Multidisciplinary System Design Optimization (MSDO) System Design Optimization (MSDO) Sensitivity...

1 © Massachusetts Institute of Technology - Prof. de Weck and Prof. WillcoxEngineering Systems Division and Dept. of Aeronautics and Astronautics

Multidisciplinary System Multidisciplinary System Design Optimization (MSDO)Design Optimization (MSDO)

Sensitivity AnalysisLecture 8

1 March 2004

Olivier de WeckKaren Willcox


Today’s TopicsToday’s Topics

• Sensitivity Analysis– effect of changing design variables– effect of changing parameters– effect of changing constraints

• Gradient calculation methods– Analytical and Symbolic– Finite difference– Adjoint methods– Automatic differentiation


Standard Problem Definition Standard Problem Definition

1

2

min ( )

s.t. ( ) 0 1,..,

( ) 0 1,..,

1,..,

j

k

ui i i

J

g j m

h k m

x x x i n

xxx

For now, we consider a single objective function, J(x).There are n design variables, and a total of mconstraints (m=m1+m2).The bounds are known as side constraints.


Sensitivity AnalysisSensitivity Analysis

• Sensitivity analysis is a key capability aside from theoptimization algorithms we discussed.

• Sensitivity analysis is key to understanding which design variables, constraints, and parameters are important drivers for the optimum solution x*.

• The process is NOT finished once a solution x* has been found. A sensitivity analysis is part of post-processing.

• Sensitivity/Gradient information is also needed by:– gradient search algorithms– isoperformance/goal programming– robust design



• How sensitive is the “optimal” solution J* tochanges or perturbations of the design variables x*?

• How sensitive is the “optimal” solution x* tochanges in the constraints g(x), h(x) andfixed parameters p ?


Sensitivity Analysis: AircraftSensitivity Analysis: Aircraft

Questions for aircraft design:

How does my solution change if I• change the cruise altitude?• change the cruise speed?• change the range?• change material properties?• relax the constraint on payload?• ...



Questions for spacecraft design:

How does my solution change if I• change the orbital altitude?• change the transmission frequency?• change the specific impulse of the propellant?• change launch vehicle?• Change desired mission lifetime?• ...


Gradient Vector Gradient Vector –– single objectivesingle objective

“How does the objective function Jvalue change as we change elementsof the design vector x?”

1

2

n

J

x

J

x

J

x

JCompute partial derivativesof J with respect to xi

i

J

xJ

Gradient vector points normalto the tangent hyperplane of J(x)

1x2x

3x


Geometry of Gradient vector (2D)Geometry of Gradient vector (2D)

0 0.5 1 1.5 20

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

x1

x 2

Contour plot

3.1

3.1

3.1

3.253.

25

3.25

3.25 3.25

3.5

3.5 3.5

3.54

4

44

5

5

Example function:1 2 1 2

1 2

1,J x x x x

x x

21 1 2

22 1 2

11

11

J

x x xJ

J

x x x Gradient normal to contours


Geometry of Gradient vector (3D)Geometry of Gradient vector (3D)2 2 21 2 3J x x x

1

2

3

2

2

2

x

J x

x

increasingvalues of J

1x2x

3x

Tangent plane1 2 32 2 2 6 0x x x

1 1 1Tox

2 2 2o

TJ

x

J=3

Example

Gradient vector points to larger values of J


Taylor Series ExpansionTaylor Series Expansion

,k k nJ wherex x

Taylor Series Expansion of Objective Function

Tangentialhyperplaneat xo

Effect of curvature(2nd derivative)at xo

0 0 0 0 01( ) ( ) ( ) ( ) ( ) ( )( ) H.O.T.

2

T TJ J J 0x x x x x x x H x x x

first order term second order term


Jacobian Matrix Jacobian Matrix –– multiple objectivesmultiple objectives

If there is more than one objective function, i.e.if we have a gradient vector for each Ji, arrange themcolumnwise and get Jacobian matrix:

1 2

1 1 1

1 2

2 2 2

1 2

z

z

z

n n n

J J J

x x x

J J J

x x x

J J J

x x x

J

1

2

z

J

J

J

J

n x z

z x 1


NormalizationNormalization

In order to compare sensitivities from differentdesign variables in terms of their relative sensitivityit is necessary to normalize:

i

J

x ox

“raw” - unnormalized sensitivity = partialderivative evaluated at point xi,o

,

( )i o

i i i

xJ J J

x x J x oo

xx

Normalized sensitivity capturesrelative sensitivity

~ % change in objective per% change in design variable

Important for comparing effect between design variables


Example: Dairy Farm ProblemExample: Dairy Farm Problem

With respect to which design variable is the

objective most sensitive?

“Dairy Farm” sample problem

L

R N

L – Length = 100 [m]N - # of cows = 10R – Radius = 50 [m]

fence

22

2 2

100 /

A LR R

F L R

M A N

C f F n N

I N M m

P I C

Parameters:f=100$/mn=2000$/cowm=2$/liter

xo

Assume that we are not at the optimal point x* !

COW COW

COW


Dairy Farm SensitivityDairy Farm Sensitivity

• Compute objective at xo

• Then compute raw sensitivities

• Normalize

• Show graphically (optional)

36.6

2225.4

588.4

P

LP

JNP

R

10036.6

13092 0.2810

2225.4 1.7( ) 13092

2.2550588.4

13092

o

oJ J

J

xx

( ) 13092oJ x

Dairy Farm Normalized Sensitivities

0 0.5 1 1.5 2 2.5

L

N

R

Des

ign

Varia

ble


Realistic Example: SpacecraftRealistic Example: Spacecraft

XY

Z

What are the design variables that are “drivers”of system performance ?

0 1 2

metersSpacecraftCAD model -60 -40 -20 0 20 40 60

-60

-40

-20

0

20

40

60

Centroid X [ m]C

entr

oid

Y [

m]

J2= Centroid Jitter on Focal Plane [RSS LOS]

T=5 sec

14.97 m

1 pixel

Requirement: J2,req=5 m

NASA Nexus Spacecraft Concept

Finite ElementModel

Simulation

“x”-domain “J”-domain


Graphical RepresentationGraphical RepresentationGraphical Representation ofJacobian evaluated at designxo, normalized for comparison.

-0.5 0 0.5 1 1.5

Kcf

Kc

fca

Mgs

QE

Ro

lambda

zeta

I_propt

I_ss

t_sp

K_zpet

m_bus

K_rISO

K_yPM

m_SM

Tgs

Sst

Srg

Tst

Qc

fc

Ud

UsRu

J1: Norm Sensitivities: RMMS WFE

Des

ign

Var

iabl

es

o 1,o 1x /J * J / x

analyticalfinite difference

-0.5 0 0.5 1 1.5

Kcf

Kc

fca

Mgs

QE

Ro

lambda

zeta

I_propt

I_ss

t_sp

K_zpet

m_bus

K_rISO

K_yPM

m_SM

Tgs

Sst

Srg

Tst

Qc

fc

Ud

UsRu

J2: Norm Sensitivities: RSS LOS

xo /J2,o * J2 / x

dist

urba

nce

varia

bles

stru

ctur

alva

riabl

esco

ntro

lva

riabl

esop

tics

varia

bles

1 2

0

1 2

u u

o

cf cf

J J

R R

JJ

J J

K K

x

J1: RMMS WFE most sensitive to:Ru - upper wheel speed limit [RPM]Sst - star tracker noise 1 [asec]K_rISO - isolator joint stiffness [Nm/rad]K_zpet - deploy petal stiffness [N/m]

J2: RSS LOS most sensitive to:Ud - dynamic wheel imbalance [gcm2]K_rISO - isolator joint stiffness [Nm/rad]zeta - proportional damping ratio [-]Mgs - guide star magnitude [mag]Kcf - FSM controller gain [-]


Analytical SensitivitiesAnalytical Sensitivities

If the objective function is known in closed form,we can often compute the gradient vector(s) in closedform (analytically, symbolically):

Example: 1 2 1 21 2

1,J x x x x

x x

21 1 2

22 1 2

11

11

J

x x xJ

J

x x x

Example

x1 = x2 =1

J(1,1)=3

0(1,1)

0J

Minimum

Analytical Gradient:

For complex systems analytical gradients are rarely available


Symbolic DifferentiationSymbolic Differentiation

• Use symbolic mathematics programs• E.g. Matlab,Maple, Mathematica

EDU» syms x1 x2EDU» J=x1+x2+1/(x1*x2);EDU» dJdx1=diff(J,x1)dJdx1 =1-1/x1^2/x2EDU» dJdx2=diff(J,x2)dJdx2 = 1-1/x1/x2^2

construct a symbolic object

difference operator


Finite Differences (I)Finite Differences (I)

Taylor Series expansion

Neglect second order and H.O.T.Solve for gradient vector

2' '' 2

2o o o o

xf x x f x xf x f x O x

xxo x+ xx- x

x xfFunction of a single variable f(x)

' o oo

f x x f xf x O x

x

Forward DifferenceApproximation to the

derivative

''

2

o o

xO x f

x x x

Truncation Error

0,x x


Finite Differences (II)Finite Differences (II)

11 1 1 1 1

11 1 1 1 1

o o o

o

J x J x J x x J xJ J

x x x x x

J(x)

11x1

ox

11J x

1oJ x

true, analyticalsensitivity

finite differenceapproximation

1x

J

11x 1

ox-

x1


Finite Differencing (III)Finite Differencing (III)

Take Taylor expansion backwards at ox x

2' '' 2

2o o o o

xf x x f x xf x f x O x

2' '' 2

2o o o o

xf x x f x xf x f x O x (1)

(2)

(1)-(2) and solve again for derivative

' 2

2o o

o

f x x f x xf x O x

x

22 '''

6

o o

xO x f

x x x

Truncation Error

Central DifferenceApproximation to the

derivative


Finite Difference OverviewFinite Difference Overview

xForward Difference

'( ) o oo

f x x f xf x

x

1st derivative 2nd derivative

2

2 2''( ) o o o

o

f x x f x x f xf x

x

xx

Central Difference2nd derivative1st derivative

'( )2

o oo

f x x f x xf x

x2

2''( ) o o o

o

f x x f x f x xf x

x


Errors of Finite DifferencingErrors of Finite DifferencingCaution: - Finite Differencing always has errors

- very dependent on perturbation size

1 2 1 21 2

1,J x x x x

x xx1 = x2 =1J(1,1)=3

0(1,1)

0J

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20123456789

10

x1

x1

10 -9 10 -8 10-7 10-610

-8

10-7

10-6

10-5

TruncationErrors ~ x

RoundingErrors ~ 1/ x

Perturbation Step Size x1

Gra

dien

t Err

or

Choice of x is critical


Perturbation Size Perturbation Size x Choicex Choice

• Error Analysis

• Significant digits (Barton 1992)

• Machine Precision

• Trial and Error – typical value ~ 0.1-1%

1/ 2

Ax f - Forward difference1/3

Ax f - Central difference

(Gill et al. 1981)

ox

theoretical function

computedvalues

~ x

A

10 qk kx xStep size

at k-th iterationq-# of digits of machinePrecision for real numbers


Computational Expense of FDComputational Expense of FD

Cost of a single objective functionevaluation of Ji

iF J

Cost of gradient vector finitedifference approximation for Jifor a design vector of length n

in F J

Cost of Jacobian finitedifference approximation with z objective functions

iz n F J

Example: 6 objectives30 design variables1 sec per simcode evaluation

3 min of CPU timefor a single Jacobianestimate - expensive !


Automatic DifferentiationAutomatic Differentiation

• Mathematical formulas are built from a finite set of basic functions, e.g. sin x, cos x, exp x

• Take analysis code in C or Fortran

• Using chain rule, add statements that generate derivatives of the basic functions

• Tracks numerical values of derivatives, does not track symbolically as discussed before

• Outputs modified program = original + derivative capability


Chain Rule exampleChain Rule example

( )

( )

u q s

s p t

quantityof interest

First compute Want to takederivatives w.r.t “t”

( )ds d

p tdt dt

Store this value numerically

Then apply chain rule

( ( )) ( )du d d ds

q s t q sdt dt ds dt

substitute

desiredsensitivity


Chain Rule Coding ExampleChain Rule Coding Example

hr = gm*eps/rho-0.5*g1*(ux*ux + vy*vy);

h_u[0] = -gm*eps/(rho*rho);

h_u[1] = -g1*ux;

h_u[2] = -g1*vy;

h_u[3] = gm/rho;

hi = (di*hr+hl)*d1;

hi_u[0] = (di*hr+hl)*d1_u[0]

+ d1*(di_u[0]*hr+di*h_u[0]);

• compute hrdifferentiate:• wrt rho• wrt ux• wrt vy• wrt eps


Adjoint MethodsAdjoint Methods

• A way to get gradient information in a computationally efficient way

• Based on theory from controls• Applied extensively in aerodynamic design and

optimization• For example, in aerodynamic shape design, need

objective gradient with respect to shape parameters andwith respect to flow parameters

– Would be expensive if finite differences are used!• Adjoint methods have allowed optimization to be used

for complicated, high-fidelity fluids problems.



Consider

where J is the cost function, w contains the N flow variables,and F contains the n shape design variables.

( , )J J w F

At an optimum, the variation of the cost function is zero:

0T T

J JJ w F

w F

N 1 n 1

N>>n



Fluid governing equations: ( , ) 0R w F

0R R

R w Fw F

We can append these constraints to the cost function using a Lagrange multiplier approach:

T

T T

T T

T T

J J R RJ

J R J R

w F w Fw F w F

w Fw w F F



T TT T

J R J RJ w F

w w F F

Choose to satisfy the adjoint equation: T

R Jw w

equivalent to one flow solve

TT

J RJ F

F FThen does not depend

on the number of flow variables

total gradient of J



“How does the optimal solution change as we change the problem parameters?”

effect on design variableseffect on objective functioneffect on constraints

Want to answer this question without having to solve the optimization problem again.Two approaches:

– use Kuhn-Tucker conditions– use feasible directions


ParametersParameters

Parameters p are the fixed assumptions.How sensitive is the optimal solution x* with respectto fixed parameters ?

Optimal solution:

x* =[ R=106.1m, L=0m, N=17 cows]TExample:

“Dairy Farm” sample problem

L

R N

fence

Fixed parameters:

Parameters:f=100$/m - Cost of fencen=2000$/cow - Cost of a single cowm=2$/liter - Market price of milk

How does x* change as parameterschange?

Maximize Profit

COW

COW COW



Recall the Kuhn-Tucker conditions. Let us assume that we have M active constraints, which are contained in the vector

*ˆ( *) ( ) 0

ˆ ( *) 0,

0,

j jj M

j

j

J g

g j M

j M

x x

x

ˆ( )g x

For a small change in a parameter, p, we require that the Kuhn-Tucker conditions remain valid:

(KT conditions)0

ddp



First, let us write out the components of the first equation:

* *ˆ( ) ( ) 0j jj M

J gx x

* *ˆ

( ) ( ) 0, 1,...,jj

j Mi i

gJi n

x xx x

Now differentiate with respect to the parameter p usingthe chain rule:

1

ni

k i

dY Y Y xdp p x p

38

© Massachusetts Institute of Technology - P

Prof. de Weck and Prof. WillcoxEngineering Systems Division and Dept. of Aeronautics and Astronautics


* *ˆ

( ) ( ) 0jj

j Mi i

gJx x

x x ˆ ( *) 0jg x

differentiate wrt p:

22

1

22

ˆ

ˆ ˆ0

nj k

jk j Mi k i k

j j j

j M j Mi i i

gJ xx x x x p

g gJx p x p x p

1

ˆ ˆ0

nj j k

k k

g g xp x p

unknowns are and jixp p


Sensitivity AnalysisSensitivity AnalysisIn matrix form we can write:

00T

A B c

B d

xn

M

Mn

22 ˆ jik j

j Mi k i k

gJA

x x x xˆ j

iji

gB

x22 ˆ j

ij Mi i

gJc

x p x pˆ j

j

gd

p

1

2

n

xp

xp

xp

x

1

2

M

p

p

p



We solve the system to find x and , then the sensitivity of the objective function with respect to p can be found:

TdJ JJ

dp px

dJJ p

dp(first-order

approximation)

px x

To assess the effect of changing a different parameter, we only need to calculate a new RHS in the matrix system.


Sensitivity Analysis Sensitivity Analysis -- ConstraintsConstraints• We also need to assess when an active constraint will

become inactive and vice versa• An active constraint will become inactive when its

Lagrange multiplier goes to zero:

jj jp p

pFind the p that makes j zero:

0j j p

j

j

p j M

This is the amount by which we can change p before the jth

constraint becomes inactive (to a first order approximation)


Sensitivity Analysis Sensitivity Analysis -- ConstraintsConstraintsAn inactive constraint will become active when gj(x)goes to zero:

( ) ( *) ( *) 0Tj j jg g p gx x x x

Find the p that makes gj zero:

( *)

( *)j

Tj

gp

g

xx x

for all j notactive at x*

• This is the amount by which we can change p before the jth

constraint becomes active (to a first order approximation)• If we want to change p by a larger amount, then the problem

must be solved again including the new constraint• Only valid close to the optimum


Lecture SummaryLecture Summary

• Sensitivity analysis– Yields important information about the design space,

both as the optimization is proceeding and once the “optimal” solution has been reached.

• Gradient calculation approaches– Analytical and Symbolic– Finite difference– Automatic Differentiation– Adjoint methods

ReadingPapalambros – Section 8.2 Computing Derivatives

Multidisciplinary System Design Optimization (MSDO) System Design Optimization (MSDO) Sensitivity...

Documents

Transcript of Multidisciplinary System Design Optimization (MSDO) System Design Optimization (MSDO) Sensitivity...