Design under uncertainty

E. Nikolaidis
Aerospace and Ocean Engineering
Department
Virginia Tech
Acknowledgments
Sophie Chen (VT)
Harley Cudney (VT)
Raphael Haftka (UF)
George Hazelrigg (NSF)
Raluca Rosca (UF)
 Outline
• Decision making problem
• Why we should consider uncertainty in
design
• Available methods
• Objectives and scope
• Comparison of probabilistic and fuzzy set
methods
• Concluding remarks
 1. Decision making problem
Noise Design 1
level (db)

Design 2
Design 3

Initial target

Cost ($)
Which design is better ?
Taxonomy of decision problems
(Keney and Raiffa, 1994)

Certainty about Uncertainty about

outcomes of actions outcomes of actions
ONE Type I problems Type II problems
ATTRIBUTE IS Approach: Approaches: Utility
SUFFICIENT Deterministic theory, fuzzy set
FOR optimization theory
DESCRIBING
AN OUTCOME
MULTIPLE Type III problems Type IV problems
ATTRIBUTES Approaches: Utility Approaches: Utility
ARE NEEDED theory, fuzzy set theory, fuzzy set
FOR theory theory
DESCRIBING
AN OUTCOME
 Types of uncertainty

Irreducible:
due to inherent Reducible: due to use
randomness Statistical:
of imperfect models
in physical due to lack
to predict
phenomena and of data for
outcomes of an action
processes modeling
uncertainty
 Preferences
• An outcome is usually described with one
or more attributes
• Preferences are defined imprecisely: no
clear sharp boundary between success and
failure
• Need a rational approach to quantify value
of an outcome to decision maker
– Utility theory
– Fuzzy sets
 2. Why we should consider
uncertainty in design
• Design parameters are uncertain -- there is
no way to make a perfectly safe design
• Ignoring uncertainty and using safety
factors usually leads to designs with
inconsistent reliability levels
 3. Available methods

• Safety factor
• Worst case scenario-convex models
• Taguchi methods
• Fuzzy set methods
• Probabilistic methods
 Probabilistic methods
• Approach
– Model uncertainties using PDF’s
– Estimate failure probability
– Minimize probability of failure and/or cost
• Advantage: account explicitly for probability of failure
• Limitations:
– Insufficient data
– Sensitive to modeling errors (Ben Haim et al., 1990)
 Fuzzy set based methods
• Possibility distributions
• Possibility of event = 1-degree of surprise
(Shackle, 1969)
• Relation to fuzzy sets (Zadeh, 1978):
X is about 10:

Possibility distribution
1

0.25 Probability distribution

8 10 12
 Fuzzy sets in structural design

• Uncertainty in mechanical vibration:

Chiang et al., 1987, Hasselman et al., 1994
• Vagueness in definition of failure of
reinforced plates (Ayyub and Lai, 1992)
• Uncertainty and imprecision in preferences
in machine design (Wood and Antonsson,
1990)
• Relative merits of probabilistic methods and
fuzzy sets may depend on:
– Amount and type of available information
about uncertainty
– Type of failure (crisp or vague)
– Accuracy of predictive models
 Important issues

• Are fuzzy sets better than probabilities in

modeling random uncertainty when little
information is available?
• How much information is little enough to
switch from probabilities to fuzzy sets?
• Compare experimentally fuzzy set and
probabilistic designs
 4. Objectives and scope
• Objectives:
– Compare theoretical foundations of
probabilistic and fuzzy set methods
– Demonstrate differences on example problems
– Issue guidelines -- amount of information
• Scope:
– Problems involving uncertainty
– Problems involving catastrophic failure Æ
clear, sharp boundary between success and
failure
 5. Comparison of probabilistic
and fuzzy set methods
• Comparison of theoretical foundations
– Axiomatic definitions
– Probabilistic and possibility-based models of
uncertainty
– Risk assessment
– Design for maximum safety
• Comparison using a design problem
 Axiomatic definitions
Probability measure, P(⋅) Possibility measure, Π(⋅)
1) P(A) ≥ 0 ∀ A∈S 1) Boundary requirements:
Π(∅)=0, Π(Ω)=1
2) Boundary requirement: 2) Monotonicity:
P(Ω)=1 ∀A, B ∈ S , if A ⊆ B,
then Π ( A) ≤ Π ( B )
3) Probability of union of 3) Possibility of union of a
events finite number of events
∀A i , i ∈ I , Ai are disjoint ∀A i , i ∈ I , Ai disjoint
I I
P ( U Ai ) = ∑ P( Ai ) Π ( U Ai ) = max i∈I (Π ( Ai ))
i =1 i∈I i =1
 Differences in axioms
• Probability measure can be assigned to the
members of a s-algebra. Possibility can be
assigned to any class of sets.
• Probability measure is additive with respect
to the union of sets. Possibility is
subadditive.
P( A) + P( AC ) = 1
Π ( A) + Π ( AC ) ≥ 1
Probability density and possibility distribution
f (x)
functions
X
≠1 Area=1

x0 x
P(X=x0)=0

ΠX(x)
Area≥1
1

x0 x

Π(X=x0)≠0
 Modeling an uncertain variable when very little
information is available
Maximum uncertainty principle: use model that maximizes
uncertainty and is consistent with data

Possibility distribution
1
Probability distribution
0.25

8 10 12

8.5
• Increase range of variation from [8,12] to [7,13]:
– Failure probability: 0.13Æ0.08
– Failure possibility: 0.50 Æ0.67
• Design modification that shifts failure zone from
[8,8.5] to [7.5,8]
– failure probability: 0.13 Æ0 (if range of
variation is [8,12])
– failure probability remains 0.08 (if range of
variations is [7,13])
• Easy to determine most conservative
possibility based model consistent with data
• Do not know what modeling assumptions
will make a probabilistic model more
conservative
• Probabilistic models may fail to predict
effect of design modifications on safety
• The above differences are due to the
difference in the axioms about union of
events
 Risk assessment: Independence of
uncertain variables
• Assuming that uncertain parameters are
independent always makes a possibility
model more conservative. This is not the
case with probabilistic models
P, P PFS=P2 if independent
PFS=P if perfectly
correlated

P, P PFS= P in both cases

where components are
independent or correlated
 A paradox
Probability-possibility consistency:
The possibility of any event should always be greater
or equal to its probability

P, P
...

PFS=1-(1-P)n PFS=P
 1 System failure
probability
P System failure
possibility
P

1
Number
of components
To ensure that failure possibility remains equal
or greater than failure possibility need to impose the condition:
∀A : P ( A) f 0, Π ( A) = 1
 Design for maximum safety
• Probabilistic design : • Possibility-based
– find d1,…, dn design:
– to minimize PFS – find d1,…, dn
– so that g‹0 – to minimize PFS
– so that g‹0

Two failure modes:

PFS=PF1+PF2-PF12 PFS=max(PF1, PF2)
 Optimality conditions
Assume PF12 small

P P PF1
PF1

PF2 PF2

d0 d d0 d

∂PF1 ∂PF 2 PF1=PF2

=−
∂d ∂d
 Comparison using a design problem

• No imprecision in defining failure

• Only random uncertainties
• Only numerical data is available about uncertainties

How to evaluate methods:

Average probability of failure
General approach for comparison
Information
about uncertainties

Incomplete
Budget
information

Optimization: Maximize Safety

Fuzzy Design Probabilistic Design

Probabilistic Analysis Probabilistic Analysis

Compare relative frequencies of failure

 m, wn2 absorber

Normalized M, wn1 original

system system
amplitude y
F=cos(wet)

Figure 2. Tuned damper system

Original SDOF system
 Failure modes
1. Excessive vibration
2. Cost > budget (cost proportional to m)
60

48
b normalized
System amplitude

natural 36

frequencies 24
(assumed
equal) 12

0
0.8 0.84 0.88 0.92 0.96 1 1.04 1.08 1.12 1.16 1.2
β
:R=0.05; : R=0.01

Figure 4. Amplitude of system vs. β, ζ=0.01

 Uncertainties
1 Normalized frequencies
2 Budget
• Know true probability distribution of budget
• Know type of probability distribution of
normalized frequencies and their mean
values, but not their scatter
• Samples of values of normalized
frequencies are available
 Design problem
• Find m
• to minimize PF (PF)

• PF=P(excessive vibration¯cost overrun)

• ½F=P(excessive vibration ¯cost overrun)

• heavy absorber, low vibration but high cost

 Estimation of variance of b

Concept of maximum uncertainty:

if little information is available, assume model
with largest uncertainty that is consistent
with the data
Comparison of ten probabilistic and ten possibility-based designs.
Three sample values were used to construct models of
uncertainties. Blue bars indicate possibility-based designs. Red
bars indicate probabilistic designs.
- Inflation factor method, unbiased estimation

0.35

0.3
Actual probability of failure

0.25

0.2

0.15

0.1

0.05

0
1 2 3 4 5 6 7 8 9 10
Data group
 sample size equal to 3,000
0.25

0.2
Actual probability of failure

0.15

0.1

0.05

0
1 2 3 4 5 6 7 8 9 10
Data group
Probabilistic approach cannot predict
design trends
0.3

Failure probability due

to excessive vibration
0.2

0.1

0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
R

distribution of frequency - U(1,0.05)

distribution of frequency - U(1,0.075)
distribution of frequency - U(1,0.1)

Figure 5. Effect of standard deviations of b1 and b2

on the probability of failure vs. R ,
b1 and b2 are equal
Comparison in terms of average failure
probability as a function of amount of
information
Sample size Best design
3
5
10
20
100
1000
Blue bullet: on average possibility is better,
red bullet: on average probability is better
 Concluding remarks
• Overview of problems and methods for
design under uncertainty
• Probabilistic and fuzzy set methods --
comparison of theoretical foundations
• Probabilistic and fuzzy set methods --
comparison using design problem
 Concluding remarks
• Important to consider uncertainties
• There is no method that is best for all problems
involving uncertainties
• Probabilistic design better if sufficient data is
available
• Possibility can be better if little information is
available
– easier to construct most conservative model consistent
with data
– probabilistic methods may fail to predict effect of
design modifications
• Major difference in axioms about union of events