Professional Documents
Culture Documents
Decision-making techniques are used to select the “best” alternatives under multiple and of-
ten conflicting criteria. Multicriteria decision making (MCDM) necessitates to incorporate
uncertainties in the decision-making process. The major thrust of this article is to extend the
framework proposed by Yager(1) for multiple decisionmakers and fuzzy utilities (payoffs).
In addition, the concept of expert credibility factor is introduced. The proposed approach is
demonstrated for an example of seismic risk management using a heuristic hierarchical struc-
ture. A step-by-step formulation of the proposed approach is illustrated using a hypothetical
example and a three-story reinforced concrete building.
KEY WORDS: Evidence theory; fuzzy; multicriteria decision making; OWA; seismic risk management;
uncertainty
78 0272-4332/10/0100-0078$22.00/1
C 2009 Society for Risk Analysis
Decision Making Under Uncertainty 79
alternatives are left unspecified), and discord or con- complete or partial information of pij (Cases 2–5)
flict (i.e., if there is a disagreement in choosing among requires the use of DS theory in conjunction with
several alternatives). Ambiguity is related to both ordered weighted averaging (OWA).(1) Further, if
nonspecificity and discord. Because of the involve- belief functions are provided by k DMs, who may
ment of human (expert) judgment for the inter- (dis)agree on the assessment, the problem requires
pretation of data and observations, the uncertain- the use of “rule of combinations,” e.g., DS(13−15)
ties become an integral part of the decision-making to determine the combined assessment of the be-
process.(8−10) lief function. The belief functions require to be dis-
Yager(1) defined decision making under uncer- counted before aggregation using the credibility of
tainty (DMUU) as a class of decision problems, when DMs. The credibility factor (dk ) can be related to the
different manifestations of a variable (x) lead to dif- experience of a DM and the confidence on his/her
ferent payoffs. In these problems, the major issue assessment. Cases 2–4 deal with uncertainties asso-
is the representation of knowledge about the vari- ciated to the state pij . For case 5, the utility value
able (x). Many uncertainty quantification methods, xij is assumed “fuzzy,” whereas the knowledge about
including probability theory, possibility theory (fuzzy the state pij is represented by a belief function. These
sets), and Dempster-Shafer (DS) theory (random cases are articulated in such a way that case 1 is a sub-
sets), have been suggested for representing knowl- set of case 2, and case 2 is a subset of case 3, and so
edge. Any appropriate use of a given uncertainty on. Therefore, case 5 becomes the most complex case
formulation in DMUU may lead to the selection of that encompasses all previous cases. These concepts
a suitable alternative using a “valuation function” are discussed in detail subsequently.
V(Ai ). The determination of V(Ai ) depends on the The remainder of the article is organized as fol-
payoffs of various criteria (xij ) involved and knowl- lows. The basic information about OWA operators
edge (pij ) of expert(s) about the state of the vari- is provided in Section 2. Section 3 provides a discus-
able.(1) The decision-making problem becomes more sion on DS theory and associated different rule of
complex when multiple experts are involved who combinations. This section also discusses a method
have different levels of credibility about their knowl- to determine the credibility factor (dk ) for an assess-
edge of the state or value of the payoffs. In this arti- ment of a DM to discount belief function (knowl-
cle, we have chosen five cases in the context of SRM, edge about the state pij ). Section 4 provides a back-
which are ranked in the increasing order of associ- ground on fuzzy set theory to deal with vagueness in
ated uncertainties:
Case 1 can be best handled using probability the utility value xij . Section 5 provides a hypothetical
(Bayesian) theory, in which the valuation function example for risk management to explain the above
or a representative value V(Ai ) is an expected value, five cases. Section 6 discusses a case study for SRM
to select the most suitable repair alternative under
i.e., E(Ai ) = nj=1 xi j pi j ; where pij is a disjoint prob-
ability distinctively known for each xij .(11) For a given uncertainty using case 5. The final section provides
alternative Ai , the basic probability axiom holds, the summary and conclusions.
pi = nj=1 pi j = 1, i.e., the sum of probabilities (also
referred as weights) is unity. The DMUU problem 2. OWA OPERATOR
becomes more involved if these probabilities are
2.1. Basic Concepts
not “disjoint” (nonspecific) (this warrants the use
of belief functions(12) ) or if the utility values xij are Yager(16) introduced a new family of aggre-
fuzzy.(1) The treatment of uncertainties related to in- gation techniques called the OWA operators to
80 Tesfamariam, Sadiq, and Najjaran
generalize the concept of “and” (minimum) and “or” Table I. Transformation of Linguistic Degree of Optimism
(maximum) logic. The OWA operators are gener- to Orness Values
alized mean operators that encompass the range of Degree of Optimism Orness
“anding” or “oring” and can incorporate the attitude
of a DM in the aggregation process. The OWA oper- Optimistic 0.9
Moderately optimistic 0.7
ator plays important role if available information is
Normative 0.5
not enough to assign disjoint pij to payoffs. Moderately conservative 0.3
An OWA operator of dimension n is a map- Conservative 0.1
ping OWA : R n → R that has an associated weight-
ing vector W = (w1 , w2 , . . . , wn )T . The require-
ments
n to be satisfied are R ∈ [0, 1], w j ∈ [0, 1], and
O’Hagan,(21) which entails the use of or ness and
j=1 w j = 1. For a given vector of payoffs for n cri-
optimizes the Disp function (entropy). These calcu-
teria (x1 , x2 , . . . , xn ), the OWA aggregation is per-
lated weights are referred to as maximum entropy
formed as follows:
(ME-) OWA weights. The objective function and as-
n
sociated constraints can be written as:
OWA(x1 , x2 , . . . , xn ) = w j bj , (2)
j=1
n
Objective function: Max(Disp) = Max − wi ln(wi )
where b j is the jth largest element in the vector i=1
(x1 , x2 , . . . , xn ), i.e., b1 ≥ b2 ≥ · · · ≥ bn . The weights 1
n
w j of OWA are not associated with any particular Subjected to: orness(W) = wi (n − i);
n−1
x j ; rather, these are associated with the ordinal posi- i=1
These weights represent the fraction of pij dis- where it can be shown that:
tributed among (x1 , x2 , x3 , x4 ) .
bel(φ) = 0; bel(x) = 1. (4b)
(3) Aggregate using Equation (2):
The upper bound, “plausibility,” is the summa-
V(A) = OW A(x1 , x2 , x3 , x4 )
tion of “basic probability assignment” of the sets Xr
= 0.46 × 0.8 + 0.28 × 0.7 + 0.16 × 0.5 that intersect with the set of interest Xl , i.e., Xr ∩ Xl
+ 0.1 × 0.45 = 0.69. = φ, and therefore it can be written as:
Therefore, based on DM’s attitude “mod- pl(Xl ) = p(Xr ). (5)
erately optimistic,” the representative value Xr ∩Xl=φ
function is 0.69.
The belief interval is an interval representing
range [plausibility – belief ], where “true” probabil-
3. DS THEORY ity may lie. A narrow belief interval represents more
precise probabilities. It can be shown that the prob-
Evidential reasoning or DS theory is named af- ability is determined uniquely if bel(Xl ) = pl(Xl );
ter Dempster(22) and Shafer,(13) which, in certain con- therefore, probability theory is applicable only where
texts, provides a generalization for Bayesian theory. all probabilities are unique and disjoint.(14) The be-
The DS theory is suited to handle nonspecific, con- lief interval of “one” implies that no information is
flicting, and ambiguous information. The DS theory available for Xl and “zero” implies that it has been
combines the rigor of probability theory with the completely confirmed by p(Xl ).
flexibility of rule-based systems.
3.3. Extensions of DS Rule of Combination results than the DS rule of combination, but provides
more specific results than the Yager rule of combina-
Zadeh(24) has highlighted the shortcoming of
tion that translates the specificity of a case, e.g., from
handling the conflict and normalization of DS rule of
{x1 , x2 } into ignorance x = {x1 , x2 , . . . , xn }. The DP
combination. To address these issues, various mod-
rule is described as:
ifications of the DS rule of combination have been
reported. The most common extensions on the DS pDP (Xl ) = p1 (Xp ) p2 (Xq )
rule of combination include Yager,(17) Smets,(25) Ina- Xp∩Xq=Xl
Xp∩Xq=φ
gaki,(26) Dubois and Prade,(15) Zhang,(27) Murphy,(28)
and more recently by Smarandache and Dezert.(29) + p1 (Xp ) p2 (Xq ). (9)
Sentz and Ferson(12) provided a comprehensive but Xp∪Xq=Xl
Xp∩Xq=φ
nonexhaustive review of these extensions. The ma-
jor differences among various rules of combination The DP rule is commutative and associative but
are handling of the conflicting mass and the closed not idempotent.
world (exhaustive) assumption of power set (e.g., tra-
ditional DS is closed over union, whereas the trans-
3.4. Discounted Evidence
ferable belief model of Smets(25) uses open world as-
sumption, i.e., “frame of discernment” is assumed The formulation described above assumes im-
nonexhaustive). A detailed discussion on this topic plicitly that all sources of information are equally
can be found in Dezert and Smarandache.(29) In this credible. The bodies of evidence obtained from dif-
article, two modified rules, Yager and Dubois and ferent sources (i.e., k DMs) need to be discounted
Prade (DP), are discussed. using a credibility factor (dk ) as:
p(Xl )dk = p(Xl ) · dk
3.3.1. Yager (Yg) Rule of Combination . (10)
p(x)dk = p(x) · dk + (1 − dk)
The rule of combination proposed by Yager(14)
The credibility factor (dk ) is constrained by 0 ≤
is similar to DS rule of combination, except that joint
dk ≤ 1, where zero represents the “fully incredible
evidence is not normalized with nonconflicting evi-
evidence,” and one represents “fully credible evi-
dence. Thus,
dence.”
pYg (Xl ) = p1 (Xp ) p2 (Xq ). (7) Yager(1) also discussed the discounting issue
Xp∩Xq=Xl in detail and suggested a credibility transformation
function. This approach discounts the evidence us-
The total conflicting evidence (K) is shifted to ig- ing dk , but distributes the remaining evidence (1 −
norance (x) and pYg (x) is given by: dk ) equally among the other elements of the frame
of discernment:
pYg (x) = p1 (Xp ) p2 (Xq ). (8)
1 − dk
Xp∩Xq=x
Xp∩Xq=φ
p(Xl )dk = p(Xl ) • dk + , (11)
g
Sentz and Ferson(12) describe this rule as an where g is the number of the focal elements in the
“epistemologically honest” interpretation of a body frame of discernment. The formulation provided in
of evidence, which does not change joint evidence Equation (10) is used for discounting of evidence.
through the normalization process by nonconflicting The assessment of pij for a single DM does not
evidence. The Yager rule of combination is commu- require discounting. However, for multiple DMs, the
tative, but not idempotent and associative. assessment of pij needs to be discounted using a cred-
ibility factor (dk ) (i.e., an expert’s weight). Kiremid-
jian(30) proposed a technique to determine dk based
3.3.2. Dubois and Prade (DP) Rule of Combination on DM’s experience qk ∈ [0, 10], and his/her confi-
Dubois and Prade(15) proposed a change in the dence fk ∈ [0, 1] in the assessment. In this study, the
DS rule of combination using disjunctive consensus. formulation is modified using a normalization factor
K β
Therefore, in the DP rule of combination, the mass Maxk=1 ( fkqk ). The normalization factor is derived
is assigned to disjunction to avoid normalization. It such that the expert’s weight for the most credible
is noted that this method may yield more nonspecific DM is 1 and for remaining DMs dk ≤ 1. Therefore,
Decision Making Under Uncertainty 83
Table II. Derivation of Credibility Factor (dk ) for β = 1 uncertainties can effectively be handled using fuzzy
β
set theory.(31)
k qk fk fkqk dk
however, the common defuzzification methods in- Table IV. Payoff (xij ) for Two Alternatives and Corresponding
clude center of area, first of maximums, last of Four Criteria
maximums, and middle of maximums. Different de- xij C1 C2 C3 C4
fuzzification techniques extract different levels of
information, and consequently are prone to rank A1 0.90 0.40 0.50 0.35
reversal.(34) In this article, generalized mean and A2 0.45 0.50 0.80 0.70
standard deviation are used to determine the rank
fuzzy numbers.(33) The definitions of generalized
mean and standard deviation are as follows:
5. RISK MANAGEMENT UNDER
b UNCERTAINTY
xμx (x) dx In this section, a hypothetical example for a risk
x̄ = a b
(13) management decision-making problem is provided,
μx (x) dx where two alternatives Ai (i = 1, 2) are being eval-
a
uated based on four criteria Cj (j = 1, 2, 3, 4). For
each alternative and corresponding criterion, payoff
values xij are provided in Table IV. Assume that
⎡ b ⎤1/2 through DM’s knowledge elicitation, bpa values and
⎢ a x 2 μx (x) dx ⎥ credibility of DMs are as provided in Table V. In the
σ =⎢
⎣ b
− [x̄]2 ⎥
⎦ , (14) following five subsections (Sections 5.1–5.5), a step-
μx (x) dx by-step approach is provided to solve these prob-
a lems.
Credibility Payoff
Case Decision maker of DM (xij ) Alternative A1 p(1 )§ Alternative A2 p(2 )§
∗
1 1 1.0 Crisp p11 = p21 = 0.50; p12 = p22 = 0.20; p13 = p23 = 0.10; p14 = p24 = 0.20
(Table IV)
2 1 1.0 Crisp p(C3 ) = 0.50 0.50 p(C4 ) = 0.20; p(C1 , C2 ) = 0.70 0.10
(Table IV)
3, 4 1 1.0, 0.6a Crisp p(C3 ) = 0.50 0.50 p(C2 , C3 ) = 0.70 0.30
(Table IV)
2 1.0, 1.0b p(C2 , C4 ) = 0.60 0.40 p(C1 , C4 ) = 0.40 0.60
3 1.0, 0.7c p(C1 , C2 ) = 0.70 0.30 p(C3 ) = 0.80 0.20
5 1 1.0 Fuzzy p(C3 ) = 0.50 0.50 p(C4 ) = 0.20; p(C1 , C2 ) = 0.70 0.10
(Table VII)
Valuation, V(A i)
As E(A1 ) > E(A2 ), thus A1 is the preferred al- 0.6
ternative. 0.4
0.2
5.2. Case 2
0.0
For each criterion Cj , the bpa p(Cj ) values may 0.0 0.2 0.4 0.6 0.8 1.0
Orness, α
not easily be discernable, and often, we may group
more than one criteria, e.g., {C1 , C2 }. As a result, Fig. 2. Sensitivity analysis based on orness values in the valuation
the p(Cj ) corresponds with {C1 , C2 } and is evaluated of two alternatives.
using decision making under ignorance (DMUI),
V(Ai ) = OWA(C1 , C2 ). The overall valuation of al- Step 4: Combine these two valuations using Equation
ternative A is: (15):
q
q
V(A) = Vj (A) p(C j ). (15) V(A1 ) = Vj (A1 ) p(C j )
j=1 j=1
Step 1: Define bpa (Table V) and the payoff values = 0.5 × 0.5 + 0.653 × 0.50 = 0.577, and
(Table IV):
q
A1 : A1 (C3 ) = <0.50> and V(A2 ) = Vj (A2 ) p(C j )
A1 () = <0.9, 0.4, 0.50, 0.35>, j=1
A2 : A2 (C4 ) = <0.7>, A2 (C1 , C2 ) = <0.45, 0.5> = 0.70 × 0.2 + 0.485 × 0.7 + 0.689 × 0.1
and A2 () = <0.45, 0.5, 0.8, 0.7>. = 0.548.
Step 2: Define decision-making attitude, e.g., mod- As V(A1 ) > V(A2 ), thus A1 is a preferred alter-
erately optimistic, thus orness = 0.7 (Table I). The native.
corresponding ME-OWA weights for n = 2, 3, and 4
In DMUI, the orness plays pivotal role. To see
are:
the impact of orness values toward the evaluation
Wn=2 = (w1 , w2 ) = (0.7, 0.3), V(A1 ) and V(A2 ), sensitivity analysis is carried out
and results are plotted in Fig. 2. At lower orness val-
Wn=3 = (w1 ,w2 ,w3 ) = (0.554, 0.292, 0.154), and ues, i.e., <0.33, V(A2 ) > V(A1 ), this implies that A2
is the preferred alternative. However, with increas-
Wn=4 = (w1 ,w2 ,w3 ,w4 ) = (0.461, 0.276, 0.165, 0.098). ing orness values, i.e., >0.38, V(A1 ) > V(A2 ), this
implies that the A1 is the preferred alternative. Be-
Step 3: The value for each alternative can be calcu- tween the orness intervals of [0.33, 0.38], there is no
lated as Vj (Ai ) = OWA(Ai (Ci )): significant different between V(A2 ) and V(A1 ), and
V1 (A1 ) = OWA(0.5) = 0.5, thus any of the two alternatives can be selected. For
orness = 0.7 (Table I), the valuation of the two alter-
V2 (A1 ) = OWA(0.9, 0.4, 0.50, 0.35) natives is V(A1 ) = 0.63 and V(A2 ) = 0.76. It implies
that A2 is a desired alternative.
= 0.461 × 0.9 + 0.276 × 0.50 + 0.165 × 0.4
+ 0.098 × 0.35 = 0.653.
5.3. Case 3
Similarly, for A2 :
Let us assume that the bpas are elicited from
V1 (A2 ) = 0.7, three DMs, who are equally credible (Table V). Once
these bpas are obtained, the bpa of the two alterna-
V2 (A2 ) = 0.485, and tives Ai (i = 1, 2) are calculated using the DS rule
of combination (Equation (6)) (Table VI). Sensitiv-
V3 (A2 ) = 0.689. ity analysis by varying orness values is carried out
86 Tesfamariam, Sadiq, and Najjaran
Table VI. DS Rule of Combination for Three Experts factor, and assigned to ignorance. The bpas for the
two alternatives after the DS rule of combination is
Alternatives
applied (Table VI). The variations in the two alter-
A1 A2 native valuations with respect to orness are provided
in Figs. 3b (DS), 3d (Yg), and 3f (DP) for different
Case 3 Case 4 Case 3 Case 4
rules of combination. It is interesting to note that the
Credibility (dk ) = 1 <> 1a 1 <> 1a discounted valuations are consistently lower than the
Subsets p(A1 ) p (A1 ) p (A2 ) p (A2 ) nondiscounted valuations. However, at orness = 1,
the impact of credibility factors diminishes.
{C1 } 0.00 0.00 0.00 0.00
{C2 } 0.38 0.27 0.00 0.00
{C3 } 0.11 0.07 0.77 0.48
5.5. Case 5
{C4 } 0.00 0.00 0.00 0.00
{C1 , C2 } 0.25 0.18 0.00 0.00 This case entails consideration of vagueness and
{C1 , C3 } 0.00 0.00 0.00 0.00
imprecision of the payoff values x̄i j . The correspond-
{C1 , C4 } 0.00 0.00 0.04 0.15
{C2 , C3 } 0.00 0.00 0.13 0.15 ing fuzzy payoffs x̄i j are provided in Table VII. For
{C2 , C4 } 0.16 0.28 0.00 0.00 x̄i j , the fuzzy number OWA (FN-OWA) is used.
{C3 , C4 } 0.00 0.00 0.00 0.00 The outlined procedure for FN-OWA is discussed
{C1 , C2 , C3 } 0.00 0.00 0.00 0.00 in Sadiq and Tesfamariam.(18) The ranking of x̄i j is
{C1 , C2, C4} 0.00 0.00 0.00 0.00
performed based on the most likely values. The val-
{C1, C3, C4} 0.00 0.00 0.00 0.00
{C2, C3, C4} 0.00 0.00 0.00 0.00 uation results for alternatives Āi (i= 1, 2) are shown
{C1, C2, C3, C4} 0.11 0.19 0.06 0.23 in Fig. 4. Figs. 4a and 4b show the impact of orness
values (using DS rule of combination) on the valua-
a Credibility factors, DM 1 = 0.6, DM 2 = 1.0, and DM 3 = 0.7. tion of alternatives A1 and A2 , respectively. Figs. 4c
and 4d provide the similar comparison for both al-
ternatives using Yager rule of combination. Finally,
Figs. 4e and 4f provide the results for DP rule of com-
(Fig. 3). For all orness values, for both DST (Fig. 3a) bination.
and Yager (Fig. 3c) rule of combinations, V(A2 ) is The major challenge in this problem is to con-
consistently higher than V(A1 ). However, for the vert the TFN into representative crisp values for fi-
Yager rule of combination, at orness = 1, the valu- nal ranking. At orness = 0, the valuation V( Ā1 ) and
ation of V(A1 ) ≈ V(A2 ). V( Ā2 ) are TFN(0.25, 0.35, 0.50) and TFN(0.21, 0.45,
For DP rule of combination, for orness ∈ [0, 7), 0.70), respectively (Figs. 4a and 4b). If the most likely
V(A2 ) is higher than V(A1 ); therefore, A2 is the value is used for comparison, the V( Ā2 ) > V( Ā1 );
preferred alternative. However, for orness ∈ (0.7, therefore, Ā2 will be the desired alternative. The
1], V(A1 ) is higher than V(A2 ), and A1 is the pre- base of the TFN gives information on the impreci-
ferred alternative that refers to as rank reversal sion (vagueness). The difference between maximum
(Fig. 3e). and minimum values is 0.25 and 0.49 for TFN Ā1 and
Ā2 , respectively. This indicates V( Ā2 ) has higher un-
certainty; however, it is the desired alternative based
5.4. Case 4
on most likely value. Thus, an alternative ranking is
The multiple DMs in the previous example were desired to handle this type of problem.
assumed to have the same credibility. However, reli- The ranking based on generalized mean
ability of the source and credibility need to be taken x̄(V( Āi )) and standard deviation σ (V( Āi )) is dis-
into consideration.(12) The expert’s credibility can be cussed below. For orness = 0, using Equations (13)
elicited by considering the experience qk and confi- and (14), the x̄(V( Āi ))and σ (V( Āi )) are computed
dence levels ck (use Equation (12)). However, it sums as:
up to one, but in our analysis we divide it by the max-
V( Ā1 ) = (0.389, 0.052)
imum value so that the DM with the highest credi-
bility will be assigned a value of 1. Assume that we
V( Ā2 ) = (0.475, 0.100).
obtain the credibility of each DM as DM 1 = 0.6,
DM 2 = 1.0, and DM 3 = 0.7. Each of the bpa as- This result indicates that A2 is the preferred
signed by the DM are discounted with the credibility alternative.
Decision Making Under Uncertainty 87
1.0 1.0
a) V(A2) V(A1) b) V(A2) V(A1)
0.8 0.8
Valuation, V(Ai)
Valuation, V(Ai)
0.6 0.6
0.4 0.4
1.0 1.0
c) V(A2) V(A1) d) V(A2) V(A1)
Valuation, V(Ai)
orness in the valuation of two alternatives
Ai (i = 1, 2) (a) DST rule of combination, 0.6 0.6
(b) DST rule of combination with
credibility factor, (c) Yager rule of 0.4 0.4
combination, (d) Yager rule of
combination with credibility factor, (e) 0.2 0.2 Yager rule of combination
Yager rule of combination
DP rule of combination, (f) DP rule of without credibility factor with credibility factor
combination with credibility factor. 0.0 0.0
0.0 0.5 1.0 0.0 0.5 1.0
1.0 1.0
e) V(A2) V(A1) f) V(A2) V(A1)
0.8 0.8
Valuation, V(Ai)
Valuation, V(Ai)
0.6 0.6
0.4 0.4
Alternatives C1 C2 C3 C4
Table VII. Fuzzy Payoffs (x̄i j ) for Two
Alternatives and Corresponding A1 [0.9, 0.9, 0.9] [0.2, 0.4, 0.45] [0.2, 0.50, 0.8] [0.25, 0.35, 0.5]
Four Criteria A2 [0.2, 0.45, 0.7] [0.4, 0.5, 0.65] [0.75, 0.8, 0.85] [0.6, 0.7, 0.75]
6. SRM FOR BUILDINGS and poor code enforcement. Most of these buildings
are currently operational and are required to be fur-
6.1. Background
ther assessed and upgraded to minimize seismic dam-
Reported damages from recent global earth- age and improve life safety. From a decision-making
quakes, 2008 Sichuan earthquake in China, 2004 perspective of a city manager, comprehensive evalu-
Sumatra earthquake in Indonesia, 2003 Bingöl earth- ations of all buildings are not economically feasible,
quake in Turkey, and 1994 Northridge earthquake and it is desirable to screen out deficient buildings.
in USA, for example, highlight vulnerability of exist- The seismic risk analysis (SRA) is used to cal-
ing buildings and importance of seismic retrofit im- culate the probability of adverse economic or social
plementation. The building vulnerability is due to effects of an earthquake or series of earthquakes.(35)
older building design codes, poor design practices, Results of the SRA are used to determine(35) (i) what
88 Tesfamariam, Sadiq, and Najjaran
level of risk is acceptable? (ii) What level of loss is risk management framework entails consideration of
acceptable? and (iii) What options are available for retrofit and/or insurance. The framework shown in
mitigating the anticipated losses? The SRM entails Fig. 5 can be used iteratively to ensure desired relia-
selection of appropriate retrofit strategy to reduce bility (or acceptable risk).
building damage(36) and/or consideration of buying The main objectives of SRM coincide with three
insurance. philosophies of new buildings design: (i) protect the
A conceptual framework for SRM is shown in structure from collapse in a future strong earthquake,
Fig. 5. The risk assessment shown in Fig. 5 includes (ii) keep damage at tolerable levels in earthquakes
consideration of site seismic hazard, building vul- of moderate intensity, and (iii) eliminate damage in
nerability, and the consequence of failure.(37) The earthquakes with relatively short return period.(38)
Decision Making Under Uncertainty 89
No
Base Shear
Base Shear
rehabilitated structure
ΔV ΔV
Fig. 6. Alternative retrofit strategies for
seismic upgrading.(38)
Δu Δu
Installation Maintenance Duration of works/ Functional Skilled labor Significance of the Significance of Significance of
cost cost disruption of use compatibility requirement need for intervention at damage risk limitation risk
(C1) (C2) (C3) (C4) (C5) foundation (C6) (C7) (C8)
Alternative C1 C2 C3 C4 C5 C6 C7 C8
A1 0.241 0.147 0.187 0.869 0.690 0.182 0.398 0.861 Table IX. Payoffs x̄i j of Four Repair
A2 0.564 0.728 0.690 0.120 0.200 0.951 0.434 0.006 Alternatives Against Eight Criteria
A3 0.117 0.255 0.192 0.443 0.087 0.186 0.723 0.509
A4 0.781 0.619 0.673 0.184 0.690 0.166 0.362 0.000
MCDM. The parameters that influence desired anchoring of individual perceptions of earthquake
retrofit selection could be grouped into “general cri- risk.(38,56,57) The experts should have greater diver-
teria” and “technical criteria.”(38) The general crite- sity in terms of experience of institutional affiliations.
ria are: consideration of cost, durability, availability Consequently, the evidence should be discounted
of workmanship, possibility for appropriate quality based on the credibility of the DM as discussed in
control, occupancy, aesthetics, preservation of archi- Section 3.4.
tectural identity, and duration of work. The techni-
cal criteria can broadly be categorized as: enhance-
6.2. Illustrative Example
ment of system ductility, enhancement of stiffness
and strength, and enhancement ductility, stiffness A MCDM hierarchical structure that incorpo-
and strength. rates both technical and general criteria is depicted in
Information for the MCDM parameters is often Fig. 7. This illustrative example is based on a three-
obtained from experts. Identification of experts and story reinforced concrete building constructed in the
information solicitation is shown to be a daunting European Laboratory for Structural Assessment.(58)
task for the earthquake damage evaluation.(55) The Four potential retrofit alternatives Ai (i = 1, 2, 3, 4)
information solicitation from experts in the ATC- are considered, A1 = confinement by glass fiber re-
13,(55) for example, is done through Delphi proce- inforced plastic, A2 = steel bracing, A3 = concrete
dure. The Delphi procedure consists of formulat- jacketing, and A4 = base isolation. For the build-
ing questionnaires, obtaining individual answers to ing under consideration, Caterino et al.(58) have un-
the questioners, iterating the questionnaires one or dertaken a structural analysis and expert elicitation
more times where the information feedback between to evaluate performance of the four retrofit alterna-
rounds and aggregating the responses by statistical tives. Payoff values xi j of each performance modifier
operations. Often solicitation of information from (Fig. 7) with respect to each alternative are summa-
experts is prone to biases, misrepresentation, and rized in Table IX. Furthermore, in this article, the
Decision Making Under Uncertainty 91
crisp payoff values xi j are fuzzified to generate fuzzy Decisionmaker 2: p(E1 ) = 0.30, p(E2 ) = 0.50,
payoff value x̄i j . For illustration, a fuzzification fac- p() = 0.20; and
tor ±15% is used, i.e., x̄i j = [0.85 xi j , xi j , 1.15 xi j ]. Let Decisionmaker 2: p(E1 ) = 0.40, p() = 0.6.
us assume that bpa are elicited from three DMs, each
having a credibility factor of 1 and are summarized
below. The bpa values for all four alternatives with Using the DST rule of combination (Equation
respect to socioeconomical factors are: (6)), the results of the four alternatives are shown
Decisionmaker 1: p(C4 ) = 0.60, p() = 0.40; in Fig. 8. Figs. 8a–8d provide TFNs of V( Ā1 ),V( Ā2 ),
Decisionmaker 2: p(C2 , C4 ) = 0.70, p() = V( Ā3 ), and V( Ā4 ). The ranking is performed using
0.30; and generalized mean x̄(V( Āi ) and standard deviation
Decisionmaker 3: p(C2 , C4 ) = 0.90, p() = σ (V( Āi ). The results are summarized below.
0.10. These results indicate that for all orness values,
V( Ā2 ) and V( Ā4 ) ranked first and second, respec-
The bpa values for all alternatives with respect to tively. However, with a change in orness, there is a
technical factors are: possible rank reversal for V( Ā1 ) and V( Ā3 ). Over-
Decisionmaker 1: p(C6 ) = 0.50, p() = 0.50; all, A2 is a preferred alternative. Sensitivity analy-
Decisionmaker 2: p(C6 , C8 ) = 0.70, p() = sis based on the fuzzification factor is performed and
0.30; and the results are summarized in Fig. 9 (orness = 0). By
Decisionmaker 3: p(C5 , C6 , C8 ) = 0.85, varying fuzzification factor from 10 to 50%, Fig. 9
p() = 0.15. shows that there is no rank reversal. However, for a
fuzzification factor of 5%, the rank of A4 is changed
The bpa values obtained for socioeconomical E1 from 2 to 1. Fig. 10 demonstrates that (orness = 0.5)
and technical E2 criteria are: for all fuzzification factors (10–50%), no rank rever-
Decisionmaker 1: p(E2 ) = 0.60, p() = 0.40; sal is observed.
92 Tesfamariam, Sadiq, and Najjaran
∗ The above table shows the results for fuzzification factor of 15%.
3. Asante-Duah DK. Hazardous Waste Risk Assessment. Ann 27. Zhang L. Representation independence and combination of
Arbor: Lewis Publishers, 1993. evidence in the Dempster–Shafer theory. Pp. 51–69 in Yager
4. Paté-Cornell ME. Global risk management. Journal of Risk RR, Kacprzyk J, Fedrizzi M (eds). Advances in Dempster–
and Uncertainty, 1996; 12:239–255. Shafer Theory of Evidence. New York: John Wiley and Sons,
5. Kahneman D, Tversky A. Prospect theory: An analysis of de- 1994.
cisions under risk. Econometrica, 1979; 47:313–327. 28. Murphy CK. Combining belief functions when evidence con-
6. Wen YK, Ellingwood BR, Veneziano D, Bracci J. Uncertainty flicts. Decision Support Systems, 2000; 29:1–9.
Modeling in Earthquake Engineering. MAE Center Project 29. Smarandache F, Dezert J. Advances and Applications of
FD-2 Report, 2003. DSmT for Information Fusion. Rehoboth, NM: American Re-
7. Klir GJ, Yuan B. Fuzzy Sets and Fuzzy Logic: Theory and search Press, 2004.
Applications. Upper Saddle River, NJ: Prentice Hall Interna- 30. Kiremidjian A. Subjective probabilities for earthquake dam-
tional, 1995. age and loss. Structural Safety, 1985; 1:309–317.
8. Grossi P, Kleindorfer P, Kunreuther H. The impact of un- 31. Zadeh LA. Fuzzy sets. Information Control, 1965; 8:338–353.
certainty in managing seismic risk: The case of earthquake 32. Ferson S, Hajagos JG. Arithmetic with uncertain numbers:
frequency and structural vulnerability. Financial Institution Rigorous and (often) best possible answers. Reliability Engi-
Center, Technical Paper #99-23, Philadelphia, PA: Wharton neering and Systems Safety, 2004; 85:135–152
School, University of Pennsylvania, 1999. 33. Chen SJ, Hwang CL. Fuzzy Multiple Attribute Decision Mak-
9. Hall JW, Le Masurier JW, Baker-Langman EA, Davis ing: Methods and Applications. New York: Springer-Verlag,
JP, Taylor CA. A decision-support methodology for 1992.
performance-based asset management. Civil Engineering 34. Prodanovic P, Simonovic SP. Comparison of fuzzy set rank-
and Environmental Systems, 2004; 21(1):51–75. ing methods for implementation in water resources decision-
10. Blockley DI, Baldwin JF. Uncertain inference in knowledge- making. Canadian Journal of Civil Engineering, 2002; 29:692–
based systems. ASCE Journal of Engineering Mechanics, 701.
1987; 113(4):467–481. 35. EERI Committee on Seismic Risk. The basics of seismic risk
11. Fagin R, Halpern JY, Megiddo N. A logic for reason- analysis. Earthquake Spectra, 1989; 5(4):675–702.
ing about probabilities. Information Systems Research, 1990; 36. Dean WE. Seismic risk and management in California. In
87(1–2):78–128. Molak V (ed). Fundamentals of Risk Analysis and Risk Man-
12. Sentz K, Ferson S. Combination of Evidence in Dempster– agement. New York: CRC Press, 1997.
Shafer Theory. SAND 2002-0835, 2002. 37. Tesfamariam S, Saatcioglu M. Seismic risk assessment of re-
13. Shafer G. A Mathematical Theory of Evidence. Princeton, NJ, inforced concrete buildings using fuzzy synthetic evaluation.
USA: Princeton University Press, 1976. Journal of Earthquake Engineering, 2008; 12(7):1157–1184.
14. Yager RR. On the Dempster–Shafer framework and new 38. Penelis GG, Kappos AJ. Earthquake-Resistant Concrete
combination rules. Information Sciences, 1987; 41:93– Structures. E & FN SPON. London, UK: Chapman and Hall,
137. 1997.
15. Dubois D, Prade H. On the combination of evidence in vari- 39. Thermou GE, Pantazopoulou SJ, Elnashai AS. Design
ous mathematical frameworks. Pp. 213–241 in Flamm J, Luisi methodology for seismic upgrading of substandard reinforced
T (eds). Reliability Data Collection and Analysis. Brussels: concrete structures. Journal of Earthquake Engineering, 2007;
ECSC, EEC, EAFC, 1992. 11(4):582–606.
16. Yager RR. On ordered weighted averaging aggregation in 40. Rodriguez M, Park R. Repair and strengthening of reinforced
multicriteria decision making. IEEE Transactions on Systems, concrete buildings for seismic resistance. Earthquake Spectra,
Man and Cybernetics, 1988; 18:183–190. 1991; 7(3):439–460.
17. Yager RR, Filev DP. Induced ordered weighted averaging 41. Sugano S. Seismic strengthening of existing reinforced con-
operators. IEEE Transactions System Man and Cybernetics, crete buildings in Japan. Bulletin of New Zealand National
1999; 29:141–150. Society of Earthquake Engineering, 1981; 14(4):209–222.
18. Sadiq R, Tesfamariam S. Probability density functions based 42. Thermou GE, Elnashai AS. Seismic retrofit schemes for RC
weights for ordered weighted averaging (OWA) operators: structures and local-global consequences. Progress in Struc-
An example of water quality indices. European Journal of Op- tural Engineering and Materials, 2006; 8(1):1–15.
erational Research, 2007; 182:1350–1368. 43. FEMA 227/228. A Benefit-Cost Model for the Seismic Reha-
19. Xu Z. An overview of methods for determining OWA weights. bilitation of Buildings, Volume 1: A User’s Manual and Vol-
International Journal of Intelligent Systems, 2005; 20:843–865. ume 2: Supporting Documentation. Washington, DC: Federal
20. Filev DP, Yager RR. On the issue of obtaining OWA operator Emergency Management Agency, 1992.
weights. Fuzzy Sets and Systems, 1998; 94:157–169. 44. Smyth AW, Altay G, Deodatis G, Erdik M, Franco G, Gülkan
21. O’Hagan M. Aggregating template or rule antecedents in real- P, Kunreuther H, Lus H, Mete E, Seeber N, Yüzügüllü Ö.
time expert systems with fuzzy set logic. Pp. 681–689 in Pro- Probabilistic benefit-cost analysis for earthquake damage mit-
ceedings of the 22nd Annual IEEE Asilomar Conference on igation: Evaluating measures for apartment houses in Turkey.
Signals, Systems and Computers, Pacific Grove, CA: IEEE Earthquake Spectra, 2004; 20(1):171–203.
and Maple Press, 1988. 45. Grossi P. Assessing the Benefits and Costs of Earthquake Mit-
22. Dempster A. Upper and lower probabilities induced by a igation. Working Paper 98-11-11, Risk Management and De-
multi-valued mapping. Annals of Statistics, 1967; 28:325–339. cisions Processes Center, Wharton School, Philadelphia, PA,
23. Klir GJ. Principles of uncertainty: What are they? Why do we 1998.
need them? Fuzzy Sets and Systems, 1995; 74:15–31. 46. Kanda J, Shah H. Engineering role in failure cost evaluation
24. Zadeh LA. Review of books: A mathematical theory of evi- for buildings. Structural Safety, 1997; 19(1):79–90.
dence. AI Magazine, 1984; 5(3):81–83. 47. Goda K, Hong HP. Optimal seismic design considering risk
25. Smets P. Data fusion in the transferable belief model. In Pro- attitude, societal tolerable risk level, and life quality criterion.
ceedings of the Third International Conference on Informa- Journal of Structural Engineering, 2006; 132(12):2027–2035.
tion Fusion, Paris, France, 2000. 48. Wilkinson S, Ying F. Life cycle costing decisions for
26. Inagaki T. Interdependence between safety-control policy and retrofitting earthquake vulnerable buildings. P. 8 in Eighth
multiple sensor scheme via Dempster–Shafer theory. IEEE U.S. National Conference on Earthquake Engineering, San
Transactions on Reliability, 1991; 40(2):182–188. Francisco, CA, 2006.
94 Tesfamariam, Sadiq, and Najjaran
49. Wen YK, Shinozuka M. Cost-effectiveness in active struc- Available at: http://www.gmu.edu/jbc/Tyler/Cowen%20on%
tural control. Engineering Structures, 1998; 20(3):216– 20cost%20benefit.doc, 1998.
221. 54. Jirsa JO. Divergent issues in rehabilitation of existing build-
50. Dodo A, Xu N, Davidson RA, Nozick LK. Optimizing re- ings. Earthquake Spectra, 1994; 10(1):95–112.
gional earthquake mitigation investment strategies. Earth- 55. ATC-13. Earthquake Damage Evaluation Data for California.
quake Spectra, 2005; 21(2):305–327. Applied Technology Council, 1985.
51. Xu N, Davidson RA, Nozick LK, Dodo A. The risk-return 56. Celsi R, Wolfinbarger M, Wald D. The effects of earthquake
tradeoff in optimizing regional earthquake mitigation in- measurement concepts and magnitude anchoring on individu-
vestment. Structure and Infrastructure Engineering, 2007; als’ perceptions of earthquake risk. Earthquake Spectra, 2005;
3(2):133–146. 21(4):978–1008.
52. Tamura H, Yamamoto K, Tomiyama S, Hatono I. Modeling 57. May PJ. Making choices about earthquake performance. Nat-
and analysis of decision-making problem for mitigating natu- ural Hazard Review, 2004; 5(2):64–70.
ral disaster risks. European Journal of Operational Research, 58. Caterino N, Iervolino L, Manfredi G, and Cosenza E. Multi-
2000; 122(2):461–468. criteria decision-making for seismic retrofitting of RC struc-
53. Cowen T. Using cost-benefit analysis to review regula- tures. Journal of Earthquake Engineering, 2008; 12(4):555–
tion. New Zealand Business Roundtable Discussion Paper. 583.