Static Optimization: An Overview: Dr. Radhakant Padhi

Lecture 23
Static Optimization: An Overview

Dr. Radhakant Padhi
Asst. Professor
Dept. of Aerospace Engineering
Indian Institute of Science - Bangalore
Topics
z
Unconstrained optimization
Constrained optimization with equality

constraints
Constrained optimization with inequality

constraints
Numerical examples
ADVANCED CONTROL SYSTEM DESIGN

Dr. Radhakant Padhi, AE Dept., IISc-Bangalore
Unconstrained Optimization
Dr. Radhakant Padhi
Asst. Professor
Static Optimization
J2 (x)
Observation for: J1 ( x )
Point 1: Local maximum
Point 2: Point of inflexion
Point 3: Local minimum
Point 4: Local maximum
Point 3: Global minimum

Point 4: Global maximum
J1(x)
1
J1(x)
2
3
J2 (x)
xa
xb
dJ ( x)
=0
At all minima/maxima:
dx
Necessary and Sufficient

Conditions for Optimality
Scalar Case:
Performance Index J (x) : An analytic function x of
Taylor series:
2
dJ
1
d
J
2
J ( x + x ) J ( x ) =
x
x
+
(
)
2
dx *
2! dx x = x
x= x
Necessary Condition:
*
If J ( x ) is a minimum irrespective of the sign of x,
d
J
then
=0
d x x = x

Sufficient Condition:
2
1
d
J
J ( x + x ) J ( x ) =
2! dx 2
( x )
+ HOT
x = x
J ( x + x ) > J ( x ) , irrespective of the sign of x
dJ
>0
2
dx x=x
2
if
d 2J
Similarly, if
dx 2
(sufficiency condition for local minimum)
< 0, it leads to a local maximum

x = x

dJ
dx
Q-1: What if
Answer:
x = x
3
d
J
1
J ( x + x ) J ( x ) =
3! dx 3
Necessary condition
Sufficient condition
d 2J
= 2
dx
=0 ?
x = x
4
d
J
1
3
(x ) +
4
4
!
dx
x= x
d 3J
dx 3
d 4J
dx 4
x =x
x =x
(x ) +
4
x = x
=0
>0
(for minimization)


dJ
d 2J
= 2
Q-2: What if
dx x= x dx
x = x
=0
but
d 3J
dx3
0?
x = x
Then x = x is a point of inflexion

J = x4
Example 1:
dJ / dx = 4x3 = 0
x = 0,0,0
d2 J
dx2
= 12x2 = 0 ,
x =0
d3 J
dx3
= 24x = 0 ,
x* =0

d4 J
dx4
= 24 > 0
x* =0
minimum
8

Example 2:
J=x
dJ / dx = 3 x 2 = 0
x = 0, 0
d 2J
dx 2
= 6 x = 0 ,
x = 0
d 3J
dx 3
=60
x* = 0
Hence, x is a point of inflexion.


Vector case
J(X )
Minimize
where
By definition,
J
X
J
x
1
xn
2 J
X 2
2 J
x 2
1
2
J
x x
n 1

2 J
x1xn
2
J
xn2
10

J ( X ) = J ( X + X )
J
= J (X ) +
X
For minimization,
2
1
T J
X + (X ) 2
2!
X
X
J ( X + X ) J ( X ) > 0
X +
X*
( irrespective of sign of X )
Necessary Condition:
J = 0
X X
2 J
X 2 > 0 (positive definite)
Remark: Further Conditions are difficult to use in practice!

11

Example 1:
1 2
J ( X ) = ( x1 + x22 )
2
Necessary Condition
J = 0
X X
0
x
1
1
= =
J
x
2 0
x
2 X*
12

2 J
X 2
X*
2 J
x 2
1
= 2
J
x2 x1
2 J
x1x2
1 0
=
0
1
J
2
x2 X *
Eigenvalues: 1,1 at X = X
2
0 is a minimum
(positive definite). So X =
>
0
2
0 point
X X


13

Example 2:
Solution:
J ( x ) = 1 x12 x22
2
J
=0
X
0
x
1
X = * =
x2 0
2 J 1 0
Eigenvalues: 1, 1
=
X 2 0 1
2 J
is neither positive definite, nor negetive definite
i.e.
2
X
Hence X = 0 is a 'saddle point'.
14
Constrained Optimization with

Equality Constraints
Dr. Radhakant Padhi
Asst. Professor
Constrained Optimization:
Equality Constraint
Problem: Minimize
Subject to
J (X )
(X )
n
f (X ) = 0
where, f ( X ) = f1 ( X )
f m ( X )
T
Solution Procedure:
Formulate an augmented cost function
J ( X , )
J ( X ) + T f ( X )

16
Constrained Optimization:
Equality Constraint
Necessary Conditions:
T
J
J f
=0
=
+
X
X X
n equations
J
= f (X ) = 0
m equations
Hence, it lead to ( n + m ) equations

with ( n + m ) variables. Solve it!
17

Equality Constraint: An Example
Minimize
Subjected to:
Solution:
1 2 2
x1 + x2
2
f ( X ) = x1 + x2 2 = 0
J (X )=
J X = 1 x12 + x22 + x1 + x2 2
J / x1 x1 + 0
1
x

1
J / x2 = x2 + = 0 , x = = 1

J / x1 + x2 2 0 2

18

Equality Constraint: Another Example
Minimize
Subject to
2
2
1 x1 x2
J ( X ) = +
2 a b
x1 + mx2 c = 0
where a, b, m, c are Constants
Solution:
x
1 x
J = + + ( x1 + mx2 c )
2 a b
2
1
2
2
2
2

19

Equality Constraint: Another Example
Solve:
0
x
J
1
+
a

x1
x2

J
0
+
2
=
=
b
x2

J
x + mx c 0
2
1

( x1 , x2 , )
2
2
ac
b mc *
c
x1 = 2
, x2 = 2
, = 2
2 2
2 2
2 2
a +m b
a +m b
a +m b
Remark: has no physical meaning. It only helps to solve the problem.

20

Equality Constraint: Sufficiency Condition
If the equation
T
2 J
f
2 I
X = 0
X
f

0
X
has only positive roots
Minimum
has only negative roots
Maximum

21
Example 1
Problem:
Solution:
J=
(x
1
2
2
1
+x
2
2
),
f ( x1 , x2 ) = x1 x2 5 = 0
J = 12 ( x 21 + x22 ) + ( x1 x2 5 )
Necessary condition:
x1 + 0
x = 0
2

x1 x2 5 0
= 52 , x1 = 52 , x2 = 52
22
Example 1
Sufficient condition:
2 J 1 0
=
,
2
X
0 1
0
1
det 0
1
1
1
The Solution
f 1
=
X 1
1
1 = 0
x1 = 52 , x2 = 52
=1> 0
is a minimum.

23
Example 1: Some Remarks

In this example, x1, x2 and do not appear in
the equation for . Moreover, the solution is the
only solution. Hence, the result is global.
In general, however will be a function of
X and . Hence, various conclusions have to
be derived on case-to-case basis.

24
Example 2
Problem:
Solution:
J = x1 x22 ,
f ( X ) = x12 + x22 1 = 0
J = x1 x2 + ( x1 + x2 1)
2
Necessary condition:
1 + 2x1 0
2 x ( 1) = 0
22 2
x1 + x2 1 0
2
2
Sufficient condition:
=
2
X
0
0 f 2 x1
,
=
2( 1) X 2 x2

25
Example 2
2
0
2x1
det 0
2 2 2x2 = 0
2x
x
2
0
1
2
x1
x2
1
-1
-1/2
0
0
1.73
-1/2
1/2
1
-3
-1
3/2
Conclusion
Maximum
Maximum
Minimum
-1/2
-1.73
3/2
Minimum

26

Inequality Constraints
Dr. Radhakant Padhi
Asst. Professor

Inequality Constraints: A nave approach
Remark: One way of dealing with inequality
constraints for the variables is as follows:
Let ximin xi xi
max
Replace: xi = xi
min
Consider
(Important for control problems)
+ ximax xi
min
sin 2 i
as a free variable.

28
Optimization with
n
R
R
Maximize
/
Minimize:
J
X
,
X
(
)
Problem:
Subject to:
g(X )
g1 ( X ) 0

g m ( X ) 0
Solution: First, introduce "slack variables" 1 , , m to convert

ineqaulity constraints to equality constraints as follows:
fg ( X , )
g1 ( X ) + 12 0
=
g m ( X ) + m2 0
Then follow the routine procedure for the equality constraints.

29
Optimization with
Augmented PI: J ( X , , ) = J ( X ) +
g
X
(
)
j j
j j
j =1
Necessary Conditions:
g j
J J m
=
+ j
= 0, i = 1, , n
xi xi j =1 xi
( n equations )
J
= g j ( X ) + 2j = 0,
j
j = 1, , m
( m equations )
J
= 2 j j = 0,
j
j = 1, , m
( m equations )

30
Optimization with
J
= g j ( X ) + 2j = 0
j
g j ( X ) = 2j
j g j = j ( j j )
J
= 2 j j = 0
j
Hence, j g j = 0
This leads to the conclusion that

either j = 0 or g j = 0
i.e.
If a constraint is strictly an inequality
constraint, then the problem can be
solved without considering it.
Otherwise, the problem can be solved
by considering it as an equality
constraint.

31
Necessary Conditions
(Kuhn-Tucker Conditions)
g j
J J m
=
+ j
= 0, i = 1, , n
xi xi j =1 xi
j g j ( X ) = 0,
j = 1, , m
( n equations )
( m equations )
For J ( X ) to be MINIMUM
For J ( X ) to be MAXIMUM
if g j ( X ) 0 then j 0
(opposite sign)
(same sign)

32
Optimization with
Inequality Constraints: Comments
z
One should explore all possibilities in the

Kuhn-Tucker conditions to arrive at an
appropriate conclusion
Kuhn-Tucker conditions are only

necessary conditions
Sufficiency check demands the concept

of convexity
33
Convex/Concave Function f (x)

z
A function is called convex, if

a straight line drawn between
any two points on the surface
generated by the function lies
completely above or on the
surface.
If the line lies strictly above the
surface, then the function is
called strictly convex.
f ( x)
x
Strictly convex
f ( x)
If the line lies below the

surface, then the function is
called a concave.
x
Strictly concave

34
Result for Local

Convexity/Concavity of f (X ) at X*
2 f
2
X X *
Definition
Strictly convex
Convex
Strictly concave
Concave
No classification
Eigenvalues
Positive definite
i > 0, i
Positive
Semi-definite
i 0, i
Negative definite
i < 0, i
Negative
Semi-definite
i 0, i
Indefinite
Some i > 0. Rest are 0

35
Conditions for which Kuhn-Tucker

Conditions are also Sufficient
Condition
J (X )
All g j ( X )
Maximum
Strictly concave
Convex
Minimum
Strictly convex
Convex

36
Example
Problem:
Minimize: J ( X ) = ( x12 + x22 )

Subject to:
Solution:
( x1 x2 ) 5
( x1 x2 ) 1
g1 ( X ) = ( x1 x2 5 ) 0
g 2 ( X ) = ( x1 + x2 + 1) 0
J = ( x12 + x22 ) + 1 ( x1 x2 5 ) + 2 ( x1 + x2 + 1)
37
Example: Kuhn-Tucker Conditions

J
= 2 x1 + 1 2 = 0
x1
1 ( x1 x2 5 ) = 0
J
= 2 x2 1 + 2 = 0
x2
2 ( x1 + x2 + 1) = 0
( x1 x2 5) 0
( x1 + x2 + 1) 0
Note: x2 = x1
1 0
2 0
All possible solutions should be investigated

38
Feasible Solution of
Kuhn-Tucker Conditions
1
1
z Case 1: 1 = 0, 2 0, Feasible: x1 = , x2 =
2
2
z
Case 2: 1 = 0, 2 = 0, Not Feasible: x1 = x2 = 0

5
5
1 0, 2 = 0, Not Feasible: x1 = , x2 =
2
2
Case 3:
Case 4: 1 0, 2 0, Not Feasible: No Solution !

39
Sufficiency condition
J ( X ) = ( x12 + x22 ) is strictly convex. g1 ( X ) , g 2 ( X ) are also convex.
Hence, the Kuhn-Tucker conditions are both Necessary and Sufficient.
2 J 2 0
Moreover,
=
> 0 and it doesnot depend on the value of X .
2
X
0 2
Hence, X * = [1/ 2 1/ 2] is the GLOBAL minimum!
T

40
References
z
T. F. Elbert, Estimation and Control Systems,

Von Nostard Reinhold, 1984.
S. S. Rao, Optimization Theory and

Applications, Wiley, Second Edition, 1984.

41

42

Static Optimization: An Overview: Dr. Radhakant Padhi

Uploaded by

Copyright:

Available Formats

Static Optimization: An Overview: Dr. Radhakant Padhi

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Static Optimization: An Overview: Dr. Radhakant Padhi

Uploaded by

Copyright:

Available Formats

Lecture 23

Static Optimization: An Overview

Constrained optimization with equality

Constrained optimization with inequality

ADVANCED CONTROL SYSTEM DESIGN

Point 3: Global minimum

Necessary and Sufficient

Necessary and Sufficient

J ( x + x ) > J ( x ) , irrespective of the sign of x

(sufficiency condition for local minimum)

< 0, it leads to a local maximum

Necessary and Sufficient

ADVANCED CONTROL SYSTEM DESIGN

Necessary and Sufficient

Then x = x is a point of inflexion

ADVANCED CONTROL SYSTEM DESIGN

Necessary and Sufficient

Hence, x is a point of inflexion.

Necessary and Sufficient

ADVANCED CONTROL SYSTEM DESIGN

Necessary and Sufficient

Remark: Further Conditions are difficult to use in practice!

Necessary and Sufficient

Necessary and Sufficient

ADVANCED CONTROL SYSTEM DESIGN

Necessary and Sufficient

Constrained Optimization with

ADVANCED CONTROL SYSTEM DESIGN

Hence, it lead to ( n + m ) equations

Constrained Optimization with

ADVANCED CONTROL SYSTEM DESIGN

Constrained Optimization with

ADVANCED CONTROL SYSTEM DESIGN

Constrained Optimization with

Remark: has no physical meaning. It only helps to solve the problem.

Constrained Optimization with

has only positive roots

has only negative roots

ADVANCED CONTROL SYSTEM DESIGN

ADVANCED CONTROL SYSTEM DESIGN

Example 1: Some Remarks

ADVANCED CONTROL SYSTEM DESIGN

ADVANCED CONTROL SYSTEM DESIGN

ADVANCED CONTROL SYSTEM DESIGN

Constrained Optimization with

Constrained Optimization with

(Important for control problems)

ADVANCED CONTROL SYSTEM DESIGN

Solution: First, introduce "slack variables" 1 , , m to convert

Then follow the routine procedure for the equality constraints.

ADVANCED CONTROL SYSTEM DESIGN

This leads to the conclusion that

ADVANCED CONTROL SYSTEM DESIGN

ADVANCED CONTROL SYSTEM DESIGN

One should explore all possibilities in the

Kuhn-Tucker conditions are only

Sufficiency check demands the concept

Convex/Concave Function f (x)

A function is called convex, if