Multidisciplinary System Design Optimization (MSDO)

Decomposition and Coupling Lecture 4

Anas Alfaris

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)

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Standard Optimization Problem

Given

x J: g:

n n n z m

x0

Optimization Engine

Solve the problem

min J ( x) s.t. g ( x) 0
3

J ( x) g ( x)

Function Evaluator

That is, find x* s.t. J( x* )

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

f ( x),

x dom( J )

dom( g )

Information Flow

A A B B
Dependent Tasks (Series) A B A A B B Independent Tasks (Parallel) A B A A B B

A B
Interdependent Tasks (Coupled) A B A A B B
Image by MIT OpenCourseWare.

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Information Flow
B C A K L J F I E D H G B C A K L J F I E D H G

Feed Forward

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Feed Backward

Information Flow
B C A K L J F I E D H G B Sequential C Parallel A K L Coupled J F I E D H G

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Advantages of Decoupling
Computation of g(x) can be very time consuming, want to divide the work and compute in parallel. For example, if x ( x1 , x2 ), where x1 Rn1 , x2 R n 2

and g( x) ( g1 ( x1 ), g 2 ( x2 ))
Then g1 and g2 can be computed in parallel. Graphically, Optimizer

Optim

x1
SS1

x2

x1

g1g 2
SS1 SS2

x2

g1
g2

SS2

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

Optim

x1
SS1

x2
u2
SS2

x1
SS1

w1 w2
SS2

x2
w2

w1

w1
w2

u1
Loop

u1

u2

x: decision variables vline: SS input w: SS outputs (constraint, cost) hline: SS output u: SS input (dependent) Computation of w1 and w2 requires an iterative method.
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Coupling
An example where such a loop happens is as follows:
min J ( x1 , x2 ) w1 s.t. w2 g1 ( x1 , g 2 ( x2 , w1 )) 0 g 2 ( x2 , g1 ( x1 , w2 )) 0

where x1 R n1 , x2

R n 2 , gi : xi ui

wi , i 1, 2

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.

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Surrogate Variables (Tearing)

Information loop can be broken by introducing surrogate variables.

min J ( x1 , x2 )

min J ( x1 , x2 ) w1 s.t. w2 g1 ( x1 , g 2 ( x2 , w1 )) 0 g 2 ( x2 , g1 ( x1 , w2 )) 0

s.t. g1 ( x1 , u1 ) 0 g 2 ( x2 , u2 ) 0 u2 g1 ( x1 , u1 ) 0 0 u1 g 2 ( x2 , u2 )

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.
10 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Numerical Example
min J1 s.t. w1 w2 where J1 J2 w1 w2 J2 0 0 x
2 1

decoupled
2 2 2

min x12 s.t. w1 w2

2 x2 ( x3 3) 2 ( x4 4) 2

x13
3 x3

3 x2 2 x5 3 x4 2 x6

0 0 0 0

( x3 3) x13
3 x3

( x4 4)

x13
3 x3

3 x2 2 x5 3 x4 2 x6

x6 x5

3 x2 2w2 3 x4 2w1

Solution:

coupled
min x12 s.t. w1
11

x (0,0, 4,3,12 13 , 24 2 3)
MATLAB 5.3
coupled: 356,423 FLOPS 4.844s uncoupled: 281,379 FLOPS 0.453s

2 x2 ( x3 3)2 ( x4 4)2

g1 ( x1 , x2 , x3 , x4 ) 0 g 2 ( x1 , x2 , x3 , x4 ) 0

w2

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 Multi-level (Distributed Design) -provide disciplinary models with design tasks -optimization at subsystem and system levels
12

non-hierarchical decomposition

hierarchical decomposition

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Single-level (Distributed Analysis) Disciplinary models provide analysis Optimization is controlled by some overseeing code or database e.g. ISight (Optimizer) iSight
System Optimizer
Shared data x J(x),g(x),h(x)
subsystem analyses Optimizer design variables constraints

Structures Local data

13

Aero Local data

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Single-level (Distributed Analysis) Optimizer objective design variables constraints

x
aerodynamic analysis

g(x) h(x)

J(x)

g(x) h(x)
structural analysis

performance analysis

14

During the optimization, the overseeing code keeps track of the values 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
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Multi-level (Distributed Design)

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.

15 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Multi-level (Distributed Design)

System level optimizer command/result command/result

command/result SS2 optimizer

SS1 optimizer

SSN optimizer

SS1 analyzer SS2 analyzer SSN analyzer

16

Subsystem black box (BB) 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.

17

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Collaborative Optimization
OPTIMIZER

Coupled
18

TARGET STATE
Uncoupled
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

19

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Collaborative Optimization
min Jsys wrt: x 0 s.t. Jk 0 target variables
x0

subproblemsk Jsys

performance analysis

x0
min J1 x

J1
2

x0
min Jk x

Jk
target local variables variables local variables
2

target local variables variables local variables

s.t. local constraints

x
computed results

s.t. local constraints

x
computed results

20

analysis for subsystem 1

analysis for subsystem k

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

CO Subsystem Level
min J1 x
target local variables variables local variables
2

s.t. local constraints

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

21

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

CO System Level
min Jsys wrt: x 0 s.t. Jk 0 target variables subproblemsk

System-level optimizer changes target variables to improve objective and reduce differences Jk
Jk=0 are called compatibility constraints compatibility constraints are driven to zero, but may be violated during the optimization

22

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.
wing area, S aspect ratio, AR

aero struct

L/D

twist angle,

wing weight, W

modified from Kroo et al. AIAA 94-4325

23

perf

range, R

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

CO Example: Aircraft Design

max R0 x0 = [R0 S0 AR0 0 L/D0 W0]T s.t. J1=0, J2=0, J3=0 x0 J1 x0 J2

x0

J3

min J1 J1= (S-S0)2 + (AR-AR0)2+ ( - 0)2 + (L/D-L/D0)2 x = [AR ]T x L/D

min J2 J2= (S-S0)2 + (AR-AR0)2 + ( - 0)2 + (W-W0)2 x = [S AR]T x W

min J3 J3= (L/D-L/D0)2+ (W-W0)2 + (R-R0)2 x = [L/D W]T x R

aero analysis
24

struct analysis

perf analysis

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Collaborative Optimization
min Jsys x0 wrt: x 0 target variables s.t. Jk 0 subproblemsk Jsys
performance analysis

x0
min J1

J1
2

x0
min Jk

Jk
target local variables variables coupling local variables variables
2

target local variables variables coupling local variables variables

y1k yk 1
x

local variables

local variables

s.t. local constraints

x
computed results

s.t. local constraints

x
computed results

25

analysis for subsystem 1

analysis for subsystem k

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 subsystem j which 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.

26

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.
27 Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Analytical Target Cascading

R11

Level 1

SS1

Y11 X11 R22

R21 Y21 X21 R32

. . . . . .

Level 2
. . .

SS2

SS3

Y22 X22 R34

. . .

R31
. . .

R33

Level NL
28
Image by MIT OpenCourseWare.

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.

29

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Analytical Target Cascading

30

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 coupling variables. Local decision variables x: 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.

31

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Hierarchical Decomposition & Multi-Domain Formulation

Decomposition

Courtesy of Anas Alfaris. Used with permission.

32

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Hierarchical Decomposition & Multi-Domain Formulation

Decomposition
L1

33

Courtesy of Anas Alfaris.Used with permission. Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Hierarchical Decomposition & Multi-Domain Formulation

Decomposition
Building System

34

Courtesy of Anas Alfaris. Used with permission. Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Hierarchical Decomposition & Multi-Domain Formulation

Formulation

35

Courtesy of Anas Alfaris. Used with permission. Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Hierarchical Decomposition & Multi-Domain Formulation

Formulation

0 0 0 0 0 1 0 0 0 0 A 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

0 0 0 0 0 1 0 0 0 0 An 1 0 0 0 0 1 0 1 0 0 1 1 0 0 0
i

1 N

ai j
i 1

V
n 1

1 N

ai j
j 1

Where: j is an indicator of the fraction of total elements to which element j provides input, i is the fraction of total elements on which element i depends, and aij is an element of a matrix that can be the DSM, a power of the DSM, or the V matrix.

36

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Hierarchical Decomposition & Multi-Domain Formulation

Formulation

Courtesy of Anas Alfaris. Used with permission.

37

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

Hierarchical Decomposition & Multi-Domain Formulation

Formulation

38

Courtesy of Anas Alfaris. Used with permission. Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

References
I.P. Sobieski and I.M. Kroo. Collaborative Optimization Using Response Surface Estimation. 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

39

Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox

MIT OpenCourseWare http://ocw.mit.edu

ESD.77 / 16.888 Multidisciplinary System Design Optimization

Spring 2010

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.