1 The essentially quadratic augmented Lagrangian function
令
ϕ
:
R
→
R
\phi:\mathbb R\rightarrow \mathbb R
ϕ:R→R 是二次连续可微的并且严格凸函数,满足
ϕ
(
0
)
=
0
,
ϕ
′
(
0
)
=
0
\phi(0)=0,\phi\prime(0)=0
ϕ(0)=0,ϕ′(0)=0 定义:
P
(
s
,
t
)
=
{
t
s
+
ϕ
(
s
)
i
f
t
+
ϕ
′
(
s
)
≥
0
min
τ
∈
R
[
t
τ
+
ϕ
(
τ
)
]
o
t
h
e
r
w
i
s
e
P(s,t)=\left \{ \begin{aligned} &ts+\phi(s) &&if\; t+\phi\prime(s)\ge 0\\ &\min_{\tau\in \mathbb R} [t\tau+\phi(\tau)] &&otherwise\end{aligned} \right.
P(s,t)=⎩⎨⎧ts+ϕ(s)τ∈Rmin[tτ+ϕ(τ)]ift+ϕ′(s)≥0otherwise
二次增广拉格朗日函数定义为:
L
1
(
x
,
λ
,
c
)
=
f
(
x
)
+
1
c
P
(
c
g
i
(
x
)
,
λ
i
)
L_1(x,\lambda,c)=f(x)+\frac{1}{c}P(cg_i(x),\lambda_i)
L1(x,λ,c)=f(x)+c1P(cgi(x),λi)
其中 c > 0 , x ∈ X , λ = ( λ 1 , . . . , λ m ) T ≥ 0 c>0,x\in X,\lambda=(\lambda_1,...,\lambda_m)^T\ge0 c>0,x∈X,λ=(λ1,...,λm)T≥0
2 The exponential-type augmented Lagrangian
令
ψ
:
R
→
R
\psi:\mathbb R\rightarrow \mathbb R
ψ:R→R 是二次可微的并且严格增函数,满足
ψ
(
0
)
=
0
\psi(0)=0
ψ(0)=0 问题是:
min
f
(
x
)
s
u
b
j
e
c
t
t
o
(
1
/
c
)
ψ
(
c
g
i
(
x
)
)
≤
0
,
i
=
1
,
.
.
.
,
m
x
∈
X
\begin{aligned} &\min\;f(x)\\ &subject\;to\;&&(1/c)\psi(cg_i(x))\le 0,i=1,...,m\\ &&&x\in X \end{aligned}
minf(x)subjectto(1/c)ψ(cgi(x))≤0,i=1,...,mx∈X
其中 c > 0 c>0 c>0 是一个参数
上述问题(P)的拉格朗日函数可以写为以下形式:
L
2
(
x
,
λ
,
c
)
=
f
(
x
)
+
1
c
∑
i
=
1
m
λ
i
ψ
(
c
g
i
(
x
)
)
,
x
∈
X
,
λ
≥
0
L_2(x,\lambda,c)=f(x)+\frac{1}{c}\sum\limits_{i=1}^m\lambda_i\psi(cg_i(x)),\quad x\in X,\lambda\ge0
L2(x,λ,c)=f(x)+c1i=1∑mλiψ(cgi(x)),x∈X,λ≥0
3 The modified barrier augmented Lagrangian
令 φ : ( − ∞ , 1 ) → R \varphi:(-\infty,1)\rightarrow \mathbb R φ:(−∞,1)→R 是二次可微分的并且严格增函数,满足 φ ( 0 ) = 0 \varphi(0)=0 φ(0)=0
我们可以把问题重新表示通过使用
φ
\varphi
φ。令
Ω
c
=
x
∈
X
∣
c
g
i
(
x
)
<
1
,
i
=
1
,
.
.
.
,
m
\Omega_c={x\in X|cg_i(x)<1,i=1,...,m}
Ωc=x∈X∣cgi(x)<1,i=1,...,m 定义 The modified barrier augmented Lagrangian 如下:
L
3
(
x
,
λ
,
c
)
=
{
f
(
x
)
+
1
c
∑
i
=
1
m
λ
i
φ
(
c
g
i
(
x
)
)
,
x
∈
Ω
c
∞
,
x
∈
X
\
Ω
x
L_3(x,\lambda,c)=\left \{\begin{aligned}&f(x)+\frac{1}{c}\sum\limits_{i=1}^m\lambda_i\varphi(cg_i(x)),&&x\in \Omega_c\\&\infty,&&x\in X\verb|\|\Omega_x\end{aligned}\right.
L3(x,λ,c)=⎩⎪⎪⎨⎪⎪⎧f(x)+c1i=1∑mλiφ(cgi(x)),∞,x∈Ωcx∈X\Ωx
其中 c > 0 , λ ≥ 0 c>0,\lambda\ge 0 c>0,λ≥0
4 Introducing the penalty function
令
ξ
:
R
→
R
\xi:\mathbb R\rightarrow \mathbb R
ξ:R→R 是一个二次可微分函数,对于所有的
t
≤
0
t\le0
t≤0 满足
ξ
(
t
)
=
0
\xi(t)=0
ξ(t)=0,对于所有的
t
>
0
t>0
t>0 满足
ξ
(
t
)
>
0
\xi(t)>0
ξ(t)>0 对于一下问题:
min
f
(
x
)
+
(
1
/
c
)
∑
i
=
1
m
ξ
(
c
g
i
(
x
)
)
s
u
b
j
e
c
t
t
o
(
1
/
c
)
ψ
(
c
g
i
(
x
)
)
≤
0
,
i
=
1
,
.
.
.
,
m
,
x
∈
X
\begin{aligned} &\min\;\; f(x)+(1/c)\sum\limits_{i=1}^m\xi(cg_i(x))\\ &subject\; to \;(1/c)\psi(cg_i(x))\le0,\;i=1,...,m,\\ &x\in X \end{aligned}
minf(x)+(1/c)i=1∑mξ(cgi(x))subjectto(1/c)ψ(cgi(x))≤0,i=1,...,m,x∈X
拉格朗日函数为:
L
4
(
x
,
λ
,
c
)
=
f
(
x
)
+
1
c
∑
i
=
1
m
[
λ
i
ψ
(
c
g
i
(
x
)
)
+
ξ
(
c
g
i
(
x
)
)
]
,
x
∈
X
,
λ
≥
0
L_4(x,\lambda,c)=f(x)+\frac{1}{c}\sum\limits_{i=1}^m[\lambda_i\psi(cg_i(x))+\xi(cg_i(x))],\quad x\in X,\lambda\ge0
L4(x,λ,c)=f(x)+c1i=1∑m[λiψ(cgi(x))+ξ(cgi(x))],x∈X,λ≥0