFCM), 其主要创新点可以概括为以下3个方面:1)在模糊聚类框架下提出了面向视角一致性的低秩约束准则. 已有的多视角模糊聚类算法大多基于成对视角之间的两两关系来构造正则化项, 忽视了多个视角的整体一致性信息. 本文在模糊聚类框架下从视角全局一致性出发引入低秩约束正则化项, 从而得到新型的低秩约束多视角模糊聚类算法.2) 在模糊聚类框架下同时考虑多视角聚类的一致性和差异性, 在引入低秩约束的同时进一步使用面向视角差异性的多视角香农熵加权策略; 在迭代优化的过程中, 通过动态调节视角权重系数来突出具有更好分离性的视角的权重, 从而提高聚类性能.3)在模糊聚类框架下首次使用交替方向乘子法(Alternating direction method of multipliers,ADMM)[15]对LR-MVEWFCM 算法进行优化求解.N D K C m x j,k j k j =1,···,N k =1,···,K v i,k k i i =1,···,C U k =[µij,k ]k µij,k k j i 在本文中, 令 为样本总量, 为样本维度, 为视角数目, 为聚类数目, 为模糊指数. 设 表示多视角场景中第 个样本第 个视角的特征向量, , ; 表示第 个视角下, 第 个聚类中心, ; 表示第 个视角下的模糊隶属度矩阵, 其中 是第 个视角下第 个样本属于第 个聚类中心的模i =1,···,C j =1,···,N.糊隶属度, , 本文第1节在相关工作中回顾已有的经典模糊C 均值聚类算法FCM 模型[17]和多视角模糊聚类Co-FKM 模型[9]; 第2节将低秩理论与多视角香农熵理论相结合, 提出本文的新方法; 第3节基于模拟数据集和UCI (University of California Irvine)数据集验证本文算法的有效性, 并给出实验分析;第4节给出实验结论.1 相关工作1.1 模糊C 均值聚类算法FCMx 1,···,x N ∈R D U =[µi,j ]V =[v 1,v 2,···,v C ]设单视角环境下样本 , 是模糊划分矩阵, 是样本的聚类中心. FCM 算法的目标函数可表示为J FCM 可得到 取得局部极小值的必要条件为U 根据式(2)和式(3)进行迭代优化, 使目标函数收敛于局部极小点, 从而得到样本属于各聚类中心的模糊划分矩阵 .1.2 多视角模糊聚类Co-FKM 模型在经典FCM算法的基础上, 文献[9]通过引入视角协同约束正则项, 对视角间的一致性信息加以约束, 提出了多视角模糊聚类Co-FKM 模型.多视角模糊聚类Co-FKM 模型需要满足如下条件:J Co-FKM 多视角模糊聚类Co-FKM 模型的目标函数 定义为7 期张嘉旭等: 基于低秩约束的熵加权多视角模糊聚类算法1761η∆∆式(5)中, 表示协同划分参数; 表示视角一致项,由式(6)可知, 当各视角趋于一致时, 将趋于0.µij,k 迭代得到各视角的模糊隶属度 后, 为了最终得到一个具有全局性的模糊隶属度划分矩阵, Co-FKM 算法对各视角下的模糊隶属度采用几何平均的方法, 得到数据集的整体划分, 具体形式为ˆµij 其中, 为全局模糊划分结果.2 基于低秩约束的熵加权多视角模糊聚类算法针对当前多视角模糊聚类算法研究中存在的不足, 本文提出一种基于低秩约束的熵加权多视角模糊聚类新方法LR-MVEWFCM. 一方面通过向多视角模糊聚类算法的目标学习准则中引入低秩约束项, 在整体上控制聚类过程中各视角的一致性; 另一方面基于香农熵理论, 通过熵加权机制来控制各视角之间的差异性.同时使用交替方向乘子法对模型进行优化求解.U 1,···,U K U U U 设多视角隶属度 融合为一个整体的隶属度矩阵 , 将矩阵 的秩函数凸松弛为核范数, 通过对矩阵 进行低秩约束, 可以将多视角数据之间的一致性问题转化为核范数最小化问题进行求解, 具体定义为U =[U 1···U K ]T ∥·∥∗其中, 表示全局划分矩阵, 表示核范数. 式(8)的优化过程保证了全局划分矩阵的低秩约束. 低秩约束的引入, 可以弥补当前大多数多视角聚类算法仅能基于成对视角构建约束的缺陷, 从而更好地挖掘多视角数据中包含的全局一致性信息.目前已有的多视角的聚类算法在处理多视角数据时, 通常默认每个视角平等共享聚类结果[11], 但实际上某些视角的数据往往因空间分布重叠而导致可分性较差. 为避免此类视角的数据过多影响聚类效果,本文拟对各视角进行加权处理, 并构建香农熵正则项从而在聚类过程中有效地调节各视角之间的权重, 使得具有较好可分离性的视角的权重系数尽可能大, 以达到更好的聚类效果.∑Kk =1w k =1w k ≥0令视角权重系数 且 , 则香农熵正则项表示为U w k U =[U 1···U K ]T w =[w 1,···,w k ,···,w K ]K 综上所述, 本文作如下改进: 首先, 用本文提出的低秩约束全局模糊隶属度矩阵 ; 其次, 计算损失函数时考虑视角权重 , 并加入视角权重系数的香农熵正则项. 设 ; 表示 个视角下的视角权重. 本文所构建LR-MVEWFCM 的目标函数为其中, 约束条件为m =2本文取模糊指数 .2.1 基于ADMM 的求解算法(11)在本节中, 我们将使用ADMM 方法, 通过交替方向迭代的策略来实现目标函数 的最小化.g (Z )=θ∥Z ∥∗(13)(10)最小化式 可改写为如下约束优化问题:其求解过程可分解为如下几个子问题:V w U V 1) -子问题. 固定 和 , 更新 为1762自 动 化 学 报48 卷(15)v i,k 通过最小化式 , 可得到 的闭合解为U w Q Z U 2) -子问题. 固定 , 和 , 更新 为(17)U (t +1)通过最小化式 , 可得到 的封闭解为w V U w 3) -子问题. 固定 和 , 更新 为Z Q U Z(20)通过引入软阈值算子, 可得式 的解为U (t+1)+Q (t )=A ΣB T U (t +1)+Q (t )S θ/ρ(Σ)=diag ({max (0,σi −θ/ρ)})(i =1,2,···,N )其中, 为矩阵 的奇异值分解, 核范数的近邻算子可由软阈值算子给出.Q Z U Q 5) -子问题. 固定 和 , 更新 为w =[w 1,···,w k ,···,w K ]U ˜U经过上述迭代过程, 目标函数收敛于局部极值,同时得到不同视角下的模糊隶属度矩阵. 本文借鉴文献[10]的集成策略, 使用视角权重系数 和模糊隶属度矩阵 来构建具有全局特性的模糊空间划分矩阵 :w k U k k 其中, , 分别表示第 个视角的视角权重系数和相应的模糊隶属度矩阵.LR-MVEWFCM 算法描述如下:K (1≤k ≤K )X k ={x 1,k ,···,x N,k }C ϵT 输入. 包含 个视角的多视角样本集, 其中任意一个视角对应样本集 , 聚类中心 , 迭代阈值 , 最大迭代次数 ;v (t )i,k ˜Uw k 输出. 各视角聚类中心 , 模糊空间划分矩阵和各视角权重 ;V (t )U (t )w (t )t =0步骤1. 随机初始化 , 归一化 及 ,;(21)v (t +1)i,k 步骤2. 根据式 更新 ;(23)U (t +1)步骤3. 根据式 更新 ;(24)w (t +1)k 步骤4. 根据式 更新 ;(26)Z (t +1)步骤5. 根据式 更新 ;(27)Q (t +1)步骤6. 根据式 更新 ;L (t +1)−L (t )<ϵt >T 步骤7. 如果 或者 , 则算法结束并跳出循环, 否则, 返回步骤2;w k U k (23)˜U步骤8. 根据步骤7所获取的各视角权重 及各视角下的模糊隶属度 , 使用式 计算 .2.2 讨论2.2.1 与低秩约束算法比较近年来, 基于低秩约束的机器学习模型得到了广泛的研究. 经典工作包括文献[16]中提出LRR (Low rank representation)模型, 将矩阵的秩函数凸松弛为核范数, 通过求解核范数最小化问题, 求得基于低秩表示的亲和矩阵; 文献[14]提出低秩张量多视角子空间聚类算法(Low-rank tensor con-strained multiview subspace clustering, LT-MSC),7 期张嘉旭等: 基于低秩约束的熵加权多视角模糊聚类算法1763在各视角间求出带有低秩约束的子空间表示矩阵;文献 [18] 则进一步将低秩约束引入多模型子空间聚类算法中, 使算法模型取得了较好的性能. 本文将低秩约束与多视角模糊聚类框架相结合, 提出了LR-MVEWFCM 算法, 用低秩约束来实现多视角数据间的一致性. 本文方法可作为低秩模型在多视角模糊聚类领域的重要拓展.2.2.2 与多视角Co-FKM 算法比较图1和图2分别给出了多视角Co-FKM 算法和本文LR-MVEWFCM 算法的工作流程.多视角数据Co-FKM视角 1 数据视角 2 数据视角 K 数据各视角间两两约束各视角模糊隶属度集成决策函数划分矩阵ÛU 1U 2U K图 1 Co-FKM 算法处理多视角聚类任务工作流程Fig. 1 Co-FKM algorithm for multi-view clustering task本文算法与经典的多视角Co-FKM 算法在多视角信息的一致性约束和多视角聚类结果的集成策略上均有所不同. 在多视角信息的一致性约束方面, 本文将Co-FKM 算法中的视角间两两约束进一步扩展到多视角全局一致性约束; 在多视角聚类结果的集成策略上, 本文不同于Co-FKM 算法对隶属度矩阵简单地求几何平均值的方式, 而是将各视角隶属度与视角权重相结合, 构建具有视角差异性的集成决策函数.3 实验与分析3.1 实验设置本文采用模拟数据集和UCI 中的真实数据集进行实验验证, 选取FCM [17]、CombKM [19]、Co-FKM [9]和Co-Clustering [20]这4个聚类算法作为对比算法, 参数设置如表1所示. 实验环境为: Intel Core i5-7400 CPU, 其主频为2.3 GHz, 内存为8 GB.编程环境为MATLAB 2015b.本文采用如下两个性能指标对各算法所得结果进行评估.1) 归一化互信息(Normalized mutual inform-ation, NMI)[10]N i,j i j N i i N j j N 其中, 表示第 类与第 类的契合程度, 表示第 类中所属样本量, 表示第 类中所属样本量, 而 表示数据的样本总量;2) 芮氏指标(Rand index, RI)[10]表 1 参数定义和设置Table 1 Parameter setting in the experiments算法算法说明参数设置FCM 经典的单视角模糊聚类算法m =min (N,D −1)min (N,D −1)−2N D 模糊指数 ,其中, 表示样本数, 表示样本维数CombKM K-means 组合 算法—Co-FKM 多视角协同划分的模糊聚类算法m =min (N,D −1)min (N,D −1)−2η∈K −1K K ρ=0.01模糊指数 , 协同学习系数 ,其中, 为视角数, 步长 Co-Clustering 基于样本与特征空间的协同聚类算法λ∈{10−3,10−2, (103)µ∈{10−3,10−2,···,103}正则化系数 ,正则化系数 LR-MVEWFCM 基于低秩约束的熵加权多视角模糊聚类算法λ∈{10−5,10−4, (105)θ∈{10−3,10−2, (103)m =2视角权重平衡因子 , 低秩约束正则项系数, 模糊指数 MVEWFCMθ=0LR-MVEWFCM 算法中低秩约束正则项系数 λ∈{10−5,10−4, (105)m =2视角权重平衡因子 , 模糊指数 多视角数据差异性集成决策函数各视角模糊隶属度U 1U 2U K各视角权重W 1W 2W kLR-MVEWFCM 视角 1 数据视角 2 数据视角 K 数据整体约束具有视角差异性的划分矩阵Û图 2 LR-MVEWFCM 算法处理多视角聚类任务工作流程Fig. 2 LR-MVEWFCM algorithm for multi-viewclustering task1764自 动 化 学 报48 卷f 00f 11N [0,1]其中, 表示具有不同类标签且属于不同类的数据配对点数目, 则表示具有相同类标签且属于同一类的数据配对点数目, 表示数据的样本总量. 以上两个指标的取值范围介于 之间, 数值越接近1, 说明算法的聚类性能越好. 为了验证算法的鲁棒性, 各表中统计的性能指标值均为算法10次运行结果的平均值.3.2 模拟数据集实验x,y,z A 1x,y,z A 2x,y,z A 3x,y,z 为了评估本文算法在多视角数据集上的聚类效果, 使用文献[10]的方法来构造具有三维特性的模拟数据集A ( ), 其具体生成过程为: 首先在MATLAB 环境下采用正态分布随机函数normrnd 构建数据子集 ( ), ( )和 ( ), 每组对应一个类簇, 数据均包含200个样本.x,y,z 其中第1组与第2组数据集在特征z 上数值较为接近, 第2组与第3组数据集在特征x 上较为接近;然后将3组数据合并得到集合A ( ), 共计600个样本; 最后对数据集内的样本进行归一化处理. 我们进一步将特征x , y , z 按表2的方式两两组合, 从而得到多视角数据.表 2 模拟数据集特征组成Table 2 Characteristic composition of simulated dataset视角包含特征视角 1x,y 视角 2y,z 视角 3x,z将各视角下的样本可视化, 如图3所示.通过观察图3可以发现, 视角1中的数据集在空间分布上具有良好的可分性, 而视角2和视角3的数据在空间分布上均存在着一定的重叠, 从而影Z YZZXYX(a) 模拟数据集 A (a) Dataset A(b) 视角 1 数据集(b) View 1(c) 视角 2 数据集(c) View 2(d) 视角 3 数据集(d) View 3图 3 模拟数据集及各视角数据集Fig. 3 Simulated data under multiple views7 期张嘉旭等: 基于低秩约束的熵加权多视角模糊聚类算法1765响了所在视角下的聚类性能. 通过组合不同视角生成若干新的数据集, 如表3所示, 并给出了LR-MVEWFCM重复运行10次后的平均结果和方差.表 3 模拟数据实验算法性能对比Table 3 Performance comparison of the proposedalgorithms on simulated dataset编号包含特征NMI RI1视角1 1.0000 ± 0.0000 1.0000 ± 0.0000 2视角20.7453 ± 0.00750.8796 ± 0.0081 3视角30.8750 ± 0.00810.9555 ± 0.0006 4视角1, 视角2 1.0000 ± 0.0000 1.0000 ± 0.0000 5视角1, 视角3 1.0000 ± 0.0000 1.0000 ± 0.0000 6视角2, 视角30.9104 ± 0.03960.9634 ± 0.0192 7视角2, 视角3 1.0000 ± 0.0000 1.0000 ± 0.0000对比LR-MVEWFCM在数据集1~3上的性能, 我们发现本文算法在视角1上取得了最为理想的效果, 在视角3上的性能要优于视角2, 这与图3中各视角数据的空间可分性是一致的. 此外, 将各视角数据两两组合构成新数据集4~6后, LR-MVEWFCM算法都得到了比单一视角更好的聚类效果, 这都说明了本文采用低秩约束来挖掘多视角数据中一致性的方法, 能够有效提高聚类性能.基于多视角数据集7, 我们进一步给出本文算法与其他经典聚类算法的比较结果.从表4中可以发现, 由于模拟数据集在某些特征空间下具有良好的空间可分性, 所以无论是本文的算法还是Co-Clustering算法、FCM算法等算法均取得了很好的聚类效果, 而CombKM算法的性能较之以上算法则略有不足, 分析其原因在于CombKM算法侧重于挖掘样本之间的信息, 却忽视了多视角之间的协作, 而本文算法通过使用低秩约束进一步挖掘了多视角之间的全局一致性, 因而得到了比CombKM算法更好的聚类效果.3.3 真实数据集实验本节采用5个UCI数据集: 1) Iris数据集; 2) Image Segmentation (IS) 数据集; 3) Balance数据集; 4) Ionosphere数据集; 5) Wine数据集来进行实验. 由于这几个数据集均包含了不同类型的特征,所以可以将这些特征进行重新分组从而构造相应的多视角数据集. 表5给出了分组后的相关信息.我们在多视角数据集上运行各多视角聚类算法; 同时在原数据集上运行FCM算法. 相关结果统计见表6和表7.NMI RI通过观察表6和表7中的和指标值可知, Co-FKM算法的聚类性能明显优于其他几种经典聚类算法, 而相比于Co-FKM算法, 由于LR-MVEWFCM采用了低秩正则项来挖掘多视角数据之间的一致性关系, 并引入多视角自适应熵加权策略, 从而有效控制各视角之间的差异性. 很明显, 这种聚类性能更为优异和稳定, 且收敛性的效果更好.表6和表7中的结果也展示了在IS、Balance、Iris、Ionosphere和Wine数据集上, 其NMI和RI指标均提升3 ~ 5个百分点, 这也说明了本文算法在多视角聚类过程中的有效性.为进一步说明本文低秩约束发挥的积极作用,将LR-MVEWFCM算法和MVEWFCM算法共同进行实验, 算法的性能对比如图4所示.从图4中不难发现, 无论在模拟数据集上还是UCI真实数据集上, 相比较MVEWFCM算法, LR-MVEWFCM算法均可以取得更好的聚类效果. 因此可见, LR-MVEWFCM目标学习准则中的低秩约束能够有效利用多视角数据的一致性来提高算法的聚类性能.为研究本文算法的收敛性, 同样选取8个数据集进行收敛性实验, 其目标函数变化如图5所示.从图5中可以看出, 本文算法在真实数据集上仅需迭代15次左右就可以趋于稳定, 这说明本文算法在速度要求较高的场景下具有较好的实用性.综合以上实验结果, 我们不难发现, 在具有多视角特性的数据集上进行模糊聚类分析时, 多视角模糊聚类算法通常比传统单视角模糊聚类算法能够得到更优的聚类效果; 在本文中, 通过在多视角模糊聚类学习中引入低秩约束来增强不同视角之间的一致性关系, 并引入香农熵调节视角权重关系, 控制不同视角之间的差异性, 从而得到了比其他多视角聚类算法更好的聚类效果.表 4 模拟数据集7上各算法的性能比较Table 4 Performance comparison of the proposed algorithms on simulated dataset 7数据集指标Co-Clustering CombKM FCM Co-FKM LR-MVEWFCMA NMI-mean 1.00000.9305 1.0000 1.0000 1.0000 NMI-std0.00000.14640.00000.00000.0000 RI-mean 1.00000.9445 1.0000 1.0000 1.0000 RI-std0.00000.11710.00000.00000.00001766自 动 化 学 报48 卷3.4 参数敏感性实验LR-MVEWFCM算法包含两个正则项系数,λθθθθλλ即视角权重平衡因子和低秩约束正则项系数, 图6以LR-MVEWFCM算法在模拟数据集7上的实验为例, 给出了系数从0到1000过程中, 算法性能的变化情况, 当低秩正则项系数= 0时, 即不添加此正则项, 算法的性能最差, 验证了本文加入的低秩正则项的有效性, 当值变化过程中, 算法的性能相对变化较小, 说明本文算法在此数据集上对于值变化不敏感, 具有一定的鲁棒性; 而当香农熵正则项系数= 0时, 同样算法性能较差, 也说明引入此正则项的合理性. 当值变大时, 发现算法的性能也呈现变好趋势, 说明在此数据集上, 此正则项相对效果比较明显.4 结束语本文从多视角聚类学习过程中的一致性和差异性两方面出发, 提出了基于低秩约束的熵加权多视角模糊聚类算法. 该算法采用低秩正则项来挖掘多视角数据之间的一致性关系, 并引入多视角自适应熵加权策略从而有效控制各视角之间的差异性,从而提高了算法的性能. 在模拟数据集和真实数据集上的实验均表明, 本文算法的聚类性能优于其他多视角聚类算法. 同时本文算法还具有迭代次数少、收敛速度快的优点, 具有良好的实用性. 由于本文采用经典的FCM框架, 使用欧氏距离来衡量数据对象之间的差异,这使得本文算法不适用于某些高维数据场景. 如何针对高维数据设计多视角聚类算法, 这也将是我们今后的研究重点.表 5 基于UCI数据集构造的多视角数据Table 5 Multi-view data constructdedbased on UCI dataset编号原数据集说明视角特征样本视角类别8IS Shape92 31027 RGB99Iris Sepal长度215023 Sepal宽度Petal长度2Petal宽度10Balance 天平左臂重量262523天平左臂长度天平右臂重量2天平右臂长度11Iris Sepal长度115043 Sepal宽度1Petal长度1Petal宽度112Balance 天平左臂重量162543天平左臂长度1天平右臂重量1天平右臂长度113Ionosphere 每个特征单独作为一个视角135134214Wine 每个特征单独作为一个视角1178133表 6 5种聚类方法的NMI值比较结果Table 6 Comparison of NMI performance of five clustering methods编号Co-Clustering CombKM FCM Co-FKM LR-MVEWFCM 均值P-value均值P-value均值P-value均值P-value均值80.5771 ±0.00230.00190.5259 ±0.05510.20560.5567 ±0.01840.00440.5881 ±0.01093.76×10−40.5828 ±0.004490.7582 ±7.4015 ×10−172.03×10−240.7251 ±0.06982.32×10−70.7578 ±0.06981.93×10−240.8317 ±0.00648.88×10−160.9029 ±0.0057100.2455 ±0.05590.01650.1562 ±0.07493.47×10−50.1813 ±0.11720.00610.2756 ±0.03090.10370.3030 ±0.0402110.7582 ±1.1703×10−162.28×10−160.7468 ±0.00795.12×10−160.7578 ±1.1703×10−165.04×10−160.8244 ±1.1102×10−162.16×10−160.8768 ±0.0097120.2603 ±0.06850.38250.1543 ±0.07634.61×10−40.2264 ±0.11270.15730.2283 ±0.02940.01460.2863 ±0.0611130.1385 ±0.00852.51×10−90.1349 ±2.9257×10−172.35×10−130.1299 ±0.09842.60×10−100.2097 ±0.03290.04830.2608 ±0.0251140.4288 ±1.1703×10−161.26×10−080.4215 ±0.00957.97×10−090.4334 ±5.8514×10−172.39×10−080.5295 ±0.03010.43760.5413 ±0.03647 期张嘉旭等: 基于低秩约束的熵加权多视角模糊聚类算法1767表 7 5种聚类方法的RI 值比较结果Table 7 Comparison of RI performance of five clustering methods编号Co-ClusteringCombKM FCMCo-FKM LR-MVEWFCM均值P-value 均值P-value 均值P-value 均值P-value 均值80.8392 ±0.0010 1.3475 ×10−140.8112 ±0.0369 1.95×10−70.8390 ±0.01150.00320.8571 ±0.00190.00480.8508 ±0.001390.8797 ±0.0014 1.72×10−260.8481 ±0.0667 2.56×10−50.8859 ±1.1703×10−16 6.49×10−260.9358 ±0.0037 3.29×10−140.9665 ±0.0026100.6515 ±0.0231 3.13×10−40.6059 ±0.0340 1.37×10−60.6186 ±0.06240.00160.6772 ±0.02270.07610.6958 ±0.0215110.8797 ±0.0014 1.25×10−180.8755 ±0.0029 5.99×10−120.8859 ±0.0243 2.33×10−180.9267 ±2.3406×10−16 5.19×10−180.9527 ±0.0041120.6511 ±0.02790.01560.6024 ±0.0322 2.24×10−50.6509 ±0.06520.11390.6511 ±0.01890.0080.6902 ±0.0370130.5877 ±0.0030 1.35×10−120.5888 ±0.0292 2.10×10−140.5818 ±1.1703×10−164.6351 ×10−130.6508 ±0.01470.03580.6855 ±0.0115140.7187 ±1.1703×10−163.82×10−60.7056 ±0.01681.69×10−60.7099 ±1.1703×10−168.45×10−70.7850 ±0.01620.59050.7917 ±0.0353R I数据集N M I数据集(a) RI 指标(a) RI(b) NMI 指标(b) NMI图 4 低秩约束对算法性能的影响(横坐标为数据集编号, 纵坐标为聚类性能指标)Fig. 4 The influence of low rank constraints on the performance of the algorithm (the X -coordinate isthe data set number and the Y -coordinate is the clustering performance index)目标函数值1 096.91 096.81 096.61 096.71 096.51 096.41 096.31 096.21 096.1目标函数值66.266.065.665.865.465.2迭代次数05101520目标函数值7.05.06.55.54.04.53.03.5迭代次数05101520迭代次数05101520目标函数值52.652.251.451.851.050.6迭代次数05101520×106(a) 数据集 7(a) Dataset 7(b) 数据集 8(b) Dataset 8(c) 数据集 9(c) Dataset 9(d) 数据集 10(d) Dataset 101768自 动 化 学 报48 卷ReferencesXu C, Tao D C, Xu C. Multi-view learning with incompleteviews. IEEE Transactions on Image Processing , 2015, 24(12):5812−58251Brefeld U. Multi-view learning with dependent views. In: Pro-ceedings of the 30th Annual ACM Symposium on Applied Com-puting, Salamanca, Spain: ACM, 2015. 865−8702Muslea I, Minton S, Knoblock C A. Active learning with mul-tiple views. Journal of Artificial Intelligence Research , 2006,27(1): 203−2333Zhang C Q, Adeli E, Wu Z W, Li G, Lin W L, Shen D G. In-fant brain development prediction with latent partial multi-view representation learning. IEEE Transactions on Medical Imaging ,2018, 38(4): 909−9184Bickel S, Scheffer T. Multi-view clustering. In: Proceedings of the 4th IEEE International Conference on Data Mining (ICDM '04), Brighton, UK: IEEE, 2004. 19−265Wang Y T, Chen L H. Multi-view fuzzy clustering with minim-ax optimization for effective clustering of data from multiple sources. Expert Systems with Applications , 2017, 72: 457−4666Wang Jun, Wang Shi-Tong, Deng Zhao-Hong. Survey on chal-lenges in clustering analysis research. Control and Decision ,2012, 27(3): 321−328(王骏, 王士同, 邓赵红. 聚类分析研究中的若干问题. 控制与决策,2012, 27(3): 321−328)7Pedrycz W. Collaborative fuzzy clustering. Pattern Recognition Letters , 2002, 23(14): 1675−16868Cleuziou G, Exbrayat M, Martin L, Sublemontier J H. CoFKM:A centralized method for multiple-view clustering. In: Proceed-ings of the 9th IEEE International Conference on Data Mining,Miami, FL, USA: IEEE, 2009. 752−7579Jiang Y Z, Chung F L, Wang S T, Deng Z H, Wang J, Qian P J. Collaborative fuzzy clustering from multiple weighted views.IEEE Transactions on Cybernetics , 2015, 45(4): 688−70110Bettoumi S, Jlassi C, Arous N. Collaborative multi-view K-means clustering. Soft Computing , 2019, 23(3): 937−94511Zhang G Y, Wang C D, Huang D, Zheng W S, Zhou Y R. TW-Co-K-means: Two-level weighted collaborative K-means for multi-view clustering. Knowledge-Based Systems , 2018, 150:127−13812Cao X C, Zhang C Q, Fu H Z, Liu S, Zhang H. Diversity-in-duced multi-view subspace clustering. In: Proceedings of the2015 IEEE Conference on Computer Vision and Pattern Recog-nition, Boston, MA, USA: IEEE, 2015. 586−59413Zhang C Q, Fu H Z, Liu S, Liu G C, Cao X C. Low-rank tensor constrained multiview subspace clustering. In: Proceedings of the 2015 IEEE International Conference on Computer Visio,Santiago, Chile: IEEE, 2015. 1582−159014Boyd S, Parikh N, Chu E, Peleato B, Eckstein J. Distributed optimization and statistical learning via the alternating direc-tion method of multipliers. Foundations and Trends in Machine Learning , 2011, 3(1): 1−12215Liu G C, Lin Z C, Yan S C, Sun J, Yu Y, Ma Y. Robust recov-ery of subspace structures by low-rank representation. IEEE1616.216.015.815.615.415.215.0目标函数值目标函数值目标函数值51015迭代次数迭代次数迭代次数 711.2011.1511.1011.0511.0010.9510.90800700600500400300200目标函数值38.638.238.438.037.837.637.437.251015205101520迭代次数 705101520(e) 数据集 11(e) Dataset 11(f) 数据集 12(f) Dataset 12(g) 数据集 13(g) Dataset 13(h) 数据集 14(h) Dataset 14图 5 LR-MVEWFCM 算法的收敛曲线Fig. 5 Convergence curve of LR-MVEWFCM algorithm图 6 模拟数据集7上参数敏感性分析Fig. 6 Sensitivity analysis of parameters on simulated dataset 77 期张嘉旭等: 基于低秩约束的熵加权多视角模糊聚类算法1769。

一种锂电池SOH估计的KNN-马尔科夫修正策略

一种锂电池SOH估计的KNN-马尔科夫修正策略

第47卷第2期自动化学报Vol.47,No.2 2021年2月ACTA AUTOMATICA SINICA February,2021一种锂电池SOH估计的KNN–马尔科夫修正策略赵光财1,2林名强1戴厚德1武骥3汪玉洁4摘要锂离子电池的健康状态(State of health,SOH)是决定电池使用寿命的关键因素.由于锂电池生产工艺、工作环境和使用习惯等的差异性导致其衰退特性具有较大差异,因此锂电池SOH难以精确估算.本文采用数据驱动的方式通过对采集的电压数据进行特征提取,使用贝叶斯正则化神经网络对锂电池SOH进行预测,同时引入KNN–马尔科夫修正策略对预测结果进行修正.实验结果证明,贝叶斯正则化算法对锂电池SOH的预测准确度较高,KNN–马尔科夫修正策略提高了预测的精确度和鲁棒性,组合预测模型对锂电池SOH的平均预测误差小于1%,与采用数据分组处理方法(Group method of data handling,GMDH)、概率神经网络(Probabilistic neural network,PNN)、循环神经网络(Recurrent neural network,RNN)的预测精度进行对比,该模型的预测精度分别提高了33.3%、48.7%和53.1%.关键词锂电池SOH,特征提取,多层前馈神经网络,贝叶斯正则化,马尔科夫链引用格式赵光财,林名强,戴厚德,武骥,汪玉洁.一种锂电池SOH估计的KNN–马尔科夫修正策略.自动化学报,2021, 47(2):453−463DOI10.16383/j.aas.c180124开放科学(资源服务)标识码(OSID):A Modified Strategy Using the KNN-Markov Chain forSOH Estimation of Lithium BatteriesZHAO Guang-Cai1,2LIN Ming-Qiang1DAI Hou-De1WU Ji3WANG Yu-Jie4Abstract The state of health(SOH)of lithium batteries is a critical factor in determining the battery s end-of-service-life.The differences of the Lithium-ion battery s production process,work environment,and use habit etc.lead to the massive differences of the battery s fade characteristics,which,in turn,inaccurate estimation of their battery s SOH. In this paper,the data-driven method was employed for experimental feature extraction.Besides,this paper presents an SOH estimation method based on the Bayesian-regularization neural network and the KNN-Markov chain used for amending the prediction results.Experimental results show that the Bayesian-regularization neural network applied to the SOH estimation could obtain superior accuracy performance,and by combining the KNN-Markov chain,the prediction accuracy(the average prediction error of SOH less than1%)could be improved.On the whole,the combined model shows good pared with the group method of data handling(GMDH),probabilistic neural network(PNN) and recurrent neural network(RNN),the prediction accuracy of the model was improved by33.3%,48.7%and53.1% respectively.Key words Lithium battery SOH,feature extraction,multilayer feedforward neural network,Bayesian regularization, Markov chainCitation Zhao Guang-Cai,Lin Ming-Qiang,Dai Hou-De,Wu Ji,Wang Yu-Jie.A modified strategy using the KNN-Markov chain for SOH estimation of lithium batteries.Acta Automatica Sinica,2021,47(2):453−463收稿日期2018-03-05录用日期2018-08-01Manuscript received March5,2018;accepted August1,2018国家自然科学基金(61501428),福建省科技攻关项目(引导性项目) (2018H0043),中国科学院科研装备研制项目(YZ201510)资助Supported by National Natural Science Foundation of China (61501428),Project of Science and Technology Department of Fujian Province(Pilot Project)(2018H0043),Research Equip-ment Development Project of Chinese Academy of Science (YZ201510)本文责任编委曹向辉Recommended by Associate Editor CAO Xiang-Hui1.中国科学院海西研究院泉州装备制造研究所晋江3622002.中国科学院大学北京1000493.合肥工业大学汽车与交通工程学院合肥2300094.中国科学技术大学信息科学技术学院合肥230026 1.Quanzhou Institute of Equipment Manufacturing,Haixi Institutes,Chinese Academy of Sciences,Jinjiang362200电池管理系统(Battery management system, BMS)是电动汽车的重要组成部分,在电动汽车电池组的状态监测、保护和电量均衡等方面起到举足轻重的作用[1−2].电池健康状态(State of health, SOH)作为电池系统的关键参数之一,其变化直接影响电池组的使用性能、可靠性及安全性.由于锂离子电池具有能量密度高、稳定性好、使用寿命长2.University of Chinese Academy of Sciences,Beijing1000493.School of Automotive and Traffic Engineering,Hefei Univer-sity of Technology,Hefei2300094.School of Information and Technology,University of Science and Technology of China, Hefei230026454自动化学报47卷等优点,已经被许多可再生能源系统用作能量存储设备.但其复杂的内部电化学结构、不确定的外部工作环境等因素使得电池的衰退过程极为复杂,即使同一规格型号的电池其衰退过程也不尽相同,这导致锂电池SOH的准确估计变得十分困难[3].因此,快速、精确地实现锂电池SOH估计一直是一个重要的研究课题.电池的SOH表示当前电池的最大可用容量占额定容量的百分比,用来度量电池的退化程度.锂电池SOH随着使用会慢慢减小,当SOH降低至额定容量的70%时,即认为锂电池寿命终止.SOH的定义如下:SOH=C currentC initial×100%(1)上式中:C current表示锂电池当前最大可用容量;C initial表示锂电池的额定容量,通常指出厂时的标称容量[4].目前已有许多专家学者采用多种不同方法对锂电池的SOH进行估计.Galeotti等[5]采用电化学阻抗谱研究锂聚合物(LiPO)电池的老化,通过拟合阻抗谱提取等效电路模型的参数再现电池放电曲线并引入电池欧姆电阻与可用容量的关系,借助证据理论评估电池的SOH.Chen等[6]基于扩散电容与SOH的相关性,采用遗传算法以实时采集的电流和电压来估计包括扩散电容在内的电池模型参数进而估计锂电池SOH.Mejdoubi等[7]提出混合估计电池荷电状态(State of charge,SOC)和SOH的思路,使用自适应观察器估计SOC的同时使用扩展卡尔曼滤波器(Extended Kalmanfilter,EKF)估计SOH,采用闭环估计策略同时结合李雅普诺夫原理保证稳定性.Liu等[8]将基于数据驱动的高斯过程回归(Gaussian process regression,GPR)方法应用于SOH估计,采用改进的组合高斯过程泛函回归(Gaussian process functional regression, GPFR)模型实现了较好的SOH预测.Moura等[9]提出了一种的基于电化学模型的自适应偏微分方程观测器来估计SOH的方法,其通过测量电压和电流来估计该模型的参数.Ng等[10]认为安时积分法对于具有高充电和高放电效率的锂电池的SOC估计是有利的.他们提出了一种基于库伦计数的方法来提高估计精度,并利用SOC来估计SOH.Lievre 等[11]利用电池电阻来量化SOH退化情况,通过电池组响应的电压和电流数据利用电池模型获得电池电阻,进而估计电池的SOH.EKF被Plett[12]引入到电池组的SOH预测中,通过电池组建模和系统参数识别,继而用EKF估计SOC,最终估算出SOH.值得一提的是,卡尔曼滤波及其衍生算法在电池SOC、SOH估计中应用十分广泛.上述方法在预测SOH方面取得了很大的成功,但在实际应用中也存在部分不足之处.基于电化学阻抗谱的SOH预测需要复杂设备测量电池阻抗谱,内阻法、化学模型法、安时积分法在预测精度上略显不足,卡尔曼滤波法在当跟踪目标长时间被遮挡时会导致目标跟踪丢失.目前很多基于机器学习的方法被用于锂电池SOH估计.Wu等[13−14]将重要性采样、数据分组处理方法(Group method of data handling,GMDH)等应用于充放电数据采样和锂电池SOH估计,取得了良好的预测效果. Klass等[15]利用电池充放电过程中的电池电流、电压和温度等参数建立基于支持向量机的电池模型,并用其对电池剩余使用寿命和瞬时电阻的估计进行预测.循环神经网络(Recurrent neural network, RNN)被Eddahech等[16]引入到锂电池SOH估计中,该方法基于等效电路方法的模型,使用RNN来预测电池性能的退化.Lin等[17]使用概率神经网络(Probabilistic neural network,PNN)估计锂电池的SOH,将放电时瞬时电压降以及开路电压作为SOH估计最重要的参数,该方法能实现较高精度的SOH预测.以上方法均可实现较好的预测效果,但SOH估计本质是回归问题,PNN更适合用在模式分类,用于SOH估计需要大量训练样本.RNN能够学习SOH随循环次数衰退的长期依赖,但在短期预测精度上可进一步提高.本文将马尔科夫链引入锂电池SOH估计中并改进为KNN–马尔科夫修正策略以提升对锂电池SOH的预测效果,该修正策略可有效减小随机扰动误差,增强模型的短期预测能力.此外,采用的贝叶斯正则化神经网络(Bayesian regularized neural network,BRNN)具有复杂度较低,同时具有良好的泛化性能.为结合两者优势,本文建立了贝叶斯正则化神经网络结合KNN–马尔科夫修正策略的组合模型实现锂电池SOH的估计,有效提高了SOH估计的准确性与鲁棒性.本文结构如下,第1节介绍预测方法原理包括贝叶斯正则化网络和马尔科夫链.第2节介绍所做的工作包括预测模型的建立、特征提取以及KNN–马尔科夫修正策略.第3节介绍实验结果,验证KNN–马尔科夫策略的有效性及所提出预测模型的准确性.第4节是本文得出的结论.1预测方法原理1.1贝叶斯正则化神经网络该神经网络为基于误差反向传播算法的多层前馈神经网络,其网络结构如图1[18]所示,包括输入层、隐藏层和输出层三层[19].设神经网络的训练样2期赵光财等:一种锂电池SOH 估计的KNN–马尔科夫修正策略455本D (x i ,t i ),i ={1,2,···,n },n 为训练样本数,W 为网络参数向量,M 为网络模型,f 为Sigmoid 激活函数,表达式为f (x )=11+e −x ,t 为网络训练目标值,神经网络的性能函数定义为均方差形式:J W=12ni =1(f (x i ,W,M )−t i )2(2)采用梯度下降法调整网络权重ωhj =ωhj −∆ωhj ,其中∆ωhj =η∂J W∂ωhj ,ωhj 为第h 层的第i 个权值,η为网络学习率.图1多层前馈神经网络结构示意图Fig.1Structure of multilayer feedforward neural network为防止神经网络训练过程中出现网络过拟合现象,采用正则化算法优化网络结构,以提高其泛化能力[20].正则化方法对网络性能函数加入正则项E W :E W=12 W 22=12mi =1ω2i(3)上式中ω为网络权值,m 为网络参数总数,目标函数于是变为F (W )=βJ W +αE W [21],即:min W,α,β(F (W ))=12βn i =1f (x i ,W,M )−t i 2+12α W 22(4)α和β为正则化系数,β侧重于减小训练误差,α侧重于减小权值规模,贝叶斯正则化算法在两者之间寻求平衡,贝叶斯正则化神经网络训练的训练过程即寻找最优化的参数使得该目标函数最小化.假定数据集和初始网络权值均服从高斯分布,由贝叶斯准则求解最大后验概率,即极小化网络训练目标函数,得到使目标函数最小的W MP 处的α和β[22]:αMP =γ2E W (W MP ),βMP =n −γ2J W (W MP )(5)其中神经网络的有效参数个数为γ=m −2αMP (tr(H −1)),m 是神经网络的参数数量,H 是目标函数在W MP 处的Hessian 矩阵α∇2E W +β∇2J W .首先确定神经网络结构,由于采用正则化算法,隐藏层神经元数量可稍大于最优的隐藏层神经元数量,同时初始化超参数α和β,利用式(5)不断估计参数α和β的新值;计算H 矩阵及有效参数个数γ,γ为网络中起减少误差作用参数数量,如果γ接近于m 时需适当增加隐藏层神经元个数.在总误差迭代过程中,当总误差没有较大改变时,网络训练收敛.1.2马尔科夫链马尔科夫过程描述了一个具有若干个状态的系统,其每个状态都可以根据固定转移概率传递给另一个状态.马尔科夫链可以看作是某些概率的状态转换过程,一个随机过程过渡到未来过程的概率只取决于当前的状态[23−24].一步转移概率矩阵和m 步转移概率矩阵P 1和P m 分别为:P 1= P (1)11...P (1)1n.........P (1)n 1...P (1)nn ,P m = P (m )11...P (m )1n .........P (m )n 1...P (m )nn(6)P (n )ij =m (n )ij /M i 为状态E i 经过n 步后转移到状态E j 的转移概率,其中m (n )ij 为训练数据中由状态E i 一步转移到E j 的次数,M i 为以E i 作为起始状态的状态个数.m 步转移的矩阵概率为一步转移矩阵经过m (m >1)步从而实现相应状态的概率矩阵.一步转移概率矩阵和m 步转移概率矩阵的关系如下:P m =[P 1]m (7)2预测模型的建立2.1预测模型的建立流程基于锂电池衰减过程复杂的非线性特征,本文采用了数据驱动的思路,对传感器获取的电池充电过程中的电压数据进行特征提取,提取具有鲁棒性及多样性的8个特征用于SOH 估计,这些特征能准确估计电池SOH.为保证SOH 估计具有较高的准确度,采用贝叶斯正则化神经网络进行SOH 估计,该网络能够防止训练过程的过拟合,具有良好的泛化能力,提高SOH 估计准确度.同时从训练数据中456自动化学报47卷学习SOH 估计误差的统计特性并用于预测SOH 估计误差,采用KNN 算法对误差值进行预处理,划分状态空间并计算状态转移矩阵,使用马尔科夫链得到预测SOH 的经验估计误差,并对神经网络SOH 预测值进行误差修正得到最终SOH 估计值.预测模型如图2所示.2.2特征提取本文采用的实验数据来自NASA 公开数据集[25]中的一组锂电池(#5、#6、#7),该组电池在室温25℃下循环进行充电、放电和阻抗测试.首先在恒流(Constant current,CC)模式下以1.5A 电流充电,直到电池电压达到4.2V.然后继续保持恒压充电(Constant voltage,CV)模式,直到电荷电流降至20mA,充电过程中电压、电流及SOC 变化曲线如图3(a)所示.本文使用电压数据并结合SOC 增量进行特征提取,基于不同SOH 电池充电电压随时间变化具有的规律性(图3(b))提取特征用作SOH 估计.首先对传感器信息进行数据预处理,根据循环次数对数据进行分组并剔除无效值.然后利用统计方法提取该组电池恒流充电过程的电压信号特征,主要包括电压均值和电压增量累积以及其d SOC /d V 特性三个方面.电压的均值反映了电压直流分量的大小,用来刻画电压幅值的大小.设电池管理系统采集到的离散电压信号表示为,则电压均值为[26]:V ave=1n ni =1|V (i )|(8)其次,电压增量累积反映SOH 变化时电压的波动大小.设M 为电压信号的增量累积,则电压增量累积公式为[26]:M =n −1 i =1|∆V (i )|=n −1 i =1|V (i +1)−V (i )|(9)再次,从电池衰退的物理容量降低的客观实际出发,即相同电压的所容纳的电量在电池衰退过程中变化明显,故d SOC/d V 可作为电池退化特征.图3(c)为不同充电循环次数下d SOC/d V 曲线变化曲线[27],在SOH 差异最明显的Peak1、Peak2处各取两个d SOC/d V 值作为神经网络输入的电池退化特征.为综合考虑特征多样性及预测的准确性,采用如图3(d)所示组合特征作为BRNN 的训练输入.2.3KNN–马尔科夫修正策略神经网络对锂电池的SOH 进行预测得到的非线性模型存在局部偏差过大,部分区域拟合效果不佳的缺点,因此本实验采用了KNN–马尔科夫方法对神经网络模型预测的SOH 进行必要的修正,从而整体上减小模型的预测误差,提高模型的短期预测精度.2.3.1KNN 误差处理假定估计误差分布为一维高斯分布,由于随机性或者数据误差产生远离其他样本点的野点,不能准确反映误差区间及分布的真实情况,如果直接从训练数据中划分马尔科夫状态将会对误差区间估计造成偏差.对于这种情况,本文采用一个基于K 近邻思路来对数据进行预处理.该方法从误差数据中心取一点开始计算,计算该点与k个近邻点的距离图2预测模型流程图Fig.2Flowchart of the proposed prediction model2期赵光财等:一种锂电池SOH估计的KNN–马尔科夫修正策略457图3特征提取Fig.3Feature extraction 的之和以及其相邻点的k个近邻的距离之和,KNN的距离计算公式为[28]:distance(X,Y)=ki=1(x i−y i)2(10)其中,X、Y为两个点,如果相邻两误差点KNN距离之和的比值大于某一阈值则相应调整这两个误差点中远离数据中心点的数据使得该点像数据中心靠拢使其满足条件,其公式为:λi=km=1distance i+1mki=1distance im(11)其中λi为第i次KNN计算的比值,km=1distance i+1m表示第i个点的k个近邻点的距离之和,ki=1distance im表示其相邻的第i+1个点的k个近邻点的距离之和.该方法不会改变误差序列排列顺序,同时选择合适的阈值不会改变误差内部数据,仅使数据两端极大或极小的偶然值不再过于偏离数据中心,使得误差的状态空间划分更能反映实际情况,从而降低由于系统不确定性或随机误差造成的干扰.2.3.2马尔科夫修正基于KNN对误差数据处理,采用马尔科夫链对神经网络预测结果进行误差修正.根据训练数据误差分布划分马尔科夫链预测的状态空间[29],本文将状态转换划分为三个状态区间,分别为S1[0, x−0.5s],S2[x−0.5s,x+0.5s],S3[x+0.5s,1],其中s为训练数据的标准差,在划分马尔科夫链状态空间前需用式(12)将数据归一化到[0,1]之间.˜x=x−x minx max−x min(12)对神经网络的误差序列进行预测修正.设残差所处状态为α(t),如当a<α(t)<b,则处于该区间中的马尔科夫预测的修正误差为c=(a+b)/2,各状态区间的修正误差构成状态误差矩阵Q,其中n 为状态区间个数,结合第1.2节马尔科夫链内容得458自动化学报47卷修正公式为:F=F g−P0·P m·Q(13)其中,F为修正后的预测值;F g值为贝叶斯正则化神经网络的预测值;P0为1×3维初始状态矩阵;P m为3×3维m步状态转移矩阵;Q指3×1维状态误差矩阵.3实验结果及分析3.1验证KNN–马尔科夫修正策略我们通过使用KNN–马尔科夫修正前后预测值对实测值的拟合程度对比对该修正策略进行实验验证,以检验该修正策略的有效性.实验中随机选取164次充放电电压电流特征数据的85%作为训练数据,15%作为测试数据,SOH真实值作为网络训练目标.采用平均绝对误差(Mean absolute error,MAE)和均方误差(Meansquare error,MSE)来评估算法SOH估计准确度,MAE反映总体预测误差的大小,MSE用来反映预测值对于真实SOH的偏离程度,MAE和MSE分别定义为:MAE=1nni=1|(x i−µ)|,MSE=ni=1(x i−µ)2n−1(14)划分状态空间前使用KNN算法对偏离误差分布的孤立点进行调整,选取阈值λ=3,从数据中心点开始,向数据两端分别进行调整,调整的结果是该假定的一维高斯分布的内部误差数值发生不发生变化或微小变化,仅分布边缘处的由于误差或偶然性出现的过大或过小数据向数据中心调整,通过调整可以使学习到的误差状态空间更接近真实情况.由上文状态划分公式计算得一个马尔科夫状态集,由数据计算得到S1[0,0.32375],S2 [0.32375,0.46865],S3[0.46865,1],由公式x= x min+˜x·(x max−x min)反归一化得到相对误差序列的三个状态分别为S1[−3.236,−1.110], S2[−1.110,−0.114],S3[−0.114,−3.539].并按此划分规则将各个预测误差划分为相应的马尔科夫状态,如表1所示.根据状态空间划分,计算得到状态转移矩阵,并对BRNN预测值进行修正.状态转移矩阵见式(15),BRNN预测值及马尔科夫修正后误差见表2.表1BRNN预测值相对误差及状态划分Table1BRNN prediction error and state division序号实测值预测值相对误差(%)归一化相对误差(%)状态10.64450.6431−0.21720.4569220.74800.7253−3.03470.1425130.96650.9646−0.19650.4496240.95020.95080.06310.4865350.58020.5695−1.84420.3194160.75560.7344−2.80570.1647170.92940.93010.07530.4879380.92220.9440 2.36390.7994390.68080.6927 1.74790.65331 100.86150.8503−1.30010.31231 110.75560.76190.83380.57063 120.64450.6431−0.21720.45692表2BRNN预测误差及马尔科夫修正误差Table2BRNN prediction and Markov correction error序号实测值预测值修正前误差(%)修正后误差(%)1 1.00000.9809 1.9090 1.444620.99420.9753 1.8847 1.420330.99410.9741 2.0029 1.538540.98780.9720 1.5855 1.121150.96650.96460.1857−0.278760.96070.9623−0.1558−0.620270.95540.9333 2.2035 1.739180.93870.9590−2.0253−0.369990.93290.9486−1.57360.0818100.92220.9440−2.1778−0.5224P1=0.710.260.030.080.710.220.240.410.35P2=0.530.380.090.160.610.030.290.500.21P3=0.430.450.130.220.570.220.290.520.19P4=0.370.480.150.250.590.200.290.520.18(15)由图4的(a)、(c)、(e)可以看出,5#、6#、7#电池马尔科夫修正后的更加接近实测值,局部放大2期赵光财等:一种锂电池SOH 估计的KNN–马尔科夫修正策略459图4马尔科夫修正过程修正前后结果对比Fig.4Comparative results of BRNN and Markov correction图显示组合模型预测准确度好于修正前,验证了马尔科夫修正的有效性.图4的(b)、(d)、(f)为5#、6#、7#电池修正前后的误差对比,从中可以看出马尔科夫修正后绝大部分误差明显减小,修正误差效果明显.如表3所示,马尔科夫修正方法在5、6、7号的电池上验证均提高了BRNN 预测精度,其中5、6、7号电池的MAE 分别降低了25.5%、17.3%、15.9%,表3KNN–马尔科夫修正结果Table 3KNN-Markov correction results编号5#电池6#电池7#电池有无修正无有无有无有MAE (%)0.470.350.520.430.440.37MSE (%2)0.490.370.520.400.480.37MSE 分别降低了26.5%、22.5%、21.8%.实验结460自动化学报47卷果表明,本文提出的KNN–马尔科夫修正方法能够获得更加精确和鲁棒的估计结果.3.2验证组合模型先进性为检验KNN–马尔科夫修正策略与BRNN 的组合预测模型(MC-BRNN)的准确度,与集中较为先进的预测方法进行对比,包括GMDH、PNN、RNN,采用的算法参数如下: BRNN:隐藏层神经元数量为25个,学习率为0.1.GMDH:最大隐藏层数为4,单层最大神经元个数为20.PNN:包括输入层、模式层、竞争层、输出层,平滑系数为0.1.RNN:单层LSTM-RNN结构,隐藏层神经元数量为70.由表4可知,与GMDH、PNN、RNN相比,MC-BRNN能够使得预测结果的MSE和MAE平均分别降低约45.8%、70.3%、79.3%和33.3%、48.7%、53.1%.通过该实验结果对比可以得出,组合模型用到的BRNN能够避免数据的过拟合现象,精度较高.对各算法的训练时间在CPU为i5-3210M (2.5GHz),4GB RAM的计算机上进行统计,得到各算法运行时间如表5所示.从表5中可以看出,PNN训练时间最短, GMDH和MC-BRNN训练时间差别不大,RNN 的训练时间较长.由于SOH变化缓慢,对算法的实时性要求不高,同时结合表3和表4可以看出, BRNN预测SOH的准确度较高,并可以通过设置较大迭代次数以增加训练时间为代价来实现最佳的预测,提高估计的准确度.因此采用MC-BRNN在估计SOH方面具有更好的效果.图5的(a)、(c)、(e)的局部放大图显示MC-BRNN与GMDH、PNN、RNN相比在SOH估计精度方面具有较为明显的优势,对真实值的拟合度更高.图5的(b)、(d)、(f)曲线所示,5#、7#电池的MC-BRNN预测的最大预测误差为3%左右, 6#电池的最大预测误差小于2.5%,误差大部分集中分布在±1%的范围内.综合以上可知,MC-BRNN组合预测模型相对于GMDH、PNN、RNN 对锂电池SOH的预测性能更好.表4各算法准确度对比Table4Comparison of accuracy of each algorithm编号5#电池6#电池7#电池算法MC-BRNNGMDH PNN RNNMC-BRNNGMDH PNN RNNMC-BRNNGMDH PNN RNNMAE(%)0.350.520.640.670.430.610.810.930.370.590.790.83 MSE(%2)0.370.670.94 1.210.400.88 1.30 2.350.370.58 1.57 1.95表5各算法的时间复杂度对比(s)Table5Comparison of time complexity of each algorithm(s)算法MC-BRNN GMDH PNN RNN5号电池 1.5101 1.11840.2994 4.5942 6号电池 1.4361 1.19480.2442 3.7446 7号电池 1.5559 1.11030.2748 4.8092平均值 1.5007 1.13990.27264.38262期赵光财等:一种锂电池SOH 估计的KNN–马尔科夫修正策略461图5MC-BRNN 、GMDH 、PNN 、RNN 预测结果对比Fig.5Comparative results of MC-BRNN,GMDH,PNN and RNN4结论本文建立了基于贝叶斯正则化神经网络与KNN–马尔科夫链的组合预测模型,该模型结合了贝叶斯正则化算法在能够有效避免过拟合的优点与KNN–马尔科夫修正方法在系统短时预测方面的优势,对锂电池SOH 具有较好的预测效果.同时,该组合预测模型对锂电池SOH 的平均预测误差小于1%,能准确估计SOH.KNN–马尔科夫链修正方法可显著提高神经网络的估计精度,该修正方法基于SOH 退化的客观规律性,可提高神经网络15%∼30%的估计精度,该方法也为其他SOH 估计方法提供了一种误差修正思路.最后,本文将组合模型与GMDH 、PNN 、RNN 进行了对比实验,经实验验证,该组合模型的预测精度相比以上两种方法分别提高了33.3%、57.2%和53.1%,同时具有良好的鲁棒性.References1Cheng K W E,Divakar B P,Wu H J,Ding K,Ho H F.Battery-management system (BMS)and SOC development for electrical vehicles.IEEE Transactions on Vehicular Tech-nology ,2011,60(1):76−882Chen Hong,Gong Xun,Hu Yun-Feng,Liu Qi-Fang,GaoBing-Zhao,Guo Hong-Yan.Automotive control:The state of the art and perspective.Acta Automatica Sinica ,2013,39(4):322−346(陈虹,宫洵,胡云峰,刘奇芳,高炳钊,郭洪艳.汽车控制的研究现状与展望.自动化学报,2013,39(4):322−346)3Scrosati B,Garche J.Lithium batteries:Status,prospects and future.Journal of Power Sources ,2010,195(9):2419−24304Andre D,Appel C,Soczka-Guth T,Sauer D U.Advanced mathematical methods of SOC and SOH estimation for lithium-ion batteries.Journal of Power Sources ,2013,224:20−275Galeotti M,Cin´a L,Giammanco C,Cordiner S,Carlo D A.Performance analysis and SOH (state of health)evalu-ation of lithium polymer batteries through electrochemical impedance spectroscopy.Energy ,2015,89:678−6866Chen Z,Mi C C,Fu Y H,Xu J,Gong X Z.Online battery state of health estimation based on genetic algorithm for electric and hybrid vehicle applications.Journal of Power Sources ,2013,240:184−1927Mejdoubi A E,Oukaour A,Chaoui H,Gualous H,Sabor J,Slamani Y.State-of-charge and state-of-health lithium-ion batteries diagnosis according to surface temperature vari-ation.IEEE Transactions on Industrial Electronics ,2016,63(4):2391−2402462自动化学报47卷8Liu D T,Pang J Y,Zhou J B,Peng Y,Pecht M.Prognostics for state of health estimation of lithium-ion batteries based on combination Gaussian process functional regression.Mi-croelectronics Reliability,2013,53(6):832−8399Moura S J,Chaturvedi N A,Krsti´c M.Adaptive partial dif-ferential equation observer for battery state-of-charge/state-of-health estimation via an electrochemical model.Jour-nal of Dynamic Systems,Measurement,and Control,2014, 136(1):01101510Ng K S,Moo C S,Chen Y P,Hsieh Y C.Enhanced coulomb counting method for estimating state-of-charge and state-of-health of lithium-ion batteries.Applied Energy,2009,86(9): 1506−151111Lievre A,Sari A,Venet P,Hijazi A,Ouattara-Brigaudet M,Pelissier S.Practical online estimation of lithium-ion battery apparent series resistance for mild hybrid vehicles.IEEE Transactions on Vehicular Technology,2016,65(6): 4505−451112Plett G L.Extended Kalmanfiltering for battery man-agement systems of LiPB-based HEV battery packs:Part1.Background.Journal of Power Sources,2004,134(2):252−26113Wu J,Wang Y J,Zhang X,Chen Z H.A novel state of health estimation method of Li-ion battery using group method of data handling.Journal of Power Sources,2016,327: 457−46414Wu J,Zhang C B,Chen Z H.An online method for lithium-ion battery remaining useful life estimation using impor-tance sampling and neural networks.Applied Energy,2016, 173:134−14015Klass V,Behm M,Lindbergh G.A support vector machine-based state-of-health estimation method for lithium-ion bat-teries under electric vehicle operation.Journal of Power Sources,2014,270:262−27216Eddahech A,Briat O,Bertrand N,Del´e tage J Y,Vinassa J M.Behavior and state-of-health monitoring of Li-ion bat-teries using impedance spectroscopy and recurrent neural networks.International Journal of Electrical Power&En-ergy Systems,2012,42(1):487−49417Lin H T,Liang T J,Chen S M.Estimation of battery state of health using probabilistic neural network.IEEE Transac-tions on Industrial Informatics,2013,9(2):679−68518Basheer I A,Hajmeer M.Artificial neural networks:fun-damentals,computing,design,and application.Journal of Microbiological Methods,2000,43(1):3−3119Lam C H,Shin F G.Formation and dynamics of modules in a dual-tasking multilayer feed-forward neural network.Physical Review E,1998,58(3):3673−367720Sun Z,Chen Y,Li X Y,Qin X L,Wang H Y.A Bayesian regularized artificial neural network for adaptive optics fore-casting.Optics Communications,2017,382:519−52721Yang Hai-Shen,Fu Hong-Zhuo.Stock index forecast based on Bayesian regularization BP neural network.Science Tech-nology and Engineering,2009,9(12):3306−3310,3318(杨海深,傅红卓.基于贝叶斯正则化BP神经网络的股票指数预测.科学技术与工程,2009,9(12):3306−3310,3318)22Li Hong-Xia,Xu Shi-Guo,Fan Chui-Ren.Long-term predic-tion of runoffbased on Bayesian regulation neural network.Journal of Dalian University of Technology,2006,46(S1): 174−177(李红霞,许士国,范垂仁.基于贝叶斯正则化神经网络的径流长期预报.大连理工大学学报,2006,46(S1):174−177)23Neal R M.Markov chain sampling methods for dirichlet pro-cess mixture models.Journal of Computational and Graph-ical Statistics,2000,9(2):249−26524Li Xin-De,Dong Qing-Quan,Wang Feng-Yu,Luo Chao-Min.A method of conflictive evidence combination based on the Markov chain.Acta Automatica Sinica,2015,41(5): 914−927(李新德,董清泉,王丰羽,雒超民.一种基于马尔科夫链的冲突证据组合方法.自动化学报,2015,41(5):914−927)25NASA.PCoE datasets:battery data set[Online],available: https:///tech/dash/groups/pcoe/prognostic-data-repository/,August22,2017.26Chen Qiao-Yan,Wei Ke-Xin.Signal Statistic analysis for electric vehicle battery SOH.Chinese Journal of Power Sources,2016,40(2):342−344(陈峭岩,魏克新.电动汽车电池SOH的信号统计分析.电源技术, 2016,40(2):342−344)27Weng C H,Feng X N,Sun J,Peng H.State-of-health mon-itoring of lithium-ion battery modules and packs via incre-mental capacity peak tracking.Applied Energy,2016,180: 360−36828Ye Tao,Zhu Xue-Feng,Li Xiang-Yang,Shi Bu-Hai.Soft sensor modeling based on a modified k-nearest neighbor re-gression algorithm.Acta Automatica Sinica,2007,33(9): 996−999(叶涛,朱学峰,李向阳,史步海.基于改进k-最近邻回归算法的软测量建模.自动化学报,2007,33(9):996−999)29Xue Peng-Song,Feng Min-Quan,Xing Xiao-Peng.Water quality prediction model based on Markov chain improving gray neural network.Engineering Journal of Wuhan Univer-sity,2012,45(3):319−324(薛鹏松,冯民权,邢肖鹏.基于马尔科夫链改进灰色神经网络的水质预测模型.武汉大学学报(工学版),2012,45(3):319−324)赵光财中国科学院海西研究院泉州装备制造研究所硕士研究生.2016年获得中国海洋大学学士学位.主要研究方向为锂电池状态估计.E-mail:zhaoguang-****************(ZHAO Guang-Cai Master stu-dent at Quanzhou Institute of Equip-ment Manufacturing,Haixi Institutes, Chinese Academy of Sciences.He received his bachelor de-gree from Ocean University of China in2016.His main research interest is states estimation of Li-ion batteries.)。

堆叠自动编码器的稀疏表示方法(Ⅲ)

堆叠自动编码器的稀疏表示方法(Ⅲ)

堆叠自动编码器的稀疏表示方法自动编码器是一种无监督学习的神经网络模型,它通过学习数据的内部表示来提取特征。

堆叠自动编码器则是由多个自动编码器叠加而成的深层网络模型。

在实际应用中,堆叠自动编码器通过学习更加抽象的特征表示,可以用于特征提取、降维和生成数据等多个领域。

在这篇文章中,我们将探讨堆叠自动编码器的稀疏表示方法,以及其在深度学习中的重要性。

稀疏表示是指在特征提取过程中,只有少数单元才被激活。

在堆叠自动编码器中,通过引入稀疏表示方法,可以让网络学习到更加鲁棒和有意义的特征。

稀疏表示可以有效地降低特征的冗余性,提高网络的泛化能力,使得网络能够更好地适应未见过的数据。

同时,稀疏表示还可以减少模型的计算复杂度,提高模型的训练效率。

因此,稀疏表示在深度学习中具有重要的意义。

在堆叠自动编码器中,稀疏表示的方法有很多种,其中最常用的方法之一是使用稀疏编码器。

稀疏编码器是一种特殊的自动编码器,它通过引入稀疏约束来学习稀疏表示。

在训练过程中,稀疏编码器会对每个隐藏单元引入稀疏性约束,使得只有少数隐藏单元被激活。

这样可以有效地提高特征的鲁棒性和泛化能力。

同时,稀疏编码器还可以使用稀疏性约束来降低特征的冗余性,提高特征的表达能力。

除了稀疏编码器,堆叠自动编码器还可以通过正则化方法来实现稀疏表示。

正则化是一种常用的方法,它可以通过引入额外的惩罚项来控制模型的复杂度。

在堆叠自动编码器中,可以通过引入L1正则化项来推动隐藏单元的稀疏性。

L1正则化项可以使得很多隐藏单元的激活值为0,从而实现稀疏表示。

通过正则化方法实现稀疏表示的堆叠自动编码器具有较好的鲁棒性和泛化能力,同时可以减少模型的计算复杂度,提高模型的训练效率。

另外,堆叠自动编码器还可以通过引入降噪自动编码器来实现稀疏表示。

降噪自动编码器是一种特殊的自动编码器,它可以通过在输入数据上添加噪声来训练模型。

在实际应用中,通过引入随机噪声,可以有效地降低模型对输入数据的敏感度,提高网络的鲁棒性。

高阶HBAM方法一般模型还可以

高阶HBAM方法一般模型还可以
Keywords: High-order neural networks; Exponential stability; Bidirectional associative memory (BAM); Time delays; Linear matrix inequality; Lyapunov fuIn recent years, Hopfield neural networks and their various generalizations have attracted the attention of many scientists (e.g., mathematicians, physicists, computer scientists and so on), due to their potential for the tasks of classification, associative memory, parallel computation and their ability to solve difficult optimization problems, see for example [1–9]. For Hopfield neural networks characterized by first-order interactions, Abu-Mostafa and Jacques [10], McEliece et al. [11], and Baldi [12] presented their intrinsic limitations. As a consequence, different architecture with high-order interactions [13–17] have been successively introduced to design neural networks which have stronger approximation property, faster convergence rate, greater storage capacity, and higher fault tolerance than lower-order neural networks; while the stability properties of these models for fixed weights have been studied in [18–21].

memetic算法matlab

memetic算法matlab

Memetic算法在解决复杂优化问题中具有重要的应用价值。

本文将介绍memetic算法的基本原理和实现流程,并结合matlab代码实例进行演示。

一、memetic算法简介1.1 memetic算法的概念memetic算法是一种结合了遗传算法和局部搜索方法的进化算法,它在进化过程中不仅利用全局搜索策略进行个体的遗传和进化,还结合了局部搜索算子对个体进行改进。

通过遗传和局部搜索的结合,memetic算法可以充分利用全局搜索的优势,又能够在局部搜索中快速收敛,从而有效地解决复杂优化问题。

1.2 memetic算法的优势memetic算法在解决复杂优化问题中具有以下优势:(1) 充分利用全局搜索和局部搜索的优势,有效平衡了探索和利用的能力,使得算法具有较强的收敛性和全局搜索能力。

(2) 通过局部搜索算子对个体进行改进,可以有效避免陷入局部最优解,提高了算法的搜索能力和解的质量。

(3) 算法的参数设置灵活,适应性强,能够适用于各种不同类型的优化问题。

1.3 memetic算法的应用领域memetic算法广泛应用于各种复杂优化问题的求解,如组合优化、路径规划、信号处理、机器学习等领域,已经成为解决复杂优化问题的重要工具。

二、memetic算法的实现流程2.1 memetic算法的基本步骤memetic算法的基本步骤包括:初始化种裙、选择操作、遗传操作、局部搜索操作、更新种裙,具体流程如下:(1) 初始化种裙:随机生成初始种裙,包括个体的编码和适应度的计算。

(2) 选择操作:根据个体的适应度值进行选择,选择优秀个体作为父代进行遗传操作。

(3) 遗传操作:通过交叉和变异等遗传操作对父代个体进行进化,生成新的子代个体。

(4) 局部搜索操作:对子代个体进行局部搜索改进,通过局部搜索算子对个体进行优化。

(5) 更新种裙:根据适应度值替换原种裙中的个体,更新种裙。

2.2 memetic算法的关键技术memetic算法的关键技术包括:适应度函数的设计、选择策略、遗传操作、局部搜索算子等。

memetic算法

memetic算法

A Memetic Algorithm for VLSI FloorplanningMaolin Tang,Member,IEEE,and Xin Yao,Fellow,IEEEAbstract—Floorplanning is an important problem in very large scale integrated-circuit(VLSI)design automation as it determines the performance,size,yield,and reliability of VLSI chips.From the computational point of view,VLSIfloorplanning is an NP-hard problem.In this paper,a memetic algorithm(MA)for a nonslicing and hard-module VLSIfloorplanning problem is presented.This MA is a hybrid genetic algorithm that uses an effective genetic search method to explore the search space and an efficient local search method to exploit information in the search region.The exploration and exploitation are balanced by a novel bias search strategy.The MA has been implemented and tested on popular benchmark problems.Experimental results show that the MA can quickly produce optimal or nearly optimal solutions for all the tested benchmark problems.Index Terms—Floorplanning,genetic algorithm(GA),local search,memetic algorithm(MA),very large scale integrated circuit(VLSI).I.I NTRODUCTIONF LOORPLANNING is important in very large scaleintegrated-circuit(VLSI)design automation as it deter-mines the performance,size,yield,and reliability of VLSI chips.Given a set of circuit components,or“modules,”and a net list specifying interconnections between the modules,the goal of VLSIfloorplanning is tofind afloorplan for the modules such that no module overlaps with another and the area of the floorplan and the interconnections between the modules are minimized.The representation offloorplans determines the size of the search space and the complexity of transformation between a representation and its correspondingfloorplan.Existing floorplan representations can be classified into two cate-gories,namely:1)“slicing representation”and2)“nonslic-ing representation.”It is commonly believed that nonslicing representations can contribute to better results than slicing representations.One of the most efficient nonslicing representations is the ordered tree(O-tree)representation proposed by Guo et al.[1]. The representation not only covers all optimalfloorplans but also has a smaller search space.For a VLSIfloorplanningManuscript received April30,2005;revised December22,2005and March2,2006.This work was supported in part by the National Natural Science Foundation of China under Grant60428202.This paper was recommended by Guest Editor Y.S.Ong.M.Tang is with Queensland University of Technology,Brisbane4001, Australia(e-mail:m.tang@.au).X.Yao is with the University of Birmingham,Birmingham B152TT,U.K., and also with the Nature Inspired Computation and Applications Laboratory, Department of Computer Science and Technology,University of Science and Technology of China,Hefei230027,China(e-mail:x.yao@). Color versions of one or more of thefigures in this paper are available online at .Digital Object Identifier10.1109/TSMCB.2006.883268problem of n modules,the search space of the representation is n!c n,wherec n=12n+12n+1nwhich is significantly smaller than otherfloorplan representa-tions.In addition,it only takes O(n)to transform between an O-tree representation and its correspondingfloorplan.More-over,the representation gives geometrical relations among modules,which is valuable for identifying meaningful build-ing blocks when designing genetic operators of evolutionary algorithms.Hence,the O-tree representation is adopted in this paper.Existing search methods for the VLSIfloorplanning prob-lem fall into two categories,namely:1)“local search”and 2)“global search.”Generally,the local search methods are efficient.However,they may not be able to produce an optimal or nearly optimal solution sometimes as their search may be trapped in a local region.A widely used global search method for VLSIfloorplanning problems is genetic algorithm(GA).GAs have been success-fully applied to solve slicing VLSIfloorplanning problems [2]–[4].For nonslicing VLSIfloorplanning,a GA has also been presented[5].Since the encoding scheme does not capture any topological information of thefloorplans,the performance of the GA is not satisfactory.For example,for the two popular benchmark problems,namely:1)“ami33”and2)“ami49,”the area usage is less than90%.In this paper,we present a memetic algorithm(MA)[6] for a nonslicing and hard-module VLSIfloorplanning prob-lem.MAs are population-based metaheuristic search methods which are inspired by Darwin’s principle of natural selection and Dawkins’notion of meme[7],defined as a unit of in-formation that reproduces itself while people exchange ideas, and they have been successfully applied on many complex problems[8]–[14].Our MA is a hybrid GA that uses an effective genetic search method to explore the search space and an efficient local search method to exploit information in the search region. The exploration and exploitation are balanced by a novel bias search strategy.The MA has been implemented and tested on popular benchmark problems for nonslicing and hard-module VLSIfloorplanning.Experimental results show that the MA can quickly produce optimal or nearly optimal solutions for all the popular benchmark problems.The remaining paper is organized as follows.Section II is the problem statement.Section III reviews related work.The MA is discussed in detail in Section IV,and empirical studies on the MA are presented in Section V.Finally,we conclude the MA in Section VI.1083-4419/$25.00©2007IEEEII.P ROBLEM S TATEMENTA module m i is a rectangular block with fixed height h i and width w i ,M ={m 1,m 2,...,m n }is a set of modules,and N is a net list specifying interconnections between the modules in M .A floorplan F is an assignment of M onto a plane such that no module overlaps with another.A floorplan has an area cost,i.e.,Area (F ),which is measured by the area of the smallest rectangle enclosing all the modules and an interconnection cost,i.e.,W irelength (F ),which is the total length of the wires fulfilling the interconnections specified by N .To minimize the costs,a module may be rotated 90◦.The cost of a floorplan F is defined as follows:cost (F )=w 1×Area (F )Area ∗+w 2×W irelength (F )W irelength ∗.(1)In the above equation,Area ∗and W irelength ∗represent theminimal area and the interconnection costs,respectively.Since we do not know their values in practice,estimated values are used.w 1and w 2are weights assigned to the area minimiza-tion objective and the interconnection minimization objective,respectively,where 0≤w 1,w 2≤1,and w 1+w 2=1.The interconnection cost is the total wire length of all the nets,and the wire length of a net is calculated by the half perimeter of the minimal rectangle enclosing the centers of the modules that have a terminal of the net on it.Given M and N ,the objective of the floorplanning problem is to find a floorplan F such that cost (F )is minimized.It should be pointed out that although the optimization problem has two objectives,multiobjective optimization techniques may not be suitable for it as the two objectives are not equally important and the weights assigned to the two objectives should be controllable by the designer to meet various requirements for the floorplanning problem.III.R ELATED W ORKA.O-Tree RepresentationA floorplan with n rectangular modules can be represented in a horizontal (vertical)O-tree of (n +1)nodes,of which n nodes correspond to n modules m 1,m 2,...,m n and one node corresponds to the left (bottom)boundary of the floorplan [1].The left (bottom)boundary is a dummy module with zero width (height)placed at x =0(y =0).In a horizontal O-tree,there exists a directed edge from module m i to module m j if and only if x j =x i +w i ,where x i is the x coordinate of the left-bottom position of m i ,x j is the x coordinate of the left-bottom position of m j ,and w i is the width of m i .In a vertical O-tree,there exists a directed edge from module m i to module m j if and only if y j =y i +h i ,where y i is the y coordinate of the left-bottom position of m i ,y j is the y coordinate of the left-bottom position of m j ,and h i is the height of m i .Fig.1(a)shows a floorplan and its horizontal O-tree representation.An O-tree can be encoded in a tuple (T,π),where T is a 2n -bit string specifying the structure of the O-tree,and πis a permutation of the nodes.For a horizontal O-tree,the tuple isA.Genetic RepresentationIn our MA,each individual in the population is an admissible floorplan represented by an O-tree and encoded in a tuple (T,π),where T is a 2n -bit string identifying the structure of the O-tree,and πis a permutation of the nodes.B.Fitness FunctionThe VLSI floorplanning is a minimization problem,and the objective is to minimize the cost of floorplan F ,i.e.,cost (F ).Thus,the fitness of an individual (T,π)in the population is defined as follows:f ((T,π))=1costF (T,π)(2)where F (T,π)is the corresponding floorplan of (T,π),and cost (F (T,π))is the cost of F defined in (1).C.Initial PopulationAn individual in the initial population is an O-tree (T,π)representing an admissible VLSI floorplan F .A constructive algorithm is designed to construct an admissible O-tree.The al-gorithm starts with randomly generating a sequence of modules π.Then,it inserts the modules into an initially empty O-tree T in the randomly generated order.When inserting a module into T ,it checks all external insertion positions for the module and inserts the module at the position that gives the best fitness.The constructive algorithm is invoked iteratively to generate an initial population of individuals.D.Genetic Operators1)Role of the Genetic Operators in the MA:The role that the genetic operators play in our MA is different from that in GAs.In GAs,crossover is used for both exploration and exploitation,and mutation is used for exploration.In our MA,however,the crossover and mutation operators are only used for exploration,or discovering new promising search regions.The crossover and mutation operators discover new promis-ing search regions by evolving memes,or units of cultural information,in a way analogous to biological evolution.Memes can mutate through,for example,misunderstanding,and two memes can recombine to produce a new meme involving ele-ments of each parent meme.It is observed that a subtree of the O-tree represents a compact placement of a cluster of modules.Hence,subtrees are used as memes in our MA.The memes are transmitted and evolved through one crossover operator and two mutation operators,which will be discussed in the following.2)Crossover:Given two parents,both of which are admis-sible floorplans represented by an O-tree,the crossover oper-ator transmits the significant structural information from two parents to a child.By recombining some significant structural information from two parents,it is hoped that better structural information can be created in the child.Fig.5.Strategy to bias the search.regions where the fitness value of the local optimum is less than v .The MA does not apply the local search method on p 2because the local search method is computationally expensive,and it takes more time for the local search method to find the optimum o 2from p 2than from p 3.(The local search method is a hill-climbing one.Thus,the better starting search points,the shorter the computation time given two starting search points on the same hill.)The threshold v is very important.It determines the balance between the exploration and exploitation of our MA.If its value is too big,then some promising search regions may be ignored,and therefore,the chance for our MA to find a global optimum is limited;if its value is too small,then our MA may waste too much time on exploiting less promising search regions,and therefore,the efficiency of our MA is decreased.It is desirable to find a threshold v such that our MA can minimize the use of the local search method without compromising its optimality.An empirical study of the threshold v can be found in Section V.To test the effectiveness of the threshold strategy,we have compared it with a random search strategy.As its name sug-gests,the random search strategy randomly picks up individu-als for exploitation.Experimental results have shown that the threshold strategy is significant.The details about the compari-son can also be found in Section V.F .Description of Our MAInitially,the MA randomly generates a population of individ-uals using the technique described above.Then,the MA starts evolving the population generation by generation.In each gen-eration,the MA uses the genetic operators probabilistically on the individuals in the population to create new promising search points (admissible floorplans)and uses the local search method to optimize them if the fitness of the admissible floorplans is greater than or equal to v .The process is repeated until a preset runtime is up.An outline of the MA is as follows:1)t :=0;2)generate an initial population P (t )of size PopSize ;3)evaluate all individuals in P (t )and find the best individ-ual best ;TABLE IS TATISTICS FOR THE T EST R ESULTS ON THE P ERFORMANCE OF THE MAFig.8.Determination of the value of the threshold v.thefitness values of the initialfloorplan and its corresponding optimizedfloorplan,respectively.There are two observations for all the benchmark problems (Fig.8).•The averagefitness value of the optimizedfloorplans from those initialfloorplans whosefitness value is less than0.6 is significantly poorer than that of the optimizedfloorplans from those initialfloorplans whosefitness is equal to or greater than0.6.•The best of the100optimizedfloorplans is generated from an initialfloorplan whosefitness value is equal to or greater than0.6.The two observations suggest that the threshold v should be set to0.6for the two benchmark problems.By doing so, the MA may ignore a number of search regions and therefore reduce computation time without compromising the optimality of solutions.C.On the Effectiveness of the Threshold StrategyTo test the effectiveness of the threshold strategy,we de-veloped two MAs,one using the threshold strategy to pick up individuals for local optimization and one randomly picking up 30%of individuals for local optimization(Since there are about 30%of individuals who are selected for local optimization in the threshold strategy,the random selection strategy also picks up30%of individuals for local optimization for a fair comparison.),and we tested on the benchmark ami33.For each MA,we repeatedly tested on the benchmark problem for ten times.Table III shows the Mann–Whitney statistical test results for the hypothesis“the cost obtained by the MA using the threshold strategy is less than or equal to the cost obtained by the MA using the random picking up strategy.”The difference between the medians of the costs obtained by the two MAs is−0.009,and the confidence interval for that is95.5%.TABLE IIIM ANN–W HITNEY T EST R ESULTS FOR THE H YPOTHESIS“T HE C OST O BTAINED BY THE MA U SING THE T HRESHOLD S TRATEGYI S L ESS T HAN OR E QUAL TO THE C OST O BTAINED BY THE MA U SING THE R ANDOM P ICKING UP S TRATEGY”[13]Y .S.Ong and A.Keane,“Meta-Lamarckian learning in memetic algo-rithms,”IEEE put.,vol.8,no.2,pp.99–110,Apr.2004.[14]N.Krasnogor and J.Smith,“A tutorial for competent memetic algorithms:Model,taxonomy,and design issues,”IEEE put.,vol.9,no.5,pp.474–488,Oct.2005.[15]H.Ishibuchi and T.Murata,“A multi-objective genetic local search algo-rithm and its application to flowshop scheduling,”IEEE Trans.Syst.,Man,Cybern.C,Appl.Rev.,vol.28,no.3,pp.392–403,Aug.1998.[16]K.-H.Liang,X.Yao,and C.Newton,“Evolutionary search of approxi-mated n -dimensional landscapes,”Int.J.Knowl.-Based Intell.Eng.Syst.,vol.4,no.3,pp.172–183,Jul.2000.[17]The MCNC Benchmark Problems for VLSI Floorplanning .[Online].Available:[18]The Analyse-it General Statistics Software for Microsoft Excel .[Online].Available:Maolin Tang (M’04)received the B.S.degree in computer science from Huazhong University of Science and Technology,Wuhan,China,the M.S.degree in computer science from Chongqing Uni-versity,Chongqing,China,and the Ph.D.degree in computer systems engineering from Edith Cowan University,Perth,W.A.,Australia.He is currently a Senior Lecturer with the Faculty of Information Technology,Queensland Univer-sity of Technology,Brisbane,Qld.,Australia.His research interests include evolutionary computa-tion,evolutionary VLSI physical design and optimization,and web-basedcomputation.Xin Yao (M’91–SM’96–F’03)received the B.Sc.degree from the University of Science and Tech-nology of China (USTC),Hefei,China,in 1982,the M.Sc.degree from the North China Institute of Computing Technology,Beijing,China,in 1985,and the Ph.D.degree from USTC,Hefei,in 1990.He was an Associate Lecturer and Lecturer from 1985to 1990with USTC while working on the Ph.D.degree.His Ph.D.work on simulated annealing and evolutionary algorithms was awarded the President’s Award for Outstanding Thesis by the Chinese Acad-emy of Sciences.In 1990,he took up a postdoctoral fellowship at the Com-puter Sciences Laboratory,Australian National University,Canberra,A.C.T.,Australia,and continued his work on simulated annealing and evolutionary al-gorithms.In 1991,he joined the Knowledge-Based Systems Group,Common-wealth Scientific and Industrial Research Organization (CSIRO)Division of Building,Construction,and Engineering,Melbourne,Vic.,Australia,working primarily on an industrial project on automatic inspection of sewage pipes.In 1992,he returned to Canberra to take up a lectureship at the School of Com-puter Science,University College,University of New South Wales,Australian Defence Force Academy,Kensington,N.S.W.,Australia,where he was later promoted to Senior Lecturer and Associate Professor.Attracted by the English weather,he moved to the University of Birmingham,Birmingham,U.K.,as a Professor (Chair)of computer science on April 1,1999.He is currently the Director of the Centre of Excellence for Research in Computational Intelli-gence and Applications,Birmingham,a Distinguished Visiting Professor of the University of Science and Technology of China,Hefei,and a Visiting Professor of three other universities.He has been invited to give more than 45invited keynote and plenary speeches at conferences and workshops worldwide.He has more than 200refereed research publications.His research has been supported by research councils,government organizations,and industry (more than £4M in the last four years).His major research interests include evolutionary computation,neural network ensembles,and their applications.Prof.Yao serves as the Editor-in-Chief of the IEEE T RANSACTIONS ON E VOLUTIONARY C OMPUTATION ,an associate editor or editorial board mem-ber of several other journals,and the Editor of the World Scientific book series on “Advances in Natural Computation.”He is also a Distinguished Lecturer of the IEEE Computational Intelligence Society.He is the recipient of the 2001IEEE Donald G.Fink Prize Paper Award for his work on evolutionary artificial neural networks.。

Memetic Algorithms遗传算法

Memetic Algorithms遗传算法

8
General Scheme of EAs
9
Pseudo-Code for typical EA
10
How to Combine EA and LS
11
Intelligent Initialization



The initial population is not given at pseudorandom but it is given according to a heuristic rule. Examples: quasi-random generator, orthogonal arrays It increases the average fitness but it decreases the diversity
19
Adaptivity + Multi-Meme


In order to properly select from the list the LS to use for the different stages of the evolution an adaptive strategy can be used If the “necessities” of the evolutionary process are efficiently encoded it is possible to use different LSs in different moments and on different individuals (or set of individuals)
12
Intelligent Variation Operators


Intelligent Crossover: finds the best combination between parents in order to generate the most performing offspring (e.g. heuristic selection of the cut point) Intelligent Mutation: tries several possible mutated individuals in order to obtain the most “lucky” mutation (e.g. bit to flip)

自适应memetic算法求解集合覆盖问题

自适应memetic算法求解集合覆盖问题

自适应memetic算法求解集合覆盖问题林耿;关健【摘要】集合覆盖问题是一个经典的NP困难的组合优化问题,有着广泛的应用背景.首先,采用动态罚函数法将集合覆盖问题等价转化为无约束的0-1规划问题.然后,基于集合覆盖问题的结构特征,设计了初始种群构造方法、局部搜索方法、交叉算子、动态变异算子和路径重连策略,提出了一个高效求解该0-1规划问题的自适应memetic算法.该算法有效平衡了集中搜索和多样化搜索.通过45个标准例子测试该算法,并将其结果与现有遗传算法进行了比较,表明该算法能够在可接受的时间内找到高质量的解,能够有效求解大规模集合覆盖问题.【期刊名称】《浙江大学学报(理学版)》【年(卷),期】2016(043)002【总页数】7页(P168-174)【关键词】集合覆盖问题;memetic算法;罚函数;局部搜索;路径重连【作者】林耿;关健【作者单位】闽江学院数学系,福建福州350108;闽江学院现代教育技术中心,福建福州350108【正文语种】中文【中图分类】O224.1;TP18集合覆盖问题(set covering problem)是一个经典的组合优化问题.它是NP困难问题[1],并且在通讯工业、后勤保障、市场销售、计算生物学等领域应用广泛.集合覆盖问题理论的重要性和应用的广泛性,引起了学者们的广泛关注,其算法大致可以分为精确算法、近似算法和启发式算法.精确算法主要基于branch-and-bound算法和branch-and-cut算法[2-3].虽然精确算法能够找到集合覆盖问题的最优解,近似算法[4-5]能够找到有质量保证的解,但是它们的求解速度都比较慢,只能求解较小规模的实例[6].启发式算法因其能够较快地找到问题的高质量解,成为学者们研究的热点[6-11].BEASLEY等[7]提出了求解集合覆盖问题的遗传算法.吴志勇等[8]提出了一个二阶段遗传算法,该算法首先对实例进行有效约简,缩小问题的规模,然后,通过遗传算法求得一个高质量的近似解.45个标准试例的测试结果表明,该算法的求解效率和求解质量高于其他遗传算法.蒋建林等[9]提出了一种改进的遗传算法.这些遗传算法求解质量较高,但由于集中搜索能力(intensification)较弱,当种群数量较大时,算法的收敛速度较慢,求解消耗的时间较长.Memetic算法是基于种群的全局搜索和基于个体的局部搜索的结合体,已被成功应用于求解各种高难度的优化问题.本文根据集合覆盖问题的特点,提出了一种自适应的memetic算法.该memetic算法通过迭代改进的局部搜索算法和路径重连策略,有效提高了算法搜索的集中能力;通过动态变异算子,提高了算法搜索的多样性.这些策略有效平衡了算法的集中搜索和多样性搜索能力.采用45个国际标准测试例子测试本文算法,并与现有遗传算法进行比较,以证明本算法的高效性.给定一个m行n列的0-1矩阵A=(aij)m×n和列的费用cj,其中aij=1表示第i行被第j列覆盖,aij=0表示第i行未被第j列覆盖,集合覆盖问题寻找一个费用总和最小的列的集合D,使得所有行都至少被D中的一列覆盖.引入n维0-1向量x=(x1,…,xn),如果第j列被选入集合D,则xj=1;否则,xj=0.则集合覆盖问题可以描述为以下0-1规划问题[6,8,10](P):,s.t. ,xj∈{0,1},j=1,2,…,n.2.1 动态罚函数的构造集合覆盖问题(P)本质上是一个带约束的0-1规划问题.精确罚函数法形式简单,已经广泛应用于将带约束的问题转化为无约束问题.如果aijxj<1,记αi(x)=1,表示第i 行未被解x覆盖,否则αi(x)=0,表示第i行已经被解x覆盖.令αi(x),表示未被解x 覆盖的行数.当α(x)=0时,x是问题(P)的可行解.记问题(P)的可行域为S={x∈{0,1}n:α(x)=0}.采用精确罚函数法,问题(P)可以转化为如下的无约束0-1规划问题(EUP):minf(x)+λα(x),其中λ>0为罚参数.精确罚函数法形式简单,但是罚参数λ比较敏感,并且难以选择合适的值.为了解决罚参数难以选择的问题,近年来,学者们提出了无参数罚函数法、自适应罚函数法等.针对带约束的连续优化问题,ALI等[12]提出了自适应动态罚函数法,此方法有效降低了罚参数对实例的敏感程度.基于集合覆盖问题的特点,构造如下动态罚函数:其中,μ>0是罚参数,U是问题(P)的上界.U的初始值可以通过贪心算法等求得.在搜索过程中,如果找到更好的可行解,则更新U.基于以上的动态罚函数,构造如下无约束0-1规划问题(DUP):minφ(x),s.t. xj∈{0,1}, j=1,2,…,n.2.2 动态罚函数的性质定理1 当μ>cmax=max{cj,j∈{1,2,…,n}}时,问题(P)和(DUP)具有相同的全局最优解.证明假设x*是问题(P)的全局最优解,则对∀x∈S,有f(x)≥f(x*).对于∀x∈{0,1}n,考虑以下2种情况:x∈S和x∉S.当x∈S时,有当x∉S时,无论f(x)≤U或是f(x)>U,由φ(x)的定义都有φ(x)>f(x),所以由式(1)与(2)得x*是问题(DUP)的全局最优解.假设x*是问题(DUP)的全局最优解,则对∀x∈{0,1}n,有φ(x)≥φ(x*).由于μ>cmax,则必有x*∈S.由φ(x)的定义得:对于∀x∈S,f(x)=φ(x)≥φ(x*)=f(x*),即x*也是问题(P)的全局最优解.由定理1可知,可以通过求解无约束的问题(DUP)来找到原问题的全局最优解.本文的基本思想是通过memetic算法求解问题(DUP),以期找到原问题高质量的解.该memetic算法采用如下的1变换邻域结构:定理2 当μ>cmax时,问题(DUP)的局部最优解必然是问题(P)的可行解.证明假设y是问题(DUP)的局部最优解,则对∀x∈N(y),有如果y∉S,则α(y)>0.故至少存在一个未选的列j,使得将该列选入后,未覆盖的行数会减少,即令yj=1后,得到新的解y′满足α(y)>α(y′).显然,y′∈N(y).又由于μ>cmax,有φ(y)>φ(y′),与式(4)矛盾.故y∈S.本文的基本思想是通过自适应memetic算法求解无约束问题(DUP)来得到问题(P)的高质量的解.下面首先介绍memetic算法的几个主要组成部分,然后给出算法的详细步骤.3.1 初始种群的生成初始种群P={x1,x2,…,xp}的构造对memetic算法的性能有很大的影响.本文采用随机贪心自适应方法来构造质量高、多样性好的初始种群.设D表示选中的列的集合,M表示未被选中的列的集合,算法的具体步骤如下:算法1:步骤1 初始化D=φ,M={1,2,…,n},xj=0,j=1,2,…,n.步骤2 记hj表示将第j列加入D后,新覆盖的行的数量.M中列j的性价比Q(j)按照式(5)计算出M中列j的性价比Q(j).步骤3 构造候选列的集合RCL={j∈M:Qmax≥Q(j)≥δ Qmax},其中1>δ>0为参数,用于平衡随机性和贪婪性.步骤4 从RCL中随机选择一列j′放入D,令M=M-{j′},xj′=1.步骤5 如果α(x)=0,停止,输出x;否则,转步骤2.重复运行以上算法p次得到初始种群.3.2 局部搜索算法文献[7-8]中的遗传算法由于缺乏局部搜索,使得算法的收敛速度较慢.针对划分问题,FIDUCCIA等[13]提出了一种高效的迭代改进搜索算法(FM),该算法的变形已经应用于求解许多划分问题.本文采用改进的FM算法作为局部搜索算法(记为SCPFM),有效提高了算法的搜索效率.该局部搜索算法SCPFM由一系列pass组成.假设x0为初始解.在每个pass的初始阶段,所有变量都能够自由翻转(即由xj变为1-xj).假设gain(j)表示翻转变量xj后的增益,即:其中x′=(x1,x2,…,xj-1,1-xj,xj+1,…,xn).SCPFM根据式(6)计算出所有能够自由翻转的变量的增益,并找出增益最大的变量(最大增益可能大于0,也可能小于0).为了增加局部搜索的多样性,SCPFM以0.8的概率翻转增益最大的变量.变量被翻转后,就锁定该变量,即在当前的pass中禁止该变量再翻转.更新自由变量的增益.记xb是该pass中到目前为止找到的最好的解.重复以上步骤,直到φ(x)-φ(xb)>ξ时,该pass停止,其中φ(x)为当前解的目标函数值,ξ>0为参数.因为当ξ比较大时,在该pass中,继续翻转自由变量,将无法找到比xb更好的解.接下来的pass以xb作为初始解.重复以上迭代,直到无法得到改进解为止.设F表示自由变量的集合,SCPFM算法的具体步骤如下:步骤1 初始化x=x0,xb=x0,F={1,2,…,n}.步骤2 根据式(6)计算所有自由变量的增益.令F′=F.步骤3 令xj=arg max{gain(i),i∈F′},随机产生σ∈(0,1),若σ≤0.8,概率将xj 翻转,即令xj=1-xj;否则,F′=F′-{j},重复步骤3.步骤4 锁定xj,即F=F-{j}.如果φ(x)<φ(xb),令xb=x.步骤5 如果φ(x)-φ(xb)>ξ,转步骤6;否则,转步骤2.步骤6 如果φ(xb)<φ(x0),令x0=xb,转步骤1;否则局部搜索算法停止,输出xb.局部搜索算法SCPFM主要有2个操作:计算自由变量的增益和寻找具有最大增益的自由变量.在每个pass的初始阶段,由式(6)计算增益,需要算出f(x)和α(x).计算f(x)和α(x)分别需要O(n)、O(mn)时间.故计算所有自由变量的增益需要O(mn)时间.本文采用桶的数据结构[13]能够在O(1)时间内找到具有最大增益的自由变量.在一个pass中,每个变量至多翻转一次,所以一个pass的时间复杂度为O(mn2).由于SCPFM通过调整参数ξ的值来提前停止pass,在一个pass中只有少数变量能够翻转,即一个pass的复杂度降低为O(mn).并且实验表明,SCPFM一般都只含有很少的pass.所以在实际求解中,SCPFM的搜索速度非常快.3.3 交叉算子与变异算子交叉算子既要继承当前种群中解的优良结构,又要保存群体的多样性.本文的交叉算子分2个步骤.从当前种群中随机选出2个解x,y作为父母,产生新解z.初始化=0,j=1,2,…,n.第1个步骤z的部分基因直接从父母继承.如果xj=yj且都等于1时,令zj=1;否则,zj=0.第2个步骤,通过贪心策略,对z进行修复.根据式(5)算出解z的所有未选列的性价比,然后,重复执行算法1中的步骤3与4,直到α(z)=0为止,并将z输出.为了提高搜索的多样性,根据集合覆盖问题的特点,引入动态变异算子,对解进行变异.假设解x选中列的集合为D,变异算子随机从D中选择mu个列删除.mu是控制变异程度的参数,在搜索过程中,随当前解的质量动态变化,详细情况将在3.5节中介绍.3.4 路径重连策略路径重连是通过建立当前解与导向解之间的连接路径,从中获得新解的一种有效的启发式搜索策略.本文采用路径重连来加强memetic算法的搜索能力.本文的路径重连以当前最好的解x*作为导向解.首先建立当前解x与x*之间的差异集合以往的路径重连策略按照某种规则,每次从Δ中选择1个变量进行翻转,直到发现更好的解或到达导向解(Δ=∅).为了更加有效地对x*附近的解进行搜索,本文的路径重连首先考虑从{1,2,…,n}中找出增益最大的变量xj,如果翻转xj,能够得到比x*更好的解,则翻转xj;否则,从Δ中选择增益最大的变量xt进行翻转.路径重连策略的具体步骤如下:算法3:步骤1 根据式(7)得到差异集Δ.步骤2 由式(6)算出所有变量的增益.设xj=arg max{gain(i),i∈{1,2,…,n}},z=(x1,x2,…,xj-1,1-xj,xj+1,…,xn),如果φ(z)<φ(x*),停止,输出z;否则,设xt=arg max{gain(i),i∈Δ},令xt=1-xt.步骤3 令Δ=Δ-{t},如果Δ≠φ,转步骤2;否则,停止,输出x*.3.5 自适应memetic算法的步骤求解集合覆盖问题的自适应memetic算法的流程如图1所示,其基本思想如下:首先通过算法1构造具有良好多样性的初始种群P={x1,x2,…,xp},并用SCPFM算法对种群中的解进一步优化.种群P中质量最好的解记为x*,令U=f(x*).其次,memetic算法通过2种方式产生新解.第1种,从P中任意选择2个解xs,xt,应用交叉算子产生解z,再通过变异,得到的解仍记为z;第2种,从P中任意选择1个解xs进行变异,得到新解z.最后,利用SCPFM算法对z进行优化,采用算法3将优化得到的解与x*做路径重连,如果找到比x*更好的解,则更新x*和U.重复以上步骤,至达到最大迭代步数,输出x*.mu是变异算子中控制变异程度的参数.在搜索的早期,为了快速提高种群解的质量,mu的取值应该较小.随着搜索的深入,种群的多样性变差,应当增加mu,开拓新的搜索空间.记xi∈P,i=1,2,…,p的变异参数分别为mu(i),i=1,2,…,p.设最小变异值为min_mu,最大变异值为max_mu,即min_mu≤mu(i)≤max_mu.初始化mu(i)=min_mu,如果通过变异和局部搜索后,得到的解比xi差,则需要增大变异参数,令mu(i)=mu(i)+1.如果在搜索的过程中,mu(i)超过所允许的最大变异值max_mu,则令mu(i)=min_mu.假设算法的最大迭代次数为G,自适应memetic算法的步骤如下:算法4:步骤1 算法1运行p次,构造初始种群P={x1,x2,…,xp}.运用SCPFM算法(算法2)分别对种群P中的解进行优化,所得解仍记为xi,i=1,2,…,p.令x*=arg min{f(xi), xi∈P}, U=f(x*),generation=1.初始化mu(i)=min_mu,i=1,2,…,p,P′=φ.步骤2 令k=1.步骤3 产生1个随机数r∈[0,1],如果r≤0.5,从P中任意选择2个解xs和xt,令,应用交叉算子和变异算子(变异参数为mu′(k))产生新解z,利用SCPFM算法对z进行优化,所得到的解仍记为z.如果φ(z)≥φ(xs)且φ(z)≥φ(xt),则令mu′(k)=mu′(k)+1.如果mu′(k)>max_mu,令mu′(k)=min_mu.步骤4 如果0.5<r≤1,从P中任意选择1个解xs,令mu′(k)=mu(s),应用变异算子(变异参数为mu′(k))产生新解z,利用SCPFM算法对z进行优化,所得到的解仍记为z.如果φ(z)≥φ(xs),则令mu′(k)=mu′(k)+1;如果mu′(k)>ma_mu,令mu′(k)=min_mu.步骤5 如果φ(z)<φ(x*)且α(z)=0,则x*=z,U=f(z),mu′(k)=min_mu.步骤6 P′=P′∪{z},采用算法3将z与x*做路径重连,如果找到比x*好的解,更新x*和U.步骤7 如果k=p,mu(i)=mu′(i),i=1,2,…,p,P=P′,P′=φ,转步骤8;否则,k=k+1,转步骤3.步骤8 令generation=generation+1,如果generation<G,转步骤2;否则,停止,输出x*和f(x*).为了检验自适应memetic算法的性能,用c语言对自适应memetic算法进行编程.本算法的运行环境为Windows 7,CPU 3.4 GHz,内存4 GB.采用OR-Library中的45个国际标准测试例子测试本文提出的算法.这些例子已广泛用于对集合覆盖问题相关算法的测试.根据前期实验,算法的参数设置如下:种群规模p=10,μ的值为x*中选中列的权总和的平均值,δ=0.8,ξ=3,min_mu=3,min_mu的值为x*中选中列的数量的3/4,G=800.利用本文提出的自适应memetic算法求解45个标准测试例子,每个例子求解10次.表1展示了10次实验获得的最好解(Best)、平均解(Avg)以及找到最好解时所用的平均时间(Time)(单位:s).为便于结果的比较分析,将遗传算法[7]、二阶段遗传算法[8]的计算结果(目标函数值C和运行时间Time)也列于表1中.遗传算法、二阶段遗传算法的计算结果引自文献[8].对比表1的结果可以看出:(1)在45个测试例子中,遗传算法、二阶段遗传算法和本文算法分别在6,44,44个例子中找到了最优解.在34个例子中,本文得到的平均解优于遗传算法.(2)从算法的求解时间看,全部例子均表明本文算法的求解时间明显低于遗传算法和二阶段遗传算法.遗传算法、二阶段遗传算法和本文算法求解的平均时间分别为1332.492,194.561和64.886 s.遗传算法、二阶段遗传算法是在CPU 3.4 GHz,2 GB内存的电脑上运行的,运行的环境与本文算法相当. 以上结果表明,本文算法优于遗传算法,能够在较短时间内找到与二阶段遗传算法可比拟的结果,大大提高了求解速度.采用例子4.1,4.2,4.3,4.4,4.5,B1、B2、B3、B4、B5、D1、D2、D3、D4、D5、cyc06、cyc07测试改进遗传算法[9]的性能,实验结果表明,除4.1外,该算法均能够找到最优解.应用本文算法时,这17个测试例子都能找到最优解.可见本文的memetic算法比改进遗传算法更为有效.集合覆盖问题本质上是一个带约束的0-1规划问题,本文通过动态罚函数法将集合覆盖问题等价转化为无约束的0-1规划问题,并提出了高效的自适应memetic算法以求解该无约束0-1规划问题.自适应memetic算法首先通过贪心自适应算法构造了具有良好多样性的初始种群,采用局部搜索算法和路径重连策略加强了其局部搜索能力,并通过动态变异算子来避免早熟收敛,提高了种群多样性.实验结果表明,所提算法较于其他算法更高效,求值速度更快.【相关文献】[1] GAREY M R, JOHNSON D S. Computers and Intractability: A Guide to the Theory ofNP-Completeness[M]. San Francisco:Freeman,1979.[2] BALAS E, CARRERA M C. A dynamic subgradient-based branch-and-bound procedure for set covering[J]. Operations Research,1996,44(6):875-890.[3] FISHER M R, KEDIA P. Optimal solution of set covering/partitioning problems using dual heuristics[J]. Management Science,1990,36(6):674-688.[4] HOCHBAUM D S. Approximation algorithms for the set covering and vertex cover problems[J]. SIAM Journal on Computing,1982,11(3):555-556.[5] GROSSMAN T, WOOL A. Computational experience with approximation algorithmsfor the set covering problem[J]. European Journal of Operational Research,1997,101(1):81-92.[6] REN Z G, FENG Z R, KE L J, et al. New ideas for applying ant colony optimization tothe set covering problem[J]. Computers & Industrial Engineering,2010,58(4):774-784. [7] BEASLEY J E, CHU P C. A genetic algorithm for the set covering problem[J]. European Journal of Operational Research,1996,94(2):392-404.[8] 吴志勇,陈韬,王红川,等.一个解决集合覆盖问题的二阶段遗传算法[J].小型微型计算机系统,2011,32(4):732-737.WU Zhiyong, CHEN Tao, WANG Hongchuan, et al. Two stage genetic algorithm for set covering problem [J]. Journal of Chinese Computer Systems,2011,32(4):732-737[9] 蒋建林,程坤,王璨璨,等.基于改进遗传算法的集合覆盖问题[J].数学的实践与认识,2012,42(5):120-126.JIANG Jianlin, CHENG Kun, WANG Cancan, et al. Improved genetic algorithm for set covering problem [J]. Mathematics in Practice and Theory,2012,42(5):120-126.[10] NAJI-AZIMI Z, TOTH P, GALLI L. An electromagnetism metaheuristic for the unicost set covering problem[J]. European Journal of Operational Research, 2010, 205(2): 290-300.[11] SUNDAR S, SINGH A. A hybrid heuristic for the set covering problem[J]. Operational Research,2012,12(3):345-365.[12] ALI M M, ZHU W X. A penalty function-based differential evolution algorithm for constrained global optimization[J]. Computational Optimization andApplications,2013,54(3):707-739.[13] FIDUCCIA C M,MATTHEYSES R M. A linear time heuristic for improving network partitions[C]// Proceedings of 19th ACM/IEEE Design Automation Conference. New Jersey: IEEE Press,1982:175-181.。

  1. 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
  2. 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
  3. 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。

A Multiobjective Memetic Algorithm Basedon Particle Swarm OptimizationDasheng Liu,K.C.Tan,C.K.Goh,and W.K.HoAbstract—In this paper,a new memetic algorithm(MA)for multiobjective(MO)optimization is proposed,which combines the global search ability of particle swarm optimization with a synchronous local search heuristic for directed localfine-tuning.A new particle updating strategy is proposed based upon the concept of fuzzy global-best to deal with the problem of premature convergence and diversity maintenance within the swarm.The proposed features are examined to show their individual and com-bined effects in MO optimization.The comparative study shows the effectiveness of the proposed MA,which produces solution sets that are highly competitive in terms of convergence,diversity,and distribution.Index Terms—Memetic algorithm(MA),multiobjective(MO) optimization,particle swarm optimization(PSO).I.I NTRODUCTIONM ANY real-world optimization problems involve opti-mizing multiple noncommensurable and often com-peting criteria that reflect various design specifications and constraints[22].Ever since the pioneering effort of Schaffer [18],many evolutionary techniques for multiobjective(MO) optimization have been proposed.A few of these algorithms in-clude the nondominated sorting genetic algorithm II(NSGAII) [4],the strength Pareto evolutionary algorithm2(SPEA2) [23],the incrementing MO evolutionary algorithm(IMOEA) [20],etc.Particle swarm optimization(PSO)[10]is a stochastic op-timization technique that is inspired by the behavior of bird flocks.Although PSO is relatively new,it has been shown to of-fer higher convergence speed for MO optimization as compared to canonical MOEAs.In recent years,thefield of MOPSO has been steadily gaining attention from the research commu-nity[2],[6],[14],[17].It is known that memetic algorithms(MAs)[15],hybridizing EAs and local search(LS)heuristics can be implemented to maintain a balance between exploration and exploitation,which is often crucial to the success of the search and optimization processes[12].While many works on such hybrids exist in the context of EAs[1],[8],[9],[13],[16],MA is rarely considered in PSO.This paper is concerned with the development of a multi-objective memetic algorithm(MOMA)within the context of PSO.In contrast to single-objective(SO)optimization,it isManuscript received May12,2005;revised October28,2005and January5,2006.This paper was recommended by Associate Editor Y.S.Ong. The authors are with the Control and Simulation Laboratory,Department of Electrical and Computer Engineering,National University of Singapore, Singapore117576(e-mail:90301345@.sg).Digital Object Identifier10.1109/TSMCB.2006.883270essential to obtain a well-distributed and diverse solution set forfinding thefinal tradeoff in MO optimization.However,the high speed of convergence in PSO often implies a rapid loss of diversity during the optimization process,which inevitably leads to undesirable premature convergence.Thus,the chal-lenge in designing a MOMA based upon PSO is to deal with the premature convergence without compromising the conver-gence speed.Apart from hybridizing LS heuristic and PSO,the existing particle updating strategy in PSO is extended to account for the requirements in MO optimization.Two heuristics are proposed to deal with the issue,including a synchronous particle LS (SPLS)and a fuzzy global-best(f-gbest)for the updating of a particle trajectory.The SPLS utilizes swarm information to perform a directedfine-tuning operation,while the second heuristic is based upon the concept of possibility[5]to deal with the problem of maintaining diversity within the swarm as well as to promote exploration in the search.The remainder of this paper is organized as follows.Some background information is provided in Section II,while de-tails of the proposed features of the MOMA are described in Section III.The proposed MAs performance on benchmarks is shown in Section IV.Section V examines the individual and combined effects of the proposed features,and Section VI presents a comparative study of the proposed MA with well-known MO optimization algorithms on a number of benchmark problems.Conclusions are drawn in Section VII.II.P RELIMINARIESA.MO OptimizationIn general,many real-world applications involve complex optimization problems with various competing specifications and constraints.Without loss of generality,we consider a minimization problem with decision space X,which is a subset of real numbers.For the minimization problem,it tends tofind a parameter set P forMinP∈XF(P),P∈R D(1)where P={p1,p2,...,p D}is a vector with D decision variables and F={f1,f2,...,f M}are M objectives to be minimized.The solution to the MO optimization problem exists in the form of an alternate tradeoff known as a Pareto optimal set. Each objective component of any nondominated solution in the Pareto optimal set can only be improved by degrading at least1083-4419/$25.00©2007IEEEone of its other objective components.A vector F a is said to dominate another vector F b,denoted asF a≺F b,iff f a,i≤f b,i∀i={1,2,...,M}and∃j∈{1,2,...,M}where f a,i≺f b,i.(2) In the total absence of information regarding the preference of objectives,a ranking scheme based upon the Pareto optimality is regarded as an appropriate approach to represent thefitness of each individual in an EA for MO optimization[7].B.Particle Swarm OptimizationA standard particle swarm optimizer maintains a swarm of particles that represent the potential solutions to the problem on hand.Each particle P i(=x i,1×x i,2×···×x i,D)embeds the relevant information regarding the D decision variables{x j,j=1,2,...,D}and is associated with afitness that pro-vides an indication of its performance in the objective space F∈R M.Its equivalence in F is denoted by F i(=f i,1×f i,2×···×f i,M),where{f k,k=1,2,...,M}are the objectives to be minimized.In essence,the trajectory of each particle is updated ac-cording to its ownflying experience as well as to that of the best particle in the swarm.The basic PSO algorithm can be described asνk+1 i,d =w×νk i,d+c1×r k1×p k i,d−x k i,d+c2×r k2×p k g,d−x k i,d(3)x k+1i,d=x k i,d+νk i,d(4)whereνki,dis the d th dimension velocity of particle i in cycle k;x k i,d is the d th dimension position of particle i in cycle k; p k i,d is the d th dimension of personal best(pbest)of particle iin cycle k;p kg,d is the d th dimension of the gbest in cycle k;wis the inertia weight;c1is the cognition weight and c2is the social weight;and r1and r2are two random values uniformly distributed in the range of[0,1].C.Performance MetricsIn order to provide a quantitative assessment for the per-formance of MO optimizer,three issues are often taken into consideration,i.e.,the distribution,the spread across the Pareto optimal front,and the ability to attain the global tradeoff [19].Comparative studies performed by researchers such as Deb et al.[4],etc.,made use of a suite of unary performance metrics pertinent to the optimization goals of proximity,diver-sity,and distribution.In this paper,three different qualitative measures are used.The metric of generational distance(GD)gives a good indi-cation of the gap between the discovered Pareto front(PF known) and the true Pareto front(PF true),which is given byGD=1n PFn PFi=1d2i1/2(5)where n PF is the number of members in PF known and d1is theEuclidean distance between the member i in PF known and itsnearest member in PF true.The metric of maximum spread(MS)measures how“well”the PF true is covered by the PF known through hyperboxesformed by the extreme function values observed in the PF trueand PF known.It is defined asMS=1MMi=1min(f maxi,F maxi)−max(f mini,F mini)F maxi−F mini2(6)where M is the number of objectives,f maxiand f miniare themaximum and minimum of the i th objective in PF known,res-pectively,and F maxiand F miniare the maximum and minimumof the i th objective in PF true,respectively.The metric of spacing(S)gives an indication of how evenlythe solutions are distributed along the discovered frontS=1n PFn PFi=1(d i−¯d )21/2/¯d ,¯d =1n PFn PFi=1d i(7)where n PF is the number of members in PF known and d i isthe Euclidean distance(in the objective domain)between themember i in PF known and its nearest member in PF known.III.MO M EMETIC PSOAs described in the Introduction,the proposed MA for MOoptimization aims to preserve population diversity forfindingthe optimal Pareto front and to include the feature of localfine-tuning for good population distribution byfilling any gapsor discontinuities along the Pareto front.In this section,theproposed features of SPLS and f-gbest will be described,andthe implementation detail of the MA will be presented.A.ArchivingIn our algorithm,elitism[21]is implemented in the form ofafixed-size archive to prevent the loss of good particles due tothe stochastic nature of the optimization process.The size ofthe archive can be adjusted according to the desired number ofparticles distributed along the tradeoff in the objective space.The archive is updated at each cycle,e.g.,if the candidatesolution is not dominated by any members in the archive,itwill be added to the archive.Likewise,any archive membersdominated by this solution will be removed from the archive.Inorder to maintain a set of uniformly distributed nondominatedparticles in the archive,the dynamic niche-sharing scheme[21]is employed.When the predetermined archive size is reached,arecurrent truncation process[11]based on niche count is usedto eliminate the most crowded archive member.B.Selection of the GbestIn MOPSO,gbest plays an important role in guiding theentire swarm toward the global Pareto front.Contrary to SOoptimization,the gbest for MO optimization exists in the form of a set of nondominated solutions,which inevitably leads to the issue of selecting the gbest.It is known that the selection of an appropriate gbest is critical for the search of a diverse and uniformly distributed solution set in MO optimization. Adopting a similar approach in[2],each particle in the swarm will be assigned a nondominated solution from the archive as gbest.Specifically,binary tournament selection of nondominated solutions from the archive is carried out inde-pendently for each particle in every cycle.Therefore,each particle is likely to be assigned different archived solution as the gbest.This implies that each particle will search along different direction in the decision space,thus aiding the exploration process in the optimization.In order to promote diversity and to encourage exploration of the least populated region in the search space,the selection criterion for reproduction is based on niche count.In the event of a tie,preference will be given to solutions lying at the extreme ends of an arbitrarily selected objective.C.Fuzzy Global Best(f-gbest)In PSO,the swarm converges rapidly within the intermediate vicinity of the gbest.However,such a high convergence speed often results in:1)the lost of diversity required to maintain a diverse Pareto front and2)premature convergence if the gbest corresponds to a local optima.This motivates the development of f-gbest,which is based on the concept of possibility measure to model the lack of information about the true optimality of the gbest.In contrast to conventional approaches,the gbest is denoted as“possibly at(d1×d2×d3···×d D),”instead of a crisp location.Consequently,the calculation of particle velocity can be rewritten asp k c,d=Np k g,d,σ(8)σ=f(k)(9)νk+1 i,d =w×νk i,d+c1×r k1×p k i,d−x k i,d+c2×r k2×p k c,d−x k i,d(10)where p kc,d is the d th dimension of f-gbest in cycle k.From(8),it can be observed that the f-gbest is characterized by anormal distribution N(p kg,d ,σ),whereσrepresents the degreeof uncertainty about the optimality of the gbest.In order to account for the information received over time that reduces uncertainty about the gbest position,σis modeled as some nonincreasing function of the number of cycles as given in(9). For simplicity,f(k)is defined asf(k)=σmax,cycles<α·max_cyclesσmin,otherwise(11)whereσmax,σmin,andαare set as0.15,0.0001,and0.4, respectively.The f-gbest should be distinguished from conventional tur-bulence or mutation operator,which applies a random pertur-bation to the particles.The function of f-gbest is toencourage Fig.1.Search region off-gbest.Fig.2.SPLS of assimilated particle along x1and x3.the particles to explore a region beyond that defined by the search trajectory,as illustrated in Fig.1.By considering the uncertainty associated with each gbest as a function of time, f-gbest provides a simple and efficient exploration at the early stage whenσis large and encourages localfine-tuning at the latter stage whenσis small.Subsequently,this approach helps to reduce the likelihood of premature convergence and guides the search towardfilling any gaps or discontinuities along the Pareto front for better tradeoff representation.D.Synchronous Particle Local Search(SPLS)The MOPSO is hybridized with an SPLS,which performs directed localfine-tuning to improve the distribution of non-dominated solutions.The issues considered in the design of the LS operator include:1)the selection of appropriate search direction;2)the selection of appropriate particles for local opti-mization;and3)the allocation of computational budget for LS.Fig.3.Flowchart of FMOPSO.SPLS is performed in the vicinity of the particles,and the procedure is outlined in the following pseudocode.Step1)Select S LS particles randomly from particle swarm. Step2)Select N LS nondominated particles with the best niche count from the archive to a selection pool. Step3)Assign an arbitrary nondominated solution from the selection pool to each of the S LS particles as gbest. Step4)Assign an arbitrary search space dimension for each of the S LS particles.Step5)Assimilation:With the exception of the assigned dimension,update the position of S LS particles inthe decision space with the selected gbest position. Step6)Update the position of all S LS assimilated parti-cles using(8)–(11)along the preassigned dimen-sion only.The operation of SPLS is illustrated in Fig.2.The rationale of selecting the least crowded particles from the archive in Step2) is to encourage the discovery of better solutions thatfill the gaps or discontinuities along the discovered Pareto front.Instead of selecting particles directly from the archive forfine-tuning purposes,the assimilation process in Step5)provides a means of integrating swarm and archive information.In contrast to random variation,Step6)exploits knowledge about the possi-ble location of gbest to probe the feasibility of the gbest position along the assigned direction.From the pseudocode,it is clear that the allocation of com-putational resource for LS is determined by S LS.A high setting of S LS allows the swarm to perform LS from different starting points,i.e.,it enables the swarm to exploit along different directions.Likewise,a small setting of S LS will restrict the LS operation.In our algorithm,the value of S LS and N LS is chosen as20and2,respectively.E.ImplementationThe proposed MA incorporating f-gbest and SPLS is named as FMOPSO.Theflowchart of FMOPSO is shown in Fig.3. In every cycle,archiving is performed after the evaluation process of particles.Then,the pbest and gbest of each particle in the swarm are updated.Note that,the pbest of each par-ticle will be updated only if a better pbest is found,and the f-gbest is implemented in place of conventional deterministic gbest.The update of particle position using f-gbest is per-formed concurrently with the SPLS.To maintain a balance between the exploration of“fly”and the exploitation of SPLS, the total number of evaluations is kept at the size of the particle swarm N pop for every cycle.Specifically,S LS par-ticles will undergo the SPLS,while the rest of the N pop−S LS particles in the swarm will be updated with the velocity calculated by(10).IV.B ENCHMARK P ROBLEM ANDFMOPSO P ERFORMANCEIn the context of MO optimization,the benchmark problems must pose sufficient difficulty to impede the ability of MOEA in searching for the Pareto optimal solutions.Deb[3]has identified several characteristics such as multimodality,convex-ity,discontinuity,and nonuniformity,which may challenge the algorithm’s ability to converge or to maintain good population diversity in MO optimization.In this paper,six benchmark problems ZDT1,ZDT4,ZDT6,FON,KUR,and POL are selected to examine the performance of the proposed MOMA. Many researchers such as the authors in[4],[17],[22],and [23]have applied these problems to examine their proposed algorithms.The definition of these test functions is summarized in Table I.The tradeoffs generated by FMOPSO for the different bench-marks in one arbitrary run are shown in Fig.4(10000eval-uations for ZDT1,FON,KUR,and POL,50000evaluations for ZDT4,and20000evaluations for ZDT6).It can be seen that FMOPSO can converge to the true Pareto front and evolve a diverse solution set for all benchmark problems, although the solutions for ZDT1and FON are not evenly distributed along the true Pareto front.Since FMOPSO has found the near-optimal Pareto front and covered the full ex-tent for ZDT1and FON,the SPLS will helpfind more non-dominated solutions on the gaps or discontinuities along the Pareto front,and the dynamic niche sharing scheme will im-prove spacing metric if given more cycles.Fig.4shows that FMOPSO can overcome the difficulties presented by these benchmarks.D EFINITION OF THE T EST PROBLEMSV .E XAMINATION OF N EW F EATURESIn this section,four versions of the algorithm (standard MOPSO,MOPSO with f-gbest only,MOPSO with SPLS only,and FMOPSO)are compared to illustrate the individual and combined effects of the proposed features.The results,with respect to the different performance metric of ZDT1(10000evaluations),ZDT4(50000evaluations),and ZDT6(20000evaluations),are summarized in Tables II,III,and IV,respec-Fig.4.Evolved tradeoffs by FMOPSO for (a)ZDT1,(b)ZDT4,(c)ZDT6,(d)FON,(e)KUR,and (f)POL.TABLE IIP ERFORMANCE OF D IFFERENT F EATURES FORZDT1tively.When there are less than three solutions in the archive,NaN is shown as the computation of spacing metric,which requires more than three solutions.It can be observed from Tables III and IV that the incor-poration of SPLS alone can greatly improve the performance of a standard MOPSO in terms of proximity,diversity,and distribution for ZDT4and ZDT6.The good performance of spacing also shows that the local search ability of SPLS helps the discovery of nondominated solutions on the gaps or dis-continuities along the Pareto front.In Table II,even though SPLS only cannot find the true Pareto front for ZDT1,it can still maintain a relatively good spacing metric.P ERFORMANCE OF D IFFERENT F EATURES FORZDT4TABLE IVP ERFORMANCE OF D IFFERENT F EATURES FORZDT6From Tables II and IV,it can be seen that the implementation of f-gbest alone helps MOPSO to discover the near-optimal Pareto front for ZDT1and ZDT6.However,the f-gbest alone cannot guarantee a good performance on spacing metric.This is expected since the enhancement of global search capability provided by f-gbest is not aimed for the uniform distribution of solutions along the Pareto front.On the other hand,the combination of f-gbest and SPLS allows the discovery of a well-distributed and diverse solution set for ZDT1,ZDT4,and ZDT6without compromising the convergence speed of the algorithm.Table III shows that f-gbest only cannot find the true Pareto front of ZDT4in 50000evaluations.The evolutionary traces shown in Fig.5reinforce such observation.From the evolu-tionary traces,it can be observed that f-gbest only can also lead the search to come closer and closer to the true Pareto front,as shown by the decreasing GD,but the improvement is too slow for ZDT4without the guide of SPLS.Without the balance between the exploration of fuzzy update and the exploitation of SPLS,f-gbest only cannot deal with the multimodality nature of ZDT4effectively.The discontinuity in the spacing metric of f-gbest only is because in those cycles there are less than three nondominated solutions in the archive.Fig.5also ascertains the exploitative nature of SPLS as reflected by the fast convergence speed relative to that without LS (f-gbestonly).Fig.5.Evolutionary trajectories in (a)GD,(b)MS,and (c)S for ZDT4.It can be noted from Table II that SPLS alone cannot over-come the difficulty of ZDT1.Further information can also be extracted by comparing the explored objective space for the algorithms at different timeline,as shown in Fig.6.From Fig.6,it is observed that the explored objective space for FMOPSO is more extensive.By taking into account the uncertainty of the information gathered about the gbest position,the fuzzy updating strategy actually prevents the evolving particles to converge upon similar regions in the search space.Therefore,in the evolution process,the explored space is extended.Standard update does not provide enough diversity for ZDT1.The fast convergence ability of SPLS leads the SPLS only algorithm to stagnate at different local Pareto front.VI.C OMPARATIVE S TUDYIn this section,the performance of FMOPSO is compared to five existing MO algorithms,including CMOPSO [2],SMOPSO [14],IMOEA [20],NSGAII [4],and SPEA2[23].All the algorithms were implemented in C++,and the simula-tions were performed on an Intel Pentium IV 2.8-GHz personal computer.Thirty simulation runs were performed for each algorithm on each test problem in order to study the statistical performance.A random initial population was created for each of the 30runs on each test problem.Note that,the number of evaluations is set to be relatively small to examine the convergence of FMOPSO.The parameter settings and indexes of the different algorithms are shown in Tables V and VI,respectively.Fig.7summarizes the statistical performance of the different algorithms.A.ZDT1From Fig.7(a)–(c),it can be observed that,except FMOPSO,other algorithms still have many solutions that are located far from the true Pareto front.It can be seen that the average performance of FMOPSO is the best among the six algorithms adopted.In addition,FMOPSO is also able to evolve a diverse solution set,as evident from Fig.7(b)and (c).At the same time,we note that the solutions are not evenly distributed alongFig.6.(Top)Explored objective space FMOPSO at generation (a)20,(b)40,(c)60,(d)80,and (e)100and (bottom)SPLS only at generation (f)20,(g)40,(h)60,(i)80,and (j)100for ZDT1.TABLE VP ARAMETER S ETTINGS OF THE D IFFERENT ALGORITHMSTABLE VII NDEXES OF THE D IFFERENT ALGORITHMSthe true Pareto front,as illustrated by a relatively large spacing metric.Since FMOPSO has covered the full extent of the true Pareto front,as illustrated by the high value of MS,and the evaluation number is set to be relatively small for ZDT1(only 10000evaluations),it will improve spacing metric if given more evaluations for the incorporated SPLS and dynamic niche sharing scheme.B.ZDT4It can be observed from Fig.7(d)that CMOPSO,SMOPSO,IMOEA,NSGAII,and SPEA2have a relatively large GD for ZDT4at the end of 50000evaluations.Besides,Fig.7(e)shows that CMOPSO and SMOPSO are unable to evolve a diverse setof solutions consistently.On the other hand,FMOPSO is able to escape the local optima of ZDT4consistently,as reflected by its low value of GD.The FMOPSO is also able to evolve a diverse and well-distributed solution set within 50000evaluations,resulting in a high value of MS and a low value of S .C.ZDT6From Fig.7(g),it can be seen that CMOPSO,NSGAII,and SPEA2failed to find the true Pareto front for ZDT6within 20000evaluations.Although the performance of SMOPSO and IMOEA on GD are better than the aforementioned three algo-rithms,they failed to evolve a diverse and well-distributed so-lution set,as shown in Fig.7(h)and (i).Overall,the FMOPSO is able to evolve a diverse and well-distributed nearly optimal Pareto front for ZDT6within 20000evaluations.D.FONIt can be observed from Fig.7(j)that most of the algorithms are able to find at least part of the true Pareto front for FON.Besides,the PSO paradigm appears to have a slight edge in dealing with the nonlinear tradeoff curve of FON,i.e.,the three PSO-based algorithms outperformed other algorithms consis-tently on GD.It can be seen from Fig.7(k)that CMOPSO and SMOPSO failed to evolve a diverse set of solutions as compared to FMOPSO.This could be due to the fast convergence of CMOPSO and SMOPSO,which may result in the loss of diversity required for covering the entire final Pareto front.E.KURThe disconnection of the Pareto front of KUR seems not a very big problem for the algorithms adopted here,as shown in Fig.7(m)–(o).All algorithms are able to find the nearly optimal Pareto front.However,FMOPSO has the best performance in MS.In addition,FMOPSO also showed competitive perfor-mance in terms of convergence and distribution.Fig.7.Statistical performance of the different algorithms.(a)GD,(b)MS, and(c)S for ZDT1.(d)GD,(e)MS,and(f)S for ZDT4.(g)GD,(h)MS,and (i)S for ZDT6.(j)GD,(k)MS,and(l)S for FON.(m)GD,(n)MS,and(o)S for KUR.(p)GD,(q)MS,and(r)S for POL.F.POLFrom Fig.7(p)–(r),it can be noted that NSGAII has the worst results in terms of convergence,diversity,and distribution.On the other hand,FMOPSO,CMOPSO,SMOPSO,and IMOEA showed competitive performance.PSO paradigm has a slight edge in the aspect of convergence.However,it should be noted that the performance of CMOPSO and SMOPSO in the aspect of S and MS,respectively,are not as good as IMOEA and FMOPSO.VII.C ONCLUSIONA new MOMA has been designed within the context of PSO.In particular,two new features in the form of f-gbest and SPLS have been proposed.The SPLS performs directed local fine-tuning,which helps to discover a well-distributed Pareto front.The f-gbest models the uncertainty associated with the optimality of gbest,thus helping the algorithm to avoid undesir-able premature convergence.The proposed features have been examined to show their individual and combined effects in MO optimization.The comparative study showed that the proposed FMOPSO produced solution sets that are highly competitive in terms of convergence,diversity,and distribution.R EFERENCES[1]E.K.Burke,P.Cowling,and P.De Causmaecker,“A memetic approachto the nurse rostering problem,”Appl.Intell.,vol.15,no.3,pp.199–214, Nov.2001.[2]C.A.C.Coello,G.T.Pulido,and M.S.Lechuga,“Handling multipleobjectives with particle swarm optimization,”IEEE put., vol.3,no.3,pp.256–279,Jun.2004.[3]K.Deb,“Multi-objective genetic algorithms:Problem difficulties andconstruction of test problem,”put.,vol.7,no.3,pp.205–230, 1999.[4]K.Deb,A.Pratap,S.Agarwal,and T.Meyarivan,“A fast and elitistmultiobjective genetic algorithm:NSGA-II,”IEEE put., vol.6,no.2,pp.182–197,Apr.2002.[5]D.Dubois and H.Prade,Possibility Theory:An Approach to Computer-ized Processing and Uncertainty.New York:Plenum,1988.[6]J.E.Fieldsend and S.Singh,“A multi-objective algorithm based uponparticle swarm optimization,an efficient data structure and turbulence,”in Proc.U.K.Workshop Comput.Intell.,Birmingham,U.K.,Sep.2002, pp.37–44.[7]C.M.Fonseca and P.J.Fleming,“Genetic algorithm for multiobjectiveoptimization,formulation,discussion and generalization,”in Proc.5th Int.Conf.Genetic Algorithms,1993,pp.416–423.[8]P.M.Franca,A.Mendes,and P.Moscato,“A memetic algorithm for thetotal tardiness single machine scheduling problem,”Eur.J.Oper.Res., vol.132,no.1,pp.224–242,2001.[9]H.Ishibuchi,T.Yoshida,and T.Murata,“Balance between genetic searchand local search in memetic algorithms for multiobjective permutation flowshop scheduling,”IEEE put.,vol.7,no.2,pp.204–223,Apr.2003.[10]J.Kennedy and R.C.Eberhart,“Particle swarm optimization,”in Proc.IEEE Int.Conf.Neural Netw.,Perth,Australia,1995,pp.1942–1948. [11]E.F.Khor,K.C.Tan,T.H.Lee,and C.K.Goh,“A study on distributionpreservation mechanism in evolutionary multi-objective optimization,”Artif.Intell.Rev.,vol.23,no.1,pp.31–56,Mar.2005.[12]N.Krasnogor and J.E.Smith,“A tutorial for competent memetic algo-rithms:Model,taxonomy and design issues,”IEEE put., vol.9,no.5,pp.474–488,Oct.2005.[13]P.Merz and B.Freisleben,“Fitness landscape analysis and memetic algo-rithms for the quadratic assignment problem,”IEEE put., vol.4,no.4,pp.337–352,Nov.2000.[14]S.Mostaghim and J.Teich,“Strategies forfinding good local guides inmulti-objective particle swarm optimization(MOPSO),”in Proc.IEEE Swarm Intell.Symp.,Indianapolis,IN,Apr.2003,pp.26–33.[15]P.Moscato,“On evolution,search,optimization,GAs and martialarts:Toward memetic algorithm,”Caltech Concurrent Comput.Prog., California Inst.Technol.,Pasadena,Tech.Rep.826,1989.[16]Y.S.Ong and A.J.Keane,“Meta-Lamarckian learning in memetic algo-rithms,”IEEE put.,vol.8,no.2,pp.99–110,Apr.2004.[17]K.E.Parsopoulous and M.N.Vrahatis,“Particle swarm optimizationmethod in multimobjective problems,”in Proc.ACM SAC,Madrid,Spain, 2002,pp.603–607.[18]J.D.Schaffer,“Multiple objective optimization with vector evaluatedgenetic algorithms,”in Proc.1st Int.Conf.Genetic Algorithms:Genetic Algorithms and Their Appl..Mahwah,NJ:Lawrence Erlbaum,1985, pp.93–100.[19]K.J.Shaw,A.L.Notcliffe,M.Thompson,J.Love,C.M.Fonseca,andP.J.Fleming,“Assessing the performance of multiobjective genetic al-gorithms for optimization of batch process scheduling problem,”in Proc.put.,1999,vol.1,pp.37–45.[20]K.C.Tan,T.H.Lee,and E.F.Khor,“Evolutionary algorithms withdynamic population size and local exploration for multiobjective opti-mization,”IEEE put.,vol.5,no.8,pp.565–588,2002.[21]K.C.Tan,E.F.Khor,T.H.Lee,and R.Sathikannan,“An evolutionaryalgorithm with advanced goal and priority specification for multiobjective optimization,”J.Artif.Intell.Res.,vol.18,pp.183–215,2003.。

相关文档
最新文档