Multidisciplinary System

Design Optimization (MSDO)

Multiobjective Optimization (I)

Lecture 14
by
Dr. Anas Alfaris

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Where in Framework ?
Objective Vector

x1
x2

Discipline A

Discipline B

Discipline C

xn
Coupling

Jz

Multiobjective
Optimization

Approximation
Methods

Optimization Algorithms
Tradespace
Exploration

(DOE)
2

Numerical Techniques
(direct and penalty methods)
Heuristic Techniques
(SA,GA, Tabu Search)

J1
J2

Sensitivity
Analysis
Coupling

Isoperformance

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Lecture Content

Why multiobjective optimization?

Example twin peaks optimization
History of multiobjective optimization
Weighted Sum Approach
Pareto-Optimality
Dominance and Pareto Filtering

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Multiobjective Optimization Problem

Formal Definition
Design problem may be formulated
as a problem of Nonlinear Programming (NLP). When
Multiple objectives (criteria) are present we have a MONLP

min J x, p

where J

s.t. g(x, p) 0
h(x, p)=0
xi , LB

xi

xi ,UB (i

x
g
1, ..., n )

h
4

J1 x
x1
g1 ( x )
h1 (x )

Jz x
xi
g m1 ( x )
hm2 (x )

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

xn
T

Multiple Objectives
The objective can be a vector J of z system responses
or characteristics we are trying to maximize or minimize

J1
J2
J3

cost [$]
- range [km]
weight [kg]

Ji

- data rate [bps]

Jz
5

- ROI [%]

Often the objective is a

scalar function, but for
real systems often we
attempt multi-objective
optimization:

J(x)

Objectives are usually

conflicting.

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Why multiobjective optimization ?

While multidisciplinary design can be associated with the
traditional disciplines such as aerodynamics, propulsion,
structures, and controls there are also the lifecycle areas of
manufacturability, supportability, and cost which require
consideration.

After all, it is the balanced design with equal or weighted

treatment of performance, cost, manufacturability and
supportability which has to be the ultimate goal of
multidisciplinary system design optimization.
Design attempts to satisfy multiple, possibly
conflicting objectives at once.
6

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Design
Decisions
Aspect Ratio
Dihedral Angle
Vertical Tail Area
Engine Thrust
Skin Thickness
# of Engines
Fuselage Splices
Suspension Points
Location of Mission
Computer
Access Door
Locations

Objectives

Example:
F/A-18 Aircraft

Speed
Range
Payload Capability
Radar Cross Section
Stall Speed
Stowed Volume

Acquisition cost
Cost/Flight hour
MTBF
Engine swap time
Assembly hours
Avionics growth
Potential
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics

Multiobjective Examples

Design
Optimization

Operations
Research

Aircraft Design
max {range}
max {passenger volume}
max {payload mass}
min {specific fuel consumption}
max {cruise speed}
min {lifecycle cost}

Production Planning
max {total net revenue}
max {min net revenue in any time period}
min {backorders}
min {overtime}
min {finished goods inventory}
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics

J1
J2
Jz

Multiobjective vs. Multidisciplinary

Multiobjective Optimization

Multidisciplinary Design Optimization

Optimization involves several disciplines

e.g. Structures, Control, Aero, Manufacturing
Issues: Human and computational infrastructure, cultural,
administrative, communication, software, computing time,
methods

All optimization is (or should be) multiobjective

Optimizing conflicting objectives

e.g., Cost, Mass, Deformation
Issues: Form Objective Function that represents designer
preference! Methods used to date are largely primitive.

Minimizing mass alone, as is often done, is problematic

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

single objective

Multidisciplinary vs. Multiobjective

single discipline
cantilever beam
m
F

costs (mfg) subject

to loading and geometry
constraint
D

l
Minimize displacement
s.t. mass and loading constraint

multiple obj.

single discipline

10

vo

multiple disciplines
support bracket
F
Minimize stamping

multiple disciplines

airfoil V
fuel
(x,y)

Maximize CL/CD and maximize

wing fuel volume for specified

vo

commercial aircraft
Minimize SFC and maximize cruise
speed s.t. fixed range and payload

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Example: Double Peaks Optimization

Objective: max J= [ J1 J2]T

J1

11

3e

x12 ( x2 1)2

3 1 x1 e

x
10 1
5

3
1

( x1 2)2 x22

(demo)

J2
5
2

x12 x22

10

3 1 x2
x2
5

x23

e x2

( x1 1) 2

x15 e x2

x12

3e

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

Engineering1Systems
2 Division and Dept. of Aeronautics and Astronautics

0.5 2 x

(2 x2 )2 x12

Double peaks optimization

Optimum for J1 alone:

Optimum for J2 alone:

x2* =
-1.5808
0.0095

x1*

=
0.0532
1.5973

J1* = 8.9280

J1(x2*)= -6.4858

J2(x1*)= -4.8202

J2* = 8.1118

Each point x1* and x2* optimizes objectives J1 and J2 individually.

Unfortunately, at these points the other objective exhibits a low
objective function value. There is no single point that simultaneously
optimizes both objectives J1 and J2 !
12

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Tradeoff between J1 and J2

Want to do well with respect to both J1 and J2
Define new objective function: Jtot=J1 + J2
max(J1)
Optimize Jtot

Result:
Xtot* =
0.8731
0.5664

tradeoff
solution
max(J1+J2)

Jtot* = 6.1439
J(xtot*)

=
3.0173 J1
=
3.1267 J2

13

max(J2)

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

History (1) Multicriteria Decision Making

Life is about making decisions. Most people attempt to make the

best decision within a specified set of possible decisions.
In 1881, Kings College (London) and later Oxford Economics
Professor F.Y. Edgeworth is the first to define an optimum for
multicriteria economic decision making. He does so for the multiutility
problem within the context of two consumers, P and :
It is required to find a point (x,y,) such that in
whatever direction we take an infinitely small step,
P and do not increase together but that, while one
increases, the other decreases.
Reference: Edgeworth, F.Y., Mathematical Psychics,
P. Keagan, London, England, 1881.

14

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

History (2) Vilfredo Pareto

Born in Paris in 1848 to a French Mother and Genovese

Father
Graduates from the University of Turin in 1870 with a
degree in Civil Engineering
Thesis Title: The Fundamental Principles of Equilibrium in
Solid Bodies

While working in Florence as a Civil Engineer from 18701893, Pareto takes up the study of philosophy and politics
and is one of the first to analyze economic problems
with mathematical tools.
In 1893, Pareto becomes the Chair of Political Economy at
the University of Lausanne in Switzerland, where he
creates his two most famous theories:
Circulation of the Elites
The Pareto Optimum

15

The optimum allocation of the resources of a society is not

attained so long as it is possible to make at least one
individual better off in his own estimation while keeping
others as well off as before in their own estimation.
Reference: Pareto, V., Manuale di Economia Politica, Societa
Editrice Libraria, Milano, Italy, 1906. Translated into English by
A.S. Schwier as Manual of Political Economy, Macmillan, New
York, 1971. Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics

History (3) Extension to Engineering

After the translation of Paretos Manual of Political Economy
into English, Prof. Wolfram Stadler of San Francisco State
University begins to apply the notion of Pareto Optimality to the
fields of engineering and science in the middle 1970s.
The applications of multi-objective optimization in engineering
design grew over the following decades.
References:
Stadler, W., A Survey of Multicriteria Optimization, or the Vector
Maximum Problem, Journal of Optimization Theory and
Applications, Vol. 29, pp. 1-52, 1979.
Stadler, W. Applications of Multicriteria Optimization in Engineering
and the Sciences (A Survey), Multiple Criteria Decision Making
Past Decade and Future Trends, ed. M. Zeleny, JAI Press,
Greenwich, Connecticut, 1984.

Ralph E. Steuer, Multicriteria Optimization - Theory, Computation

and Application, 1985
16

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Notation and Classification

Traditionally - single objective constrained optimization:

max J

f x

s.t. x S

f x
S

J objective function
feasible region

If f(x) linear & constraints linear & single objective = LP

If f(x) linear & constraints linear & multiple obj. = MOLP
If f(x) and/or constraints nonlinear & single obj.= NLP
If f(x) and/or constraints nonlinear & multiple obj.= MONLP
Ref: Ralph E. Steuer, Multicriteria Optimization - Theory, Computation
and Application, 1985

17

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Design Space vs Objective Space

Domain

Range
J(xtot*)

Xtot*

A function f which may (but does not necessarily)

associate a given member of the range of f with
more than one member of the domain of f.
18

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Formal Solution of a MOO Problem

Trivial Case:
There is a point x* S that simultaneously optimizes
all objectives J i , where 1 i z
Such a point almost never exists - i.e. there is no point
that will simultaneously optimizes all objectives at once
Two fundamental approaches:
Scalarization
Approaches
max U J1 , J 2 ,

, Jz

s.t. J i

1 i

x S

19

fi x

Pareto
Approaches
J i1
z

J i2

and J i1

i
J i2 for

at least one i

Preferences
Preferences
includedupfront
included a posteriori
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics

SOLP versus MOLP

Optimal solution
if only J1 considered

S
x1*

x2

c1
max J1

cx

max J 2

c2 x

s.t.

constraints

What is the optimal

solution of a MOLP?

x S

c2

x2*

Optimal solution
if only J2 considered

x1
20

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Weighted-Sum Approach
Each objective i is multiplied by a strictly positive scalar

Rz

x2

k
i

0,

i 1

x1
S

c3

c1

x3*

1
2
c
+
c
1
2

c3

constraints

21

criterion
cone

Solve the
composite or
WSLP:

max TCx x S
c2
x2
x1

Strictly convex
combination
of objectives

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Group Exercise: Weights (5 min)

We are trying to build the optimal automobile
There are six consumer groups:
-G1: 25 year old single (Cannes, France)
-G2: family w/3 kids (St. Louis, MO)
-G3: electrician/entrepreneur (Boston, MA)
-G4: traveling salesman (Montana, MT)
-G5: old lady (Rome, Italy)
-G6: taxi driver (Hong Kong, China)

Objective Vector:
J1: Turning Radius [m]
J2: Acceleration [0-60mph]
J3: Cargo Space [m3]
J4: Fuel Efficiency [mpg]
J5: Styling [Rating 0-10]
J6: Range [km]
J7: Crash Rating [poor-excellent]
J8: Passenger Space [m3]
J9: Mean Time to Failure [km]

Assignment: Determine

, i =19

9
i

1000

i 1
22

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

What are the scale factors sfi?

Scaling is critical in multiobjective optimization

J i J i sfi
Scale each objective by sfi:
Common practice is to scale by sfi = Ji*
Alternatively, scale to initial guess J(xo)=[1..1]T
Multiobjective optimization then takes place in
a non-dimensional, unit-less space
Recover original objective function values by
reverse scaling
Saab 9-5
Example:

J1=range [sm]
J2=fuel efficiency [mpg]

sf1=573.5 [sm]
sf2=36 [mpg]
Suzuki Swift

23

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Weighted Sum: Double Peaks

J tot

J1

J 2 where

[0,1]

=0.05

Demo:
24

At each setting of we solve a new single

objective optimization problem the underlying
function changes at each increment of
Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics

Weighted Sum Approach (II)

=0

Weighted sum
finds interesting,
solutions but misses
many solutions of
interest.

=1
25

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Weighted Sum (WS) Approach

weight

Min(J2 /sf2)

J MO

miss this
concave region

i 1

J*i+1

sfi

Ji

1> 2

scale
factor

convert back to SOP

LP in J-space
easy to implement
2> 1
scaling important !
Pareto
weighting determines
front
which point along PF is
J*i
J-hyperplane
found
misses concave PF
Max(J1 /sf1)
PF=Pareto Front(ier)
26

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Properties of optimal solution

x* optimal if J x*
for x*

S and for x

J x

(maximization)

x*

This is why multiobjective optimization is also

sometimes referred to as vector optimization
x* must be an efficient solution

x S is efficient if and only if (iff) its objective vector

(criteria) J(x) is non-dominated
A point x S is efficient if it is not possible to move feasibly
from it to increase an objective without decreasing at least
one of the others
27

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Dominance (assuming maximization)

Let

J1 , J 2

Rz

be two objective (criterion) vectors.

Then J1 dominates J2 (weakly) iff

J1
More
precisely:

J 2 and J1

Ji

J2

Jz

J i1

J i2

and J i1

J i2 for at least one i

Also J1 strongly dominates J2 iff

More
precisely:
28

J1
J2

J i1

J i2

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Set Theory
J
J*
D

Set Theory:

x S
J, J* Z

A solution must be feasible

ND

Z , ND

ND Z
ND

A solution is either dominated or non-dominated

but cannot be both at the same time

J* x *
29

ND

All dominated and non-dominated

solutions must be feasible
All feasible solutions are either non-dominated
or dominated
Pareto-optimal solutions are non-dominated

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Dominance versus Efficiency

Whereas the idea of dominance refers to vectors in
criterion space J, the idea of efficiency refers to points
in decision space x.
Can use this criterion as a Pareto Filter if the design
space has been explored (e.g. DoE).

30

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Dominance - Exercise
max{range}
min{cost}
max{passengers}
max{speed}
#1

#2

7587
321
112
950

6695
211
345
820

[km]
[$/km]
[-]
[km/h]
#3

#4

Multiobjective
Aircraft Design

#5

3788 8108 5652

308
278 223
450
88
212
750
999 812

#6

#7

6777 5812
355 401
90
185
901 788

#8
7432
208
208
790

Which designs are non-dominated ? (5 min)

Non-dominated Designs: #1, #2, #3, #4 and #8
Dominated Designs: #5, #6, #7
31

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Dominance - Exercise
Algorithm for extracting non-dominated solutions:
Pairwise comparison
#1
7587
321
112
950

>
>
<
>

#2

Score
#1

Score
#2

#1

6695
211
345
820

1
0
0
1

0
1
1
0

7587
321
112
950

2
Neither #1 nor #2
dominate each other

vs

>
<
>
>

#6

Score
#1

Score
#6

6777
355
90
901

1
1
1
1

0
0

0
0
vs

Solution #1 dominates
solution #6

In order to be dominated a solution must

have ascore of 0 in pairwise comparison
32

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Domination Matrix
Shows which solution dominates which other
solution (horizontal rows) and (vertical columns)

Row
1
2

3
4
Solution 2 dominates
Solution 5

5
6
7

k-th column indicates

8
by how many solutions
the k-th solution is dominated 1 2 3 4 5 6 7 8

Column

0 0 0 0 1 1 2 0

1
2
0
0
0
0
0
1

j-th row
indicates
how many
solutions
j-th solution
dominates

Solution 7 is dominated
by Solutions 2 and 8

Non-dominated solutions have a zero in the column

33

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Double Peaks: Non-dominated Set

Filtered the
Full Factorial
Set: 3721
Non-dominated
set approximates
Pareto frontier:
79 points (2.1%)
34

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Lifecycle Cost [B$]

Simulation Results - Satellites

10

Iridium actual
Iridium simulated
Globalstar actual

10

10
Global
35

Pareto
Front

Globalstar simulated

10

10

10

10

de Weck, O. L. and Chang D., Architecture Trade

Methodology for LEO Personal Communication Systems
Capacity Cs [# of duplex channels]
, 20th International Communications Satellite Systems
Massachusetts Institute of Technology Conference,
- Prof. de Weck
andAIAA-2002-1866,
Prof. Willcox Montral,
Paper No.
Canada, May
2002.
Engineering Systems Division and Dept.Qubec,
of Aeronautics
and12-15,
Astronautics

Pareto-Optimal vs ND
min(J2)

True
Pareto
Front

Obtain
different
points for
different weights

Approximated
Pareto Front
D

ND

max (J1)
All pareto optimal points are non-dominated
Not all non-dominated points are pareto-optimal
Its easier to show dominatedness than non-dominatedness !!!
36

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

PO

Lecture Summary
A multiobjective problem has more than one optimal solution
All points on Pareto Front are non-dominated
Methods:
Weighted Sum Approach (Caution: Scaling !)
Pareto-Filter Approach
Methods for direct Pareto Frontier calculation next time:
AWS (Adaptive Weighted Sum)
NBI (Normal Boundary Intersection)

The key difference between multiobjective optimization

methods can be found in how and when designer
preferences are brought into the process.
. More in next lecture
37

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

Remember .
Pareto Optimal means ..

Take from Peter to pay Paul

38

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

Engineering Systems Division and Dept. of Aeronautics and Astronautics

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.