ArticlePDF Available

Abstract

To solve the problem that the surrogate is usually limited to a single type, a multi-objective surrogate-based optimization method based on general improvement decomposition strategy is proposed. In this method, the model prediction value information is fully used to construct the general improvement multi-objective decomposition criterion and the general improvement R2 indicator criterion, thus expanding the selection scope of surrogate models in multi-objective surrogate-based optimization methods. The two proposed criteria can achieve an adaptive balance between global exploration and local exploitation with random uniform weights. Comparison results show that the proposed method has good optimization performance under limited simulation conditions, and the Pareto front has certain advantages in convergence, diversity, and spatial distribution. Compared with similar methods, the proposed method has the following advantages: 1) it is suitable for many different surrogate-based optimization methods because the uncertainty of the model prediction is unnecessary, 2) its implementation is simple and computational complexity is low, which can significantly improve the optimization efficiency of expensive black-box problems.
39 6 Vol.39 No.6
20246Control and Decision Jun. 2024
基于广义改进分解策略的多目标代理优化方法
林成龙,马义中,肖甜丽
(南京理工大学 经济管理学院,南京 210094)
:为解决多目标代理优化方法中代理模型选择单一问题,提出基于广义改进函数分解策略的多目标代理
优化方法.该方法充分利用模型预测信息构建广义改进多目标分解准则和广义改进 R2 指标准则,有效拓展多目
标代理优化中代理模型的选择空间.所提两种准则通过随机均匀权重实现全局探索和局部搜索能力的自适应平
.研究结果表明,所提方法在有限仿真条件下拥有良好的寻优性能, Pareto 前沿在收敛性、多样性及空间分
布性方面均具有一定优势.相比同类方法,该方法具有优势: 1) 不需要模型预测不确定性信息,适用于基于不同种
类代理模型的代理优化方法; 2) 实现简单且计算复杂度低,能够有效提升昂贵黑箱问题优化效率.
关键词:昂贵多目标优化;代理模型;多目标代理优化方法;广义改进多目标分解准则;广义改进R2指标准则
中图分类号:N945.12O212.6 文献标志码:A
DOI: 10.13195/j.kzyjc.2022.2106
引用格式:林成龙,马义中,肖甜丽.基于广义改进分解策略的多目标代理优化方法[J]. 控制与决策, 2024, 39(6):
1829-1839.
Multi-objective surrogate-based optimization method based on general
improvement decomposition strategy
LIN Cheng-long, MA Yi-zhong, XIAO Tian-li
(School of Economics and ManagementNanjing University of Science and TechnologyNanjing 210094China)
Abstract: To solve the problem that the surrogate is usually limited to a single type, a multi-objective surrogate-based
optimization method based on general improvement decomposition strategy is proposed. In this method, the model
prediction value information is fully used to construct the general improvement multi-objective decomposition criterion
and the general improvement R2 indicator criterion, thus expanding the selection scope of surrogate models in
multi-objective surrogate-based optimization methods. The two proposed criteria can achieve an adaptive balance
between global exploration and local exploitation with random uniform weights. Comparison results show that the
proposed method has good optimization performance under limited simulation conditions, and the Pareto front has
certain advantages in convergence, diversity, and spatial distribution. Compared with similar methods, the proposed
method has the following advantages: 1) it is suitable for many different surrogate-based optimization methods because
the uncertainty of the model prediction is unnecessary, 2) its implementation is simple and computational complexity is
low, which can significantly improve the optimization efficiency of expensive black-box problems.
Keywords: expensive multi-objective optimizationsurrogate modelmulti-objective surrogate-based optimization
methodgeneral improvement multi-objective decomposition criteriongeneral improvement R2 indicator criterion
0
多目标优化问题(multi-objective optimization
problems, Mops) 在现代工程中广泛存在,但目标间的
冲突性使得求解困难[1-2].随着计算机技术的发展,
精度仿真试验取代实物试验成为主流,但其需要耗
费大量的时间.例如:翼型优化设计时, 1 000 次仿真
需要 6个月[3] ;福特汽车碰撞试验单次仿真需要36
120 小时, 1 000 次试 验将超过 4[4] .此外,
的封闭性和复杂的底层代码使得输入和输出形成黑
箱关系,这使得传统进化算法不再适用[1-4].这类涉及
黑箱且 具有耗时 特性的 Mops, 被称为昂贵多目标
化问题.从贝叶斯优化观点看,自适应试验设计策略
可进一步丰富样本信息,对于弥补数据缺陷具有重要
意义[2-3,5].
多目标代理优化(multi-objective surrogate-based
optimization, MSBO) 方法中常用的代理模型有多项
收稿日期:2022-12-05录用日期:2023-04-03.
基金项目:国家自然科学基金项目(71931006, 71871119, 72171117).
通讯作者. E-mail: yzma-2004@163.com.
1830 39
式响 应面 (polynomial response surface, PRS)[2]
项式混沌展开(polynomial chaos expansions, PCE)[5]
最小二乘支持向量回归(least square support vector
regression, LSSVR)[6] 径向基函数 (radial basic
function, RBF)[7] Kriging 模型[1-4] ., Kriging
模型具备对未试验点预测不确定性度量能力,大多
数自适应试验设计准则基于该特性进行构造.例如,
Jones [8] 构造了期望改进(expected improvement,
EI) ,并将其应用于无约束黑箱优化问题;
Schonlau [9] 开发了广义EI准则,及适用于约束优化
问题 的约束 EI (constrained EI, CEI) 准则; Alexandrov
[10] 通过调整参数阈值,利用预测统计特性构造了
较低置信下界(lower confidence bound, LCB)
.上述 准则难以 直接应 用于 Mops, 开发 用于
MSBO方法的试验设计准则十分有必要.
MSBO 研究主要分两类.一是将Mops 转换为单
,标准 .
: Knowels[11] (weighted sum
approach) 和切比雪夫方法 (Tchebycheff approach)
混合来实现Mops 转换的混合 ParEGO ; Zhang
[12] Mops 分解 (MOEA decomposition, MOEA/D)
为多个子问题, 出的基 MOEA/D 的高效全局
(efficient global optimization based MOEA/D, M/D-
EGO) 方法.二是直接建立多目标模型,匹配多目标
. : Svenson [13]
(Monte Carlo, MC) 近似计算构造的最大最小距离
(maximin distance EI, EIm) ; Hupkens
[14] Couckuyt [15] 构造的快速期望改进超立
(expected hypervolume improvement, EHVI) ;
Zhan [16] 通过线性加和构建的期望改进矩阵
(expected improvement matrix, EIM) ; Han [17]
基于 EI 准则构造 的期望改进 R2 指标 (R2 indicator of
EI, EIR2) 准则.此外, Chugh [18] 构造 Kriging
参考向量引导的进化算法(Kriging assisted reference
vector guided evolutionary algorithm, KRVEA)
参考向量引导快速收敛,但在两目标Mops 中效果较
.目前 MSBO 存在如下问题: 1) 多目标准则积分计
算产生新的耗时问题[1,11-13]; 2) 变量数多或大样本时,
Kriging 建模引发的维度魔咒[1-2,4].因此,如何提升设
计准则的计算效率和Kriging建模效率是值得研究的
问题.
对于准则耗时问题, Zhan [16] 所提 EIM 准则使
应用性大大增强,但仍需预测不确定信息; Zhou[19]
基于 SVR 模型进行自适应设计,但仿真过程过于耗
.,本文结合切比雪夫方法提出基于广义改
进分解策略的多目标代理优化(MSBO based general
improvement decomposition strategy, GIDS-MSBO)
. GIDS-MSBO 方法适用于不同建模方法,且设
准则仅需预测值及非支配前沿点即可,具有复杂度
低、用性强的优.主要创新如: 1) 构建广义改
进多目标分解(general improvement multi-objective
decomposition, GIMD) 广 R2
(general improvement R2 indicator, GIR2) 准则,实现高
效的空间探索; 2) 基于 GIMD GIR2 准则和不同代
理模型,构建适用于昂贵黑箱问题的多种MSBO
; 3) 所构造广义改进自适应试验设计准则实现简
,计算复杂度,有助 MSBO 方法的广泛应用及
推广.
1基础理论知识
1.1 昂贵多目标优化问题
昂贵多目标优化问题[1,12] 可以表示为
min {f1(x), . . . , fm(x)}.
s.t. gi(x)0, i = 1,2, . . . , r;
xlxxr,xRd.(1)
其中:fm(x) m个目标,m2;gi(x) i个约
;x为输入变量;Rdd维实数空间;xlxr为左、
边界向量.
1[11,16] xx,f(x)/=f(x)
, Mops 满足 fi(x)fi(x), 称点 x 配点 x,
x<x.x组成 的元 素集 合称 Pareto 优化
解集 (Pareto optimal set, PoS). PoS 中元素x所对应的
标向 f(x)组成的集合称为Pareto 沿 (Pareto
frontier, PF), PF ={f(x)|xPoS}. PF 中的点称
为非支配前沿点.
1.2 多目标评价指标
超体积 (hypervolume, HV) 及反转世代距离
(inverted generational distance, IGD) 是度量Pareto
沿近似效果的综合评价指标[12,16],其形式如下:
HV(PF,RP) = Le(|S|
i=1
Vi),(2)
IGD(PF,PF) =
vPF
min d(v, PF)
|PF|.(3)
其中: Le Lebesgue 测度,|S|表示非支配点数,Vi
表示 i个非支配 前沿点 与参考 (reference point,
RP) 构成的超体积; PFPF分别表示优化获得的
Pareto 前沿和真实Pareto 前沿, min d(v , PF)表示 PF
中点vPF中点的最小欧氏距离.
6林成龙 :基于广义改进分解策略的多目标代理优化方法 1831
1.3 Kriging模型
给定数据= [X,y], Kriging[1-2] 假设存在关系
y(x) = µ+z(x).(4)
其中:µ全局 势项;z(x) E[z(x)] = 0,协方
Cov(z(xi), z(xj)) = σ2
zR(xi,xj|θ)的高斯随机过
.参数µσ2
z的估计为
ˆµ= (1TR11)11TR1y,(5)
ˆ
σ2
z=1
n(y1ˆµ)TR1(y1ˆµ).(6)
其中:R= [R(xi,xj|θ)]i,j ,1i, j n,1为由元素 1
构成的列向量.预测均值µˆy及方差s2
ˆy
µˆy(x) = ˆµ+r(x)R1(y1ˆµ),
s2
ˆy(x) = ˆσ2
z[1r(x)TR1r(x) + (1 1TR1r)2
1TR11].
1.4 期望改进准则
Jones [8-9] 假定响应Y(x)N(µˆy, s2
ˆy),Y(x)
相比最小值ymin 的改进I(x) = max(0, ymin Y(x)),
I(x)求期望得
EI(x) = sˆy[τ(x)Φ[τ(x)] + ϕ[τ(x)]], sˆy0.(7)
其中:τ(x)=(ymin µˆy(x))/sˆy(x);Φ(·)ϕ(·)
准正态分布的累积分布和概率密度函数.由式 (7)
, EI 准则计算需预测不确定信息,故仅能提供预测
值的代理模型无法直接使用EI准则.
2基于广义改进分解策略的MSBO方法
2.1 广义改进多目标分解策略
切比雪夫法[11-12] 是非线性多目标聚合方法,
给定权重偏好及参考点信息下进行聚合,其表示为
gtc(x|λ) = max{λr
i|fiz
i|}.(8)
其中:λ= [λr
1, λr
2, . . . , λr
m]T为均匀权重向量,且满足
m
i=1
λr
i= 1, λr
i0;z
i=min{fi(x)|x}(i=
1,2, . . . , m)Mops中第i个目标的最小值.
对于非支配点(f2
1, f 2
2)及目标响应y= [y1,y2]T,
易知 f2
1min y1, f2
2min y2.(f2
1, f 2
2)构造
进均值 函数 I(x)[4],I1(x) = max(f2
1ˆy1(x),0)
I2(x) = max(f2
2ˆy2(x),0). Kriging 型的插值
性使得(f2
1, f 2
2)相对(ˆy1,ˆy2)必存在改进.对于其他模
,建模的回归性质使样本点位于曲线或曲线两侧.
因此,若存在k个非支配点,可得广义改进函数向量
GI(x) = [GIj
1(x),GIj
2(x), . . . , GIj
m(x)]T.(9)
其中: GIj
i(x) = max(fj
iˆyi(x),0),fj
i(j= 1,2, . . . , k)
为第 i个目标的第 j个非支配点,ˆy(x)为预测,采用
Kriging建模时为预测均值.
非支配点对应的改进函数I1(x)I2(x)对应广
义改进(general improvement, GI) 坐标系 (GI1, GI2).
对坐标系内改进Mops进行切比雪夫分解得
GItc(x|λ) = m
max
i=1 {λr
i[GIj
iPF
i]}, j = 1,2, . . . , k.
其中:λr
i表示目标分解赋予第i目标的随 权向,
m
i=1
λr
i= 1; PF
i=min fj
i, fj
iPF. 选取当前 kGI
坐标系提供的改进最小的切比雪夫方向进行优化,
GIMD准则可写为
GIMD(x, λ, PF) = |PF|
min
j=1 GItc(x|λ).(10)
2.2 广义改进R2指标
R2 指标是评价候选解集综合性能的评价指
[17].在给定参考点集 c、效用函数 u(x)及效用函
数对应的权重向量λ, R2 写为
R2(U, PF) = 1
|U|
|U|
i=1
max
xPF{u(x)},(11)
其中|U|是效用集中U的元素数.
构建广义改进R2指标函数
GIR2(x, λ, PF) =
1
|U|
|U|
k=1
|PF|
min
j=1
m
max
i=1 {λk
i(GIj
i(x)PF
i)},
其中 λk
i是第 k个权均匀分量在第 i个非支配点处的
权系数.
2.3 PF近似度量及终止准则
Binois [20] 基于条件帕累托集 (conditional
Pareto front set, CPFs)和模型后验信息对真实 PF 进行
估计,并构造 Vorob’ev 期望和方差度量其不确定性.
CPFs是一种随机模拟方法,其可写为
YCS(x) = ˆ
Y(x)+[S(x)ˆ
S(x)].(12)
其中:ˆ
Y(x)ˆ
S(x)为依试验数据和由非条件仿真数
据建立 Kriging 模型;YCS(x)S(x)分别为条件模
拟的黑箱函数和依后验信息仿真的黑箱函数. MSBO
方法中,样本更新使得后验信息更准确,故可采用条
件估 计的近似 PF 度量收敛性. Zhang [21] 构造相
HV (relative HV, RHV) 值评 MSBO 方法的收
,其表达式为
RHV = 1 HVest
HVref
.(13)
其中: HVest 优化解 集的 HV ; HVref CPFs 计算
HV . RHV 越小,表明 MSBO 方法的综合性能
越好.试验过程中,选取RHV ε作为终止条件,设定
阈值ε= 0.01.
1832 39
2.4 计算复杂度比较
MSBO 方法中的经典多目标准则总结见表1.
1可知, EHVI EIm准则均需m重积分,计算困.
EIM准则和 EIR2 准则仅需 1维积分,可有效提升计算
效率,但依赖预测不确定信息;所提 GIR2 GIMD
则不需积分和预测不确定信息,可适用于多种代理模
.选取表2中函数进行验证,设定输入d= 6,训练样
本数n= 10d.计算EIm准则和GIR2准则在非支配前
沿点k= 1002005001 000时的运算时间,见图1.
1不同准则计算复杂度比较
方法 准则 代理模型 积分 计算方法
文献 [15] EHVI Kriging mWFG方法
文献 [13] EImKriging mm= 2,解析
m > 2, MC
文献 [16] EIM Kriging 1 线性聚合
文献 [17] EIR2 Kriging 1 R2 聚合
本文 GIMDGIR2 KrigingSVR... 0 TCR2 聚合
2测试函数信息
TF m区间 dRF 特征
Z1 2 [0,1]d6 (11, 11)
Z2 2 [0,1]d6 (11, 11) 非凸
Z3 2 [0,1]d6 (11, 11) 凸且不连续
D2 3 [0,1]d6 (2.5, 2.5, 2.5) 非凸
D5 3 [0,1]d6 (2.5, 2.5 ,2.5) 非凸且退化
D7 3 [0,1]d6 (40, 40, 40) 混合、不连续、多模态
4 8 100
0
2
4
6
8
10
t/ (1 0 s)
3
2 6
非支配前沿点个数/ 1 02
DTL Z2
DTL Z5
DTL Z7
ZDT 1
ZDT 2
ZDT 3
4 8 100
0.5
2.5
t/ s
2 6
(b) GIR2准则
1.0
1.5
2.0
(a) EIm准则
非支配前沿点个数/ 1 02
DTLZ2
DTLZ5
DTLZ7
ZDT 1
ZDT 2
ZDT 3
1非支配前沿点数量与计算时间关系
1表明,采用解析 (m= 2) 和蒙特卡洛数值近
(m > 2) 计算的 EIm准则计算时间均远多于所提
GIR2 准则,GIMD 准则计算则更为简单.因此,
GIMDGIR2 有助于进一步提升计算效率,并有利
于更多代理模型在MSBO方法中的使用.
2.5 多目标GIDS-MSBO方法
MSBO方法的实现步骤如下.
step 1: 始化 (设定). 始样数为 n=
10d+ 1;终止条件,最大迭代次数T=n+ 100,初始
迭代Iteration = 1.
step 2: .
(maximin Latin hypercube sampling, mLHS) 得到设计
X,功能性评估获得响应Y.
step 3: 模型构建/刷新.构建/刷新代理模型.
step 4: 信息获取.依据代理模型获取m个目标
的自适应信息 ˆy及非支配前沿PF.
step 5: . ,
GIMD/GIR2策略引导后续序贯填充.
step 6: .满足终止条件,输出 PoS PF, 转入
step 7; 否则, 大化 GIR2/GIMD 获取单个新设计点,
Iteration =Iteration + 1,转入step 2.
step 7: 策分.根据决策者偏好,设定目标权
,对获取的PF进行决策分析.
step 8: 方法终止.输出最优设计参数.
3实验与结果分析
3.1 测试函数信息
选取 ZDT (ZT) DTLZ (DZ) 系列函数作为测试
函数[11,16],特征信息见表 2.
3.2 实验参数设定
初始实验设计,采用 mLHS 抽取 n= 10d
.选择显著性水平α= 0.05 Wilcoxon 秩和检验
方法进行评价.符号分别表示统计结果
优于、无显著差别及劣于参考方法.GIDS-MSBO
框架中,选取 LSSVRKrigingRBFPCE PRS 作为
代理模型,参数设定见表3.
3代理模型构建信息
代理模型 建模特征 程序实现
PRS 2 阶多项式 自编
Kriging corrgauss, reploy0 DACE 程序包[22]
RBF 高斯基函数 RBF 程序包[7]
PCE 正交基 Lengegare, 阶数 p= 4 PCE 程序包[5]
LSSVR 二次损失函数,径向基函数 LSSVR 程序包[6]
3.3 代理模型比较
将表3中模型与 GIMDGIR2 准则结,并对表2
中函数优化.统计 30 次独 立随机实验 HVIGD 的均
值和方差.Kriging +GIR2Kriging +GIMD 组合
作为比较基础,在显著性水平α= 0.05 条件下进行
Wilcoxon秩和检验,结果见表 4及表 5.
6林成龙 :基于广义改进分解策略的多目标代理优化方法 1833
4不同代理建模方法作用下GIR2-MSBO方法优化结果比较
tests indicators LSSVR RBF PRS PCE Kriging
ZT1
HV 120.149 (0.482) 120.306 (0.532) 120.543 (0.028)120.145 (0.269) 120.505 (0.241)
IGD (1e-4) 21.901 (6.793) 28.536 (15.343) 35.159 (6.503) 26.020 (4.721)20.738 (8.017)
ZT2
HV 120.227 (0.026) 118.333 (3.824) 120.233 (0.026)120.234 (0.022) 120.236 (0.037)
IGD (1e-4) 60.946 (11.196)65.883 (73.982) 56.338 (13.013) 58.634 (8.976)50.379 (19.358)
ZT3
HV 126.371 (2.822) 119.317 (3.931) 126.920 (2.743) 121.788 (2.625) 127.423 (2.408)
IGD (1e-4) 51.914(18.827) 113.209(34.307) 58.255 (15.762) 68.352 (24.092) 29.255 (16.554)
DT2
HV 14.781 (0.024)14.780 (0.028) 14.421 (0.056) 14.699 (0.039) 14.759 (0.048)
IGD (1e-4) 24.027 (1.577) 23.064 (1.697) 23.630 (1.142)24.488 (1.949) 21.758 (1.452)
DT5
HV 12.884 (0.050) 12.923 (0.038) 12.501 (0.086) 12.924 (0.035)12.895 (0.071)
IGD (1e-4) 24.753 (5.206) 22.484 (4.059)26.546 (3.164) 23.887 (4.540) 18.737 (4.120)
DT7
HV (1e+4) 5.863 (0.016) 5.732 (0.126) 5.858 (0.006) 5.853 (0.013) 5.907 (0.092)
IGD (1e-4) 34.491 (7.346)52.628 (16.241) 37.039 (9.177) 37.785 (5.353) 33.455 (13.721)
rank //0 / 5 / 7 0 / 5 / 7 1 / 2 / 9 0 / 3 / 9
5不同代理建模方法作用下的GIMD-MSBO方法统计值比较
tests indicators LSSVR RBF PRS PCE Kriging
ZT1
HV 120.243 (0.381) 120.294 (0.463) 120.595 (0.016)120.235 (0.304) 120.608 (0.105)
IGD (1e-4) 13.541 (3.912) 29.412 (9.885) 22.512 (4.979) 12.746 (2.905) 9.972 (2.297)
ZT2
HV 120.300 (0.006)120.264 (0.143) 120.260 (0.010) 120.287 (0.013) 120.293 (0.011)
IGD (1e-4) 13.186 (2.551)14.250 (3.695) 25.589 (3.147) 16.195 (4.488) 17.408 (4.739)
ZT3
HV 124.630 (2.797) 120.826 (3.898) 124.220 (2.682)122.710 (3.041) 127.344 (3.116)
IGD (1e-4) 47.215 (16.107) 83.451 (33.073) 50.121 (12.123)55.006 (22.493) 27.361 (29.887)
DZ2
HV 14.907 (0.017) 14.949 (0.107) 14.410 (0.057) 14.896 (0.017) 14.951 (0.011)
IGD (1e-4) 17.948 (1.553) 14.632 (1.065) 23.431 (1.071) 19.348 (1.713) 13.163 (0.879)
DZ5
HV 13.048 (0.018) 13.090 (0.009)12.488 (0.073) 12.988 (0.025) 13.095 (0.013)
IGD (1e-4) 12.251 (1.808) 8.137 (0.994)27.098 (3.289) 17.235 (3.797) 7.244 (1.269)
DZ7
HV (1e+4) 5.864 (0.065) 5.812 (0.074) 5.867 (0.050)5.792 (0.054) 5.959 (0.082)
IGD (1e-4) 25.322 (4.459)28.842 (8.398) 24.729 (4.460) 32.207 (5.408) 19.107 (8.959)
rank //2 / 0 / 10 1 / 1 / 10 0 / 0 /12 0 / 2 / 10
4 : Kriging +GIR2 24
11 ,表明该组合优化性能最好; LSSVR +
GIR2 RBF +GIR2 组合均有5次与 Kriging +GIR2
组合无显著差异, LSSVR RBF
GIR2 作用下的性能弱于Kriging , PRS
PCE模型.
由表 5可知: Kriging +GIMD 获取最优值个数
,表明该组合是解决昂贵多目标优化问题的首选;
LSSVR +GIMD 组合在解决非凸的ZDT2 问题性能
,RBF +GIMD 策略 稳健优于 Kriging +
GIMD 组合,表明非凸问题中 LSSVR 模型和 RBF
型更优.
对比表 4和表 5: Kriging 模型作用下的结果
最优,其次为 LSSVR ,最差结果为 PRS .
ZDT3DTLZ7 测试问题, GIMD-MSBO 方法获取
HV IGD 统计 值均弱 GIR2-MSBO 方法,其余
GIMD-MSBO 方法更具优势, GIR2-
MSBO 方法更适合离散Pareto 前沿问题,Pareto
沿较为连贯时应优先考虑GIMD-MSBO方法.
选取 KrigingLSSVRPCERBF PRS 进行建
,并在 GIMD 准则作用下优化 6种测试函数. 30
独立实验HV平均值的迭代收敛曲线见图2.
1834 39
40 80 1200
HV
20 60
100
105
110
115
120
125
100
迭代次数
(a) ZDT 1
40 80 1200
HV
20 60
90
100
110
120
100
迭代次数
(b) ZDT 2
115
120
10 2 0 30
40 80 1200
HV
20 60
100
110
120
130
100
迭代次数
(c) ZDT 3
40 80 1200
HV
20 60
13.8
15.0
100
迭代次数
(d) DTL Z2
14.2
14.6
40 80 1200
HV
20 60
12.2 10 0
迭代次数
(e) DTL Z5
12.4
12.6
12.8
13.0
13.2
40 80 1200
HV / 10 4
20 60
4.8 100
迭代次数
(f) DTL Z7
5.2
5.6
6.0
LSSVR P CE P RSRBF
Krigi ng
118
120
515 25
2GIMD-MSBO方法中HV指标的变化趋势
2: ZDT1ZDT2 优化中,Kriging
LSSVR 作用下收敛速度最快,精度相;对于 ZDT3
问题, Kriging 收敛速度较慢,但获取 HV 最高,
明其寻优能力更;对比 DTLZ 结果可知, Kriging
RBF DTLZ2 DTLZ5 中的优化结果和收敛速度
致相 , LSSVR PCE , PRS ;对于
DTLZ7 问题, Kriging 的效果最好,但收敛速度差别不
.
3.4 不同MSBO方法比较
3.4.1 优化结果比较
T= 100 + n为终止条件,选取表现最差的
PRS +GIR2 (L) 和最好的Kriging +GIMD (B) 合作
为参考,ParEGO 方法、M/D-EGO 法、EIM-EGO
方法、KRVEA方法及 EIR2-EGO 方法进行比较. 30
独立实验的HVIGD指标结果见表6.
6不同MSBO方法与所提MSBO方法比较
tests indicators (L)ParEGO (B) (L)M/D-EGO (B) (L)EIM-EGO (B) (L)KRVEA (B) (L)EIR2-EGO (B)
ZT1
HV 120.620 (0.020) 119.660 (0.779) 120.631 (0.009) 117.091 (0.997) 120.568(0.030)
IGD (1e-4) 9.748 (1.555) 34.411 (14.674) 11.309 (3.077) 83.330 (25.162) 29.000 (8.140)
ZT2
HV 120.268 (0.103) 116.928 (2.963) 120.204 (0.171) 111.938 (1.437) 120.061 (0.269)
IGD (1e-4) 14.608 (9.547) 60.432 (34.107) 41.994 (16.921) 127.155 (30.522) 88.000 (12.000)
ZT3
HV 127.503 (1.088) 123.218 (4.142) 127.646 (1.466) 109.882 (4.099) 127.034 (1.632)
IGD (1e-4) 21.153 (7.045) 63.526 (28.153) 16.997 (8.611) 202.738 (41.333) 33.000 (14.000)
DZ2
HV 14.806 (0.027) 14.684 (0.050) 14.983 (0.006) 14.988 (0.102) 14.516 (0.160)
IGD (1e-4) 18.496 (1.199) 22.422 (1.847) 11.166 (7.634) 9.913 (0.994) 23.000 (1.733)
DZ5
HV 12.983 (0.028) 12.732 (0.061) 13.062 (0.015) 13.071 (0.113) 12.715 (0.065)
IGD (1e-4) 11.589 (1.520) 26.319 (3.931) 9.806 (1.662) 6.785 (0.783) 26.963 (4.485)
DZ7
HV (1e+3) 59.720 (0.003) 59.146 (0.409) 59.531 (0.247) 59.706 (0.001) 59.316 (0.534)
IGD (1e-4) 16.786 (1.678) 17.659 (3.601) 13.387 (1.602) 11.440 (1.352) 41.000 (13.000)
rank //11 / 0 / 1 5 / 0 / 7 3 / 3 / 6 1 / 0 / 11 10 / 1 / 1 6 / 0 /6 6 / 0 / 6 5 / 0 / 7 4 / 1 / 7 0 / 0 /12
由表 6 : ZDT 函数, PRS+GIR2 方法
M/D-EGO 方法、KRVEA 方法 EIR2-EGO 方法;
对于 DTLZ 系列函数, KRVEA 方法表现最优, Kriging
+GIMD 方法与EIM-EGO 方法结果相当,其余方法
均劣于 Kriging +GIMD .上述结果表明,提两
种策略具有广泛的应用潜.选取 30 次随机实验
6林成龙 :基于广义改进分解策略的多目标代理优化方法 1835
(a) RRS + GIR2 (b) EIM-E GO (c) K ri ging + GI MD
0.4 0.8
0
f2
0.2 0.6
0
0.2
0.4
0.6
0.8
1.0
1.0
ZDT1
f1
rea l PF
opt imal PF
0.4 0.8
0
f2
0.2 0.6
0
0.2
0.4
0.6
0.8
1.0
1.0
ZDT2
f1
rea l PF
opt imal PF
0.4 0.8
0
f2
0.2 0.6
-0.5
0
0.5
1.0
1.5
1.0
ZDT3
f1
-1.0
rea l PF
opt imal PF
0.4 0.8
0
f2
0.2 0.6
0
0.2
0.4
0.6
0.8
1.0
1.0
ZDT3
f1
rea l PF
opt imal PF
0.4 0.8
0
f2
0.2 0.6
-0.5
0
0.5
1.0
1.0
ZDT3
f1
-1.0
rea l PF
opt imal PF
1.0
0.4 0.8
0
f2
0.2 0.6
0
0.2
0.4
0.6
0.8
1.0
ZDT1
f1
rea l PF
opt imal PF
0.4 0.8
0
f2
0.2 0.6
0
0.2
0.4
0.6
0.8
1.0
1.0
ZDT2
f1
rea l PF
opt imal PF
0.4 0.8
0
f2
0.2 0.6
0
0.2
0.4
0.6
0.8
1.0
1.0
ZDT1
f1
rea l PF
opt imal PF
0.4 0.8
0
f2
0.2 0.6
0
0.2
0.4
0.6
0.8
1.0
1.0
ZDT2
f1
rea l PF
opt imal PF
3最大HV值对应的Pareto近似前沿
1.5
(a) PRS + GIR2 (b) EIM-E GO (c) Kri ging + GI MD
f2
f1
f3
1.5
1.0
0.5
0
0.5 1.0
000.5 1.0
DTLZ2
f2
f1
f3
1.5
1.0
0.5
0
DTLZ2
0.5 1.0
0
00.5 1.0
f2
f1
f3
1.5
1.0
0.5
0
DTLZ2
0.5 1.0
000.5 1.0
f2
f1
f3
1.5
1.0
0.5
0
0.5 1.0
0
1.5 1.0
0
0.5
DTLZ5
f2
f1
f3
1.0
0.6
0.4
0
0.2 0.6
00
0.5
DTLZ5
0.8
0.2
0.4 f2
f3
1.0
0.5
0
0.2 0.6
0
DTLZ5
0.4 f1
0
0.5
f2f1
f3
0
0.5
1.0 1.0
0
0.5
DTLZ7
2
4
6
f2f1
f3
1.0
0.5
00
1.0
0.6
2
6
10
8
4
0.20.4
0.8
DTLZ7
f2f1
f3
0.5
00
0.5
DTLZ7
2
4
1.0
1.0
6
4最大HV值对应的Pareto近似前沿
1836 39
HV最大值时的PRS+GIR2Kriging +GIMD EIM-
EGO方法优化结果绘制优化Pareto前沿(optimal PF),
并与真实Pareto前沿(real PF)进行比较,结果见图3.
3: ZDT1 ZDT2 , 3 MSBO
方法均获得了理想的PF, 表明 GIDS-MSBO 方法具
有良好的收敛性及广泛的适用性.对于 ZDT3 问题,
EIM-EGO 方法 果最 , Kriging +GIMD 果与
, PRS +GIR2 策略 现较 . 面选 30
实验中DTLZ 系列函数最大HV 值时的PRS +GIR2
Kriging +GIMD EIM-EGO 方法优化结果绘制
Pareto 前沿,见图 4. :蓝点表 real PF, 红点表示
optimal PF.
由图 4: PRS+GIR2 结果与真实Pareto 前沿
具有较大差距,表明该方法有效但收敛速度较慢;
Kriging +GIMD 方法获取PF 与真实前沿十分接近,
表明所提GIMS-MSBO 方法具有良好的优化能力和
收敛性. 合表 5 6的结果可知,同代理模型与
GIMDGIR2 准则组合,均可实现复杂Mops 的求解,
且具有良的综合性.此外, EIM-EGO 方法,
所提 GIDS-MSBO 方法不需要积分计算且适用于不
同代理模型,具有更强的应用潜力.
3.4.2 优化成本比较
设定 RHV 0.01 为终止条件,选取 EImEIR2
EIM 准则与所提GIMDGIR2 准则优化表 2中函.
Kriging LSSVR 作为代理模型,GIMD GIR2
准则进行结合.统计优化终止时评价次数的均值和
标准差,见表 7. 数据加粗表示最优结果,大于 100 (>
100)表示最大迭代次数超过100.
7不同准则作用下评价次数的均值(标准差)比较
测试函数
Kriging 模型 LSSVR 模型
EImEIR2 EIM GIMD GIR2 GIMD GIR2
ZDT1 9.47 (14.13) 9.10 (6.74) 8.90 (6.11) 2.73 (1.57) 7.70 (5.87) 6.63(18.17) 25.67 (40.64)
ZDT2 13.87 (19.53) 36.10 (20.65) 24.40 (26.47) 2.47 (1.70) 14.17 (17.80) 2.13 (0.43) 6.23 (3.91)
ZDT3 65.30 (31.51) 78.77 (26.64) 63.60 (29.64) 75.73 (29.19) 60.03 (28.37) 75.73 (29.18) 53.10(40.34)
DTLZ2 49.47 (8.45) >100 45.27 (8.32) 70.60 (10.30) 80.03 (26.73) 36.93 (15.18) 77.36 (22.80)
DTLZ5 85.97 (15.13) > 100 80.07 (14.51) 57.23 (13.97) > 100 91.33 (11.48) > 100
DTLZ7 49.40 (39.68) 52.43 (33.93) 36.80 (31.21) 49.77 (40.00) 37.30 (39.30) 85.83 (31.00) 98.47 (13.88)
由表7对比可知,所提 GIMDGIR2 准则获取了6
个最优值中的 5, EIM 准则仅获得DTLZ7函数的最
优均值,EIM 准则结果与Kriging 建模下的GIR2
则差异很小.因此,所提两种准则具有广义性,且具有
计算成本更低的优势.
3.5 案例及结果分析
3.5.1 谈判问题
慕昆 [23] ,以买方效用 ub(x)
方效用 us(x)目标,将采购所涉及的关键属性价格
x1(万元)保修期 x2() 发货期 x3()作为输,
并设定企业对谈判款项的偏好情况,选取最大化自身
效用的供应商进行交易,具体信息见表8.
8买卖双方基本参数设置
谈判属性
买方 ub卖方 us
wb
j议价 vxj
bws
j议价 vxj
s
价格 x10.4 [1,8] 8x1
810.7 [3,10] x110
10 3
保修期 x20.3 [12,36] x212
36 12 0.2 [6,24] 24 x2
24 6
决策期 x30.3 [1,7] 7x3
710.1 [3,14] x33
14 3
虑谈 ,设定 目标 权重 w1=w2
进行决策.Kriging (KG) LSSVR (LR) 作为 GIR2-
MSBO GIMD-MSBO 方法的代理模型,EIM-
EGO ParEGO 方法.绘制 30 实验最大
用时的散点图,见图5,其中文献[23]结果为参考线.
5表明:不同 MSBO 方法最优结果均可实现谈
判问题的高效优化; GIR2-MSBO GIMD-MSBO
法均可实现对谈判问题的高效求解,表明所提方法具
有广泛的适用性及优异的综合性.选取 30 次独
实验中获取的最优决策方案进行比较,结果见表9.
9不同MSBO方法获取谈判问题设计优化结果对比
MSBO 优化解 效用 ub效用 usHV
文献 [23] (7.70, 23.93, 3.09) 0.361 6 0.471 8
ParEGO (8.00, 24.00, 3.00) 0.350 0 0.500 0 2.58 (4.00e-3)
EIM-EGO (8.00, 12.00, 3.25) 0.350 0 0.500 0 2.59 (9.91e-5)
LR+GIR2 (8.00,24.00, 3.00) 0.350 0 0.500 0 2.59 (2.07e-4)
KG+GIR2 (8.00, 23.73, 3.00) 0.346 6 0.503 0 2.59 (1.60e-2)
LR+GIMD (7.99, 24.00, 3.00) 0.350 4 0.499 3 2.59 (1.83e-5)
KG+GIMD (8.00, 23.81, 3.00) 0.347 6 0.502 2 2.59 (2.04e-5)
6林成龙 :基于广义改进分解策略的多目标代理优化方法 1837
10 20
515
0
0.2
0.4
0.6
0.8
1.0
30
(a) LSS VR + G IR2
仿真次数
效用
仿真次数
25 10 20
515
0
0.2
0.4
0.6
0.8
1.0
30
效用
25
仿真次数
10 20
515
0
0.2
0.4
0.6
0.8
1.0
30
效用
25
买方参考效用
卖方参考效用
买方效用
卖方效用
(b) LSS VR + G IMD (c) EIM -E GO
10 20
515
0
0.2
0.4
0.6
0.8
1.0
30
(d) Kri gi ng + GIR2
仿真次数
效用
25 10 20
515
0
0.2
0.4
0.6
0.8
1.0
30
(e) Kri gi ng + GIMD
仿真次数
效用
25 10 20
515
0
0.2
0.4
0.6
0.8
1.0
30
(f) Par EG O
仿真次数
效用
25
5不同MSBO方法获得的最优结果
由表9可知,上述MSBO 方法均可实现高效优化.
与曹慕昆等[23] 结果相比略差的原因: 1) NSGA-II
法进行了2 000 次实验; 2) 曹慕昆等[23] 构建让步策略,
且谈判在优化前沿的基础上进行了15 ,而所提方
法在假设买卖双方平等的基础上进行决策.
3.5.2 四杆桁架问题(four bar truss problem, FBTP)
四杆桁架设计问题[24] 将节点 A1A2A3A4
横截面积看作设计变量,要求同时最小化该桁架结构
设计问题节点的体积和位移.
6GIMD +Kriging (GM-Krig) 组合在RHV
0.01 , FBTP 案例后验信息的CPFs. 由图 6,
GM-Kriging获取 CPFs 具有良好的 PF 近似能力,故采
RHV 作为收敛判定准则.统计终止条件满足时
所需评估次数的均值和标准差,并设定目标权重为
w1=w2= 0.5,采用最大化目标 f=w1f1s+w2f2s
进行决策,结果见表10.
f1/ 10 3
f2
-0.01
0.01
0.03
0
0.02
0.04
0.05
0.06
1.2 1.61.4 1.8 2. 0 2. 2 2. 4 2. 6
CPF s1
CPF s2
CPF s3
CPF s4
CPF s5
CPF s6
CPF s7
CPFs8
CPFs9
CPFs1 0
CPFs11
CPFs1 2
CPFs1 3
CPFs1 4
CPF s1 5
CPF s1 6
CPF s1 7
CPF s1 8
CPF s1 9
CPF s2 0
opt imal PF
6GM-Krig组合满足终止条件时FBTP案例的CPFs
10 不同方法获取FBTP案例结果对比(RHV 0.01)
优化方法 优化解 f1/103f2评价次数
NSGA-II (1, 3, 1.414, 1.918) 1.511 0 > 1 000
ParEGO (1, 3, 1.414, 1.910) 1.510 0 35.50 (40.00)
EIM-EGO (1, 3, 1.414, 1.924) 1.514 0 8.03 (10.10)
LSSVR+GIR2 (1, 3, 1.414, 2.065) 1.511 0 1.93 (0.25)
Kriging+GIR2 (1, 3, 1.414, 1.980) 1.524 0 2.07 (0.25)
LSSVR+GIMD (1, 3, 1.414, 1.924) 1.513 0 2.37 (0.67)
Kriging+GIMD (1, 3, 1.414, 1.951) 1.518 0 2.70 (1.18)
由表 10 , NSGA-II 需要大量的评价次数,
使得求解成本高昂.对比其余 MSBO 法可知,
GIMDGIR2 则的 MSBO 方法具有更快的收敛速
度和稳健.综上,所提 GIMDGIR2 准则适用于昂
Mops问题求解,且综合性能优异.
3.5.3 传动机构齿轮组(gear train, GT)
传动 构齿 组设 [24] 是带离散
变量的最小最大问题,4个齿轮及主动轴和从动轴
构成.传动机构齿轮组设计以每个齿轮上的齿x1
x2x3x44个参数 (取整)作为设计变,在给定参
考齿轮比为1/6.931的基础上,以最小化齿轮的最大
尺寸和齿轮比误差为目标.
7为满足终止条件时, GM-Krig 合获 GT
案例的CPFs. 7 , GM-Krig 组合获取的非
支配前沿点与估计的PF 性较 , GIMD-
MSBO 方法具有良好的综合性能.对不同MSBO
法进行比较,并统计终止条件满足时所需评估次数的
均值和标准差,结果见表11.
1838 39
f1
f2
CPF s1
CPF s2
CPF s3
CPF s4
CPF s5
CPF s6
CPF s7
CPF s8
CPF s9
CPF s1 0
CPF s11
CPF s1 2
CPF s1 3
CPF s1 4
CPF s1 5
CPF s1 6
CPF s1 7
CPF s1 8
CPF s1 9
CPF s2 0
opt im al PF
-10 -505
0
20
40
60
80
7GM-Krig组合在满足终止条件时GT案例的CPFs
11 不同方法获取GT 案例结果对比 (RHV 0.01)
优化方法 优化解 (圆整后)f1f2/10 评价次数
NSGA-II (12, 12, 12, 13) 0.612 1.296 > 1 000
ParEGO (18, 13, 19, 17) 0.296 1.912 82.83 (21.98)
EIM-EGO (12, 12, 13, 12) 0.588 1.290 82.50 (22.49)
LSSVR +GIR2 (12, 12, 12, 12) 0.732 1.200 98.26 (14.97)
Kriging +GIR2 (12, 12, 12, 13) 0.612 1.296 83.07 (21.51)
LSSVR +GIMD (12, 12, 12, 12) 0.732 1.200 99.47 (8.40)
Kriging +GIMD (12, 12, 12, 12) 0.732 1.200 77.27 (27.57)
由表 11 结果可知, GT 问题中GIMD-MSBO 方法
和选取LSSVR 建模的GIR2-MSBO 方法获得了相同
的最优决策,且与NSGA-IIEIM-EGO 方法优化结果
相近.相比 NSGA-II 方法需要大于1 000 的评价次数,
MSBO 方法均仅需要100 以内的评价即可.对比 7
种方法的优化结果, Kriging +GIMD 组合最优,其次
EIM-EGO 方法和ParEGO . , GIR2-
MSBO 方法和GIMD-MSBO 方法可实现昂贵多目标
优化问题的快速求解,适用于多种代理模型,具有计
算复杂度低、实用性强的优势.
4
GIDS-MSBO 策略在未试验点选取过程中兼顾
索和 , 现了 Mops 问题
的高效序贯优化.研究结果表明: GIDS-MSBO
具有 于实现、计算复 度低适用性强 优点.
GIMD GIR2 计算 需模预测,MSBO 方法
中更多代理模型应用创造了条件,对进一步基于预测
值开发新准则具有重要的指导意义.
GIDS 策略在MSBO 方法应用中存在巨
大的发展空间和应用潜力. :
(weighted sum approach) ,基于惩罚的交叉边界
方法 (penalty-based boundary intersection approach)
聚合方法的拓展研;开发 更多 有广义性的试
设计策略,诸如使用最大Pareto 前沿误差(maximum
Pareto front error, MPFE) 来取代R2 指标等.
参考文献(References)
[1] 林成龙,马义中,刘丽君,.基于双重权约束期望改
进策略的多目标并行代理优化方法[J]. 控制与决策,
2022, 37(12): 3149-3159.
(Lin C L, Ma Y Z, Liu L J, et al. Multi-objective parallel
surrogate-based optimization based on dual weighted
constraint expectation improvement strategy[J]. Control
and Decision, 2022, 37(12): 3149-3159.)
[2] Forrester A I J, Sóbester A, Keane A J. Engineering
design via surrogate modelling: A practical guide[M].
Chichester: J. Wiley, 2008.
[3] Koziel S, Leifsson L. Multi-level CFD-based airfoil
shape optimization with automated low-fidelity model
selection[J]. Procedia Computer Science, 2013, 18:
889-898.
[4] 林成龙,马义中,肖甜丽.基于均值改进控制策略的昂
贵约束并行代理优化算法[J]. 控制理论与应用, 2021,
38(6): 707-718.
(Lin C L, Ma Y Z, Xiao T L. Expensive constraints
parallel surrogate-based optimization algorithm based on
mean improvement control strategy[J]. Control Theory &
Applications, 2021, 38(6): 707-718.)
[5] Xiu D B. Numerical methods for stochastic
computations: A spectral method approach[M].
Princeton: Princeton University Press, 2010.
[6] Brabanter K D, Karsmakers P, Ojeda F, et al.
LS-SVMlab toolbox user’s guide: Version 1.7[R].
Leuven: Katholieke Universiteit Leuven, 2011.
[7] Sarra S A. The Matlab radial basis function toolbox[J].
Journal of Open Research Software, 2017, 5(1): 8.
[8] Jones D R, Schonlau M, Welch W J. Efficient
global optimization of expensive black-box functions[J].
Journal of Global Optimization, 1998, 13(4): 455-492.
[9] Schonlau M, Welch W J, Jones D R. Global versus
local search in constrained optimization of computer
models[M]. Hayward: Institute of Mathematical
Statistics, 1998: 11-25.
[10] Alexandrov N M, Dennis J E Jr, Lewis R M, et al.
A trust-region framework for managing the use of
approximation models in optimization[J]. Structural
Optimization, 1998, 15(1): 16-23.
[11] Knowles J. ParEGO: A hybrid algorithm with on-line
landscape approximation for expensive multiobjective
optimization problems[J]. IEEE Transactions on
Evolutionary Computation, 2006, 10(1): 50-66.
[12] Zhang Q F, Liu W D, Tsang E, et al. Expensive
6林成龙 :基于广义改进分解策略的多目标代理优化方法 1839
multiobjective optimization by MOEA/D with Gaussian
process model[J]. IEEE Transactions on Evolutionary
Computation, 2010, 14(3): 456-474.
[13] Svenson J, Santner T. Multiobjective optimization of
expensive-to-evaluate deterministic computer simulator
models[J]. Computational Statistics & Data Analysis,
2016, 94: 250-264.
[14] Hupkens I, Deutz A, Yang K F, et al. Faster exact
algorithms for computing expected hypervolume
improvement[C]. International Conference on
Evolutionary Multi-Criterion Optimization. Cham:
Springer, 2015: 65-79.
[15] Couckuyt I, Deschrijver D, Dhaene T. Fast calculation of
multiobjective probability of improvement and expected
improvement criteria for Pareto optimization[J]. Journal
of Global Optimization, 2014, 60(3): 575-594.
[16] Zhan D W, Cheng Y S, Liu J. Expected improvement
matrix-based infill criteria for expensive multiobjective
optimization[J]. IEEE Transactions on Evolutionary
Computation, 2017, 21(6): 956-975.
[17] Han D, Zheng J R. A kriging model-based expensive
multiobjective optimization algorithm using R2 indicator
of expectation improvement[J]. Mathematical Problems
in Engineering, 2020, 2020: 1-16.
[18] Chugh T, Jin Y C, Miettinen K, et al. A surrogate-assisted
reference vector guided evolutionary algorithm
for computationally expensive many-objective
optimization[J]. IEEE Transactions on Evolutionary
Computation, 2018, 22(1): 129-142.
[19] Zhou X J, Jiang T, Zhou Z H, et al. Sequential ε-support
vector regression based online robust parameter
design[J]. Computers & Industrial Engineering, 2021,
158: 107391.
[20] Binois M, Ginsbourger D, Roustant O. Quantifying
uncertainty on Pareto fronts with Gaussian process
conditional simulations[J]. European Journal of
Operational Research, 2015, 243(2): 386-394.
[21] Zhang J X, Ma Y Z, Yang T, et al. Estimation of the
Pareto front in stochastic simulation through stochastic
kriging[J]. Simulation Modelling Practice and Theory,
2017, 79: 69-86.
[22] Lophaven S N, Nielsen H B, Søndergaard J. DACE A
matlab Kriging toolbox-Version 2.0[R]. Lyngby:
Informatics and Mathematical modelling, The Technical
University of Denmark, 2002.
[23] 曹慕昆,王刚,张奇志.动态时间帕累托最优电子采购
多属性谈判模型[J]. 系统工程学报, 2021, 36(1): 1-12.
(Cao M K, Wang G, Zhang Q Z. Dynamic time-dependent
Pareto optimal multi-attribute negotiation model
in electronic procurement[J]. Journal of Systems
Engineering, 2021, 36(1): 1-12.)
[24] Mohamed A T, Vimal S. ε-constraint heat transfer search
(ε-HTS) algorithm for solving multi-objetive engineering
design problems[J]. Journal of Computational Design and
Engineering, 2018, 4: 104-119.
作者简介
林成龙 (1989), ,博士生,从事质量工程、智能优化
算法等研究, E-mail: cllin0814@163.com;
马义中 (1964), ,教授,博士生导师,从事质量管理
与质量工程等研究, E-mail: yzma-2004@163.com;
肖甜丽 (1988), ,博士生,从事质量工程、可靠性优
化等研究, E-mail: 1017973062@qq.com.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Most of the multiobjective optimization problems in engineering involve the evaluation of expensive objectives and constraint functions, for which an approximate model-based multiobjective optimization algorithm is usually employed, but requires a large amount of function evaluation. Aiming at effectively reducing the computation cost, a novel infilling point criterion EIR2 is proposed, whose basic idea is mapping a point in objective space into a set in expectation improvement space and utilizing the R2 indicator of the set to quantify the fitness of the point being selected as an infilling point. This criterion has an analytic form regardless of the number of objectives and demands lower calculation resources. Combining the Kriging model, optimal Latin hypercube sampling, and particle swarm optimization, an algorithm, EIR2-MOEA, is developed for solving expensive multiobjective optimization problems and applied to three sets of standard test functions of varying difficulty and comparing with two other competitive infill point criteria. Results show that EIR2 has higher resource utilization efficiency, and the resulting nondominated solution set possesses good convergence and diversity. By coupling with the average probability of feasibility, the EIR2 criterion is capable of dealing with expensive constrained multiobjective optimization problems and its efficiency is successfully validated in the optimal design of energy storage flywheel.
Article
Full-text available
In this paper, an effective ∊-constraint heat transfer search (∊-HTS) algorithm for the multi-objective engineering design problems is presented. This algorithm is developed to solve multi-objective optimization problems by evaluating a set of single objective sub-problems. The effectiveness of the proposed algorithm is checked by implementing it on multi-objective benchmark problems that have various characteristics of Pareto front such as discrete, convex, and non-convex. This algorithm is also tested for several distinctive multi-objective engineering design problems, such as four bar truss problem, gear train problem, multi-plate disc brake design, speed reducer problem, welded beam design, and spring design problem. Moreover, the numerical experimentation shows that the proposed algorithm generates the solution to represent true Pareto front. Highlights A novel multi-objective optimization (MOO) algorithm is proposed. Proposed algorithm is presented to obtain the Pareto-optimal solutions. The multi-objective optimization algorithm compared with other work in the literature. Test performance of proposed algorithm on MOO benchmark/design engineering problems.
Article
Full-text available
Radial Basis Function (RBF) methods are important tools for scattered data interpolation and for the solution of Partial Differential Equations in complexly shaped domains. The most straight forward approach used to evaluate the methods involves solving a linear system which is typically poorly conditioned. The Matlab Radial Basis Function toolbox features a regularization method for the ill-conditioned system, extended precision floating point arithmetic, and symmetry exploitation for the purpose of reducing flop counts of the associated numerical linear algebra algorithms.
Conference Paper
Full-text available
This paper is about computing the expected improvement of the hypervolume indicator given a Pareto front approximation and a predictive multivariate Gaussian distribution of a new candidate point. It is frequently used as an infill or prescreening criterion in multiobjective optimization with expensive function evaluations where predictions are provided by Kriging or Gaussian process surrogate models. The expected hypervolume improvement has good properties as an infill criterion, but exact algorithms for its computation have so far been very time consuming even for the two and three objective case. This paper introduces faster exact algorithms for computing the expected hypervolume improvement for independent Gaussian distributions. A new general computation scheme is introduced and a lower bound for the time complexity. By providing new algorithms, upper bounds for the time complexity for problems with two as well as three objectives are improved. For the 2-D case the time complexity bound is reduced from previously O(n3logn)O(n^3 \log n) to O(n2)O(n^2). For the 3-D case the new upper bound of O(n3)O(n^3) is established; previously O(n4logn)O(n^4 \log n). It is also shown how an efficient implementation of these new algorithms can lead to a further reduction of running time. Moreover it is shown how to process batches of multiple predictive distributions efficiently. The theoretical analysis is complemented by empirical speed comparisons of C++ implementations of the new algorithms to existing implementations of other exact algorithms.
Article
The existing online Robust Parameter Design (RPD) focuses on updating the optimal setting from the perspective of Bayesian and adopting the low-order polynomial model as the response surface. Polynomial model, especially the low-order, has poor capability for fitting data, particularly for nonlinear data. Adopting an inaccurate model as a response surface will result in obtaining an unreliable recommended optimal setting. In this paper, a machine-learning method to RPD is proposed to recompute the optimal control factor settings based on the optimal control factor settings in the former stage and the online measurements of the noise factors in the current stage. A single response model approach to RPD is adopted. In the proposed method, the response surface is constructed using the sequential ∊-Support Vector Regression (SVR) model, which combines the control factor settings and the online measurements of the noise factors as the input factors and regard the specified quality characteristic as the output factor. When the new measurement of the noise is available, the online SVR can be updated based on the current control factor settings and the observed quality characteristic; and then the RPD process can be redone according to the updated response surface (constructed by SVR model); and eventually the updated control factor settings can be obtained. The proposed methodology is confirmed using three examples, which illustrate that the underlying optimal settings can be found by utilizing the proposed online RPD strategy.
Article
Multi-objective optimization of simulated stochastic systems aims at estimating a representative set of Pareto optimal solutions and a common approach is to rely on metamodels to alleviate computational costs of the optimization process. In this article, both the objective and constraint functions are assumed to be smooth, highly non-linear and computationally expensive and are emulated by stochastic Kriging models. Then a novel global optimization algorithm, combing the expected hypervolume improvement of approximated Pareto front and the probability of feasibility of new points, is proposed to identify the Pareto front (set) with a minimal number of expensive simulations. The algorithm is suitable for the situations of having disconnected feasible regions and of having no feasible solution in initial design. Then, we also quantify the variability of estimated Pareto front caused by the intrinsic uncertainty of stochastic simulation using nonparametric bootstrapping method to better support decision making. One test function and an (s, S) inventory system experiment illustrate the potential and efficiency of the proposed sequential optimization algorithm for constrained multi-objective optimization problems in stochastic simulation, which is especially useful in Operations Research and Management Science.
Article
Many engineering design optimization problems contain multiple objective functions all of which are desired to be minimized, say. This paper proposes a method for identifying the Pareto Front and the Pareto Set of the objective functions when these functions are evaluated by expensive-to-evaluate deterministic computer simulators. The method replaces the expensive function evaluations by a rapidly computable approximator based on a Gaussian process (GP) interpolator. It sequentially selects new input sites guided by values of an "improvement function" given the current data. The method introduced in this paper provides two advances in the interpolator/improvement framework. First, it proposes an improvement function based on the "modified maximin fitness function" which is known to identify well-spaced non-dominated outputs when used in multiobjective evolutionary algorithms. Second, it uses a family of GP models that allows for dependence among output function values but which permits zero covariance should the data be consistent with this model. A closed-form expression is derived for the improvement function when there are two objective functions; simulation is used to evaluate it when there are three or more objectives. Examples from the multiobjective optimization literature are presented to show that the proposed procedure can improve substantially previously proposed statistical improvement criteria for the computationally intensive multiobjective optimization setting.
Article
Multi-objective optimization algorithms aim at finding Pareto-optimal solutions. Recovering Pareto fronts or Pareto sets from a limited number of function evaluations are challenging problems. A popular approach in the case of expensive-to-evaluate functions is to appeal to metamodels. Kriging has been shown efficient as a base for sequential multi-objective optimization, notably through infill sampling criteria balancing exploitation and exploration such as the Expected Hypervolume Improvement. Here we consider Kriging metamodels not only for selecting new points, but as a tool for estimating the whole Pareto front and quantifying how much uncertainty remains on it at any stage of Kriging-based multi-objective optimization algorithms. Our approach relies on the Gaussian process interpretation of Kriging, and bases upon conditional simulations. Using concepts from random set theory, we propose to adapt the Vorob’ev expectation and deviation to capture the variability of the set of non-dominated points. Numerical experiments illustrate the potential of the proposed workflow, and it is shown on examples how Gaussian process simulations and the estimated Vorob’ev deviation can be used to monitor the ability of Kriging-based multi-objective optimization algorithms to accurately learn the Pareto front.
Article
The first graduate-level textbook to focus on fundamental aspects of numerical methods for stochastic computations, this book describes the class of numerical methods based on generalized polynomial chaos (gPC). These fast, efficient, and accurate methods are an extension of the classical spectral methods of high-dimensional random spaces. Designed to simulate complex systems subject to random inputs, these methods are widely used in many areas of computer science and engineering.The book introduces polynomial approximation theory and probability theory; describes the basic theory of gPC methods through numerical examples and rigorous development; details the procedure for converting stochastic equations into deterministic ones; using both the Galerkin and collocation approaches; and discusses the distinct differences and challenges arising from high-dimensional problems. The last section is devoted to the application of gPC methods to critical areas such as inverse problems and data assimilation.Ideal for use by graduate students and researchers both in the classroom and for self-study,Numerical Methods for Stochastic Computationsprovides the required tools for in-depth research related to stochastic computations.The first graduate-level textbook to focus on the fundamentals of numerical methods for stochastic computations Ideal introduction for graduate courses or self-study Fast, efficient, and accurate numerical methods Polynomial approximation theory and probability theory included Basic gPC methods illustrated through examples.