寻找鞍点 min x max y f ( x , y ) \min_{x}\max_{y} f(x,y) minxmaxyf(x,y)的一个著名算法
输入:
(
α
,
ϵ
)
(\alpha,\epsilon)
(α,ϵ)-relaxed proximal oracle
O
(
z
)
\mathcal{O}(z)
O(z) for gradient mapping
g
g
g, distance generating
r
r
r.
参数:
迭代次数
K
K
K
输出:
点
z
ˉ
K
\bar{z}_{K}
zˉK,满足
E
[
G
a
p
(
z
ˉ
)
]
≤
α
Θ
K
+
ϵ
\mathbb{E}[Gap(\bar{z})]\leq \frac{\alpha\Theta}{K}+\epsilon
E[Gap(zˉ)]≤KαΘ+ϵ.
初始化:
z
0
=
arg
min
r
(
z
)
z_{0}=\arg\min r(z)
z0=argminr(z)
for
k
=
1
,
2
,
.
.
.
,
K
k = 1,2,..., K
k=1,2,...,K do:
z
k
−
1
/
2
=
O
(
z
k
−
1
)
\quad z_{k-1/2}=\mathcal{O}(z_{k-1})
zk−1/2=O(zk−1)
z
k
=
P
r
o
x
z
k
−
1
α
(
g
(
z
k
−
1
/
2
)
)
\quad z_{k}=Prox^{\alpha}_{z_{k-1}}(g(z_{k-1/2}))
zk=Proxzk−1α(g(zk−1/2))
return z ˉ K = 1 K ∑ k = 1 K z k − 1 / 2 \bar{z}_{K}=\frac{1}{K}\sum_{k=1}^{K}z_{k-1/2} zˉK=K1∑k=1Kzk−1/2