03 分层Kriging代理模型理论的相关推导
分层Kriging(Hierarchical Kriging, HK)模型是一种简单实用的变可信度代理模型,分两层或多层建立代理模型,本文以两层模型为例对分层代理模型理论进行推导。
3.1 代理模型问题的基本描述
对于一个有
m
m
m个设计变量的优化问题,在设计空间中同时进行高可信度分析和低可信度抽样,以建立所谓变可信度模型,变可信代理模型在达到相同近似精度的条件下,可以显著提高建立代理模型的效率。设高、低可信度分析程序的抽样位置分别为:
{
S
1
=
[
x
1
(
1
)
x
1
(
2
)
…
x
1
(
n
1
)
]
T
∈
R
n
1
×
m
S
2
=
[
x
2
(
1
)
x
2
(
2
)
…
x
2
(
n
2
)
]
T
∈
R
n
2
×
m
(1)
\begin {cases} \mathbf S_1=[\mathbf x_1^{(1)} \ \ \mathbf x_1^{(2)} \ \ \ldots \ \ \mathbf x_1^{(n_1)}]^T\in\R^{n_1\times m} \\ \mathbf S_2=[\mathbf x_2^{(1)} \ \ \mathbf x_2^{(2)} \ \ \ldots \ \ \mathbf x_2^{(n_2)}]^T\in \R^{n_2 \times m} \end {cases} \tag{1}
{S1=[x1(1) x1(2) … x1(n1)]T∈Rn1×mS2=[x2(1) x2(2) … x2(n2)]T∈Rn2×m(1)
上式中,下标 “1” 和 “2” 分别代表高、低可信度,例如
n
1
n_1
n1 和
n
2
n_2
n2 分别代表高、低可信度样本点数(可合理假设
n
2
≫
n
1
n_2\gg n_1
n2≫n1 )。相应的目标函数或约束函数响应值为:
{
y
1
=
[
y
1
(
1
)
y
1
(
2
)
…
y
1
(
n
1
)
]
T
∈
R
n
1
y
2
=
[
y
2
(
1
)
y
2
(
2
)
…
y
2
(
n
2
)
]
T
∈
R
n
2
(2)
\begin {cases} \mathbf y_1=[y_1^{(1)} \ \ y_1^{(2)} \ \ \ldots \ \ y_1^{(n_1)}]^T\in\R^{n_1} \\ \mathbf y_2=[y_2^{(1)} \ \ y_2^{(2)} \ \ \ldots \ \ y_2^{(n_2)}]^T\in\R^{n_2} \end {cases} \tag{2}
{y1=[y1(1) y1(2) … y1(n1)]T∈Rn1y2=[y2(1) y2(2) … y2(n2)]T∈Rn2(2)
分层代理模型首先在低可信度样本数据
(
S
2
,
y
2
)
(\mathbf S_2,y_2)
(S2,y2) 的基础上建立 Kriging 模型
y
^
2
\hat y_2
y^2 ,然后以
y
^
2
\hat y_2
y^2 为全局趋势模型,在高可信度样本数据集
(
S
1
,
y
1
)
(\mathbf S_1,y_1)
(S1,y1) 基础上建立所需的代理模型
y
^
1
\hat y_1
y^1。
3.2 HK 模型及其预估值
3.2.1 首先针对低可信度样本点建立Kriging代理模型
首先基于低可信度数据来建立低可信度代理模型以辅助预测,针对低可信度模型假设高斯静态随机过程:
Y
l
f
(
x
)
=
β
0
,
l
f
+
Z
l
f
(
x
)
(3)
Y_{lf}(\mathbf x)=\beta_{0,lf}+Z_{lf}(\mathbf x) \tag{3}
Ylf(x)=β0,lf+Zlf(x)(3)
参照 Kriging 模型建立过程,可得低可信度模型表达式:
y
^
l
f
(
x
)
=
β
0
,
l
f
+
r
l
f
T
(
x
)
R
l
f
−
1
(
y
s
,
l
f
−
β
0
,
l
f
F
1
)
(4)
\hat y_{lf}(\mathbf x)=\beta_{0,lf}+\mathbf r_{lf}^T(\mathbf x)\mathbf R_{lf}^{-1}(\mathbf y_{s,lf}-\beta_{0,lf}\mathbf F_1) \tag{4}
y^lf(x)=β0,lf+rlfT(x)Rlf−1(ys,lf−β0,lfF1)(4)
上式中,
β
0
,
l
f
=
(
F
1
T
R
l
f
−
1
F
1
)
−
1
F
1
T
R
l
f
−
1
y
s
,
l
f
\beta_{0,lf}=(\mathbf F_1^T\mathbf R_{lf}^{-1}\mathbf F_1)^{-1}\mathbf F_1^T\mathbf R_{lf}^{-1}\mathbf y_{s,lf}
β0,lf=(F1TRlf−1F1)−1F1TRlf−1ys,lf ;
R
l
f
∈
R
n
l
f
×
n
l
f
\mathbf R_{lf}\in\R^{n_{lf}\times n_{lf}}
Rlf∈Rnlf×nlf 是现有的低可信度样本点组成的相关矩阵;
F
1
\mathbf F_1
F1 是全为1的列向量;
r
l
f
∈
R
n
l
f
\mathbf r_{lf}\in \R^{n_{lf}}
rlf∈Rnlf 是任意点
x
\mathbf x
x 与已知样本点间的相关矢量。
3.2.2 建立分层代理模型
不同于Cokriging 模型,将低可信度数据建立的代理模型作为趋势项,将高可信度函数视为下述高斯静态随机过程的具体实现:
Y
(
x
)
=
β
0
y
^
l
f
(
x
)
+
Z
(
x
)
(5)
Y(\mathbf x)=\beta_0 \hat y_{lf}(\mathbf x)+Z(\mathbf x) \tag {5}
Y(x)=β0y^lf(x)+Z(x)(5)
将
y
^
l
f
(
x
)
\hat y_{lf}(\mathbf x)
y^lf(x) 乘以缩放系数
β
0
\beta_0
β0 作为全局趋势模型,
Z
(
x
)
Z(\mathbf x)
Z(x) 是均值为零、方差为
σ
2
\sigma ^2
σ2 的静态随机过程。在设计空间不同位置处,这些随机变量存在一定的相关性(协方差),表示为:
C
o
v
[
Z
(
x
)
,
Z
(
x
′
)
]
=
σ
2
R
(
x
,
x
′
)
(6)
Cov[Z(\mathbf x),Z(\mathbf x')]=\sigma^2R(\mathbf x, \mathbf x') \tag{6}
Cov[Z(x),Z(x′)]=σ2R(x,x′)(6)
上式中,
R
(
x
,
x
′
)
R(\mathbf x,\mathbf x')
R(x,x′) 是相关函数,它是
x
\mathbf x
x 和
x
′
\mathbf x'
x′ 间欧几里德距离的函数,将高可信度函数视为已知高可信度样本点的线性加权,即:
y
^
(
x
)
=
ω
T
y
s
(7)
\hat y(\mathbf x)=\boldsymbol \omega^T\mathbf y_s \tag{7}
y^(x)=ωTys(7)
上式中,
ω
=
[
ω
(
1
)
ω
(
2
)
…
ω
(
n
)
]
T
\boldsymbol \omega=\begin {bmatrix} \omega^{(1)} & \omega^{(2)} & \ldots \omega^{(n)} \end {bmatrix}^T
ω=[ω(1)ω(2)…ω(n)]T 是高可信度样本点的加权系数向量,用
Y
s
=
[
Y
(
1
)
Y
(
2
)
…
Y
(
n
1
)
]
\mathbf Y_s=\begin {bmatrix} Y^{(1)} & Y^{(2)} & \ldots & Y^{(n_1)} \end {bmatrix}
Ys=[Y(1)Y(2)…Y(n1)] 代替
y
s
=
[
y
(
1
)
y
(
2
)
…
y
(
n
)
]
\mathbf y_s=\begin {bmatrix}y^{(1)} & y^{(2)} &\ldots & y^{(n)} \end {bmatrix}
ys=[y(1)y(2)…y(n)] 。寻找最优加权系数
ω
\boldsymbol \omega
ω 使得均方差最小:
M
S
E
[
y
^
(
x
)
]
=
E
[
(
ω
T
Y
s
−
Y
(
x
)
)
2
]
(8)
MSE[\hat y(\mathbf x)]=E[(\boldsymbol \omega^T\mathbf Y_s-Y(\mathbf x))^2] \tag{8}
MSE[y^(x)]=E[(ωTYs−Y(x))2](8)
并满足无偏约束:
E
[
ω
T
Y
s
]
=
E
[
Y
(
x
)
]
(9)
E[\boldsymbol \omega^T \mathbf Y_s]=E[Y(\mathbf x)] \tag{9}
E[ωTYs]=E[Y(x)](9)
对上式化简得:
以下推导过程要注意,应将 y ^ l f ( x ) \hat y_{lf}(\mathbf x) y^lf(x) 视为由低可信度代理模型得出的值,代入到高可信度模型后则应视为常值(此常值在设计空间不同点处的值不同)
∑ i = 1 n 1 ω ( i ) y ^ l f ( x 1 ( i ) ) = y ^ l f ( x ) (10) \sum_{i=1}^{n_1}\omega^{(i)}\hat y_{lf}(\mathbf x_1^{(i)}) =\hat y_{lf}(\mathbf x) \tag{10} i=1∑n1ω(i)y^lf(x1(i))=y^lf(x)(10)
采用拉格朗日乘数法,令:
L
(
ω
,
μ
)
=
E
[
(
ω
T
Y
s
−
Y
(
x
)
)
2
]
+
μ
(
∑
ω
(
i
)
y
^
l
f
(
x
1
(
i
)
)
−
y
^
l
f
(
x
)
)
(11)
L(\boldsymbol \omega, \mu)=E[(\boldsymbol \omega^T\mathbf Y_s-Y(\mathbf x))^2] + \mu(\sum \omega^{(i)}\hat y_{lf}(\mathbf x_1^{(i)})-\hat y_{lf}(\mathbf x) ) \tag {11}
L(ω,μ)=E[(ωTYs−Y(x))2]+μ(∑ω(i)y^lf(x1(i))−y^lf(x))(11)
令:
∂
L
(
ω
,
μ
)
∂
ω
(
i
)
=
2
E
[
(
ω
T
Y
s
−
Y
(
x
)
)
Y
s
(
x
1
(
i
)
)
]
+
μ
y
^
l
f
(
x
1
(
i
)
)
=
2
{
∑
j
=
1
n
1
ω
(
j
)
[
β
0
2
y
^
l
f
(
x
1
(
i
)
)
y
^
l
f
(
x
1
(
j
)
)
+
σ
2
R
(
x
1
(
i
)
,
x
1
(
j
)
)
]
−
[
β
0
2
y
^
l
f
(
x
)
y
^
l
f
(
x
1
(
i
)
)
}
+
σ
2
R
(
x
1
(
i
)
,
x
)
]
+
μ
y
^
l
f
(
x
1
(
i
)
)
=
2
σ
2
∑
j
=
1
n
1
ω
(
j
)
R
(
x
1
(
i
)
,
x
1
(
j
)
)
−
2
σ
2
R
(
x
1
(
i
)
,
x
)
+
μ
y
^
l
f
(
x
1
(
i
)
)
=
0
(12)
\begin {align} \frac {\partial L(\boldsymbol \omega,\mu)}{\partial \omega^{(i)}} &= 2E[(\boldsymbol \omega^T\mathbf Y_s-Y(\mathbf x))Y_s(\mathbf x_1^{(i)})] +\mu\hat y_{lf}(\mathbf x_1^{(i)}) \\ &= 2\{\sum_{j=1}^{n_1}\omega^{(j)}[\beta_0^2\hat y_{lf}(\mathbf x_1^{(i)})\hat y_{lf}(\mathbf x_1^{(j)})+\sigma^2R(\mathbf x_1^{(i)},\mathbf x_1^{(j)})]-[\beta_0^2\hat y_{lf}(\mathbf x)\hat y_{lf}(\mathbf x_1^{(i)})\}+\sigma^2R(\mathbf x_1^{(i)},\mathbf x)]+\mu \hat y_{lf}(\mathbf x_1^{(i)}) \\ &= 2\sigma^2\sum_{j=1}^{n_1}\omega^{(j)}R(\mathbf x_1^{(i)},\mathbf x_1^{(j)})-2\sigma^2R(\mathbf x_1^{(i)},\mathbf x)+\mu\hat y_{lf}(\mathbf x_1^{(i)}) \\ &=0 \end {align} \tag{12}
∂ω(i)∂L(ω,μ)=2E[(ωTYs−Y(x))Ys(x1(i))]+μy^lf(x1(i))=2{j=1∑n1ω(j)[β02y^lf(x1(i))y^lf(x1(j))+σ2R(x1(i),x1(j))]−[β02y^lf(x)y^lf(x1(i))}+σ2R(x1(i),x)]+μy^lf(x1(i))=2σ2j=1∑n1ω(j)R(x1(i),x1(j))−2σ2R(x1(i),x)+μy^lf(x1(i))=0(12)
则:
{
∑
j
=
1
n
1
ω
(
j
)
R
(
x
1
(
i
)
,
x
1
(
j
)
)
+
μ
y
^
l
f
(
x
1
(
i
)
)
2
σ
2
=
R
(
x
1
(
i
)
,
x
)
∑
i
=
1
n
1
ω
(
i
)
y
^
l
f
(
x
1
(
i
)
)
=
y
^
l
f
(
x
)
(13)
\begin {cases} \sum_{j=1}^{n_1}\omega^{(j)}R(\mathbf x_1^{(i)},\mathbf x_1^{(j)}) + \frac {\mu \hat y_{lf}(\mathbf x_1^{(i)})}{2\sigma^2}=R(\mathbf x_1^{(i)},\mathbf x) \\ \sum_{i=1}^{n_1}\omega^{(i)}\hat y_{lf}(\mathbf x_1^{(i)})=\hat y_{lf}(\mathbf x) \end {cases} \tag{13}
{∑j=1n1ω(j)R(x1(i),x1(j))+2σ2μy^lf(x1(i))=R(x1(i),x)∑i=1n1ω(i)y^lf(x1(i))=y^lf(x)(13)
写成矩阵形式为:
[
R
F
F
T
0
]
[
ω
μ
^
]
=
[
r
y
^
l
f
(
x
)
]
(14)
\begin {bmatrix}\mathbf R & \mathbf F \\ \mathbf F^T & 0\end {bmatrix}\begin {bmatrix} \boldsymbol \omega \\ \hat \mu \end {bmatrix} =\begin {bmatrix} \mathbf r \\ \hat y_{lf}(\mathbf x) \end {bmatrix} \tag{14}
[RFTF0][ωμ^]=[ry^lf(x)](14)
上式中:
F
=
[
y
^
l
f
(
x
1
(
1
)
)
,
…
,
y
^
l
f
(
x
1
(
n
1
)
)
]
T
,
μ
^
=
μ
/
(
2
σ
2
)
R
=
(
R
(
x
1
(
i
)
,
x
1
(
j
)
)
)
i
,
j
∈
R
n
1
×
n
1
,
r
=
(
R
(
x
1
(
i
)
,
x
)
)
i
∈
R
n
1
(15)
\begin {align} & \mathbf F=[\hat y_{lf}(\mathbf x_1^{(1)}),\ldots,\hat y_{lf}(\mathbf x_1^{(n_1)})]^T, \ \ \ \ \hat \mu=\mu /(2\sigma^2) \\ & \mathbf R =(R(\mathbf x_1^{(i)},\mathbf x_1^{(j)}))_{i,j}\in \R^{n_1 \times n_1}, \ \ \ \ \mathbf r=(R(\mathbf x_1^{(i)},\mathbf x))_i \in \R^{n_1} \end {align} \tag {15}
F=[y^lf(x1(1)),…,y^lf(x1(n1))]T, μ^=μ/(2σ2)R=(R(x1(i),x1(j)))i,j∈Rn1×n1, r=(R(x1(i),x))i∈Rn1(15)
则:
[
ω
T
μ
^
]
=
(
[
R
F
F
T
0
]
−
1
[
r
y
^
l
f
(
x
)
]
)
T
=
[
r
y
^
l
f
(
x
)
]
T
[
R
F
F
T
0
]
−
1
(16)
\begin {bmatrix} \boldsymbol \omega^T & \hat \mu \end {bmatrix}= \left(\begin {bmatrix} \mathbf R & \mathbf F \\ \mathbf F^T & 0 \end {bmatrix}^{-1} \begin {bmatrix} \mathbf r \\ \hat y_{lf}(\mathbf x) \end {bmatrix} \right)^T=\begin {bmatrix} \mathbf r \\ \hat y_{lf}(\mathbf x) \end {bmatrix}^T \begin {bmatrix} \mathbf R & \mathbf F \\ \mathbf F^T & 0 \end {bmatrix}^{-1} \tag{16}
[ωTμ^]=([RFTF0]−1[ry^lf(x)])T=[ry^lf(x)]T[RFTF0]−1(16)
则设计空间任意未知点
x
\mathbf x
x 处预测的响应值为:
y
^
(
x
)
=
[
ω
T
μ
^
]
[
y
s
0
]
=
[
r
y
^
l
f
(
x
)
]
T
[
R
F
F
T
0
]
−
1
[
y
s
0
]
(17)
\hat y(\mathbf x) =\begin {bmatrix} \boldsymbol \omega^T & \hat \mu \end {bmatrix} \begin {bmatrix} \mathbf y_s \\ 0 \end {bmatrix} =\begin {bmatrix} \mathbf r \\ \hat y_{lf}(\mathbf x) \end {bmatrix}^T \begin {bmatrix}\mathbf R & \mathbf F \\ \mathbf F^T & 0 \end {bmatrix}^{-1}\begin {bmatrix} \mathbf y_s \\ 0 \end {bmatrix} \tag {17}
y^(x)=[ωTμ^][ys0]=[ry^lf(x)]T[RFTF0]−1[ys0](17)
由分块矩阵求逆公式,有:
[
R
F
F
T
0
]
−
1
=
[
R
−
1
−
R
−
1
F
(
F
T
R
−
1
F
)
−
1
F
T
R
−
1
R
−
1
F
(
F
T
R
−
1
F
)
−
1
(
F
T
R
−
1
F
)
−
1
F
T
R
−
1
−
(
F
T
R
−
1
F
)
−
1
]
(18)
\begin {bmatrix} \mathbf R & \mathbf F \\ \mathbf F ^T & 0\end {bmatrix} ^{-1} = \begin{bmatrix} \mathbf R^{-1}-\mathbf R^{-1}\mathbf F(\mathbf F^T \mathbf R^{-1}\mathbf F)^{-1}\mathbf F^{T}\mathbf R^{-1} & \mathbf R^{-1} \mathbf F(\mathbf F^T \mathbf R^{-1} \mathbf F)^{-1} \\ (\mathbf F^T \mathbf R^{-1} \mathbf F)^{-1}\mathbf F^T\mathbf R^{-1} & -(\mathbf F^T\mathbf R^{-1} \mathbf F)^{-1}\end {bmatrix} \tag{18}
[RFTF0]−1=[R−1−R−1F(FTR−1F)−1FTR−1(FTR−1F)−1FTR−1R−1F(FTR−1F)−1−(FTR−1F)−1](18)
将上式代入式 (16) 并化简得:
y
^
(
x
)
=
[
r
T
R
−
1
−
r
T
R
−
1
F
(
F
T
R
−
1
F
)
−
1
F
T
R
−
1
+
y
^
l
f
(
x
)
(
F
T
R
−
1
F
)
−
1
F
T
R
−
1
]
y
s
=
β
0
y
^
l
f
(
x
)
+
r
T
(
x
)
R
−
1
(
y
s
−
β
0
F
)
(19)
\begin {align} \hat y(\mathbf x) &= [\mathbf r^T\mathbf R^{-1} - \mathbf r^T \mathbf R ^{-1}\mathbf F(\mathbf F^T\mathbf R^{-1}\mathbf F)^{-1}\mathbf F^T\mathbf R^{-1} +\hat y_{lf}(\mathbf x)(\mathbf F^T\mathbf R^{-1}\mathbf F)^{-1}\mathbf F^T \mathbf R^{-1}]\mathbf y_s \\ &=\beta_0\hat y_{lf}(\mathbf x)+\mathbf r^T(\mathbf x)\mathbf R^{-1}(\mathbf y_s -\beta_0\mathbf F) \end {align} \tag{19}
y^(x)=[rTR−1−rTR−1F(FTR−1F)−1FTR−1+y^lf(x)(FTR−1F)−1FTR−1]ys=β0y^lf(x)+rT(x)R−1(ys−β0F)(19)
上式中:
β
0
=
(
F
T
R
−
1
F
)
−
1
F
T
R
−
1
y
s
\beta_0=(\mathbf F^T\mathbf R^{-1}\mathbf F)^{-1}\mathbf F^T\mathbf R^{-1}\mathbf y_s
β0=(FTR−1F)−1FTR−1ys 是放缩系数,
V
H
K
=
R
−
1
(
y
s
−
β
0
F
)
V_{HK}=\mathbf R^{-1}(\mathbf y_s-\beta_0\mathbf F)
VHK=R−1(ys−β0F) 只与已知样本点数据有关,可以在模型训练结束后一次性计算并存储。之后,预测任意
x
\mathbf x
x 处的响应值只需要计算
r
(
x
)
\mathbf r(\mathbf x)
r(x) 与
V
H
K
\mathbf V_{HK}
VHK 间的点乘。
分层Kriging模型给出的预估值的均方差为:
M S E [ y ^ ( x ) ] = E [ ( ω T Y s − Y ( x ) ) 2 ] = E [ ( ω T Y s − Y ( x ) ) ( ω T Y s − Y ( x ) ) ] = σ 2 ( ω T R ω − 2 ω T r + 1 ) = σ 2 [ ω T ( R ω + F μ ^ ) − ω T F μ ^ − 2 ω T r + 1 ] = σ 2 [ ω T r − ω T F μ ^ − 2 ω T r + 1 ] = σ 2 [ 1 − ω T r − ω T F μ ^ ] = σ 2 [ 1 − [ ω T μ ^ ] [ r y l f ( x ) ] ] = σ 2 [ 1 − [ r y l f ( x ) ] T [ R F F T 0 ] − 1 [ r y l f ( x ) ] ] = σ 2 { 1.0 − r T R − 1 r + [ r T R − 1 F − y ^ l f ( x ) ] ( F T R − 1 F ) − 1 [ r T R − 1 F − y ^ l f ( x ) ] T } (20) \begin {align} MSE[\hat y(\mathbf x)] & =E[(\boldsymbol \omega ^T\mathbf Y_s-\mathbf Y(\mathbf x))^2]\\ & =E[(\boldsymbol \omega ^T\mathbf Y_s-\mathbf Y(\mathbf x))(\boldsymbol \omega ^T\mathbf Y_s-\mathbf Y(\mathbf x))]\\ & = \sigma ^2(\boldsymbol \omega^T\mathbf R \boldsymbol \omega - 2\boldsymbol \omega^T\mathbf r + 1) \\ & =\sigma^2[\boldsymbol \omega^T(\mathbf R\boldsymbol \omega+\mathbf F\hat \mu)-\boldsymbol \omega^T\mathbf F\hat \mu - 2\boldsymbol \omega^T\mathbf r + 1] \\ & = \sigma ^2[\boldsymbol \omega^T\mathbf r-\boldsymbol \omega^T\mathbf F\hat\mu-2\boldsymbol \omega^T\mathbf r + 1] \\ & =\sigma ^2[1-\boldsymbol \omega^T \mathbf r-\boldsymbol \omega^T \mathbf F\hat\mu] \\ &=\sigma ^2[1-[\boldsymbol \omega ^T \ \ \hat \mu]\begin {bmatrix}\mathbf r \\ \boldsymbol y_{lf}(\mathbf x) \end {bmatrix}]\\&= \sigma ^2[1-\begin {bmatrix} \mathbf r \\ y_{lf}(\mathbf x)\end {bmatrix}^T\begin{bmatrix} \mathbf R &\mathbf F \\ \mathbf F^T & 0 \end {bmatrix}^{-1}\begin {bmatrix}\mathbf r \\ \boldsymbol y_{lf}(\mathbf x) \end {bmatrix}] \\ &=\sigma^2 \{ 1.0-\mathbf r^T\mathbf R^{-1}\mathbf r+[\mathbf r^T\mathbf R^{-1}\mathbf F-\hat y_{lf}(\mathbf x)](\mathbf F^T\mathbf R^{-1}\mathbf F)^{-1}[\mathbf r^T\mathbf R^{-1}\mathbf F-\hat y_{lf}(\mathbf x)]^T \} \end {align} \tag {20} MSE[y^(x)]=E[(ωTYs−Y(x))2]=E[(ωTYs−Y(x))(ωTYs−Y(x))]=σ2(ωTRω−2ωTr+1)=σ2[ωT(Rω+Fμ^)−ωTFμ^−2ωTr+1]=σ2[ωTr−ωTFμ^−2ωTr+1]=σ2[1−ωTr−ωTFμ^]=σ2[1−[ωT μ^][rylf(x)]]=σ2[1−[rylf(x)]T[RFTF0]−1[rylf(x)]]=σ2{1.0−rTR−1r+[rTR−1F−y^lf(x)](FTR−1F)−1[rTR−1F−y^lf(x)]T}(20)
分层Kriging 模型的相关函数,模型训练过程与Kriging模型十分相似,可以参考之前的博客文章(链接: link),在此不再赘述。
参考文献
[1] HAN Z H , Görtz S. Hierarchical Kriging Model for Variable-Fidelity Surrogate Modeling[J]. AIAA Journal, 2012, 50(9):1885-1896.