1 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Multidisciplinary System
Design Optimization (MSDO)
Design Space Exploration
Lecture 5
18 February 2004
Karen Willcox
2 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Today’s Topics
Design of Experiments Overview
Full Factorial Design
Parameter Study
One at a Time
Latin Hypercubes
Orthogonal Arrays
Effects
DoE Paper Airplane Experiment
3 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Design of Experiments
A collection of statistical techniques providing a systematic
way to sample the design space
Useful when tackling a new problem for which you know
very little about the design space.
Study the effects of multiple input variables on one or more
output parameters
Often used before setting up a formal optimization problem
– Identify key drivers among potential design
variables
– Identify appropriate design variable ranges
– Identify achievable objective function values
Often, DOE is used in the context of robust design. Today
we will just talk about it for design space exploration.
4 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Design of Experiments
Design variables = factors
Values of design variables = levels
Noise factors = variables over which we have no control
e.g. manufacturing variation in blade thickness
Control factors = variables we can control
e.g. nominal blade thickness
Outputs = observations (= objective functions)
Factors
+
Levels
“Experiment”
Observation
(Often an analysis code)
5 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Matrix Experiments
Each row of the matrix corresponds to one experiment.
Each column of the matrix corresponds to one factor.
Each experiment corresponds to a different combination of
factor levels and provides one observation.
Expt No. Factor A Factor B Observation
1A1B1 K
1
2A1B2 K
2
3A2B1 K
3
4A2B2 K
4
Here, we have two factors, each of which can take two levels.
6 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Full-Factorial Experiment
Specify levels for each factor
Evaluate outputs at every combination of values
– complete but expensive!n factors
l levels
l
n
observations
Factor
Expt
No.
AB
1 A1 B1
2 A1 B2
3 A1 B3
4 A2 B1
5 A2 B2
6 A2 B3
7 A3 B1
8 A3 B2
9 A3 B3
2 factors, 3 levels each:
l
n
= 3
2
= 9 expts
4 factors, 3 levels each:
l
n
= 3
4
= 81 expts
7 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Fractional Factorial Experiments
Due to the combinatorial explosion, we cannot
usually perform a full factorial experiment
So instead we consider just some of the possible
combinations
Questions:
– How many experiments do I need?
– Which combination of levels should I
choose?
Need to balance experimental cost with design
space coverage
8 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Fractional Factorial Design
Initially, it may be useful to look at a large number of
factors superficially rather than a small number of
factors in detail:
1112
2212
12
,
,
,
nnn
fll
fll
fll
#
#
11121314
22122324
33132334
,,,,
,,,,,
,,,,,
fllll
fllll
fllll
!
!
!
vs.
many levels
many factors
9 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
DoE Techniques Overview
TECHNIQUE COMMENT EXPENSE
(l=# levels, n=# factors)
Full factorial
design
Evaluates all possible
designs.
l
n
-grows
exponentially with
number of factors
Orthogonal arrays
Don’t always seem to
work - interactions?
Moderate – depends
on which array
One at a time
Order of factors?
1+n(l-1) - cheap
Latin hypercubes
Not reproducible,
poor coverage if
divisions are large.
l - cheap
Parameter study
Captures no
interactions.
1+n(l-1) - cheap
10 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Parameter Study
Specify levels for each factor
Change one factor at a time, all others at base level
Consider each factor at every level
Factor
Expt
No.
ABCD
C1 D1
D1
D1
D1
D1
D1
D1
D2
D3
C1
C1
C1
C1
C2
C3
C1
C1
4 A1 B2
5 A1 B3
6 A1 B1
7 A1 B1
8 A1 B1
1 A1 B1
2 A2 B1
3 A3 B1
9 A1 B1
n factors
1+n(l-1)
evaluations
l levels
4 factors, 3 levels each:
1+n(l-1) =
1+4(3-1) = 9 expts
Baseline : A1, B1, C1, D1
11 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Parameter Study
Select the best result for each factor
Factor
Expt
No.
AB C D
Observation
D1
K
1
K
2
K
3
K
4
K
5
K
6
K
7
K
8
K
9
D1
D1
D1
D1
D1
D1
D2
D3
C1
C1
C1
C1
C1
C2
C3
C1
C1
4 A1 B2
5 A1 B3
6 A1 B1
7 A1 B1
8 A1 B1
1 A1 B1
2 A2 B1
3 A3 B1
9 A1 B1
1. Compare K
1
, K
2
, K
3
? A*
2. Compare K
1
, K
4
, K
5
? B*
3. Compare K
1
, K
6
, K
7
? C*
4. Compare K
1
, K
8
, K
9
? D*
“Best design” is
A*,B*,C*,D*
Does not capture interaction between variables
12 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
One At a Time
Change first factor, all others at base value
If output is improved, keep new level for that factor
Move on to next factor and repeat
Factor
Expt
No.
ABCD
C1 D1
D1
D1
D1
D1
D1
D1
D2
D3
C1
C1
C1
C1
C2
C3
C*
C*
4 A* B2
5 A* B3
6 A* B*
7 A* B*
8 A* B*
1 A1 B1
2 A2 B1
3 A3 B1
9 A* B*
n factors
l levels
1+n(l-1)
evaluations
4 factors, 3 levels each:
1+n(l-1) =
1+4(3-1) = 9 expts
Result depends on order of factors
13 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Parameter Study vs. One at a Time
Parameter study:
– Chances are you will not actually evaluate the
“best design” as part of your original experiment
– “Best design” is chosen by extrapolating each
factor’s behavior, but interactions are not
considered
One at a Time:
– The “best design” is a member of your matrix
experiment
– Some interactions are captured, even though the
result depends on the order of the factors
14 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Latin Hypercubes
Divide design space uniformly into l divisions for
each factor
Combine levels randomly
– specify l points
– use each level of a factor only once
e.g. two factors, four levels each:
A
A1 A2 A3 A4
B
B1
B2
B3
B4
Results not repeatable
Can have poor coverage
15 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Orthogonal Arrays
Specify levels for each factor
Use arrays to choose a subset of the full-
factorial experiment
Subset selected to maintain orthogonality
between factors
n factors
subset of l
n
evaluations
l levels
Does not capture all interactions, but is
efficient
Experiment is balanced
16 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Orthogonal Arrays
D3C2B1A24
D1C3B2A25
D2C1B3A26
D2C3B1A37
D3C1B2A38
D1
D3
D2
D1
D
Factor
C3B3A13
C2B3A39
C2B2A12
C1B1A11
CBA
Expt
No.
Factor
Expt
No.
AB C
1 A1 B1 C1
2 A1 B2 C2
3 A2 B1 C2
4 A2 B2 C1
L
4
(2
3
)
4 expts
3 factors
L
9
(3
4
)
9 expts
4 factors
3 levels
2 levels
17 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Orthogonality
Notice that for any pair of columns, all combinations of factor
levels occur and they occur an equal number of times.
This is the balancing property.
In general, the balancing property is sufficient for orthogonality.
There is a formal statistical definition of orthogonality, but we will
not go into it here.
Factor
D
D1
D2
D3
4 A2 B1 C2 D3
5 A2 B2 C3 D1
6 A2 B3 C1 D2
7 A3 B1 C3 D2
8 A3 B2 C1 D3
D1
Expt
No.
AB C
1 A1 B1 C1
2 A1 B2 C2
3 A1 B3 C3
9 A3 B3 C2
All of the combinations
(1-1, 1-2, 1-3, 2-1, 2-2,
2-3, 3-1, 3-2, 3-3)
occur once for each
pair of columns.
L
9
(3
4
)
18 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Effects
Once the experiments have been performed, the results
can be used to calculate effects.
The effect of a factor is the change in the response as
the level of the factor is changed.
– Main effects: averaged individual measures of
effects of factors
– Interaction effects: the effect of a factor
depends on the level of another factor
Often, the effect is determined for a change from a minus
level (-) to a plus level (+) (2-level experiments).
19 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Effects
Consider the following experiment:
– We are studying the effect of three factors on the price of
an aircraft
– The factors are the number of seats, range and aircraft
manufacturer
– Each factor can take two levels:
Factor 1: Seats 100<S1<150 150<S2<200
Factor 2: Range (nm) 2000<R1<2800 2800<R2<3500
Factor 3: Manufacturer M1=Boeing M2=Airbus
20 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Main Effects
Expt
No.
Seats
(S)
Range
(R)
Mfr
(M)
Price
(observation)
1 S1 R1 M1 P
1
2 S1 R1 M2 P
2
3 S1 R2 M1 P
3
4 S1 R2 M2 P
4
5 S2 R1 M1 P
5
6 S2 R1 M2 P
6
7 S2 R2 M1 P
7
8 S2 R2 M2 P
8
L
8
(2
3
)
(full factorial
design)
The main effect of a factor is the effect of that factor on the
output averaged across the levels of other factors.
21 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Main Effects
Question: what is the main effect of manufacturer? i.e. from
our experiments, can we predict how the price is affected by
whether Boeing or Airbus makes the aircraft?
Expt
No.
Seats
(S)
Range
(R)
Mfr
(M)
Price
(observation)
1 S1 R1 M1 P
1
2 S1 R1 M2 P
2
3 S1 R2 M1 P
3
4 S1 R2 M2 P
4
5 S2 R1 M1 P
5
6 S2 R1 M2 P
6
7 S2 R2 M1 P
7
8 S2 R2 M2 P
8
21 43 65 87
(P -P )+ (P -P )+ (P -P )+ (P -P )
4
main effect of
manufacturer
=
expts 1 and 2 differ only
in the manufacturer
22 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Main Effects – Another Interpretation
123 45678
+ + +
8
PPPPPPPP
m
overall mean
response:
135 7
1
4
M
PPPP
m
avg over all expts
when M=M1 :
effect of mfr
level M1
=
1M
mm
Effect of factor level can be defined for multiple levels
effect of mfr
level M2
=
2M
mm
main effect
of mfr
=
21MM
mm
Main effect of factor is defined as
difference between two levels
NOTE: The main effect should be interpreted individually only if
the variable does not appear to interact with other variables
23 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Main Effect Example
Expt
No.
Aircraft
Seats
(S)
Range
(R)
Mfr
(M)
Price
($M)
1 717
A318-100
737-700
A319-100
737-900
A321-200
737-800
A320-200
S1 R1 M1 24.0
2 S1 R1 M2 29.3
3 S1 R2 M1 33.0
4 S1 R2 M2 35.0
5 S2 R1 M1 43.7
6 S2 R1 M2 48.0
7 S2 R2 M1 39.1
8 S2 R2 M2 38.0
Sources:
Seats/Range data: Boeing Quick Looks
Price data: Aircraft Value News
Airline Monitor, May 2001 issue
100<S1<150 150<S2<200
2000<R1<2800 2800<R2<3500
M1=Boeing M2=Airbus
24 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Main Effect Example
overall mean price = 1/8*(24.0+29.3+33.0+35.0+43.7+48.0+39.1+38.0)
= 36.26
mean of experiments with M1 = 1/4*(24.0+33.0+43.7+39.1)
= 34.95
mean of experiments with M2 = 1/4*(29.3+35.0+48.0+38.0)
= 37.58
Main effect of Boeing (M1) = 34.95 – 36.26 = -1.3
Main effect of Airbus (M2) = 37.58 – 36.26 = 1.3
Main effect of manufacturer = 37.58 – 34.95 = 2.6
Interpretation?
25 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Interaction Effects
We can also measure interaction effects between factors.
Answers the question: does the effect of a factor depend on
the level of another factor?
e.g. Does the effect of manufacturer depend on whether we
consider shorter range or longer range aircraft?
The interaction between manufacturer and range is defined
as half the difference between the average manufacturer
effect with range 2 and the average manufacturer effect with
range 1.
mfr u range
interaction
avg mfr effect
with range 1
avg mfr effect
with range 2
=
-
2
26 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Interaction Effects
Expt
No.
Seats
(S)
Range
(R)
Mfr
(M)
Price
($M)
1 S1 R1 M1 24.0
2 S1 R1 M2 29.3
3 S1 R2 M1 33.0
4 S1 R2 M2 35.0
5 S2 R1 M1 43.7
6 S2 R1 M2 48.0
7 S2 R2 M1 39.1
8 S2 R2 M2 38.0
range R1 : expts 1,2,5,6
range R2 : expts 3,4,7,8
21 65
(-)+ (-)
2
PP PP
avg mfr effect
with range 1
(29.3-24.0)+ (48.0-43.7)
4.8
2
avg mfr effect
with range 2
43 87
(-)+ (-)
2
PP PP
(35.0-33.0)+ (38.0-39.1)
0.45
2
mfr u range
interaction
0.45 4.8
2.2
2
Interpretation?
27 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Interpretation of Effects
seats
manufacturer
Main effects are
the difference
between two
averages
range
seats
range
mfr
S2
S1
1
2
3
4
5
6
7
8
R1
R2
1
2
3
4
5
6
7
8
M2
M1
1
2
3
4
5
6
7
8
main effect
of mfr
Expt
No.
Seats
(S)
Range
(R)
Mfr
(M)
1 S1 R1 M1
2 S1 R1 M2
3 S1 R2 M1
4 S1 R2 M2
5 S2 R1 M1
6 S2 R1 M2
7 S2 R2 M1
8 S2 R2 M2
8642 7531
(+++)-(+++)
4
PPPP PPPP
Adapted from Fig 10.2 Box, Hunter & Hunter
28 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Interpretation of Effects
Expt
No.
Seats
(S)
Range
(R)
Mfr
(M)
1 S1 R1 M1
2 S1 R1 M2
3 S1 R2 M1
4 S1 R2 M2
5 S2 R1 M1
6 S2 R1 M2
7 S2 R2 M1
8 S2 R2 M2
seats u range
mfr u seats
Interaction effects are
also the difference
between two averages,
but the planes are no
longer parallel
-
seats
range
mfr
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
8541 7632
(+++)-(+++)
4
PPPP PPPP
mfr u range
interaction
Adapted from Fig 10.2 Box, Hunter & Hunter
mfr u range
1
2
3
4
5
6
7
8
29 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Design Experiment
Objective: Maximize Airplane Glide Distance
Design Variables:
Weight Distribution
Stabilizer Orientation
Nose Length
Wing Angle
Three levels for each design variable.
Experiment courtesy of Prof. Eppinger
30 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Design Experiment
Full factorial design : 3
4
=81 experiments
We will use an L
9
(3
4
) orthogonal array:
Expt
No.
Weight
A
Stabilizer
B
Nose
C
Wing
D
1 A1 B1 C1 D1
D2
D3
4 A2 B1 C2 D3
5 A2 B2 C3 D1
6 A2 B3 C1 D2
7 A3 B1 C3 D2
8 A3 B2 C1 D3
D1
2 A1 B2 C2
3 A1 B3 C3
9 A3 B3 C2
31 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
Design Experiment
Things to think about ...
Given just 9 out of a possible 81 experiments,
can we predict the optimal airplane?
Do some design variables seem to have a
larger effect on the objective than others
(sensitivity)?
Are there other factors affecting the results
(noise)?
32 ? Massachusetts Institute of Technology - Prof. de Weck and Prof. Willcox
Engineering Systems Division and Dept. of Aeronautics and Astronautics
References
Phadke, : Quality Engineering Using Robust Design,
Prentice Hall, 1995
Box, G.; Hunter, W. and Hunter, J.: Statistics for
Experimenters, John Wiley & Sons, 1978.