Multidisciplinary System Design Optimization (MSDO) Decomposition and Coupling Lecture 4
Anas Alfaris
1
© Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Today’s 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)
2
© Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Standard Optimization Problem Given
x
n
J:
n
g:
n
*
x0
x
z m
Optimization Engine
J ( x) g ( x)
x
Solve the problem
min J ( x) s.t. g ( x) 0 3
That is, find x* s.t. J( x* )
Function Evaluator
f ( x),
x dom( J )
dom( g )
© Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Information Flow
A
A
B
B
Dependent Tasks (Series)
Independent Tasks (Parallel)
Interdependent Tasks (Coupled)
A B A A B B
A B A A B B
A B A A B B
A
B
Image by MIT OpenCourseWare.
4
© Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
5
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 Backward
Feed Forward
Information Flow
© 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 • Sequential C • Parallel A • K • L • Coupled J • F • I • E • D • H • G •
6
© 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
x1
g1g 2 SS1
7
Optim
x2
SS2
g1 g2
x1
x2
SS1
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,
Optim
Optimizer
x1 SS1
u1
8
x2
w1 w2 SS2
u2
w1 w2
x1 SS1
u1
x2 u2 SS2
Loop
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
w1 w2
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 Newton’s method) in order to obtain w1 and w2, which can be time consuming. Want a way to return to the situation of decoupled constraints.
9
© 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 )
s.t.
w1 s.t. w2
g1 ( x1 , u1 ) 0
g1 ( x1 , g 2 ( x2 , w1 )) 0 g 2 ( x2 , g1 ( x1 , w2 )) 0
g 2 ( x2 , u2 ) 0 u2
g1 ( x1 , u1 )
0
u1 g 2 ( x2 , u2 )
0
• 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
J2
s.t. w1
0
w2
0
where J1
decoupled 2 1
x
2 2
x
2
J2
( x3 3)
( x4 4)
w1
x13
x23 2w2
w2
x33
x43 2w1
2
coupled min x12
x22 ( x3 3)2 ( x4 4)2
s.t. w1
g1 ( x1 , x2 , x3 , x4 ) 0
w2
g 2 ( x1 , x2 , x3 , x4 ) 0
11
min x12
x22 ( x3 3) 2 ( x4 4) 2
s.t. w1
x13
x23 2 x5
0
w2
x33
x43 2 x6
0
x13
x23 2 x5
x6
0
x33
x43 2 x6
x5
0
Solution:
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
© 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 Optimizer design variables constraints
System Optimizer Shared data
x
13
Structures
Aero
Local data
Local data
J(x),g(x),h(x)
subsystem analyses
© Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Single-level (Distributed Analysis) Optimizer objective design variables constraints x
g(x) h(x)
aerodynamic analysis
14
x
J(x)
performance analysis
x
g(x) h(x) structural analysis
• 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
SS1 optimizer
command/result SS2 optimizer
command/result
SSN optimizer
…… SS1 analyzer
16
SS2 analyzer
SSN analyzer
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
x0 min J1 x
local variables
s.t. local constraints x
20
computed results
analysis for subsystem 1
0
subproblemsk Jsys
2
performance analysis
Jk
x0
J1
target local variables variables
target variables
x0
min Jk x
target local variables variables local variables
s.t. local constraints x
computed results
analysis for subsystem k © Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
2
CO – Subsystem Level min J1 x
target local variables variables
2
local variables
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
target variables 0
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 L/D
aero
twist angle,
modified from Kroo et al. AIAA 94-4325 23
struct
wing weight, W
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
min J1 J1= (S-S0)2 + (AR-AR0)2+ ( - 0)2 + (L/D-L/D0)2 x = [AR ]T x
L/D
aero analysis 24
x0
J2
min J2 J2= (S-S0)2 + (AR-AR0)2 + ( - 0)2 + (W-W0)2 x = [S AR]T x
struct analysis
W
J3
x0
min J3 J3= (L/D-L/D0)2+ (W-W0)2 + (R-R0)2 x = [L/D W]T x
R
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
x0
J1
target local variables variables
min J1
coupling local variables variables
x
local variables
s.t. local constraints x
25
computed results
analysis for subsystem 1
2
Jk
x0
2
min Jk
y1k yk 1
performance analysis
target local variables variables coupling local variables variables
x
local variables
s.t. local constraints x
computed results
analysis for subsystem k © Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
2
2
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
Y11
SS1
Level 1
X11
R21 Y21
SS2
Level 2 . . .
R22
R31 . . .
SS3
X21 R32
R33 . . .
Y22
. . .
X22 R34 . . .
Level NL 28
Image by MIT OpenCourseWare.
© Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Analytical Target Cascading • ATC’s 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. • ƐR and Ɛy: 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
An n 1
36
A2
1 0 0 0 0 1 0 0 0 0
0 0 0 0 0 1 0 0 0 0
4
V
1 0 0 0 0 0 0 1 0 0 0 1 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
j
1 0 0 0 0 1 0 1 0 0 1 1 0 0 0
i
1 N
1 N
N
ai j i 1
N
ai j j 1
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
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.
© 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.