Stoch Envir Res and Risk Ass (2004) 18: 324–338
DOI 10.1007/s00477-004-0187-3
ORIGINAL PAPER
E. Kentel
M. M. Aral
Probabilistic-fuzzy health risk modeling
Abstract Health risk analysis of multi-pathway expo-
sure to contaminated water involves the use of mecha-
nistic models that include many uncertain and highly
variable parameters. Currently, the uncertainties in these
models are treated using statistical approaches. How-
ever, not all uncertainties in data or model parameters
are due to randomness. Other sources of imprecision
that may lead to uncertainty include scarce or incom-
plete data, measurement error, data obtained from ex-
pert judgment, or subjective interpretation of available
information. These kinds of uncertainties and also the
non-random uncertainty cannot be treated solely by
statistical methods. In this paper we propose the use of
fuzzy set theory together with probability theory to
incorporate uncertainties into the health risk analysis.
We identify this approach as probabilistic-fuzzy risk
assessment (PFRA).
Based on the form of available information, fuzzy set
theory, probability theory, or a combination of both can
be used to incorporate parameter uncertainty and vari-
ability into mechanistic risk assessment models. In this
study, tap water concentration is used as the source of
contamination in the human exposure model. Ingestion,
inhalation and dermal contact are considered as multiple
exposure pathways. The tap water concentration of the
contaminant and cancer potency factors for ingestion,
inhalation and dermal contact are treated as fuzzy
variables while the remaining model parameters are
treated using probability density functions. Combined
utilization of fuzzy and random variables produces
membership functions of risk to individuals at different
fractiles of risk as well as probability distributions of risk
for various alpha-cut levels of the membership function.
E. Kentel Æ M. M. Aral (&)
Multimedia Environmental Simulations Laboratory,
School of Civil and Environmental Engineering,
Georgia Institute of Technology, Atlanta,
Georgia, USA
e-mail: mustafa.aral@ce.gatech.edu
Tel.: +/) 404–8942243
Fax: +/) 404–894-5111
Ó Springer-Verlag 2004
The proposed method provides a robust approach in
evaluating human health risk to exposure when there is
both uncertainty and variability in model parameters.
PFRA allows utilization of certain types of information
which have not been used directly in existing risk
assessment methods.
Keywords Human health risk
Probability theory

Fuzzy set theory
Exposure
Uncertainty
Variability
1 Introduction
Risk assessment has been used to quantify the human
health impacts due to exposure to contaminated water
via multiple exposure routes such as ingestion (drink-
ing), inhalation (breathing volatilized contaminants
during showering) and dermal contact (contact of con-
taminated water with skin, for example while shower-
ing). The goal of risk assessment is to estimate the
severity and likelihood of harm to human health from
exposure to a substance or activity that under plausible
circumstances can cause harm to human health. Quan-
titative risk characterization involves evaluating expo-
sure estimates against a benchmark of toxicity, such as a
cancer slope factor. Risk is calculated by multiplying the
cancer potency factor (slope factor) of the toxic sub-
stance by the dose an individual receives.
Uncertainties in risk estimates may arise from many
different sources such as measurement or estimation of
parameters, natural variability in individual’s response,
variability in environmental concentration of toxicants
over time and space and unverifiable assumptions in
dose-response models or extrapolations of the results of
these models. For example, variability in individuals is
not only due to response. Air intake or amount of water
consumed per day may vary from person to person.
Variability in environment is not only associated with
concentration of toxicant in the medium, but also vola-
tilization of contaminant from tap water to air as it varies
as a function of temperature. In risk assessment studies,

analysts always acknowledge the existence of these
uncertainties but can only provide a point estimate of
risk that may represent the ‘‘best’’ estimate. In recent
years, the probabilistic risk assessment (PRA) studies
have become popular in analyzing this variability. PRA
is the general term for risk assessment that uses proba-
bility theory and probability models to represent the
likelihood of different risk levels in a population (i.e.,
variability) or to characterize the uncertainty in risk
estimates (Maxwell et. al. 1998; Maxwell and Kastenberg
1999; Ma 2002; Ma 2000; Ma et. al. 2002). Assuming all
uncertainty is due to randomness, PRA uses probability
distributions for one or more variables in a risk equation
in order to quantitatively characterize variability and/or
uncertainty in the outcome. The output of a PRA is a
probability distribution of risk that reflects the combi-
nation of the input probability distributions. Using this
approach, if the input distributions represent variability
in a probabilistic sense, then the output risk distribution
may provide information on the variability in risk in the
population of concern. If the input distributions reflect
uncertainty in a probabilistic sense, then the output risk
distribution may provide information about the uncer-
tainty in the risk estimate (U.S. EPA 2001). However,
only rarely there are sufficient data available to estimate
and characterize accurately the probability distribution
of the input variables. Moreover, if uncertainty is not due
to randomness or if the available information is in the
form of an expert judgment or subjective interpretations
of system parameters or imprecisely defined boundaries
of parameters, than probabilistic analysis may not be
sufficient to represent the true nature of uncertainty. In
such cases, fuzzy set theory can be used to incorporate
uncertainty into the computational models. In this study,
we propose a framework for probabilistic-fuzzy risk
assessment (PFRA), which will enable the analyst to use
other types of information (i.e., other than statistical
data) such as expert knowledge or fuzzy information in
combination with PRA in health risk assessment studies.
The proposed approach integrates utilization of proba-
bility theory and fuzzy arithmetic in treating variability
and uncertainty.
As defined in U.S. EPA (2001), variability is the true
heterogeneity or diversity that characterizes an exposure
variable or response in a population. Further study (e.g.,
increasing sample size) will not reduce variability, but it
can provide greater confidence, that is lower uncertainty,
in quantitative characterization of variability. Variabil-
ity maybe associated with parameters related to indi-
vidual behavior patterns and individual variations in
dose-response sensitivity to risk agents. On the other
hand, uncertainty is somewhat different than variability
and is defined as lack of knowledge about specific vari-
ables, parameters, models, or other factors (U.S. EPA
2001). Examples include limited data regarding the
concentration of a contaminant in an environmental
medium, lack of information on cancer potency factors.
In this study, in order to demonstrate the proposed
approach in a simple manner, we will only consider the
325
cancer potency factors (indicators of toxicological re-
sponse of the individual to a specific risk agent) associ-
ated with different exposure pathways and tap water
toxicant concentration as non-random but uncertain
parameters. We will assume that these uncertain
parameters can be represented as fuzzy variables. The
other parameters of the risk model will be treated as
random variables and the associated probability distri-
butions of these variables, as they are defined in the
literature, will be used. However, we emphasize that the
proposed approach is not limited to these selections. As
it will become apparent, any combination of fuzzy and
random variables can be analyzed using the approach
described here.
2 Methodology
Our approach for calculating risk estimates involve a
multi-pathway exposure, an individual health risk model
and a framework for evaluating the membership func-
tion of health risk as a function of variability and
uncertainty.
2.1 Health risk assessment
Health risk assessment process consists of four main
steps: Source/release assessment, exposure assessment,
dose-response assessment and risk characterization. The
details of each step can be found in Louvar and Louvar
(1998) or Cohrssen and Covello (1989) and will not be
repeated here. In this study we consider the contami-
nated tap water, obtained through some environmental
modeling, as the source of contamination and we will
only discuss exposure assessment, dose-response assess-
ment and risk characterization computations.
Exposure assessment techniques estimate or directly
measure the quantities of concentrations of risk agents
received by individuals, populations, or ecosystems. In
this study, we will carry out the risk assessment analysis
only from the individual risk perspective; so the main
concern is health risk estimates of an individual as a
result of exposure to contaminated tap water. The risk is
defined to be the increased individual probability of
developing cancer during a specified time interval as a
result of such an exposure.
Exposure model proposed by U.S. EPA (2001) for
cancer risk model (assuming that the ingestion is the
only route of intake) is:
C.IR.EF.ED
CDI ¼
BW:AT ð1Þ
Risk ¼ CDI.CSF
where C is the concentration of a chemical in an expo-
sure medium (e.g., mg/L), IR is the ingestion rate (e.g.,
L/day for water, mg/day for soil, etc.), EF is the expo-
sure frequency (days/year), ED is exposure duration
326

(years), BW is the body weight (kg), AT is averaging
time (equal to 70 years  365 days/year), CDI is the
chronic daily intake of the chemical (mg/kg-day) and
CSF is the cancer slope factor (linear low-dose cancer
potency factor) for the chemical (mg/kg-day))1.
In this study we assume that the source is the con-
taminated tap water and routes of exposure are inges-
tion, inhalation and dermal contact. Risk equations for
these specific routes are provided in Appendix.
Exposure analysis is used to estimate the daily intake,
while epidemiological studies are used to determine the
dose-response relationship. Most of the variables that
occur in chronic daily intake formula are uncertain and/
or variable. Thus, using single-value estimates for these
variables may result in significant errors. In PRA ap-
proach a probability distribution is specified for one or
more of these variables. For the majority of PRA
studies, it is expected that probability distributions, ra-
ther than a single value, will be used to characterize
inter-individual variability, which refer to true hetero-
Table 1 Parameters of risk calculations
geneity or diversity in a population (U.S. EPA 2001).
Thus, variability in daily intake, for example, can be
characterized by combining multiple sources of vari-
ability in exposure, such as ingestion rate, inhalation
rate, water use rates, skin surface area, skin permeabil-
ity, body weight, etc. In this study most of the param-
eters occurring in daily intake formulas for ingestion,
inhalation and dermal contact routes are also considered
to be variable parameters and probability density func-
tions are used instead of single value estimates (Table 1).
However, tap water concentration and cancer potency
factors associated with ingestion, inhalation and dermal
contact may not exibit only variability as defined above.
Depending on the form of the available information, it
may be better to interpret them as uncertain parameters.
Based on this assumption, one has to use a method other
than the probabilistic approach when incorporating
them into the mechanistic models.
To estimate the responses of populations exposed to a
given dose of contaminant, risk analysts conduct mathe-

Parameter Symbol Units Type of Uncertainty Variable Type Value/Distribution*

Averaging time AT days none constant 25500 (70 years)

Exposure duration ED years none constant 30
Exposure frequency EF d/year none constant 350
Ingestion rate per IR/BW L/d/(kg body wt) variable random lognormal ~ (3.3e–2, 1.3e–2)
body weight
Water use rate Wi L/hr
Shower Ws L/hr variable random lognormal ~ (480,160)
Bathroom Wb L/hr variable random lognormal ~ (40,15)
House Wh L/hr variable random lognormal ~ (40,15)
Transfer efficiency from TEi –
water to air
Shower TEs – variable random triangular ~ (0.1, 0.5, 0.9)
Bathroom TEb – variable random triangular ~ (0.1, 0.43, 0.9)
House TEh – variable random triangular ~ (0.1, 0.43, 0.9)
Air exchange rate VRi m3/hr
Shower VRs m3/hr variable random uniform ~ (4–20)
Bathroom VRb m3/hr variable random uniform ~ (10–100)
House VRh m3/hr variable random uniform ~ (300–1200)
Exposure time ETi hr/d
Shower ETs hr/d variable random lognormal ~ (0.13, 0.09)
Bathroom ETb hr/d variable random lognormal ~ (0.32, 0.21)
House ETh hr/d variable random uniform ~ (8–20)
Inhalation rate per unit IH/BW m3/d/(kg body wt) variable random lognormal ~ (0.39, 0.5)
body weight
Skin surface area available SA/BW m2/kg variable random lognormal ~ (0.39, 0.5)
per body weight
Chemical-specific dermal PC m/hr variable random uniform ~ (0.004–0.01)
permeability constant
Fraction of skin in contact F – variable random unirorm ~ (0.4–0.9)
with water
Cancer potency factor CPF (kg.day)/mg
Ingestion CPFingestion (kg.day)/mg uncertain fuzzy triangular: (0.08, 0.11, 0.14)
Inhalation CPFinhalation (kg.day)/mg uncertain fuzzy triangular: (0.0014, 0.0018,
0.0022)
Dermal CPFdermal (kg.day)/mg uncertain fuzzy triangular: (0.1, 0.14, 0.18)
Contaminant (PCE) CW mg/L uncertain fuzzy triangular: (0.012, 0.015,
concentration in tap water 0.018)

All parameters other than fuzzy variables are from McKone and
Bogen (1991)
*
For lognormal and normal distributions, the values given in par-
entheses represent the arithmetic mean and standard deviation,
respectively. For uniform distributions the numbers in parentheses
represent the minimum and maximum values in the distribution.
For triangular distributions (probability density functions), the
numbers in parentheses represent the minimum, likeliest and
maximum values, respectively

matical extrapolations. Determining how much of a
hazard the absorbed substance poses is a complicated
problem. Clear-cut relations are rare since epidemiologi-
cal studies are not very sensitive in detecting health effects
from relatively low levels of exposure (Hattis and Ken-
nedy 1986). As explained in detail in Hattis and Kennedy
(1986), the problem is that the rates of specific illnesses
from a given hazard often must be several times above
average before one can conclude that they aren’t simply
random fluctuations. Moreover, in great majority of the
cases where the epidemiological evidence is incomplete or
ambiguous, using animal studies to make projections
make more sense. However, such studies suffer from their
own serious uncertainties. In order to ensure that the toxic
effects will appear at statistically significant levels, the
animals are generally exposed to high concentrations of
chemical. A dose-response curve is fitted to the resulting
data obtained from these experimental animals. The dose-
response curve is used to assess the probability of humans
developing cancer to this chemical, but most of the time,
at much lower levels. Using such high doses on experi-
mental animals can complicate the interpretation of re-
sults in various ways. Thus, the different mathematical
models produce widely varying results. Other serious
difficulty is in the process of interpreting animal studies
and extrapolating them to humans. Because animals and
humans metabolize substances differently, the level of the
test chemical that reaches various parts of the animal and
the human body can vary widely. Hence, humans and
animals may suffer from different health effects. As the
National Research Council, which often advises the fed-
eral government on scientific issues, points out in a con-
gressionally commissioned report that correcting for
these difficulties are not easy because researchers often
lack enough information about human and animal sys-
tems (Risk Assessment in the Federal Government:
Managing the Process 1983). Thus assessing toxicity
values – which defines the quantitative relationship be-
tween dose and response – of risk agents on humans is full
of uncertainties, and the information available is not en-
ough to provide clear-cut values. Based on this observa-
tion, cancer potency factors may be treated more
appropriately as non-random but uncertain variables and
using single value estimates or probabilistic estimates for
this parameter may not be the proper interpretation. In an
application, the analyst must gather all the available
information about cancer potency factors and include this
information – even if it is not complete – into risk
assessment studies in some other manner. We believe that
as an initial approximation, modeling these parameters as
fuzzy variables may reduce the associated interpretation
errors and may lead to more informed decisions.
Contaminant concentration in the tap water is con-
sidered as the toxicant source in this study. Calculation
of tap water concentration requires detailed studies as
well, which has its own uncertainty structure. Contam-
inant concentration within the medium, which is used as
the source of tap water, must be determined through
detailed environmental modeling studies or measure-
327
ments. Depending on the source of tap water (i.e.,
groundwater, surface water or combination of both)
groundwater modeling, surface water modeling or
combined modeling of groundwater and surface water
may be required. It is not our intention here to provide
groundwater or surface water modeling studies to esti-
mate tap water concentration. However, we are aware of
the fact that contaminant concentration in the tap water
is another uncertain parameter and must be treated as
such in its own uncertainty structure. In this study,
contaminant concentration in tap water is also treated as
a fuzzy variable. Measured contaminant concentrations
in tap water or expert knowledge may help the analyst to
assign membership function for this fuzzy variable.
To summarize, our goal here is to provide a frame-
work which enables the risk analyst to make use of all
available information. Avaliable information can be in
the form of expert judgement (e.g., subjective informa-
tion of system boundaries, intuitive information), and
incomplete or ambiguous information (e.g., imprecisely
defined boundaries of parameters, linguistic informa-
tion). We believe that risk assessment studies will benefit
from inclusion of these types of information into health
risk estimates. Fuzzy set theory can be used to incor-
porate such information into the computational models.
Therefore, in this study we will define the cancer potency
factors associated with ingestion, inhalation and dermal
contact and contaminant concentration in tap water in
terms of fuzzy variables. We will treat most of the other
parameters of the mechanistic models described above in
terms of appropriate probability distributions, as is
commonly done in PRA approach.
2.2 Variability and uncertainty
As discussed previously, variable parameters are those
which can be measured but which have different values at
different locations, or which may differ from person to
person. Uncertain variables are the ones for which there
exist measurement errors or incomplete information. In
Table 1 a list of all of the parameters used in risk cal-
culations are shown, along with the definition of the type
of variability/uncertainty associated with them and how
they are treated (random or fuzzy) in this study. The
corresponding distributions used to represent that vari-
ability/uncertainty (probability distribution functions for
random variables and membership functions for fuzzy
variables) are also given in the last column of Table 1.
Characterizing variability in exposure parameters
In PRA a probability distribution is specified for all of
the variable parameters (see Table 1). Monte Carlo
analysis (MCA) is the most widely used probabilistic
method in PRA. MCA is a probabilistic method that
uses computer simulation to combine multiple proba-
bility distributions in a risk equation. The details of this
328
method are explained in U.S. EPA’s Risk Assessment
Guidance for Superfund Report (U.S. EPA, 2001). In
our study we will use MCA together with fuzzy interval
analysis to carry out PFRA analysis.
Characterizing uncertainty in exposure parameters
In this study, all uncertain parameters are treated as
fuzzy variables. Thus, fuzzy arithmetic and interval
analysis will be used to integrate uncertainty due to these
parameters into the PFRA study. Namely, contaminant
concentration in tap water, and cancer potency factors
associated with ingestion, inhalation and dermal contact
are considered as fuzzy variables. The membership
functions of these fuzzy variables are given in Fig. 1. For
ease of calculations triangular membership functions are
used for membership functions of fuzzy variables. The
PFRA method is not restricted to this membership
function choice.
Fuzzy set theory was first introduced by L. A. Zadeh
in 1965 (Zadeh 1965) and have been applied in various
fields, such as decision making, control theory and ex-
pert systems. Basic definitions of fuzzy sets and fuzzy
arithmetic along with some engineering applications can
be found in Zimmerman (1985), Ross (1995) and Ka-
ufmann and Gupta (1985). Objects of fuzzy set theory
are fuzzy sets whose boundaries are not precise. The
 
Risk ¼ CW: f ðV1ing ; . . . ; Vning Þ:CPFing þ gðV1inh ; . . . ; Vminh Þ:CPFinh þ qðV1der ; . . . ; Vkder Þ:CPFder
# # # #
fuzzy random fuzzy
membership in a fuzzy set is not a matter of acceptance
or rejection (i.e., membership function for crisp set is a
binary pair {0, 1}, 1 representing membership and 0
representing non-membership) but rather a matter of
degree. Fuzzy sets have the capability to express gradual
transitions from membership to nonmembership and
vice versa.
The membership concept provides us not only with a
meaningful and powerful representation of measure-
ment uncertainties, but also with a meaningful repre-
sentation of vague concepts expressed in natural
language. Thus, the significance of fuzzy variables is that
they facilitate gradual transition between states and,
consequently possess a natural capability to express and
deal with observation and measurement uncertainties.
Traditional variables, which we may refer to as crisp
variables, do not have this capability. Although the
definition of states by crisp sets is mathematically cor-
rect, it is unrealistic in the face of unavoidable uncer-
tainty for certain applications. A measurement that falls
into a close neighborhood of each precisely defined
border between states of a crisp variable is taken as
evidential support for only one of the states, in spite of
the inevitable uncertainty involved in this decision (Klir
and Yuan 1995).
One of the most important concepts of fuzzy sets is
the concept of an apha-cut, ac. Given a fuzzy set A de-
fined on X and any number a 2 [0, 1], the a-cut, Aa, is
the crisp subset of X consisting of all of the elements of
X for which membership grades in A are greater than or
equal to the specified value of a. The membership con-
cept is shown in Fig. 1a. To summarize, the membership
value of a real number reflects the ‘‘likeliness’’ of
occurrence of that number; the alpha-cut level reflects
the different sets of numbers with a given minimum
likeliness (Kaufmann, 1985).
3 Probabilistic-fuzzy risk assessment analysis
The procedure for PFRA starts with the identification of
fuzzy or random variables in a risk model. The form of
information available to the analyst determines whether
a parameter will be treated as a fuzzy or a random
variable. Thus, the initial step of PFRA is data collec-
tion and evaluation. Once all available information is
gathered and data evaluated, appropriate probability
density functions and membership functions can be
specified for variable and uncertain parameters, respec-
tively.
The risk model we propose is as follows:
# # # ð2Þ
random fuzzy random fuzzy
where f ðV1ing ; . . . ; Vning Þ represents the function involving
random variables occurring in the risk formula associ-
ated with ingestion route, Equation [A1] (Appendix), n
is the number of random variables occurring in Equa-
tion [A1] (i.e. ingestion rate per body weight),
gðV1inh ; . . . ; Vminh Þ represents the function involving ran-
dom variables occurring in the risk formula associated
with inhalation route, Equation [A2], m is the number of
random variables occurring in Equation [A2] and
qðV1der ; . . . ; Vkder Þ represents the function which involves
random variables occurring in the risk formula associ-
ated with dermal contact route, Equation [A5], k is the
number of random variables occurring in Equation [A5].
In this study, as indicated in Table 1, some of the var-
ibles occurring in f, g and q functions are defined as
constants. However, the solution methodolgy will not
alter

if they are chosen as random variables. Here,
Viing ; Vjinh ; Vkder are vectors that contain the sampled
values of these parameters from the probability distri-
butions selected for these parameters, respectively
(Table 1).
As can be seen in Equation [2], the risk equation is
composed of various fuzzy and random variables. In this
 329

Fig. 1a Membership function

of contaminant concentration
in tap water Fig. 1b. Mem-
bership function of CPF
associated with ingestion
Fig. 1c. Membership func-
tion of CPF associated with
inhalation Fig. 1d. Member-
ship function of CPF associ-
ated with dermal contact

study, the PFRA analysis is carried out for two different
scenarios. In the first scenario, it is assumed that only
the contaminant concentration in the tap water is trea-
ted as a fuzzy variable and cancer potency factors
associated with ingestion, inhalation and dermal contact
are chosen constant values. This would represent a
simpler case, where only the environmental modeling
outcome is assumed to be uncertain. In the second sce-
nario, cancer potency factors associated with all three
routes together with the contaminant concentration in
330
Fig. 1 (Contd)
the tap water are treated as uncertain variables thus
represented in terms of fuzzy membership functions. For
both cases, most of the terms in functions f, g, q are
modeled as random variables while some are defined as
constants.
3.1 Scenario one
In this case only the contaminant concentration in the
tap water is treated as a fuzzy variable which has a tri-
angular membership function, Figure 1a. The total risk
equation for this case can be given as:
 331

 
Risk ¼ CW : f ðV1ing ; . . . ; Vning Þ:CPFing þ gðV1inh ; . . . ; Vminh Þ:CPFinh þ qðV1der ; . . . ; Vkder Þ:CPFder
# # # # # # # ð3Þ
fuzzy random constant random constant random constant
  h i
Monte Carlo Analysis is used together with interval to as Riskalower
c
cdf
and Riskaupper
c
in equation [6]
cdf
analysis to generate cumulative distribution functions of respectively. Thus, the multiplication operation used in
risk for upper and lower limits of each alpha-cut level. Equation [6] represents multiplication of a cdf with a
PRFA study is conducted with 5000 and 10000 Monte constant. The collection of all of these risk cdfs (i.e., 21
Carlo simulations. Since the results obtained are very cdfs corresponding to lower and upper limits of each ac)
similar, it is decided to proceed with 5000 Monte Carlo will be referred to as ½Riskcdf . The cdfs corresponding to
simulations for the rest of the study. The procedure to lower and upper limits of all alpha-cut levels are gen-
conduct PFRA is given below. erated. However, to provide a clear and simple presen-
The term inside the square parenthesis in Equation [3] tation, cdfs for the lower and upper limits of only 0.0,
(this term will be referred as P from now on) involves 0.5 and 1.0 alpha-cut levels are given in Fig. 2.
only random variables and constants. 5000 Monte Carlo Using these cumulative distribution functions of risk,
simulations are conducted to generate sets of random ½Riskcdf , the membership function of the total risk to a
variables by sampling from the probability distribution specific fractile of individuals at risk can now be gener-
functions of the associated variables provided in Table 1. ated. This concept is explained in Fig. 2. A horizontal
P ¼ f ðV1ing ; . . . ; Vning Þ:CPFing þ gðV1inh ; . . . ; Vminh Þ:CPFinh line cutting through ½Riskcdf curves are drawn and risk
values corresponding to upper and lower limits of each
þ qðV1der ; . . . ; Vkder Þ:CPFder ð4Þ ac are read. For example, a and e are risk values having
The cumulative distribution function of P, ½P cdf can be membership values of zero for 50th fractile. Similarly, all
generated. of the other risk values and associated membership
The next step, is to include the uncertainty due to values are read from the graph and used to plot the fuzzy
fuzziness associated with the contaminant concentration risk membership function corresponding to 50th fractile.
in the tap water, CW. Figure 3 depicts the membership function of total risk to
The total risk equation, Equation [3], being a individuals at 30th, 60th and 90th fractiles of risk.
monotonic function will allow us to use interval analysis In PFRA, all variable parameters are modeled as
to carry out fuzzy calculations. Interval analysis involves random variables (in this specific case study, some are
discretizing the membership domain of the fuzzy vari- taken as constants but the proposed method allows all of
able into a specified number of alpha-cut, ac, levels. these parameters to be modeled as random variables).
First, an alpha-cut interval, in this case 0.1, is selected Here the assumption is that none of the variables are
(i.e., 0.0:0.1:1.0). The lower and upper bounds of the correlated with one another and an independent distri-
intervals for each ac for fuzzy variable CW (this concept bution over all individuals at risk is determined for each
is depicted in Figure 1a) can be given as: of these variable parameters; which we believe is an
acceptable assumption. A set of random variables is
CWac ¼ ½CWalower
c
; CWaupper
c
 8 ac ¼ 0 : 0:1 : 1:0 ð5Þ sampled from these distributions to construct an indi-
vidual. This individual is not associated with any par-
Fuzzy arithmetic is carried out for each ac level in ticular individual but is representative of a particular
accordance with interval analysis. This procedure is risk fractile in the population. Detailed explanation of
summarized below and detailed explanations of interval the particular risk fractile concept can be found in
analysis can be found in Moore (1979). Maxwell et. al. (1998).
The lower and upper values of CW (i.e.,
CWalower
c
and CWaupper
c
) are used together with the previ-
ously calculated cumulative distribution function (cdf), 3.2 Scenario Two
½P cdf , to calculate two cumulative distribution functions
of risk corresponding to each ac level (i.e., a total of 21 In this case all the variable parameters are modeled as
cumulative distribution functions for risk are calcu- random variables as before and all the uncertain
lated). parameters (contaminat concentration and cancer po-
  tency factors) are modeled as fuzzy variables. The risk
Riskalower
c
¼ ½P cdf :CWalower
c
equation associated with this case is given in Equation
h icdf 8 ac ¼ 0 : 0:1 : 1:0 ð6Þ
ac ac [2]. A similar procedure to the one used in Scenario 1 is
Riskupper ¼ ½P cdf :CWupper
cdf followed for this case. Lower and upper limits of each
Multiplying ½P cdf with a constant, either fuzzy variable for each ac level are determined. Again
CWalower
c
or CWaupper
c
, results in shifting the cumulative 5000 Monte Carlo simulations are conducted to generate
distribution function CWalowerc
or CWaupper
c
units, and random variables by sampling from the probability
these new cumulative distribution function are referred distribution functions of associated variables. Since the
332

Fig. 2 The cdfs of risk for 0.0,

0.5 and 1.0 ac levels for Sce-
nario 1

Fig. 3 The membership

function of risk to individuals
at 30th, 60th and 90th
fractiles of risk, Scenario 1
 333

functions associated with ingestion, f ðV1ing ; . . . ; Vning Þ, factors associated with ingestion, inhalation and dermal
inhalation, gðV1inh ; . . . ; Vminh Þ and dermal contact, contact are provided in McKone and Bogen (1991) as
qðV1der ; . . . ; Vkder Þ, are functions of random variables and constants. We used these values as the peak values (i.e.,
constants, 5000 different values for these three functions value which has a membership value of one), then we
are generated: assigned a triangular distributions for these fuzzy vari-

 ables according to our interpretation of uncertainty in
½f  ¼ f Viing i ¼ 1; 2; 3; . . . ; n these parameters. Since our goal here is to present the


½g ¼ g Vjinh j ¼ 1; 2; 3; . . . ; m ð7Þ framework of PFRA analysis, the particular selection of
½q ¼ qðVlder Þ l ¼ 1; 2; 3; . . . ; k the support base of an uncertain variable is not impor-
tant and can be changed as one interprets the data.
As stated before the total risk equation, Equation [2], is Again, since we are working with a hypothetical case, we
a monotonic function and this will allow us to use again assigned an arbitrary triangular membership
interval analysis to carry out fuzzy calculations. The function for the contaminant concentration in the tap
lower and upper values of contaminant concentration, water. Any kind of membership function is possible for
CWalower
c
and CWaupper
c
, and

cancer
ac potency
factorsasso- fuzzy variables but for the sake of simplicity we pre-
ac
ciated with ingestion, CPFing lower and CPFing upper , ferred to work with triangular distributions here. While
ac ac
inhalation, ðCPFinh Þlower and ðCPFinh Þupper , and dermal conducting toxicant and site specific PFRA studies the
contact, ðCPFder Þalower
c
and ðCPFder Þaupper
c
, corresponding analyst should collect all the available information
to each ac are used together with the previously calcu- about these uncertain parameters and assign the most
lated ½f , ½g and ½q results to calculate two risk values appropriate membership functions to them as they
for each of the 5000 Monte Carlo simulation results (i.e., interpret the uncertainty.
one for lower and one for upper limit of each alpha-cut The proposed risk model is provided in Equation [2].
of the membership functions of contaminant concen- First, we conducted simulations for a simple case, Sce-
tration and cancer potency factors associated with nario 1, in which only contaminant concentration in the
ingestion, inhalation and dermal contact): tap water is treated as a fuzzy variable, Equation [3]. In

  
ac 
Riskalower
c
¼ CWalower
c
: ½f : CPFing lower þ½g:ðCPFinh Þalower
c
þ½q:ðCPFder Þalower
c

h i 
ac  8ac ¼ 0 : 0:1 : 1:0 ð8Þ

Riskaupper
c
¼ CWaupper
c
: ½f : CPFing upper þ½g:ðCPFinh Þaupper
c
þ½q:ðCPFder Þaupper
c

Then acfor upperh andac lower

i limits of each ac level,
Risklower and Riskupper are used to generate cumu-
hlative distribution
i functions of risk, Riskalower
c
cdf
and
Riskaupper
c
, respectively. Thus, a total of 21 cumula-
cdf
tive distribution functions for risk are calculated. The
cumulative distribution functions corresponding to
lower and upper limits of 0.0, 0.5 and 1.0 ac levels are
given in Fig. 4.
The cumulative distribution functions corresponding
to lower and upper limits of each alpha-cut level are
used to generate membership functions of total risk to
individuals at 30th, 60th and 90th fractiles of risk are
calculated similar to process described earlier and they
are given in Fig. 5. In Fig. 5, results of Scenario 1 are
also shown for comparison purposes.
4 Results
4.1 Interpretation of results with fuzzy set theory
In this study we used the probability density functions
provided by McKone and Bogen (1991) for the variable
parameters. They are summarized in Table 1. For fuzzy
variables we used triangular membership functions
which are given in Fig. 1. The values for cancer potency
this case, cancer potency factors associated with inges-
tion, inhalation and dermal contact are chosen as con-
stants. Variable parameters (see Table 1) are treated as
random variables. We should note here that some of the
variable parameters are also chosen as constants (i.e.,
exposure duration, exposure frequency) but the proce-
dure will be the same if they are treated as random
variables as well. Probabilistic-fuzzy risk assessment
study is conducted as explained in section 3.1 to generate
cumulative density functions for lower and upper limits
of each ac. Then, these cumulative density functions are
used to form membership functions of risk to individuals
at certain fractiles of risk.
Scenario 1 uses contaminant concentration in tap
water as a fuzzy variable so the resultant risk is also a
fuzzy variable. The membership functions of risk to
individuals at 30th, 60th and 90th fractiles of risk are
given in Fig. 3. As can be seen from Fig. 3, the mem-
bership functions of risk to individuals at different
fractiles of risk have triangular distributions. Triangular
membership functions of risk can be interpreted as risk
to individuals at a certain fractile of risk being around
the peak value (i.e., value corresponding to a member-
ship function of 1.0). A membership value of one reflects
the most likely value for the variable. If the fuzzy set
corresponding to the risk to the 90th most highly ex-
posed person is called Rf f
90 , then R90 can be defined as
334

Fig. 4 Cdfs of risk for 0.0, 0.5

and 1.0 ac levels for
Scenario 2

Fig. 5 Comparison of mem-

bership functions of risk to
individuals at certain fractiles
of risk for Scenario 1 and
Scenario 2

‘‘around 1.5  10)4’’ (see Figure 3). The support of the
membership function provides the possible ranges of
the risk for the individuals at the corresponding fractile.
The risk value corresponding to a membership value of
1.0 is the most likely risk to occur for the associated
fractile. The possible range of risk to individuals at 90th
fractile is 4.05  10)5 to 2.43  10)4 and the most likely
outcome is 1.5  10)4(see Fig. 3). Risk values outside
4.05  10)5 – 2.43  10)4 range have zero membership
values (i.e., no membership) and 1.5  10)4 has a
membership value of one (i.e., full membership). Since
the membership values indicate the degree to which a
risk value is a member of the fuzzy set Rf 90 , values outside
4.05  10)5 – 2.43  10)4 range are not members of Rf 90
and 1.5  10)4 is a full member of Rf 90 . As can be seen
from Figure 3, the uncertainty (i.e., support of the
membership function being wide implies that the
uncertainty is larger) in added risk to individuals at 90th
fractile is larger than that of 30th fractile. This is due to
the shapes of the cumulative distribution functions used
to generate membership functions. As can be seen from
Fig. 2, cumulative distribution functions have higher
slopes up to around 0.8 and then they level off. Thus, the
rate of change of risk for lower fractiles is smaller than
those of higher fractiles.
The triangular membership function obtained for risk
for a certain fractile (Fig. 2) is symmetric for Scenario 1.
This is due to the fact that only contaminant concen-
tration is modeled as a fuzzy variable, whose member-
ship function is taken as a symmetric triangular
distribution. The rest of the risk equation (i.e., P term
defined in section 3.1) results in a single number for each
of the 5000 Monte Carlo simulations and then these
results are used to form ½P cdf . For each fractile one risk
value can be obtained from this cdf. Thus, the fuzzy
arithmetic carried out to generate the membership
function of risk for a certain fractile involves multiplying
the lower and upper bounds of a fuzzy number at a
certain ac level with a constant (i.e., the risk value for
that certain fractile). This yields a risk membership
function similar to the one used for contaminant con-
centration in tap water.
In the second scenario, PFRA is applied to generate
membership function of risk to individuals at certain
fractiles of risk using Equation [2]. In this scenario,
contaminant concentration in tap water and cancer po-
tency factors associated with ingestion, inhalation and
dermal contact are treated as fuzzy variables while most
of the variable parameters are treated as random vari-
ables (Table 1). The corresponding cumulative distri-
bution curves for Scenario 2 are given in Fig. 4 and
comparison of membership functions of risk for certain
fractiles for Scenario 1 and Scenario 2 is given in Fig. 5.
As can be seen from Fig. 5, when cancer potency factors
associated with ingestion, inhalation and dermal contact
are also treated as fuzzy variables, the uncertainty in risk
increases (i.e., compare support base of triangular dis-
tributions corresponding to the same fractile for two
different scenarios). The support base of risk for indi-
335
viduals at 90th fractile of risk for only contaminant
concentration fuzzy case is 4.05  10)5 – 2.43  10)4
while for contaminant concentration and all cancer po-
tency factors fuzzy case it is 2.94  10)5 – 2.93  10)4.
For the later case the uncertainty in the resulting risk is
higher due to the fact that uncertainty in more param-
eters (i.e., cancer potency factors) are included into the
study. Another observation from Fig. 5 is that the peak
values corresponding to same fractile of risk for both
scenarios are almost the same. Ideally, they should be
the same, however 5000 Monte Carlo simulations are
used in the calculation of these peak values. For two
different scenarios two different sets of 5000 Monte
Carlo simulations are generated and they produce
slightly different values for most likely added risk values.
The shapes of the membership functions of risk for a
certain fractile for Scenario 2 are not similar to the
shapes of input fuzzy variables (i.e., contaminant con-
centration in tap water and cancer potency factors
associated with ingestion, inhalation and dermal con-
tact) any more. This is due to the fact that, in Scenario 2,
several of the variables are defined as fuzzy variables and
the risk equation involves multiplication and addition
operations of these fuzzy variables. Thus, the shape of
the membership function for risk at a certain fractile is
not symmetric but skewed for this case. The selection of
membership functions for the fuzzy variables and also
the probability distribution functions of the random
variables will impact the shape of the membership
function of risk obtained for a certain fractile. The shape
of the membership function may also represent valuable
information to the analyst. There are several methods to
defuzzify fuzzy variables (i.e., convert a fuzzy variable
into a point estimate). One of the most popular methods
is the center of gravity method (Ross 1995). For exam-
ple, if the shape of the membership function of risk is
skewed to the left or right of the peak and if the fuzzy
risk is deffuzified using the center of gravity method, a
crisp risk value which will be higher or lower depending
on the orientation of the skewness of the triangle. Thus,
the deffuzified value of risk will be either smaller (or
larger) than the most likely risk value of the fuzzy risk
(i.e., value corresponding to a membership value of one).
This would indicate that a lower (or higher) risk value is
possible when compared to the risk corresponding to the
membersip value of one, which also corresponds to the
risk one would obtain from PRA approach for that
fractile.
Membership functions can also be used to interpret
the results from possibility point of view. Possibility
theory is briefly explained in the following section.
4.2 Interpretation of results with possibility theory
Possibility theory
Possibility theory is a measure-theoretic counterpart of
fuzzy set theory based upon the standard fuzzy opera-
336
tions. It provides us with appropriate tools for pro-
cessing incomplete information expressed in terms of
fuzzy propositions; consequently, it plays a major role in
fuzzy logic and approximate reasoning (Klir and Yuan
1995).
The theory of possibility is related to the theory of
fuzzy sets by defining the concept of a possibility dis-
tribution as a fuzzy restriction which acts as an elastic
constraint on the values that may be assigned to a var-
iable. More specifically (Zadeh, 1978):
Let F be a fuzzy subset of a universe of discourse U
which is characterized by its membership function lF ,
with the grade of membership lF ðuÞ, interpreted as the
compatibility of u with the concept labeled F.
Let X be a variable taking values in U, and let F act as
a fuzzy restriction, R(X), associated with X. Then the
proposition ‘‘X is F’’ which translates into
RðXÞ ¼ F ð9Þ
which associates a possibility distribution, PX , with X
which is postulated to be equal to R(X):
PX ¼ R(X) ð10Þ
Correspondingly, the possibility distribution function
associated with X (or the possibility distribution func-
tion of PX ) is denoted by pX and is defined to be
numerically equal to the membership function of F:
4
pX ¼ lF ð11Þ
Thus, pX ðuÞ, the possibility that X = u, is postulated
to be equal to lF ðuÞ. In this way, X become a fuzzy
variable which is associated with the possibility distri-
bution, PX in much the same way as a random variable
is associated with a probability distribution.
Interpretation
In our example, in the terminology of possibility defined
above, for individuals at 90th fractile risk for Scenario 2,
F is the fuzzy subset defined as ‘‘around 1.4  10)4’’
(Fig. 5) and the variable X is the risk for individuals at
90th fractile risk. Then the proposition ‘‘risk for indi-
viduals at 90th fractile is around 1.4  10)4’’ translates
into
Rðrisk for individuals at 90th fractile of riskÞ
¼ ‘‘around 1:4  104 ’’ ð12Þ
which associates a possibility distribution, PRisk , with
concentration which is postulated to be equal to RðRiskÞ,
i.e., PRisk ¼ RðRiskÞ. Thus for our example, the possi-
bility distribution associated with risk is denoted by pRisk
and is equal to the membership function of F, Fig. 5.
From Fig. 5, we can conclude that for Scenario 2, for
individuals at 90th fractile risk, the possibility that risk is
7.7  10)5 or 2.09  10)4, given that ‘‘risk to individu-
als at 90th fractile risk is around 1.4  10)4’’ is 0.5.
Another significant result that can be derived using
possibility theory is that, the possibility that risk will be
smaller than 2.94  10)5 and greater than 2.93  10)4
(this is the support of the membership function corre-
sponding to Scenario 2, for the individual at 90th frac-
tile, see Fig. 5) is zero, thus the probability is also zero.
Similarly, we can assign possibility values for all the risk
values covered by the domain of fuzzy restriction
‘‘around 1.4  10)4’’. The results obtained from possi-
biliy theory application may be used in approximate
reasoning studies. This topic is not discussed in this
paper since it is a completely different subject matter.
5 Conclusion
In health risk assessment studies, it is very important to
include all available information into the mathematical
models. Tradiationally, the available information is
interpreted in a probabilistic sense and probability the-
ory has been used to integrate this information into
mathematical models. However, we believe that the
form of the information must guide the analyst in
deciding which mathematical tool to use. The proba-
bility theory or possibility theory or a combination of
the two maybe more appropriate for a specific study.
Probability theory is a very strong and well established
mathematical tool to treat variability. It has certain in-
put requirements and whenever these requirements are
met, probability theory will provide powerful results.
However, it is clear that not all uncertainties in data or
model parameters are random, and can be treated by
statistical approaches alone. Another source of impre-
cision that may lead to uncertainty is scarce or incom-
plete data, measurement error or data obtained from
expert judgment or subjective interpretation of available
information. These kinds of uncertainties and also the
non-random uncertainty cannot be treated solely by
statistical methods. Thus, usefullness and applicability
of other mathematical tools, such as fuzzy set theory or
possibility theory should be explored. We believe that
depending on the form of the available information,
health scientists may also make use of these methods as
demonstrated in this study.
While conducting health risk assessment studies the
first step is to collect all information that is available.
Depending the type of collected information, variables
of the risk equation should be modeled as crips, random
or fuzzy variables. The proposed probabilistic-fuzzy
approach allows the analysist to include uncertainties in
model parameters in the desired fashion (i.e., allows
combining crip, random and fuzzy variables in the risk
equation).
In this study, we proposed an approach to include
two different kinds of uncertainties into risk assessment
models using probability theory and fuzzy set theory in
combination. Treating cancer potency factors associated
with ingestion, inhalation and dermal contact together
with the contaminant concentration in tap water as
fuzzy variables allowed us to include uncertainties due to

reasons other than randomness into the risk assessment
model. Fuzzy set calculations resulted in membership
values of risk for individuals at a certain fractile of risk.
Instead of a single probability distribution of risk as
provided by PRA, PFRA approach provides the prob-
ability distributions of risk for various ac levels. If the
appropriate information is available, utilization of
PFRA may provide result which may help the decision
maker to make more informed decisions. We, believe
that the proposed risk assessment model provides a
better approach than the existing ones.
Representation of uncertainty using fuzzy or ran-
dom variables will impact the form of the uncertainty
in the calculated risk. The membership function of risk
for a certain fractile may provide significant informa-
tion for the analyst. For example, the possibility of
occurrence of risk values having zero membership
values for a specific fractile are zero, while the risk
value with a membership value of one is the most likely
risk. The shape and the support base of the risk pro-
vides extra information about the resulting uncertainty
which is a combined effect of the random and fuzzy
input variables. For example, the uncertainty associ-
ated with a risk that has a small support base is
respectively smaller than that of a risk which has a
larger support base. For simplicity purposes triangular
membership functions are used in this study; however,
membership functions for fuzzy variables do not need
to be triangular. Since the membership functions of the
input parameters are choosen as triangular distribu-
tions the resulting fuzzy risks have triangular distribu-
tions. If other membership functions are used for the
input variables, the procedure to conduct PFRA will
not change but the shape of resulting fuzzy risk will
change. The shape of the membership function depends
on the available information about the fuzzy variable.
Application of the proposed PFRA using different
membership values for the input variables and sensi-
tivity of the resulting fuzzy risks might be interesting
topics for further studies.
Appendix
In this study we assume that the source is the contami-
nated tap water and routes of exposure are ingestion,
inhalation and dermal contact. Risk equations for these
specific routes are given as follows (U.S. EPA, 1989):
Ingestion of chemicals in drinking water:
CW.IR.EF.ED
Riskingestion ¼ ðCPFing Þ ðA1Þ
BW.AT
where CW is chemical concentration in tap water (mg/
L), IR is the ingestion rate (L/day), EF is the exposure
frequency (days/year), ED is exposure duration (years),
BW is the body weight (kg), AT is averaging time (equal
to 70 years  365 days/year), and CPFing is the cancer
potency factor associated with ingestion (mg/kg-day))1.
337
Inhalation of airborne chemicals:
CA.IH.ET.EF.ED
Riskinhalation ¼ ðCPFinh Þ ðA2Þ
BW.AT
where CA is chemical concentration in air (mg/m3), IH
is the inhalation rate (m3/hr), ET is the exposure time
(hours/day) and CPFinh is the cancer potency factor
associated with inhalation (mg/kg-day))1.
A modified risk equation due to indoor inhalation of
vaporized chemicals is given by McKone and Bogen
(1991). The concentration of vaporized chemicals is
estimated assuming a linear transfer function between
tap water and air in various locations in the home. The
contaminant vapor concentration in indoor air can be
estimated from:
Wi :TEi
ACi ¼ CW ; i ¼ s; b; h ðA3Þ
VRi
where ACi is the contaminant concentration in indoor
air (mg/m3), Wi is the tap water use rate (1/hr), TEi is the
transfer efficiency from tap water to air (unitless), VRi is
the air exchange rate (mg/m3), and i takes the subscripts
s, b, and h, for the shower, bathroom and house,
respectively. Thus the modified risk equation due to
inhalation of airborne chemicals is:
Riskinhalation ¼ ½ðACs :ETs Þ þ ðACb :ETb Þ
IH.EF.ED
þ ðACh :ETh Þ: ðCPFinh Þ ðA4Þ
BW.AT
where ETi is the exposure time (hr/d) in shower, bath-
room and house for i = s, b and h respectively.
Dermal contact with chemicals in water (modified
according to McKone and Bogen (1991)):
CW.SA.F.PC.ETs .EF.ED.CF
Riskdermal ¼ ðCPFder Þ
BW.AT
ðA5Þ
where SA is skin surface area available (cm2), F is the
fraction of skin in contact with water, PC is chemical-
specific dermal permeability constant (cm/hr), CF is the
volumetric conversion factor for water (1 lt/1000 cm3)
and CPFder is the cancer potency factor associated with
dermal contact (mg/kg-day))1.
References
Cohrssen JJ, Covello VT (1989) Risk Anaysis: a guide to principles
and methods for analyzing health and environmental risks, The
National Technical Information Service
Hattis D, Kennedy D (1986) Assessing risks from health hazards:
an imperfect science. Technol Rev 89(4): 60–71
Kaufmann A, Gupta MM (1985) Introduction to fuzzy arithmetic:
theory and application. Van Nostrand Reinhold Co., New
York
Klir GJ, Yuan B (1995) Fuzzy sets and fuzzy logic: theory and
application. Prentice-Hall PTR, Upper Saddle River, NJ
Louvar JF, Louvar BD (1998) Health and environmental risk
analysis: fundamentals with applications, Prentice Hall PTR,
Upper Saddle River, NJ
338
Ma H-W (2000) The incorporation of stochasticity in risk anaysis
and management: a case study. Stochastic Environ Res Risk
Assess 14:195–206
Ma H-W (2002) Stochastic multimedia risk assessment for a site
with contaminated groundwater. Stochastic Environ Res Risk
Assess 16:464–478
Ma H-W, Wu K-Y, Ton C-D (2002) Setting information priorities
for remediation decisions at a contaminated-groundwater site.
Chemosphere 46:75–81
Maxwell RM, Kastenberg WE (1999) Stochastic environmental
risk analysis: an integrated methodology for predicting cancer
risk from contaminated groundwater. Stochastic Environ Res
Risk Assess 13:27–47
Maxwell RM, Pelmulder SD, Tompson AFB, Kastenberg WE
(1998) On the development of a new methodology for
groundwater-driven health risk assessment. Water Resources
Res 34(4):833–847
McKone TE and Bogen KT (1991) Predicting the uncertainties in
risk assessment. Environ Sci Technol 25(10):1674–1681
Moore RE (1974) Methods and applications of interval analysis.
Siam, Philadelphia
Risk Assessment in the Federal Government: Managing the Pro-
cess (1983) National Academy Press, Washington DC
Ross TJ (1995) Fuzzy logic with engineering applications.
McGraw-Hill, New York
US EPA (1989) Risk assessment guidance for superfund, vol. I.
Human Health Evaluation Manual (Part A), EPA/540/1–89/
002, Office of emergency and remedial response, Washington,
DC
US EPA (2001) Risk assessment guidance for superfund (RAGS),
Vol. III – Part A, Process for conducting probabilistic risk
assessment, EPA 540-R-02–002, Office of emergency and
remedial response, Washington, DC
Zadeh LA (1965) Fuzzy Sets. Inf Control 8:338–353
Zadeh LA (1978) Fuzzy sets as a basis for a theory of possibility.
Fuzzy Sets Sys 1(1):3–28
Zimmerman HJ (1985) Fuzzy set theory and its applications.
International Series in Management Science/Operations Re-
search, Kluwer.Nijhoff Publishing, Boston