Beruflich Dokumente
Kultur Dokumente
Anas Alfaris
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)
x J: g:
n n n z m
x0
Optimization Engine
min J ( x) s.t. g ( x) 0
3
J ( x) g ( x)
Function Evaluator
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.
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
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
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
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.
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
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
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
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
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 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
SS1 optimizer
SSN optimizer
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
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
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
20
CO Subsystem Level
min J1 x
target local variables variables local variables
2
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
21
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
aero struct
L/D
twist angle,
wing weight, W
perf
range, R
x0
J3
aero analysis
24
struct analysis
perf analysis
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
y1k yk 1
x
local variables
local variables
25
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.
Level 1
SS1
Level 2
. . .
SS2
SS3
R31
. . .
R33
Level NL
28
Image by MIT OpenCourseWare.
29
30
Targets T: Values set by parent subspaces to be matched by the corresponding quantities from child subspaces.
R and y: allowable compatibility tolerance.
31
32
33
Courtesy of Anas Alfaris.Used with permission. Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
34
Courtesy of Anas Alfaris. Used with permission. Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
35
Courtesy of Anas Alfaris. Used with permission. Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
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
37
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
For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.