(problem decomposition)
TRANSCRIPT
-
7/27/2019 (Problem Decomposition)
1/40
1 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Anas Alfaris
Decomposition and Coupling
Lecture 4
Multidisciplinary System
Design Optimization (MSDO)
-
7/27/2019 (Problem Decomposition)
2/40
2 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Todays Topics
Information Flow and Coupling
MDO frameworks
Single-Level (Distributed analysis)
Multi-Level (Distributed design)
Collaborative Optimization
Analytical Target Cascading
(Hierarchical Decomposition & Multi-Domain Formulation)
-
7/27/2019 (Problem Decomposition)
3/40
3 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Standard Optimization Problem
Given
min ( )s.t. ( ) 0
J xg x
Solve the problem
* *
That is, find s.t. J( ) ( ), dom( ) dom( )x x f x x J g
Optimization Engine
Function Evaluator
x( )J x
( )g x
0x*
x
mn
zn
n
g
J
x
:
:
-
7/27/2019 (Problem Decomposition)
4/40
4 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Information Flow
A
B
A B
A
B
A
B
A B
A
B
AA
B
A B
B
A
B
A B
A
B
Dependent Tasks (Series) Independent Tasks (Parallel) Interdependent Tasks (Coupled)
Image by MIT OpenCourseWare.
-
7/27/2019 (Problem Decomposition)
5/40
5 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Information Flow
C
D
A
B
G
H
E
F
K
L
I
J
C DAB GHEFK L IJ
FeedForward
FeedBackward
-
7/27/2019 (Problem Decomposition)
6/40
6 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Information Flow
C
D
A
B
G
H
E
F
K
L
I
J
C DAB GHEFK L IJ
Sequential
Parallel
Coupled
-
7/27/2019 (Problem Decomposition)
7/40
7 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Advantages of Decoupling
Computation ofg(x) can be very time consuming, want
to divide the work and compute in parallel.
For example, if 1 21 2 1 2
1 1 2 2
( , ), where ,
and g( ) ( ( ), ( ))
n nx x x x x
x g x g x
Then g1 and g2can be computed in parallel. Graphically,
Optimizer
SS1 SS2
1x 1g 2g 2x SS1
SS2
Optim1x 2x
1g
2g
RR
-
7/27/2019 (Problem Decomposition)
8/40
8 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Coupling
The decoupled constraints assumption is not general. Subsystems
can be coupled and loops can arise. For example,
Optimizer
SS1 SS2
1x 2x
1u 2u
2w1w
x: decision variables
w: SS outputs (constraint, cost)
u: SS input (dependent)
SS1
SS2
Optim
vline: SS input
hline: SS output
1w
2w1u
1x
2u
2x
1
w
2w
Loop
Computation ofw1 and
w2 requires an iterative method.
-
7/27/2019 (Problem Decomposition)
9/40
9 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Coupling
An example where such a loop happens is as follows:
1 2
1 1 1 2 2 1
2 2 2 1 1 2
min ( , )
( , ( , )) 0s.t.
( , ( , )) 0
J x x
w g x g x w
w g x g x w
1 21 2where , , : , 1,2n n i i i ix x g x u w i
w1
and w2
satisfy coupled relations at each optimization iteration.
At each constraint evaluation, nonlinear equations must be solved
(e.g. by Newtons method) in order to obtain w1
and w2, which can
be time consuming.
Want a way to return to the situation of decoupled constraints.
R R
-
7/27/2019 (Problem Decomposition)
10/40
10 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Surrogate Variables (Tearing)
Information loop can be broken by introducing surrogate variables.
1 2
1 1 1 2 2 1
2 2 2 1 1 2
min ( , )
( , ( , )) 0s.t.
( , ( , )) 0
J x x
w g x g x w
w g x g x w
u1
and u2
are decision variables acting as the inputs to
g1(SS1) and g2 (SS2). Introducing surrogate variables
breaks information loop but increases the number of
decision variables.
1 2
1 1 1
2 2 2
2 1 1 1
1 2 2 2
min ( , )s.t.
( , ) 0
( , ) 0
( , ) 0
( , ) 0
J x x
g x u
g x u
u g x u
u g x u
-
7/27/2019 (Problem Decomposition)
11/40
11 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Numerical Example
1 2
1
2
2 2
1 1 2
2 2
2 3 4
3 3
1 1 2 2
3 3
2 3 4 1
min
s.t. 0
0
where
( 3) ( 4)
2
2
J J
w
w
J x x
J x x
w x x w
w x x w
2 2 2 2
1 2 3 4
1 1 1 2 3 4
2 2 1 2 3 4
min ( 3) ( 4)
s.t. ( , , , ) 0
( , , , ) 0
x x x x
w g x x x x
w g x x x x
coupled
2 2 2 2
1 2 3 4
3 3
1 1 2 5
3 3
2 3 4 6
3 3
1 2 5 6
3 33 4 6 5
min ( 3) ( 4)
s.t. 2 0
2 0
2 0
2 0
x x x x
w x x x
w x x x
x x x x
x x x x
decoupled
Solution:
1 23 3(0,0, 4,3,12 , 24 )x
MATLAB 5.3
coupled: 356,423 FLOPS 4.844s
uncoupled: 281,379 FLOPS 0.453s
-
7/27/2019 (Problem Decomposition)
12/40
12 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Single-level and Multi-Level Frameworks
Single-level
(Distributed Analysis)-disciplinary models provide
analysis
-all optimization done at
system level
non-hierarchical
decomposition
Multi-level
(Distributed Design)
-provide disciplinary
models with design tasks-optimization at
subsystem and system
levels
hierarchical
decomposition
-
7/27/2019 (Problem Decomposition)
13/40
13 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Single-level (Distributed Analysis)
Disciplinary models provide analysis
Optimization is controlled by some overseeing codeor database
e.g. ISight (Optimizer)
SystemOptimizer
Shared data
Local data
Structures
Local data
Aero
Optimizer
design variables
constraints
iSight
x J(x),g(x),h(x)
subsystemanalyses
-
7/27/2019 (Problem Decomposition)
14/40
14 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Optimizer
objectivedesign variables
constraints
x J(x)
performance
analysis
aerodynamic
analysisstructural
analysis
x g(x) h(x) x
g(x)
h(x)
During the optimization, the overseeing code keeps track of thevalues of the design variables and objective
The values of the design variables are changed according to
the optimization algorithm
Disciplinary models are asked to evaluate constraints/objective
Single-level (Distributed Analysis)
-
7/27/2019 (Problem Decomposition)
15/40
15 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Multi-level Optimization methods distribute decision
making throughout the system
Subsystem level models are provided with design
tasks
Optimization is performed at a subsystem level in
addition to the system level
Provide some autonomy to design groups and
reduces communication requirements.
Multi-level (Distributed Design)
-
7/27/2019 (Problem Decomposition)
16/40
16 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Multi-level (Distributed Design)
System level optimizer
SS1
optimizer
SS2
optimizer
SSN
optimizer
SS1
analyzer
SS2
analyzer
SSN
analyzer
command/resultcommand/result
command/result
Subsystem
black box (BB)
-
7/27/2019 (Problem Decomposition)
17/40
17 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Collaborative Optimization
Collaborative Optimization (CO)
disciplinary teams satisfy local constraints while
trying to match target values specified by a
system coordinator
preserves disciplinary-level design freedom.
CO is used typically to solve discipline-based
decomposed system optimization problems.
-
7/27/2019 (Problem Decomposition)
18/40
18 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Collaborative Optimization
OPTIMIZER
TARGET STATECoupled
Uncoupled
-
7/27/2019 (Problem Decomposition)
19/40
19 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Collaborative Optimization
Two levels of optimization:
A system-level optimizer provides a set of targets.
These targets are chosen to optimize the system-level
objective function
A subsystem optimizer finds a design that minimizes the
difference between current states and the targets.
Subject to local constraints
-
7/27/2019 (Problem Decomposition)
20/40
20 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Collaborative Optimization
0
k
min
wrt: target variabless.t. 0 subproblems
sys
k
J
J
x
2
1
target local
variables variables
local variables
local constraints
min
s.t.
-J
x
0x
1
J
2target local
variables variables
local variables
local constraints
min
s.t.
-kJ
x
0xkJ
analysis for
subsystem 1
analysis for
subsystem k
x xcomputedresults
computed
results
performance
analysis
0x
sysJ
-
7/27/2019 (Problem Decomposition)
21/40
21 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
CO Subsystem Level
2
1 target localvariables variables
local variables
local constraints
min
s.t.
-J
x
The subsystem optimizer modifies local variables to
achieve the best design for which the set of local
variables and computed results most nearly matches the
system targets
The local constraints must also be satisfied
-
7/27/2019 (Problem Decomposition)
22/40
22 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
CO System Level
System-level optimizer changes target variables toimprove objective and reduce differences Jk
Jk=0are called compatibility constraints
compatibility constraints are driven to zero, but may
be violated during the optimization
0
k
min
wrt: target variables
s.t. 0 subproblems
sys
k
J
J
x
-
7/27/2019 (Problem Decomposition)
23/40
23 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
CO Example: Aircraft Design
Consider a simple aircraft design problem:
maximize range for a given take-off weight by choosing
wing area, aspect ratio, twist angle, L/D, and wing weight.
aero
struct
perfmodified from Kroo et al. AIAA 94-4325
wing area, S
aspect ratio,AR
twist angle,
range, R
L/D
wing weight,
W
-
7/27/2019 (Problem Decomposition)
24/40
-
7/27/2019 (Problem Decomposition)
25/40
25 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Collaborative Optimization
0
k
min
wrt: target variables
s.t. 0 subproblems
sys
k
J
J
x
1
2
2
coupling local
variables variable
target local
variables variables
local variables
local constraints
s
min
s.t.
-
-J
x
0x
1J 0xkJ
analysis for
subsystem 1
x xcomputedresults
computed
results
performance
analysis
0x
sysJ
1ky
1ky
analysis for
subsystem k
2
2coupling local
variables variab
target local
variables variables
local variables
local constra
le
in s
s
t
min
s.t.
-
-kJ
x
-
7/27/2019 (Problem Decomposition)
26/40
26 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Collaborative Optimization
x0 = system-level target variable values
x = subsystem local variables
yij = coupling functions
yij =outputs of subsystemjwhich are needed as inputs to
subsystem i.
Coupling equations must also be satisfied, so coupling
variables are included in subsystem objective.
Used to reduce the number of system-level parameters.
-
7/27/2019 (Problem Decomposition)
27/40
27 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Analytical Target Cascading
ATC was initially developed as a product development
tool to cascade system-level product targets through a
hierarchy of design groups
ATC is typically used to solve object-based decomposed
system optimization problems
The ATC paradigm is based on hierarchical organizational
and analysis structures
ATC approach is to take a high-level system analysis and
use more detailed subsystem analyses at the lower levels.
-
7/27/2019 (Problem Decomposition)
28/40
28 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Analytical Target Cascading
SS1
SS2 SS3
R11
Y11
X11
R21 R22
R34R33R32R31
Level 2
Level NL
Level 1
Y22
X22
Y21
X21
.
.
..
.
.
.
.
.
.
.
.
.
.
.
Image by MIT OpenCourseWare.
-
7/27/2019 (Problem Decomposition)
29/40
29 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Analytical Target Cascading
ATCs mathematical formulation is similar to CO
although they were developed with different motivations.
Bottom level problems have the most design freedom.
Many possible solutions can exist that both match
targets while satisfying local design constraints.
At higher levels design freedom is progressively
reduced, until it is a minimum at the top level.
-
7/27/2019 (Problem Decomposition)
30/40
30 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Analytical Target Cascading
-
7/27/2019 (Problem Decomposition)
31/40
31 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Analytical Target Cascading
Linking variables y: Quantities that are input to more than one
subspace. These could be either shared variables or couplingvariables.
Local decision variablesx: Variables that a particular subspace
determines the value of.
Responses R: Values generated by subspaces required as inputs
to respective parent subspaces.
Targets T: Values set by parent subspaces to be matched by the
corresponding quantities from child subspaces.
R and y: allowable compatibility tolerance.
Hierarchical Decomposition &
-
7/27/2019 (Problem Decomposition)
32/40
32 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Hierarchical Decomposition &
Multi-Domain Formulation
Decomposition
Courtesy of Anas Alfaris. Used with permission.
Hierarchical Decomposition &
-
7/27/2019 (Problem Decomposition)
33/40
33
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Decomposition
L1
Hierarchical Decomposition &
Multi-Domain Formulation
Courtesy of Anas Alfaris.Used with permission.
Hierarchical Decomposition &
-
7/27/2019 (Problem Decomposition)
34/40
34
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Building System
Decomposition
Hierarchical Decomposition &
Multi-Domain Formulation
Courtesy of Anas Alfaris. Used with permission.
Hierarchical Decomposition &
-
7/27/2019 (Problem Decomposition)
35/40
35
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Formulation
Hierarchical Decomposition &
Multi-Domain Formulation
Courtesy of Anas Alfaris. Used with permission.
Hierarchical Decomposition &
-
7/27/2019 (Problem Decomposition)
36/40
36 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
A
0 0 0 0 0
1 0 0 0 0
1 0 0 0 0
0 0 1 0 0
0 1 0 0 0
A2
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
1 0 0 0 0
1 0 0 0 0
A3
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
V An
n 1
4
0 0 0 0 0
1 0 0 0 0
1 0 0 0 0
1 0 1 0 0
1 1 0 0 0
j1
Nai j
i 1
N
i
1
Nai j
j 1
N
Where:
j is an indicator of the fraction oftotal elements to which element j
provides input,
i is the fraction of total elements
on which element idepends, and
aij is an element of a matrix that
can be the DSM, a power of the
DSM, or the V matrix.
Formulation
Hierarchical Decomposition &
Multi-Domain Formulation
Hierarchical Decomposition &
-
7/27/2019 (Problem Decomposition)
37/40
37 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Formulation
Hierarchical Decomposition &
Multi-Domain Formulation
Courtesy of Anas Alfaris. Used with permission.
Hierarchical Decomposition &
-
7/27/2019 (Problem Decomposition)
38/40
38
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Formulation
Hierarchical Decomposition &
Multi-Domain Formulation
Courtesy of Anas Alfaris. Used with permission.
-
7/27/2019 (Problem Decomposition)
39/40
39 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
References
I.P. Sobieski and I.M. Kroo. Collaborative Optimization Using Response SurfaceEstimation. AIAA Journal Vol. 38 No. 10. Oct 2000.
Erin J. Cramer et al. Problem Formulation for Multidisciplinary Optimization.
SIAM Journal of Optimization. Vol. 4, No. 4 pp. 754-776, Nov 1994.
Kim, H.M., Michelena, N.F., Papalambros, P.Y., and Jiang, T., "Target Cascading
in Optimal System Design," Transaction of ASME: Journal of Mechanical
Design, Vol. 125, pp. 481-489, 2003
-
7/27/2019 (Problem Decomposition)
40/40
MIT OpenCourseWarehttp://ocw.mit.edu
ESD.77 / 16.888 Multidisciplinary System Design OptimizationSpring 2010
For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.
http://ocw.mit.edu/http://ocw.mit.edu/termshttp://ocw.mit.edu/termshttp://ocw.mit.edu/