Sie sind auf Seite 1von 52

Semidefinite Programming

Application:
Truss Topology Design
Jean-Christophe Lilot
Ting Ting Ren
CAS735, Friday 2
nd
& 9
th
April 2004
Outline
Basic TTD problem (Assignment 2)
Limitation of the basic TTD
Standard Static TTD model
SemiDefinite Reformulation Primal P
Robust model leads to Universal model Pr
Simplification Dl of the Dual D of Pr
Comparison of Pr
+
, dual of Dl, and Pr
Basic TTD model:
Presentation
Given:
A set of tentative bars
A set of fixed nodes
An upper bound of the quantity of material
A load F to apply to the structure.
The property of the material
Design the most stable truss topology
design
Basic TTD model:
Formulation
min
2
1
) ( = v f t Compl
T
f
f v t A t s = ) ( . .
e s

=
n
i
i
t
1
0 > t
Limitations of the model
No geographical limitations
Not enough material limitations
Can this model be called robust?
Geographical limitation
h
A B
3
Material limitation
What we have
Upper and lower bounds on the total
amount of material
What we need
Upper bound on the bar diameter
Lower bound on the bar diameter
Robustness: Problem
Compliance = 1.000
200
*
f
f =
*
f
Compliance = 8.4
Robustness: Solutions
Take into account several load
scenarios
Take into account all small occasional
loads that may apply to the system
Robustness:
New problems
Mathematical point of view
Engineering point of view
Robustness:
New solutions
The set of force F can be the
ellipsoidal envelope of the original
finite set of loads and a small ball
We use a two stage scheme.
First we dont include the ellipsoidal
envelope to get the important nodes
Then we treat the set of nodes actually
used by the preliminary truss as our new
nodal set
Robustness: Results
Euclidean norm of the ball not exceeding 10% of
the norm of the original force.
Old Design New Design
Compliance always less or equal to 1.03 within the
elipsoid.
Compliance with respect to the original load: 1.0024 so
only 0.24% larger than the optimal compliance

Standard Static TTD
Model: S.1.
The set kinematically admissible
displacement is a polyhedral set:


The system of linear inequalities

satisfies the Slater condition:
{ } | |
qm m
M R r Rv R v e s e = | v
r Rv s
r R t s < - v v . .
Standard Static TTD
Model: S.2.
The set T of admissible designs is given by
a single linear constraint on the vector t,
namely upper and lower bounds on t
i
and
an upper bound on the total material
ressource



with given parameters
)
`

s s s e = =

=
+
n
i
i i i i i n
t t S t t t t T
1
1
, , | ) ,..., ( e
e < < < s

=
n
i
i i i
1
, 0
Standard Static TTD
Model: S.3.
The set F of loading scenarios is either
a finite set (multiload structural
design)


or an ellipsoid (robust structural design)
{ }
k
f f F ,...,
1
=
{ } ] [ 1 |
mk T
M Q u u Qu f F e s = =
Standard Static TTD
Model: S.4.


Exclusion of rigid body motions of the
ground structure
n i t whenever b t b
i
n
i
T
i i i
,..., 1 , 0 0
1
= >

=

Standard Static TTD Model:
The variation principal
Proposition: [variationnal description of compliance]
Consider a ground structure (n, m, B
1
,, B
n
) along with a load ,
and let be a truss. Let us associate with these data the
quadratic form (The potential energy of the closed system)

of . The compliance is finite if and only if the form
below bounded on , and whenever it is the case, one has
m
R f e
( )
n
S t
+
e
m
R ve
v f v t A v v C
T T
f t
= ) (
2
1
) (
,
m
R
) ( min ) (
,
v C t Compl
f t
v
f
=
) (t Compl
f
Standard Static TTD Model:
The simple lemma
Lemma:
Let be a quadratic form on with symmetric

positive semidefinite matrix A. Then:
(i) The form below is bounded if and only if it attains its minimum;
(ii) The form attains its minimum if and only if the equation
is solvable, and if it is the case, the set of minimizers of the
form is exactly the set of solutions to the equation.
(iii) The minimum value of the form, if exists, is equal to
being (any) solution to .
v b Av v v A
T T
=
2
1
) (
m
R
b Av =
v v b
T
,
2
1

b Av =
Standard Static TTD Model:
The simple Lemma proof
(i) There are two possibilities:
(a) b is orthogonal to Ker A
if it is below bounded then it attains its
minimum
(b) b has a nonzero projection b onto Ker A
below unbounded
(ii) Since the form is convex and smooth, its
minimizers are exactly the same as its
critical points
(iii)
v b v b v b v A v b Av v
T T T T T
2
1
2
1
) ( = = =
Errata

Outline: Not Pl, (Pr+)

5
th
slide: Its not the primal (Pr) but
the dual (Dl)


Standard Static TTD
Model
With:
1. A ground structure, i.e.,
The space of virtual displacement along with its closed
convex subset of kinematically admissible displacements,
A collection .
2. A set of admissible designs
3. A set of loading scenarios,
Find which minimize the worst compliance:
m
R
V
{ }
n i i
b
,..., 1 =
{ } i S t t t T
i n
e =
+
, | ) ,..., (
1
m
R F c
T t e
T t v b t b v v f t Compl
n
i
T
i i i
T T
V v F f
F
e
(



=
e e
| min
2
1
sup sup ) (
1
SDO reformulation I
( ) ( )

)

s
(

e =

=
s e
+
t t v b t b v v f t Compl R R S f t C
n
i
T
i i i
T T
r Rv R v
f
m
n
m
1
:
2
1
sup ) ( | , ,
We need a SDR of the compliance:
SDO reformulation II
Proposition:
Let
( )
n
n
S t t t
+
e = ) ,..., (
1
and
m
R f e then the inequality t s ) (t Compl
f
is satisfied if and only if there exists a nonnegative vector of the
dimension q equal to the number of linear inequalities defining the set
( )
|
|
|
.
|

\
|
+
+
=

=
n
i
T
i i i
T
T T T
b t b R f
R f r
f t A
1
2 2
, , ,

t
t
of virtual displacements
such that the matrix:

is positive semidefinite.
Thus the epigraph of (regarded as a function of
and ) admits the SDR:
) (t Compl
f
( )
n
S t
+
e
m
R f e
0
,..., 1 , 0
0
2 2
1
>
= >
|
|
|
.
|

\
|
+
+

t
n i t
b t b R f
R f r
i
n
i
T
i i i
T
T T T

SDO reformulation III


Proposition proof:
t s
(


s e
v t A v v f
T T
r Rv R v
m
) (
2
1
sup
:

=
=
n
i
T
i i i
b t b t A
1
) ( t s ) (t Compl
f
if
(


e
v t A v v f t
T T
R v
f
m
) (
2
1
sup ) , ( |
) (
, ) ( ) (
2
1
. . 0
t Compl then
R v Rv r v t A v v f t s if
f
m T T T
>
e s + > -
t
t
) ( Rv r ) ( Rv r
T
+
SDO reformulation III
Proposition proof:
Lagrange duality:
Under the slater condition from S.1.
for properly chosen is exactly
the supremum of


) , ( | t
f
0 >
v t A v v f
T T
) (
2
1

| | ( ) v R f v t A v r v Q
some for t t Compl
T
T T T
f f
t
| t t
+ + + =
> > >
) (
2
1
) (
0 ) , ( ) (
Which is non negative
SDO reformulation III
The simple Lemma
Lema:
A quadratic inequality with a (symmetric) nxn matrix A
0 2 > + + c x b Ax x
T T
is trivially true is valid for all - if and only if the matrix
n
R xe
|
|
.
|

\
|
A b
b c
T
is positive semidefinite.
PSD is f t A if only and if v Q ) , , , ( 0 ) ( t >
SDO reformulation of the
multiload TTD problem
k l e
n i t d
t c
n i t b
k l
b t b R f
R f r
a
t s
l
i i i
n
i
i
i
n
i
T
i i i l
T
l
T
l
T
l l
T
,..., 1 , 0 ) (
,..., 1 , ) (
) (
,..., 1 , 0 ) (
,..., 1 , 0
2 2
) (
. .
min
1
1
= >
= s s
s
= >
=
|
|
|
.
|

\
|
+
+

=
=

t
t

Second Model:
The ellipsoid
n i t
t
n i t
b t b Q
Q I
t s
i i i
n
i
i
i
n
i
T
i i i
T
k
,..., 1 ,
,..., 1 , 0
0
2
. .
min
1
1
= s s
s
= >
|
|
|
.
|

\
|

=
=

e
t
t

{ } | |
mk T
M Q u u Qu f F e s = = 1 |
Robust Model (Pr):
Universal model!
0 ) (
,..., 1 , ) (
) (
,..., 1 , 0 ) (
,..., 1 , 0
] [
] [ 2
) (
. .
min
1
1
>
= s s
s
= >
=
|
|
|
.
|

\
|
+
+ + +

=
=
z e
n i t d
t c
n i t b
K l
b t b E z
E z D z D I
a
t s
i i i
n
i
i
i
n
i
T
i i i l l
T
l l l l p

e
c
c t
t

Link between the models


Multiload model:


Robust model:


, , 1 k K p = = ( )
q q k
R R z e = ... ,...,
1

l T
l l l
l T
l l
R f E z r D z D c + = + = + , 2
, 1 , = = K k p ? = z
Q E zI D z D
l k l l
= = = + , 0 , 2 c
The primal (Pr)
Dimension of (Pr):
Multiload design of M-node structure
Design dimention is O(M
2
)
K big LMIs with row size of 2M or 3M
Example: 9*9 Nodes (15*15 nodes)
M=81 (M=225)
3200 vars (25100 vars)
k LMIs of row size 160 or 240 (450 or 675)
From Primal (Pr)
to Dual (D
ini
) (step 0)
| |
| |
| |
| |
| |
| |
| |
| |

= = + +
= + +
=
e >
e >
e = >
e =
e e e =
|
|
.
|

\
|
+




=
+
=
=
+ +
= =
+
n i I b b
V D
Tr
R
R
R n i
S n i
M V S S K l
V
V
t s
V E D Tr
K
l
d i i i i l
T
i
K
l
l l l l
K
l
l
N
i i i i
D
i i
mp
l
m
l
p
l
l l
T
l l
k
l
n
i
i i i i l
T
l l l
,..., 1 , 0
0 2
1 ) ( 2
0
0
, ,..., 1 , 0 ,
,..., 1 , 0
, , ,..., 1 , 0
. .
) 2 ( max
1
1
* *
1
1 1
1
o o t |
q c o
o
q q

o o o o
t t
| o
|
o
e o o o |

{ } { } q t o o | o , , , , , , ,
1 1
n
i
i i i
K
l
l l l
V
=
+
=
With design variables:
Simplification:
From (D
ini
) to (D)

(step 1)

| |
| |
| |
| |
| |
| |
| |
| |

= = + +
= + +
=
e >
e >
e = >
e =
e e e =
|
|
.
|

\
|
+




=
+
=
=
+ +
= =
+
n i I b b
V D
Tr
R
R
R n i
S n i
M V S S K l
V
V
t s
V E D Tr
K
l
d i i i i l
T
i
K
l
l l l l
K
l
l
N
i i i i
D
i i
mp
l
m
l
p
l
l l
T
l l
k
l
n
i
i i i i l
T
l l l
,..., 1 , 0
0 2
1 ) ( 2
0
0
, ,..., 1 , 0 ,
,..., 1 , 0
, , ,..., 1 , 0
. .
) 2 ( max
1
1
* *
1
1 1
1
o o t |
q c o
o
q q

o o o o
t t
| o
|
o
e o o o |

{ } { } q t o o | o , , , , , , ,
1 1
n
i
i i i
K
l
l l l
V
=
+
=
With design variables:
Simplification:
From (D
ini
) to (D)

(step 1)
| |
| |
| |

= +
s +
=
>
= >
=
|
|
.
|

\
|
+ + +
+
=
=
=
+
= =
+



) ( ,..., 1 ,
) ( 0 2
) ( 1 ) ( 2
) ( 0
) ( ,..., 1 , 0 ,
) ( ,..., 1 , 0
. .
) 2 ( min
1
1
* *
1
1 1
f n i I b b
e V D
d Tr
c
b n i
a K l
V
V
t s
V E D Tr
d i i
K
l
i l
T
i
K
l
l l l l
K
l
l
i i
l l
T
l l
k
l
n
i
i i i i l
T
l l l
o o |
c o
o

o o
|
o
e o o o |

{ } { } o o | o , , , , ,
1 1
n
i
i i
K
l
l l l
V
=
+
=
With design variables:
Eliminating and , we get: q
{ }
n
i
i
1 =
t
Simplification:
From (D) to (D)

(step 2)
1. (D) is feasible
2. We can add the following constraint without
changing the opt value:
ending up with a new dual (D).
3. Now, note that if
is a feasible sol. to (D), then the collection

is also a feasible sol. with the same objective:
From the LMI (D.a) by the Schur complement Lema it
follows that , so that replacing by
we preserve validity for the LMIs (D.f) and (D.a)
) ( ,..., 1 , 0 g K l
l
= o
{ } { } { } { } ( ) o o | | o o , , , ,
1 1 1 1
n
i i
K
l l
K
l l
K
l l
V V
=

= = =
= = = =
{ } ( ) o o o | o | o , , ) , ( ) , ( , ,
1
1
K
l
T
l l l l
V V V V V
=

= =
l l
V | o | ) , (
l
| ) , ( V
l
o |
Simplification:
From (D) to (D)

(step 2)
| |
| |
| |

= +
s +
=
>
= >
=
|
|
.
|

\
|
+ + +
+
=
=
=
+
= =
+



) ( ,..., 1 ,
) ( 0 2
) ( 1 ) ( 2
) ( 0
) ( ,..., 1 , 0 ,
) ( ,..., 1 , 0
. .
) 2 ( min
1
1
* *
1
1 1
f n i I b b
e V D
d Tr
c
b n i
a K l
V
V
t s
V E D Tr
d i i
K
l
i l
T
i
K
l
l l l l
K
l
l
i i
l l
T
l l
k
l
n
i
i i i i l
T
l l l
o o |
c o
o

o o
|
o
e o o o |

{ } { } o o | o , , , , ,
1 1
n
i
i i
K
l
l l l
V
=
+
=
With design variables:
Simplification:
From (D) to (D)

(step 2)
| |
| |
| |

=
= +
s +
=
>
= >
+ + +
+
=

=
=
+
= =
+



) ( ,..., 1 , 0
) ' ( ,..., 1 ,
) ( 0 2
) ( 1 ) ( 2
) ( 0
) ( ,..., 1 , 0 ,
. .
) 2 ( min
1
1
1
* *
1
1 1
g K l
f n i I b V V b
e V D
d TR
c
b n i
t s
V E D Tr
l
d i i
K
l
i
T
l l l
T
i
K
l
l l l l
K
l
l
i i
k
l
n
i
i i i i l
T
l l l

o
o o o
c o
o

o o
e o o o |
{ } { } { } o o o o , , ,
1 1 1
n
i
i
K
l
l
K
l
l
V V
=

= =
= = = With design variables:
Eliminating , we get:
{ }
K
l
l
1 =
|
Simplification:
From (D) to (Dl)

(step n)
| |
| |

s +
=
>
= >
=
|
|
.
|

\
|
+
+ + +



=
=
+
+
= =
+
0 2
1 ) ( 2
0
,..., 1 , 0 ,
,..., 1 , 0
) ( ) (
) ( ) (
. .
) 2 ( min
1
* *
1
1 1
K
l
l l l l
K
l
l
i i
d i i i
T
i
k
l
n
i
i i i i l
T
l l l
V D
Tr
n i
N i
I V B
V B A
t s
V E D Tr
c o
o

o o
o o
o
e o o o |

{ } { } { } R R M V V S
n
i
i
K
l
p m
l
K
l
p
l
e e = e = e =
=

= =
o o o o , , ,
1 1
.
1
With design variables:
The Dual (Dl)
Both Primal and Dual are strictly
feasible both (Pr) and (Dl) are
solvable with equal optimal values and
with bounded level set.
The optimal values are negation of
each other.
Case of a simple bound:
From (Dl) to (Dl
sb
)

| |
| |

s +
=
>
= >
=
|
|
.
|

\
|
+
+ + +



=
=
+
+
= =
+
0 2
1 ) ( 2
0
,..., 1 , 0 ,
,..., 1 , 0
) ( ) (
) ( ) (
. .
) 2 ( min
1
* *
1
1 1
K
l
l l l l
K
l
l
i i
d i i i
T
i
k
l
n
i
i i i i l
T
l l l
V D
Tr
n i
N i
I V B
V B A
t s
V E D Tr
c o
o

o o
o o
o
e o o o |

{ } { } { } R R M V V S
n
i
i
K
l
p m
l
K
l
p
l
e e = e = e =
=

= =
o o o o , , ,
1 1
.
1
With design variables:
Case of a simple bound:
From (Dl) to (Dl
sb
)
| |

>
=
s +
=
|
|
.
|

\
|
+ +

=
=
=
0
1 ) ( 2
0 2
,..., 1 , 0
) (
) ( ) (
. .
) 2 ( min
1
1
* *
1

o
c o

o
e o |
K
l
l
K
l
l l l l
d i
T
i
k
l
l
T
l l l
Tr
V D
N i
I V B
V B A
t s
V E D Tr

{ } { } R M V V S
K
l
p m
l
K
l
p
l
e e = e =
= =
o o , ,
1
.
1
With design variables:
What we got!!!
Size
(assuming k = 10)
(Pr) (Pr)
M=9*9 (15*15)
(Dl
sb
) (Dl
sb
)
M=9*9 (15*15)
Design dimension 0.5 M
2
3200 (25100) 2 k M 810 (2250)
# and
size of LMIs
k
4 m
2
10
236000 (1800000)
0.5 M
2

k
2
3200 (25100)
100
# of linear
constraints
0.5 M
2
3200 (25100)

K+1 11
M = number of nodes
k = number of load scenarios
m = 2M in plan and 3M in space

(Pr) in O(M
6
)
(Dl
sb
) in O(k
3
M
3
)

Back to the Primal: (Pr
+
)
| | | | | | | |
0 ) (
,..., 1 , ) (
) ( ) (
,..., 1 , ) ( ) (
,..., 1 , 0
2
) (
. .
min
1 1
1
1
1 1
1 11
1 1 11
>
= + =
s
= s s
=
|
|
|
|
|
|
|
|
|
|
|
.
|

\
|
+ +

= =
=
z e
k l E z q b d
t Tr c
n i t Tr b
K l
t q
t q
t q
t q
q q q q D z D I
a
t s
l l
n
i
S
s
l
is is
n
i
i
i i i
n
l
nS
n
l
n
l
S
l
T
l
nS
T
l
n
T
l
S
T
l
l l p
c
e

t
t





Equivalence between
(Pr) and (Pr
+
)
Proposition:
A collection is a feasible solution to (Pr) if and only if it can
be extended by properly chosen
to a feasible solution to (Pr
+
)
{ } S s n i K l q
l
is
,..., 1 , ,..., 1 , ,..., 1 | = = =
{ } ( ) t , ,
1
z t
n
i i =
From (Pr
+
) to (Pr)
If part:

Let a collection
be a feasible sol to (Pr
+
)
{ } { } ( ) S s n i K l q z t
l
is
n
i i
,..., 1 , ,..., 1 , ,..., 1 | , , ,
1
= = =
=
t
K l s
| | | | 0 2 2
1 1
>
(

+ + + + +

= =
y b t b y y z E x x D z D I x
n
i
S
s
T
is i
T
is
T
T
l i
T
l l p
T
c t
K l
b t b E z
E z D z D I
n
i
T
i i i l l
T
l l l l p
,..., 1 , 0
] [
] [ 2
1
=
|
|
|
.
|

\
|
+
+ + +

c
c t
We want to prove the validity of the LMI (Pr.a):
For . We want:
From (Pr
+
) to (Pr)
Using (Pr
+
.d): we get:
| | | | 0 2 2
1 1
>
(

+ + + + +

= =
y b t b y y z E x x D z D I x
n
i
S
s
T
is i
T
is
T
T
l i
T
l l p
T
c t
k l E z q b
l l
n
i
S
s
l
is is
,..., 1 ,
1 1
= + =

= =
c
| |
| | | | y b y y t y y q x x D z D I x
y b t b y y q b x x D z D I x
T
is is
n
i
S
s
is i
T
is
n
i
S
s
is
T
l
is
T
l l p
T
n
i
S
s
T
is i is
T
T
n
i
S
s
l
is is
T
l l p
T
= + + + + =
(

+
(

+ + +


= = = =
= = = =
, 2 2
2 2
1 1 1 1
1 1 1 1
t
t
{ }
s i is
y
,
Which is the value of the quadratic form with the matrix (Pr
+
.a) at the
vector comprising x and , and therefore is non negative.
From (Pr) to (Pr
+
)
Only if part:

Let a be a feasible sol to (Pr). Let us fix
and set
{ } ( ) t , ,
1
z t
n
i i =
| |
(

= > + + + +

= =
n
i
S
s
T
is i is
T T
l
T
l l p
T
b t b t A y t A y y f x x D z D I x
1 1
) ( , 0 ) ( 2 2t
K l l s s 1 ,
l l l
E z f + =c
p
R xe
M
R y e
x f y t A
l
= ) (
x Y y
l
=
x x f x Y t A
l l
= , ) (
l l
f Y t A = ) (
For every the quadratic form of :


is nonnegative, i.e., the equation is solvable for every x.
We can chose its solution to be linear in x:
Then i.e.
From (Pr) to (Pr
+
)
Let set

then
| |
i is
T
l
T
l
is
t b Y q =
l l l
n
i
S
s
T
is i is
n
i
S
s
l
is is
f Y t A Y b t b q b = = =

= = = =
) (
1 1 1 1
k l E z q b
l l
n
i
S
s
l
is is
,..., 1 ,
1 1
= + =

= =
c Which proves (Pr
+
.d):
With
l l l
E z f + =c
From (Pr) to (Pr
+
)
{ }
is
y
p
R xe
| | | | , 0 2 2 }) { , (
1 1 1 1
> + + + + =

= = = =
n
i
S
s
is i
T
is
n
i
S
s
is
T
l
is
T
l l p
T
is
y t y y q x x D z D I x y x F t
We need to verify that for every collection of vectors of
appropriate dimension and for every we have:
Given x, let set
x Y b y
l
T
is is
=
*
minimize : the partial derivative with respect to is
We will prove the min of in is non negative. }) { , (
is
y x F } {
is
y
} {
*
is
y
) , ( x F
is
y
| | 0 2 2 2
*
= = + x Y b t x Y b t y t x q
l
T
is i l
T
is i is i
l
is
We need to prove that
| | | | , 0 2 2 }) { , (
1 1 1 1
*
> + + + =

= = = =
n
i
S
s
l
T
is i is
T
l
T
n
i
S
s
l
T
is
T
l
is
T
l l p
T
is
x Y b t b Y x x Y b q x x D z D I x y x F t
A More Robust Problem?
Back to the antenna design
Basic Problem
Linear Programming
A little bit more Robust:
Second Order Conic
Optimization
Even more Robust:
Semidefinite Optimization
A More Robust Problem?
What about the TTD?
Basic Problem
Second Order Conic
Optimization
A little bit more Robust:
Semidefinite Optimization
Even more Robust:
Must not be easily solvable!
Thank you for your attention
Have a good week-end!
References:
Course TextBook:
BenTal and A. Nemirovskii: Convex Optimization in
Engineering, SIAM Publications, 2001

Course slides: http://www.cas.mcmaster.ca/~coe774/

Das könnte Ihnen auch gefallen