Introduction
Mathematical Optimization
The basic form:
m
i
n
i
m
i
z
e
:
f
0
(
x
)
s
u
b
j
e
c
t
t
o
:
f
i
(
x
)
≤
b
i
minimize: f_{0}(x)\\ subject\ to: f_{i}(x)\le b_{i}
minimize:f0(x)subject to:fi(x)≤bi
.
x
=
(
x
1
,
.
.
.
,
x
n
)
:
o
p
t
i
m
i
z
a
t
i
o
n
v
a
r
i
a
b
l
e
s
x=(x_{1},...,x_{n}): optimization\ variables
x=(x1,...,xn):optimization variables
.
f
0
:
R
n
→
R
:
o
b
j
e
c
t
i
v
e
f
u
n
c
t
i
o
n
f_{0}: R^{n}\to R: objective\ function
f0:Rn→R:objective function(可以看作是用来寻找最优解的成本)
.
f
i
:
R
n
→
R
,
i
=
1
,
.
.
.
,
m
:
c
o
n
s
t
r
a
i
n
t
f
u
n
c
t
i
o
n
f_{i}: R^{n}\to R, i=1,...,m: constraint\ function
fi:Rn→R,i=1,...,m:constraint function
optimal solution
x
∗
x^{*}
x∗ has smallest value of
f
0
f_{0}
f0 among all vectors that satisfy the constraints.
Application:
. portfolio optimization(投资组合优化)
. device sizing in electronic circuits
. data fitting(数据拟合)
General optimization problem:
. very difficult to solve
. methods involve some compromise,
e
.
g
.
e.g.
e.g., very long computation time, or not always finding the solution.
Exception: certain problem classes can be solved efficiently and reliably
. least-squares problem
. Linear programming
. Convex Optimization
Least-square
The basic form:
m
i
n
i
m
i
z
e
:
∥
A
x
−
b
∥
2
2
=
∑
i
=
1
k
(
a
i
T
x
−
b
i
)
2
,
A
∈
R
k
×
n
minimize:\Vert Ax-b\Vert^{2}_{2}=\sum_{i=1}^{k}(a_{i}^{T}x-b_{i})^{2},\ A\in R^{k\times n}
minimize:∥Ax−b∥22=i=1∑k(aiTx−bi)2, A∈Rk×n
Solving least-squares problems:
. analytical solution:
x
∗
=
(
A
T
A
)
−
1
A
T
b
x^{*}=(A^{T}A)^{-1}A^{T}b
x∗=(ATA)−1ATb
. reliable and efficient algorithms and softwares
. computation time proportional to
n
2
k
n^{2}k
n2k(n is the size of X and K is the number of rows; In other words, K is the number of examples and n is the number of features)
. a mature technology
Using least-squares:
. least-squares problems are very easy to recognize
. a few standard techniques increase flexibility(
e
.
g
.
,
i
n
c
l
u
d
i
n
g
w
e
i
g
h
t
s
,
a
d
d
i
n
g
r
e
g
u
l
a
r
i
z
a
t
i
o
n
t
e
r
m
s
e.g., including weights, adding regularization terms
e.g.,includingweights,addingregularizationterms)
Linear Programming
The basic form:
m
i
n
i
m
i
z
e
:
c
T
x
s
u
b
j
e
c
t
t
o
:
a
i
T
x
≤
b
i
,
i
=
1
,
2
,
.
.
.
,
m
minimize:\ c^{T}x\\ subject\ to:\ a_{i}^{T}x\le b_{i},\ i=1,2,...,m
minimize: cTxsubject to: aiTx≤bi, i=1,2,...,m
Solving linear program:
. no analytical formula for solution
. reliable and efficient algorithm and software
. computation time proportional to
n
2
m
n^{2}m
n2m if
m
≥
n
m\ge n
m≥n
. a mature technology
Using linear programming
. not as easy to recognize as least-squares problems
. a few standard tricks used to convert problems into linear programs (
e
.
g
.
e.g.
e.g., problems involving
l
1
−
l_{1-}
l1− or
l
∞
−
l_{\infty-}
l∞−norms, piecewise-linear functions)
Application
Convex Optimization
The basic form:
m
i
n
i
m
i
z
e
:
f
0
(
x
)
s
u
b
j
e
c
t
t
o
:
f
i
(
x
)
≤
b
i
,
i
=
1
,
2
,
.
.
.
,
m
minimize: f_{0}(x)\\ subject\ to: f_{i}(x)\le b_{i},\ i=1,2,...,m
minimize:f0(x)subject to:fi(x)≤bi, i=1,2,...,m
. objective and constraint functions are convex:
f
i
(
α
x
+
β
y
)
≤
α
f
i
(
x
)
+
β
f
i
(
y
)
,
α
+
β
=
1
0
≤
α
≤
1
a
n
d
0
≤
β
≤
1
f_{i}(\alpha x+\beta y)\le \alpha f_{i}(x)+\beta f_{i}(y),\ \alpha+\beta=1\\ 0\le \alpha \le1 \ and\ 0\le\beta \le1
fi(αx+βy)≤αfi(x)+βfi(y), α+β=10≤α≤1 and 0≤β≤1
(The least-squares and linear programming is special classes of convex problem. linear problem is like a bowl-shaped thing.If you slice it, you will get ellipsoids(椭圆) and it kind of goes down like that.)
Example
注:1、当最小二乘法使用了包含权重系数的时候,我的理解是:先利用基础的最小二乘法,求得所有的
P
j
P_{j}
Pj,然后将所有的P_{j}中接近或者超过
P
m
a
x
P_{max}
Pmax的
w
j
w_{j}
wj设置的很高,然后再求的这一次的
P
j
P_{j}
Pj。同理,按照这样的iteration,求得合适的
P
j
P_{j}
Pj,最后使得所有的
P
j
P_{j}
Pj在
P
m
a
x
/
2
P_{max}/2
Pmax/2附近“fluctuation”。
2、利用线性规划的求解方法,与上面所提到的切比雪夫估计所使用的方法一致,即式子中包含了
l
∞
范
数
l_{\infty}范数
l∞范数。
Course Goals
注:对于第二条,不需要自己对写代码实现,可以直接调用现有的Matlab中的CVX函数库