Application of Bayesian network to the probabilistic risk assessment

of nuclear waste disposal
Chang-Ju Lee
a,
*
, Kun Jai Lee
b,1
a
Korea Institute of Nuclear Safety, P.O. Box 114, Yuseong, Daejeon 305-600, South Korea
b
Korea Advanced Institute of Science and Technology, 373-1 Kuseong, Yuseong, Daejeon 305-701, South Korea
Received 24 April 2004; accepted 16 March 2005
Available online 25 May 2005
Abstract
The scenario in a risk analysis can be defined as the propagating feature of specific initiating event which can go to a wide range of
undesirable consequences. If we take various scenarios into consideration, the risk analysis becomes more complex than do without them.
A lot of risk analyses have been performed to actually estimate a risk profile under both uncertain future states of hazard sources and
undesirable scenarios. Unfortunately, in case of considering specific systems such as a radioactive waste disposal facility, since the behaviour
of future scenarios is hardly predicted without special reasoning process, we cannot estimate their risk only with a traditional risk analysis
methodology. Moreover, we believe that the sources of uncertainty at future states can be reduced pertinently by setting up dependency
relationships interrelating geological, hydrological, and ecological aspects of the site with all the scenarios. It is then required current
methodology of uncertainty analysis of the waste disposal facility be revisited under this belief.
In order to consider the effects predicting from an evolution of environmental conditions of waste disposal facilities, this paper proposes a
quantitative assessment framework integrating the inference process of Bayesian network to the traditional probabilistic risk analysis. We
developed and verified an approximate probabilistic inference program for the specific Bayesian network using a bounded-variance
likelihood weighting algorithm. Ultimately, specific models, including a model for uncertainty propagation of relevant parameters were
developed with a comparison of variable-specific effects due to the occurrence of diverse altered evolution scenarios (AESs). After providing
supporting information to get a variety of quantitative expectations about the dependency relationship between domain variables and AESs,
we could connect the results of probabilistic inference from the Bayesian network with the consequence evaluation model addressed. We got
a number of practical results to improve current knowledge base for the prioritization of future risk-dominant variables in an actual site.
q 2005 Published by Elsevier Ltd.
Keywords: Scenario; Probabilistic risk assessment; Bayesian network; Probabilistic inference; Causal dependency; Likelihood weighting algorithm;
Uncertainty propagation; Waste disposal
1. Introduction
The notion of ‘risk’ represents a possibility of potential
harm against human beings. Over past two decades, we have
provided this notion in many different ways to assess any
hazards from daily activities. Kaplan and Garrick provided a
conceptual guidance on the risk analyses by introducing
general notions of risk [1,2]. According to their definition,
the concept of ‘sets of triplets’, i.e. the scenario, the
likelihood, and the consequence are emerged. Actually, it is
well known that a formal decomposition of risk can be given
by three queries—what can happen; how likely things are to
happen; and what are the end point’s measures from sets of
occurrences [2]. This has become a standard approach for
predicting the risk.
A nuclear risk generally relates with regular emissions of
radioactive nuclides or possible accidental releases of
radioactivity [3]. In a nuclear aspect, the risk has been
defined as an exceeding expectation of the magnitude of
undesirable radioactive releases, e.g. the product of
probability and consequence of an accident [4]. The end
point’s measures of the nuclear risk prediction may be
expressed with some measures such as core damage
Reliability Engineering and System Safety 91 (2006) 515–532
www.elsevier.com/locate/ress
0951-8320/$ - see front matter q 2005 Published by Elsevier Ltd.
doi:10.1016/j.ress.2005.03.011
* Corresponding author. Tel.: C82 42 868 0149; fax: C82 42 861 1700.
E-mail addresses: cjlee@kins.re.kr (C.-J. Lee), kunjailee@kaist.ac.kr
(K.J. Lee).
1
Tel.: C82 42 869 3818; fax: C82 42 869 3810.
frequency, exposure dose of humans, or the number of
casualty. They can be used to interpret the various
contributors to the risk with which are compared and
ranked [5]. We can identify, these measures with limiting
(or regulatory) indicators which affect about humans or
environment.
The results of a deterministic approach are usually
judged in view of specific measures known as performance
criteria. The performance criteria used to facilitate com-
parison with specific conservative regulatory criteria.
Unfortunately, we are not sure that the performance criteria
produce best-estimate values under incomplete analysts’
knowledge bases. On the other hand, a probabilistic
approach in quantifying the risk, generally known as a
probabilistic risk assessment (PRA), has been used to
provide a probability distribution of the end point’s
measures in various fields such as nuclear reactor operation,
environmental forecasting, space accident prediction, etc.
[6–8]. The subjective probability distribution represents an
uncertainty against the impact of such a lack of knowledge
or stochastic processes. The uncertainty must be practically
treated in the estimation of the end point’s measures. The
depth of uncertainty depends upon our total state of
knowledge; upon all the evidence, data, and experience
with similar courses of actions in the past [1]. Therefore, a
number of factors affect a variation of the risk results.
Especially, long periods of evolution in natural environment
facing diverse stochastic processes have influence on the
uncertainty [9]. Evolution examples coming from stochastic
processes in the environment involve potential changes of
weather, hydraulic profile, geological form, and ecological
condition, and so on. Since the prediction for such an
evolution is extremely difficult, a knowledge base might not
be easily provided. Due to both lack of the knowledge and
stochastic features, the potential risks to future generations
may be subjected to complex uncertainty propagation
assessment via PRA techniques.
Diverse approaches for categorizing uncertainty have set
forth last two decades. Most risk analysts [10–12] agree that
all the variables used in the risk analysis models contain
both epistemic and aleatory uncertainty because of their
unique nature. In estimating, the variables in a model
domain (called here as ‘domain variables’), therefore, we
had to consider a system-specific feature originating from a
very broad range of aleatory and epistemic uncertainties,
besides an issue of numerical solution error due to the
mathematical modelling approximations and temporal
discretization in simulations, as denoted by Oberkampf
[13]. The issue of numerical solution error has been
generally ignored in risk analysis, which is also beyond
our concerns. Recently, some remarkable studies have been
conducted, which focussed on uncertainty propagation with
emphasis on the separation principle for both uncertainty
types in every variables. This approach has been widely
introduced by Hoffman [14], Hofer [15], Siu [16], and
Helton [17] etc., and known as so-called ‘divide and
conquer’ concept.
A lot of risk analyses have been performed to actually
estimate a risk profile from both uncertain future states of
hazard sources and undesirable scenarios. These risk
prediction would require a systematic process for the
scenarios or accident sequences. In addition, a comprehen-
sive risk prediction needs a complete listing of scenarios and
estimation of their frequencies, as well as rankings of
occurrence probability of the scenarios. The scenarios will
be formed by combining both initiating events and
mitigating capabilities. Therefore, the scenario in the risk
analysis can be defined as a propagating feature of specific
initiating event which can go to a wide range of undesirable
consequences [18], which answers to the question, ‘What
can go wrong?’
If we imagine an initial stage of scenario, a hazard source
itself can be a trigger for specific end point’s measures.
Various hazard sources exist in diverse natural or man-made
systems. In terms of treating scenarios, we can classify
states of systems with two—active and passive. First, the
state of system is called ‘active’ if an event can be directly
controlled according to its happening simultaneously,
whose risk has been easily assessed using a traditional
PRA technique. The operating risk from nuclear power
plants is categorized into this case. In this case, through a lot
of operating experience, as well as by taking the support of
standardized and enhanced evaluation techniques, the risk
from scenarios can be systematically predicted. Next, on the
other hand, the state of system is called ‘passive’ if any
event cannot be directly controlled according to its
happening simultaneously, whose risk then scarcely has
assessed via the traditional PRA technique. The future risk
from radioactive waste disposal facilities will be categor-
ized into this case. In this case, with natural hazards being
considered, the risk under any of the scenarios should be
estimated.
With a view to preventing exposure or unacceptable risk
to human health, which includes an adverse impact to the
natural environment, a waste disposal facility has to be
isolated from natural circumstances. Furthermore, it is just
necessary to demonstrate that waste disposal treatment is
reliable and that both man and environment can be
adequately protected. The risk assessment of a disposal
site includes specific models for addressing the major
pathways of radionuclides ill-affected human beings. Since
the system performance in the passive state is very sensitive
depending on a variety of scenarios, a temporal evolution by
change in both external and internal environmental
conditions becomes a fundamental factor in estimating the
risk. However, since the release of risk-dominant radio-
nuclides from a disposal site is expected to occur after
hundreds or even thousands years, usually there are no
direct methods of assuring the credibility of the evolution.
The risk profile coming from combination of likelihood
and consequences of future transport and exposure depends
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 516
on aleatory uncertainty, as well as epistemic uncertainty. In
addition, sources of the uncertainty at future states can be
reduced by pertinently setting up dependency relationships
interrelating geological, hydrological, and ecological
aspects of the site with all the scenarios. It is then obvious
that current methodology of uncertainty analysis of a waste
disposal site should be revisited with these diverse
considerations. The uncertainties arisen from different
evolutions should be adequately propagated throughout
the analysis. Interpreting the significance of results of a PRA
in view of the uncertainties is also important if the PRA
results are applied to make meaningful risk-informed
decisions. Subjective probability distributions on the
numerical results are given by uncertainty analysis and are
used to represent the confidence level where a risk criterion
is being met. While it is important to characterize the overall
uncertainty, it is equally worth understanding well what
factors or variables mainly propagate the uncertainty [19].
It is required to explicitly consider distribution functions of
random variables that define both aleatory and epistemic
uncertainty associated with each variable in a scenario
evaluation. In some variables, only epistemic uncertainty
may be treated [20–24].
After a survey of relevant studies [9,18,20,25,26], we
found most scenarios developed for a disposal site, without
considerations on above-mentioned dependency treatment
required, were not practically linked to the PRA. Further-
more, as to the evolution modelling for stochastic domain
variables under adversely-environmental conditions in
waste repositories, scenario analyses up to the present
have not given a specific and realistic approach. These
findings notice us not to lose sight of how dominating
domain variables contribute to the uncertainty of the risk in
relation with diverse complex phenomena and/or scenarios.
Therefore, the first and the second terms used in ‘sets of
triples’ of risk definition are delineated in detail in this
paper. The last term of ‘sets of triples’, i.e. the consequence,
may be easily identified using deterministic evaluation
models. In order to consider the effects owing to the
evolution of environmental conditions of waste disposal
facilities, this paper proposes a quantitative assessment
framework combining an inference process of Bayesian
network with a traditional risk analysis.
The remainder of this paper consists of four sections.
Section 2 provides the methodology framework for using
Bayesian network in the study. The probabilistic inference
algorithm employed is explained in Section 3 in detail.
A deterministic model of radioactive waste disposal,
including an explanation of the dependency between
scenarios and domain variables is presented in Section 4.
This section provides an overview of scenario categoriz-
ation, as well as characteristics of the domain variables.
A major portion of Section 4 is devoted to the development
of the evaluation models in relation with the information
given in Section 3. This section also provides some
illustrative results and discussion. The last section summar-
izes and concludes the study.
2. Framework development for the risk assessment using
Bayesian networks
2.1. Overview
Bayesian network also known as belief networks,
probabilistic networks, Markov random fields, and causal
networks, is a new concept for reasoning complex uncertain
problems, where network means a graphical model. In
utilizing Bayesian networks, it is general to communicate
qualitative assumptions about cause-effect relationships and
to derive causal inferences from a combination of diverse
assumptions. Bayesian networks have been applied to
diverse technical or societal problems, including medical
diagnosis (e.g. PathFinder, QMR [Quick Medical Refer-
ence]), map learning, real-time monitoring, language
understanding, pattern classification, forecasting, heuristic
search (e.g. Microsoft troubleshooter), etc. [27]. In addition,
in terms of providing decision analysis and expert systems
using artificial intelligence (AI), they have been popularly
utilized [28].
Since Bayesian network constitutes a model of reasoning
process in a system, answers to a variety of queries are
required. For example, a system analyst may want knowing
whether or not it is worthwhile to reinforce engineering
protection features in order to defence a random hazard (e.g.
an initiating event). The use of Bayesian networks helps to
answer such the queries even when no experiment about the
effects of reinforced protection is available.
The reasoning processes can be operated by propagating
probabilistic evidence in any directions, where various
forms of reasoning [29] such as prediction, abduction, and
‘explaining away’ are made: for example, if a forward
direction of reasoning from a cause (i.e. a predecessor) to an
effect (i.e. a descendant) is occurred, then prediction is
formed; if a backward direction of reasoning from an effect
(i.e. a descendant) to a cause (i.e. a predecessor) is occurred,
then abduction is formed; on the other hand, if we have
more than two causes and one of those causes is found, then
that reduces the likelihood of causes of the others, so called
it as ‘explaining away’. As these reasoning processes with
useful evidence are made, Bayesian network can provide a
basis to perform uncertainty propagation of query variables.
In order to build overall reasoning processes for Bayesian
networks, following three steps are generally engaged:
† Structuring the graphical models for causality
representation.
† Learning the structure or defining parameters of the
query variables.
† Estimating the probability of the query variables by any
inference algorithm.
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 517
As explained by Heckerman [30], Bayesian networks
have been explicitly applied as a popular graphical
modelling language to represent the causality under
uncertain relational situations in a real world. However, as
addressed in the above, a whole reasoning process is
deemed to be a complex task in itself. Unfortunately, since a
practical implementation of Bayesian networks requires
complex modelling of the inter-relationships between
various initiating events and engineered mitigating features
which diversify the risk profile, they are far from popular in
the nuclear PRA. Accordingly, after considering the
causality from the initiating events to the engineered
features, which will be explained in the following sections,
we propose a conceptual framework to apply Bayesian
networks to the nuclear PRA. In the PRA adopting Bayesian
networks, it is kept in mind that an initiating event simply
means a conditional initiation for any occurrence of the
scenarios.
2.2. Representation of the cause-effect dependency
In this section, we suggest a special technique for
applying the general reasoning concept of Bayesian net-
works to the PRA of a nuclear facility. Since our approach
mainly focuses on connecting scenarios (i.e. causes) to
consequences (i.e. their effects), we are mainly interested in
the direct representation of the real circumstance, not in a
full reasoning process of Bayesian networks.
A structure of Bayesian network is generally represented
by an annotated directed acyclic graph (DAG) G, where
some nodes denote variables of interest, and arcs, i.e.
linking lines, realize a probabilistic and causal dependence
between random variables that they connect. Usually, arcs
are drawn from a cause to an effect. Incomplete DAGs
encode expert knowledge in the form of missing arrows.
Therefore, the DAGs are regarded as a substantial language
for communicating conditional independence assumptions,
and for representing qualitative causal inferences [25].
Since the DAGs represent high dimensional uncertainty
distributions, construction of Bayesian networks helps to
capture engineer’s intuitive understanding of complex
systems or problems. Once the network is configured,
subsequent computations are pursued by symbolic manipu-
lation of probabilistic expressions. In general, after assign-
ing a probability (or its distribution) to a node conditional on
that node’s predecessors, the joint probability (or its
distribution) over the whole network is determined [26].
Presented with a step-by-step decomposition of the
network structure for solving a problem (e.g. for describing
the performance of a system), Fig. 1 shows an overall
process for making DAG G which finally gives a causality
between random variables of E denoting a cause (input like
an occurrence of any initiating event) and R denoting an
effect (output like a response of given problem). As a first
step for representing G, we can figure a problem to a simple
structure as outlined in Fig. 1(a). However, it seems that we
cannot provide any practical reasoning for the problem only
with this structure. If we want resolving the causality (i.e.
relationship of cause and effect) in detail, then we have to
make a further expansion in G. With this desire, the
structure in G is divided into two modules providing with
decomposing paths—‘from E to V’ and ‘from V to R’,
shown with dotted boxes in Fig. 1(b). A random variable V
denotes a trigger (a kind of filtering interaction) reflecting
the fact that the input E cannot always produce the output
R. In other words, by introducing the upper module (i.e. a
trigger) in the decomposition of Fig. 1(b), we can consider a
possible interaction to hinder a direct impact from
insignificant causes. The trigger V remains in a hidden
state. Moreover, we can represent this idea by expanding the
upper module in Fig. 1(b). Fig. 1(c) shows a result using
AND gate and NOT operator, where M denotes an
inhibitory indicator (or mechanism) that prevents any
cause E from giving an effect R and thus has two states—
true (on) and false (off). As a last step for figuring G, we
Fig. 1. The causality represented in four decomposition steps: (a) a basic
DAG simply representing a cause and an effect; (b) more extended DAG
consisting of 2 modules; (c) module 1: a detailed DAG between E and V
using a concept of M; (d) module 2: a detailed DAG between V and R with
n-elements of X. (Note: All the symbols in a node represent random
variables. E denotes an occurrence of any initiating event while R denotes a
response of given problem. V denotes a trigger or a filtering interaction
which reflects the fact that the cause E cannot always produce the effect R.
M denotes an inhibitory indicator or mechanism that prevents any cause E
from giving an effect R. X
k
is an actual system performance indicator
corresponding to a domain variable k under the dependency of initiating
events).
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 518
have to fully represent the lower module of Fig. 1(b) as
shown in Fig. 1(d). In the figure, we also introduce some
random variables which can give a system response R
following after the instantiation (it commonly means that a
random variable becomes a true state) of triggering;
therefore, let a random variable X
k
an actual system
performance indicator corresponding to a domain variable
k under the dependency of initiating events.
With the above-mentioned decomposition steps, we can
fully represent the causal relationships from the occurrence
of initiating events to the system response. In the general
graphical representation it is noted that the circle nodes
denote continuous random variables, squares denote
discrete; furthermore, a clear node indicates a hidden state
and a shaded represents an observed. The states of unshaded
nodes in the figures, therefore, cannot be observed in an
actual circumstance.
2.3. Modelling considerations for building
a network structure
In a graphical structure adopted in Bayesian networks
causality plays a great role. Practically, when identifying
direct physical connections we should consult a graphical
structure. According to this consideration, Bayesian net-
works take advantage of expressing a graphical description
of the dependencies between random variables. It is also
noted that when we make a network structure, the nodes
must always be numbered in topological order, i.e. ancestors
before descendants, to express the causality. Using this
variable ordering and a concept of conditional independen-
cies, we can usually configure the network structure.
We always want making a graphical structure predict
responses and infer causes from evidence. To do this,
Bayesian network needs a corresponding database and a set
of explicit assumptions about our prior probabilistic
knowledge of a domain. It is noted that almost all the
observed things can be represented with the evidence, i.e.
available information about the states of nodes. By way of
interfacing structure model to data, all kinds of evidences
can be propagated within the structure.
With this modelling concept, we attempted to represent a
more general structure model for application of a nuclear
PRA. Fig. 2 shows a fully-specified Bayesian network
corresponding with an actual inference problem. For the
purpose of showing an illustration, a multiply-connected
network with four query nodes is outlined in Fig. 2. The
network structure has a characteristic to encode conditional
independencies among random variables. Therefore, in the
selected graphical models, a missing edge explicitly means
that there is an independent relationship between random
variables. Absence of a direct link between any E and R, for
example, reveals that there is no direct influence of initiating
events on the final system status. In describing Bayesian
networks, we use a lower-case letter to represent any single
state of a random variable and an upper-case letter as a set of
state of a random variable. We also write xZk, for example,
to denote that variable x is in state k. When we observe every
state of a variable in any set X, we call it a state of X.
The logic of the structure provided in Fig. 2 is explained
in detail. E
i
in a root node represents the happening of an
initiating event i which affects physical domain variables of
given system. As a descendant node is affected by different
trigger nodes, diverse connections from any jth node of V
(V
j
) to any kth node of X (X
k
) exist depending on their causal
characteristics. For instance, in some variables there is only
one connection, but in others, 3 or 4 connections are
possible. As far as the response R from an instantiation of X
k
is concerned, we may define it as an adverse impact on the
system or any other consequential measure of the system.
Direct inference of the network structure shown in Fig. 2
is not easy to solve and requires numerous probability
distributions because a causal complexity exists between
both ancestors and descendants. For the purpose of
simplifying the inference after structuring of the model,
we may take up the concept of a noisy-OR gate [28,31],
specified by a specific function incorporating independen-
cies between nodes relevant to the converging path. Let’s
take an example for expressing this case. As denoted in
Fig. 1(c), if an individual trigger (V
j
following E
i
) occurs
together with its corresponding inhibitory indicator M, the
instantiation of their descendant (X
k
) in Fig. 1(d) is given
only that input is present, i.e. P
j
ZP(X
k
íV
j
only), where the
expression P(X
k
íV
j
only) means a conditional probability of
kth X node given only V
j
node. In addition, because the
process which prevents the input signal from being
sufficient is independent of the others, we can easily assume
that the causal mechanism in the connection path from V
j
to X
k
is mutually independent. Therefore, if this assumption
is accepted, the noisy-OR gate model can be selected.
Fig. 2. A Bayesian network with fully treating causal relationships in PRA
problems (As an illustrative purpose, four Es, Vs, and Xs are configured as
nodes, respectively. Three R nodes are also provided for the states of the
end point’s measure).
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 519
In the noisy-OR gate model, each input variable has a
probability of being sufficient to cause an output variable.
Given any combination of the input, the probability of a
descendant is expressed with a form of minimal cut upper
bound (using a principle of inclusion–exclusion) as
Pr(X
k
íV
1
; V
2
; .; V
n
) Z1 K(1 Kp
V
1
)(1 Kp
V
2
).(1 Kp
V
n
)
(1)
where the left expression specifies the probability distribution
of X
k
conditional on its parents of V
1
, V
2
, ., V
n
; n denotes a
maximum number of V node; p
Vj
means the input probability
of the parents V
j
to X
k
given an arbitrary number j, jZ1,2,.,n.
It is well known that the principle of inclusion–exclusion can
compensate the combined belief of outcome by the degree of
belief that is common to a different set of causes. The main
advantage of the noisy-OR gate model is that the number of
parameters in the variables of interest may be reduced
proportionally to the number of causes (n), while it is
exponentially increased (2
n
with binary state) in the
application of other general gate types. As a consequence,
the noisy-OR gate simplifies knowledge acquisition, saves
memory space, and allows evidence propagation in time
proportional to the number of parents of each variable [32].
When modelling a real world, we often encounter
situations where an effect has many potential causes.
However, if we practically consider the actual occurrence
of initiating events, since it may be reasonable that multiply-
occurring initiating events are relatively rare under the time
span considered, we can reduce Eq. (1) to an equation with
first order probability. Furthermore, with adopting assump-
tion about mutually exclusive situation for preventing dual
initiating events, the above reduced equation can be further
reduced to a single term but disjunctive element on the
possibility of each initiating event. As a result, we can
simplify the structure of Bayesian networks with one root
node. Given the response R shown in Fig. 3, only one end
effect is also proposed for the purpose of illustrative
application.
3. Approaches for a probabilistic inference in Bayesian
network
As denoted by Henrion [28], the notion of probability is
known for the best way to actualize uncertain beliefs, and its
role in Bayesian network is to provide a belief basis about
qualitative encoding of a network structure. Overlooking
based on our available knowledge base, expressing our
beliefs for the random variables’ states in a term of
probability, is quite a demanding task. In solving various
real problems, most values of the probability would be
obtained by engineering judgment rather than by insufficient
empirical data.
3.1. Representation of variable states
Generally, random variables have been regarded as a
matter of state. In an ideal case, they can have two simple
states—true and false. However, this is not fit for a general
case. Actually, a node can be gone to diverse states
representing continuous values.
As discussed by Heckerman [31], it is well acknowl-
edged that, in complex networks, parents’ inputs should be
carefully treated so as not to induce an impractical problem.
Suppose n parents X
1
, X
2
,., X
n
bearing on a descendant R
can have m states, respectively. According to the charac-
teristic of Bayesian networks, we must specify the
probability distribution of R conditional on every state of
its parents. Thus, we have to specify m
n
probability
distributions for representing this causality. Let’s take
another example for continuous random variable; the
value of root nodes given in Fig. 3 might be a Poisson-
distributed probability with parameters depending on their
occurrence frequencies.
Any random variable is represented by either a
continuous scalar or a set of discrete values. In Bayesian
networks, however, it is usual that a lot of variables such
as a decision variable are inherently discrete. Even
though a variable representing time-dependent situations
may be inherently continuous, it can be also treated as
discrete in the simulation for the convenience of
computation.
If a set YZ{Y
1
, Y
2
,.,Y
n
} is composed of n nodes of
discrete random variable Y
i
, iZ1,2,.,n and given a
specified DAG G in a Bayesian network, a joint probability
distribution over the set can be represented by the product of
the individual probability distributions over each Y
i
conditional on its parents. In other words, the causal
independence relationship in a graphical model informs us
how a global high-dimensional probability distribution over
all the variables can be factored into a product of local
probability distributions over lower dimensional spaces.
Fig. 3. Asimplified DAG of Bayesian network for considering non-multiple
initiating events in PRA (For the purpose of showing an illustration, n nodes
for random variable Xs and one node for a state of the endpoint’s measure
are provided).
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 520
If we denote the parents’ set of any ith variable Y
i
in G by
Pa
i
ZlPa
i
1
; .; Pa
i
m
i
¦ where m
i
is the number of parents of
Y
i
, and assign their values with a corresponding probability
set f
i
Zlf
i
1
; .; f
i
m
i
¦, Bayesian network defines a prob-
ability distribution over the possible assignments of values
to all the variables as
Pr(YíG) Z
Y
n
iZ1
Pr(Y
i
Zy
i
íPa
i
1
Zf
i
1
; .; Pa
i
m
i
Zf
i
m
i
; G):
(2)
Eq. (2) implies that, given the parents’ sets and G, each
variable Y
i
is conditionally independent of the other parents,
which implicitly provides a well-known definition of
‘d-separation (directional separation)’ [33] in Bayesian
network model.
3.2. Conditional probability estimation
In Bayesian networks, coming into existence of any
states of random variables is termed an instantiation. The
probability of each instantiation can be simply calculated as
the product of the conditional probabilities of its prede-
cessors. Therefore, we need specifying a complete joint
probability distribution (JPD) over all the random variables.
Full JPD over a given domain can be provided with a
combination of conditional independencies in network
structure and local probability models. From the rules of
probability, if we will apply Fig. 3 as an example, we have
the conditional probability as
Pr(xír) Z
Pr(x; r)
Pr(r)
Z
P
e;v
Pr(e; v; x; r)
P
e;v;x
Pr(e; v; x; r)
: (3)
In the above equation, Pr(e,v,x,r) is the joint probability
distribution of a set of random variables e, v, set of x, and r
determined from the overall network structure. In reality,
since this distribution may not be feasible to estimate, it is
desirable to efficiently utilize an alternative approach with
the consideration of conditional independence assumption.
The expression of Eq. (3) can be expanded in a way that
reflects the structure of Bayesian network itself [29]. It is
also noted that all the classical and logical operators can be
brought when defining unknown parameters in Bayesian
networks. In the case of Fig. 3, if n-element Xs are identified
in network, the JPD can be expressed, based on the chain
rule of probability:
Pr(e
i
; v
j
; x; r) ZPr(e
i
íq
e
)Pr(v
j
íe
i
; q
v
)Pr(ríx; q
r
)
Y
k
Pr(x
k
ív
j
; q
x
k
)
(4)
As explained previously, lowercase letters are commonly
used here to represent particular instantiations of variables.
A bold lowercase letter x represents a set of instantiations of
X elements. Each q shown in right side terms of Eq. (4)
denotes the parameter of each query node. We can state that
the variables of X
k
are conditionally independent of V
because the dependence between X
i
and X
j
is conditional on
the certainty of V, where isj. If we expand the Eq. (3) by
applying Eq. (4), we can obtain an explicit form of the
conditional probability:
Pr(xír)
Z
Pr(ríx;q
r
)
P
e
Pr(e
i
íq
e
)
P
v
Pr(v
j
íe
i
;q
v
)
Q
k
Pr(x
k
ív
j
;q
x
k
)
P
e
Pr(e
i
íq
e
)
P
v
Pr(v
j
íe
i
;q
v
)
P
x
Pr(ríx;q
r
)
Q
k
Pr(x
k
ív
j
;q
x
k
)
(5)
In functional and logical determinations, the state values
of the descendant node are wholly determined by the values
of the parent nodes. Under the circumstances of diverging
connections in Fig. 3, for example, where any of the paths
from one parent node V to descendant nodes Xs remain
unchanged, information can be transmitted from any X
k
node to the other X
k
/ node through the parent V unless V is
instantiated. On the other hand, in case of converging
connections where any of the paths from parent nodes Xs to
descendant node R also remain unchanged, if nothing is
known about R, then R’s parents (Xs) are being inter-
independent so that no information can be transmitted
between them.
3.3. Simulation algorithm for a probabilistic inference
In performing a traditional PRA, a variety of sensitivity
analysis are generally used to examine the importance of
various parameters and assumptions for risk modelling,
including specific risk attributes. On the other hand, in
Bayesian networks it is possible to use a learning process
supported by data on the structure and parameters of
variables. In a learning process with fixed graphical
structure, a parameter fitting problem is exposed and it
can be resolved with available reasoning process considered
[34]. Many structure identification methods or learning
algorithms are also used to optimize a structure, which is a
process for structure learning. However, since we let a fixed
structure in the study, methodology adoption for structure
learning is out of scope.
A key point of practical reasoning is to take the evidence
propagation by computing the answer probabilistically for
particular queries about the domain, which is generally
referred to as a probabilistic inference. Therefore, in order to
obtain an adequate solution on the reasoning with parameter
fitting, it is necessary to consider an appropriate probabil-
istic inference algorithm [35]. The probabilistic inference
focuses on estimating the posterior distributions of random
variables when relevant evidences are observed. There are
two kinds of probabilistic inferences; observing the effect(s)
of a given model and trying to estimate the causes is termed
a diagnostic reasoning; observing the root(s) of a given
model and trying to forecast the effect is termed a predictive
reasoning. Both forms of reasoning can be selected in
Bayesian networks. The probabilistic inference should give
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 521
an adequate answer to a query in a reasonable amount of
time. It has been frequently provided to query problems
within the specified bounds.
A lot of efficient methodology or algorithms for evidence
propagation in Bayesian networks have been proposed [36].
Various exact inference algorithms have been also devel-
oped for the probabilistic inference [37]. However, since an
exact inference solution in real-world networks seems to be
impractical, approximate probabilistic inferences have been
widely surveyed by many researchers. A wide variety of
approximate inference algorithms for Bayesian networks
share many common traits. They have realized utilizing
stochastic simulation-based, search-based, and variational-
based techniques [38], and so on. As a special case, search-
based approximation algorithms, which search for high
probability configurations through a space of possible
values, have appeared as an alternative to the overall
algorithms in Bayesian networks with extreme probabilities
[39]. Variational-based inference methods, meanwhile, use
a special technique, which exploits the law of large numbers
to approximate large sums of random variables by their
means [40].
One of the categories in approximate inference
algorithms, stochastic simulation-based algorithms such
as direct sampling, rejection sampling, likelihood weight-
ing algorithm, and Markov Chain Monte Carlo algorithm
(sometimes Gibbs sampling) are popularly used, although
they depend on the characteristics given a problem.
Advantage of the Monte Carlo algorithms includes their
simplicity of implementation and theoretical guarantees
of convergence. Disadvantages of the Monte Carlo
approach are that it can be slow to converge and be
hard to diagnose its convergence. The accuracy of
stochastic simulation-based algorithms generally depends
on the size of samples, irrespective of the structure of
Bayesian network. However, these algorithms have a
little disadvantage, as they converge very slowly if there
are extreme probabilities in the conditional probability
tables (CPTs) [37].
Finding a sufficient and effective inference algorithm
in static Bayesian networks is desirable with a computing
efficiency of simulation considered. Based on the
consideration for characteristics of the study, it may be
reasonable to apply a simple simulation-based algorithm.
However, since general marginalization using joint
probabilities requires exponential time in the probabilistic
inference, we had to consider more efficient methods. To
meet the above demands, we finally selected the best out
of a lot of inference algorithms, i.e. likelihood weighting
algorithm, emerged as a variation of stochastic logic
simulation for the purpose of giving an algorithm speed-
up. Instead of randomly choosing a value, the algorithm
takes values with conditional probability as a likelihood
weight. For instance, in the likelihood weighting, each
sample has a weight, and a final ratio is used to compute
the probability. The following explicit advantages have
given to the algorithm:
† It can be applicable to any type of nodes, discrete or
continuous.
† It avoids the inefficiency of rejection sampling by
generating only events that are consistent with the
evidence. That is, it fixes the values on evidence
variables and samples only query variables left.
† Each event generated is consistent with an evidence.
Also, it weights each sample by the product of the
conditional probabilities of the evidence variables,
given its parents.
Furthermore, it is noted that the algorithm displays
much faster convergence time than a simple stochastic
simulation, but is still slow under extremely small
conditional probability values. For applying to Bayesian
networks with extreme conditional probabilities, Dagum
[38] proposed a bounded-variance likelihood weighting
algorithm to provide efficient probabilistic inference in
polynomial time. This updated algorithm is adopted in the
study.
The weights in a network make local assertions about
the relationship between neighbouring nodes. Inference
algorithms turn these local assertions into global assertions
about the relationships between nodes. In the following
discussions, let E denote a set of observed node, Y the
nodes not contained in E. It is remarkable that a
probabilistic inference problem can be easily solved if
we have a JPD of the random variables involved. To
provide an inference probability Pr(YZyíEZe) which is an
extended expression of similar term as denoted in Eq. (3),
we must generate the relative approximation of two
probabilities Pr(YZy,EZe) and Pr(EZe), respectively,
with a relative error 3. In order to make these approxi-
mations, the likelihood weighting algorithm utilizes a
technique to decompose the full joint probability Pr(YZy,
EZe) into a path probability distribution p(y,e) and a
weight distribution u(y,e) presented by
p(y; e) Z
Y
Y
i
2Y
Pr[Y
i
íPa(Y
i
\E)]
YZy;EZe
; (6)
and
u(y; e) Z
Y
E
i
2E
Pr[E
i
íPa(E
i
)]
YZy;EZe
; (7)
where the expressions are conditioned, for example
Q
Pr[E
i
íPa(E
i
)]
YZy;EZe
, to denote an instantiation of their
arguments, in this case Y to y, and E to e; also, Y
i
\E denotes
ith arbitrary node which is contained in a set Y but not
included in the evidence subset E. Next, in order to use the
algorithm, let’s define a new binary random variable c(y,e)
that is equal to ‘true’ if YZy instantiates the node X to x,
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 522
and is equal to ‘false’ otherwise:
c(y; e) Z
True; if X Zx in y;
False; if Xsx in y:
(
Then the expectation of the random variable
c(y,e)u(y,e) becomes
E[c$u] Z
X
y2U
c(y; e)p(y; e)u(y; e) ZPr[Y Zy; E Ze];
where a sample space U for Y contains all the instantiations
of YZy. Also the expectation of weight distribution is
represented by
E[u] Z
X
y2U
p(y; e)u(y; e) Z
X
y2U
c(y; e)p(y; e)u(y; e)í
c(y;e)ZTrue
C
X
y2U
c(y; e)p(y; e)u(y; e)í
c(y;e)ZFalse
ZPr[EZe]:
Therefore, if we sample the distribution p(y,e) in
sufficient counts, the limiting ratio of c(y,e)u(y,e) and
u(y,e) can converge to a desired inference probability. The
simulation of the algorithm uses a random number
generator to uniformly select values for given subset of
evidence nodes and subsequently propagates them to pick
values for the other query nodes. Statistics are then
maintained by the corresponding values that the nodes
take out, and this process gives a desired answer. In
obtaining the simulation output, a key feature seems to
make a practical estimation of the query variable whose
parameters are an unknown function even under the
simulation inputs.
4. An example application to an environmental transfer
model of the nuclear facility
The application involves an example for estimating the
risk in a nuclear facility, particularly in association with a
radiological waste disposal facility. Near surface disposal is
one of the preferred options for comparatively large volumes
of low and intermediate-level wastes, which arise during
nuclear power plant operation, and also for wastes arising
from radionuclide applications in hospitals and research
establishments. Especially, the disposal system is a good
domain example for this study because its performance
potentially depends on a variety of scenarios. The end point’s
measure of interest can be simply defined as a peak dose to a
person drinking from a biosphere well. It is noted that the
likelihood of a peak dose can be also represented as an end
point’s measure resulting from the scenario evaluation.
4.1. Model inputs
Release pathways strongly depend on both a location and
design of a repository. Once the repository has been
determined, release pathway models should be assembled
so that they can represent radionuclide transport from the
repository to the environment and ultimately to the human
being. A normal exposure pathway in the post-closure
repository is generally defined by the non-accidental
leaching of radionuclides from the repository to human
beings. Key exposure pathways can be defined with the
consideration of both a likelihood and significance.
However, with respect to the risk concept, it may be unable
to identify the key exposure pathway because both high
likelihood and a great consequence cannot generally come
together. Several key factors such as delay times from
different engineered barriers, a transport of radionuclides
through groundwater in the geological formation, and
inventory uses by a critical human group are considered in
the model of exposure pathways.
4.1.1. Radionuclide release characteristics
A simple, hypothetic site with an engineered vault design
is adopted to the environmental transfer model. Wastes are
normally disposed at or near the ground surface, generally
not below 50 m in depth at near surface disposal repository.
In general, both natural and engineered systems form a
series of multiple and potentially independent barriers
against the release of waste [41]. The engineered barriers, as
well as waste containers and surrounding materials, and
backfills within the vault, can also contribute to the
repository performance. Various processes in the under-
ground environment (i.e. geosphere) such as dispersion,
retention by geologic media, and holdup in sub-layer aquifer
usually relate to the release amount of radioactive materials
to accessible environment (i.e. biosphere). Also, the release
timing is greatly affected by the assumptions concerning
occurrence time of breach of the engineered barriers.
Therefore, unavoidable uncertainties remain unsolved in
the time when waste packages are breached and in the way
how the radionuclides are released.
Source terms describe the release of radionuclides from
the repository to the pathways through the geosphere and far
field, which can ultimately go to the biosphere. From a
geochemical standpoint, there are several considerations in
defining the source terms; the rate and extent of container
degradation which determines when and how much waste is
exposed to the leaching and the other release mechanisms;
the rate how much radionuclides are released into the media
where they are transported; and local retardation of nuclide
movement on container debris, backfill materials, or other
engineered barriers etc. Altogether, these considerations are
important in the determination of a delay time of release. In
other words, in spite of the engineered features, a normal
delay of the source terms will occur because a delay from
breach of waste container and a relatively-slow leaching are
usually expected.
A basic mechanism governing the release in underground
media is the mass transfer in a porous medium where pore
water is a main agent for radionuclide transfer. After an
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 523
elapse of a delay period, the groundwater system becomes
an important part of pathway for radionuclide transfer. The
degree of saturation of medium, flow of the groundwater
system, and chemical properties of radionuclides may be
key uncertain elements in determining the release rate.
4.1.2. Deterministic pathway model
We adopted a deterministic model in the pathway
analysis, which was prepared by the National Council on
Radiation Protection and Measurements (NCRP) [42]. Since
the main objective of the NCRP model is to give a simple
means in predicting the results for screening analyses,
conservative approaches and parameters with effective dose
factors for 826 radionuclides were proposed in many
environmental transport pathways. The model gives a
screening technique that can be employed to demonstrate
compliance with environmental standards or other admin-
istratively-set reference levels against releases of radio-
nuclides to the biosphere. The screening dose assessment
gives a defensible basis for making regulatory decisions
because resulting dose is likely to be overestimated rather
than underestimated.
Only a groundwater ingestion pathway is considered in
this paper. The others such as direct irradiation, inhalation,
and soil and vegetable ingestion pathways are not addressed
because any given pathway model has no problem for
fulfilling the objective of the study. According to the NCRP
model, with only one time burial assumed, the average
yearly fraction of the initial inventory of the parent
radionuclide X
0
left from time t
delay
to t
delay
CT
av
is derived
as
X
0
Z
[1 Ke
K(l
L
Cl
r
0
)T
av
]e
Kt
delay
l
r
0
T
av
(l
L
Cl
r
0
)
(8)
where
l
L
leach rate (y
K1
)
l
r
0
radiological decay constant of parent (y
K1
)
T
av
averaging time
t
delay
delay time for release (y)
The leach rate defined in the NCRP model is
l
L
Z
I
R H n
(9)
where
R retardation coefficient (dimensionless)
I groundwater infiltration rate (m y
K1
)
H thickness of the layer of buried soil (m)
n soil porosity (dimensionless).
The considerable domain variables in Eq. (9) are the
infiltration rate and the retardation coefficients for each
radionuclide of interest. For nuclides with significant
progeny, average fractions of each progeny appeared from
the parent nuclide over whole lifetime is also provided in
Ref. [42]. With consideration of the screening factor for the
groundwater ingestion pathway, we can calculate the
screening dose rate D
i
, i.e. annual committed effective
dose of radionuclide i as:
D
i
ZM
0;i
$l
L
T
0
U
DW
V
X
N
iZ0
X
i
$DF
ing;i
(10)
where
M
0,i
initial source term inventory (Bq)
T
0
maximum leaching period (y)
l
L
leach rate of the parent nuclide (y
K1
)
U
DW
consumption of drinking water, taken to be
800 L y
K1
V dilution volume, i.e. total volume of water in the
aquifer (L y
K1
)
DF
ing,i
ingestion dose factor for radionuclide i (Sv Bq
K1
)
X
i
decay correction factor (dimensionless)
Basic site data are adopted from the study conducted with
the IAEA research program on the safety assessment of near
surface radioactive waste disposal facilities [43]. Thus, we
get a lot of system information; the repository is located in a
stable geological formation with an approximate depth of
3 m; the length of the facility becomes perpendicular to
the direction of groundwater flow; all the soil layers are
composed of saturated zone, which implicitly means the
properties of unsaturated soil layer such as partition
coefficient or bulk density are the same as those of saturated
soil layer; the geosphere is also characterized as an
unconfined sand aquifer with defined properties. Although
this adoption would intentionally limit the work scope, yet
there is sufficient realism to identify the feasibility of
modelling approach suggested in this paper.
4.1.3. Scenario classification
The combination of events, features, and processes
causing diverse natural phenomena can be taken as a
scenario. Sometimes aggregation of events and correspond-
ing special situations can go into scenarios. We introduce
the concept of ‘altered evolution scenarios (AESs)’ here as
one of extreme and special types of the scenarios. For
convenience’ sake, AESs are defined as unusual happenings
that can cause significant alteration or momentary changes
of underground geochemical, hydrological, and/or mechan-
ical properties which are directly linked with the system
performance. Therefore, significant initiating events and
concurrent responses resulting in adverse disturbance of the
systems concerned can go to the major AESs. In this paper,
disturbance means an abnormal state departed from the
normal state. Any scenarios leading to similar consequences
can be grouped into the same category of AES.
Referring to previous scenario studies [44,45], it may be
enough to classify the scenarios with four major AES
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 524
categories, as outlined in Table 1, for applying to the waste
disposal repository. In general, the AESs has a stochastic
behaviour, and may depend on a time-dependent character-
istic [46].
(1) Geological AES. Earthquake or tectonic displacements
(i.e. faults) have been regarded to be critical in the
geological stochastic analysis. It is noted that large
faults causing a direct release of radionuclides are
expected to be improbable because of the plastic
behaviour and self-sealing capabilities of clay. Never-
theless such the faults with low or intermediate level
may be responsible for local disturbances in the
properties of geological barrier, so result in altering a
normal evolution of the system. The primary hazard
considered is the rupture of waste canisters induced by
the faults. Observing many historical surface fault
ruptures, McGuire [47] indicates that often the map-
view width of the fault zone during an earthquake is not
restricted to a narrow zone along the primary fault but a
fault zone that is wide up to kilometers. He also notes
that changes in ground stress following an earthquake
can induce pore pressure changes and groundwater
flow, as well as water table changes. It is assumed that
the geological AES can, therefore, strongly influence
two variables of the system, i.e. delay time for release
and dilution volume. We can include a human intrusion
scenario in this AES.
(2) Climatological AES. The AES representing a climatic
variation can cause medium level perturbations in the
geosphere. The major consequences in this case will
probably be a reduction of infiltration rate or aquifer
flow. It is well acknowledged that the infiltration rate is
highly dependent upon the weather; spatial and temporal
variations (soil erosion) in precipitation have the greatest
effect on the infiltration rate. Another possible conse-
quence will be the variation of aquifer flow, which
becomes a considerable increase or decrease in the
aquifer characteristic, ultimately changing the release
quantity of radionuclides. Therefore, we need determin-
ing how much the infiltration rate or an aquifer flow can
change from a nominal value due to the AES.
(3) Surface hydrological AES. A lot of phenomena such as
freshwater sediment transport and deposition, natural
thermal effects, and site flooding have the potential to
change regional hydrological properties over time
scales of hundreds years. The phenomena are combined
with the normal evolution pathway creating a lumped
scenario, surface hydrological AES. This AES may
greatly influence the groundwater recharge and dis-
charge. The quantity changes of a water table and/or an
aquifer can also occur resulting from the disturbances of
local surface hydrological behaviour due to this AES.
Adverse effects of this case result in a considerable
decrease in the dilution volume, causing an increase in
the release quantity of radionuclides.
(4) Ecological AES. Vegetative progression is expected to
occur near surface disposal after a loss of waste control.
This causes two effects; a biotic intrusion is consider-
ably increased, even though its likelihood depends on
root depths; also a water budget can be changed.
Progression toward trees, for instance, may lead to
larger canopy hold-up, which can alter the amount of
water available for infiltration. Progression may tend to
either increase or decrease the infiltration [48]. The
extreme water consumption by floras can also cause to
reduce the width of a water table or an aquifer. The
soil/liquid partition coefficients of radionuclides can be
derived from soil-to-plant concentration ratios for
specific plants and types of soil. Sheppard and Thibault
[49] defined a partition coefficient correlation of leafy
vegetables based on plant-to-soil concentration ratios.
Consulting this correlation, we assume that the AES
indirectly influence the change of plant-to-soil concen-
tration ratios, ultimately the soil partition coefficients.
4.1.4. Consideration of domain variables
Nominal parameters of the domain variables are mainly
adopted from available references [43,50], as given in
Table 2. Adopted values for the parameters depend on
assigned subjective distributions. ‘Fixed but unknown’
distributions represent the epistemic uncertainty about
characteristics at a particular site unless site-specific
information is available for the corresponding parameters.
To develop a subjective distribution of the parameters in the
model, reported available data must be implicitly inter-
preted. Uncertainty in a variable is also treated as ‘aleatory’
when the distribution is determined from a random process.
Table 1
Example compilation of events, features, and processes for the classification of AESs
Geological AES (AES-1) Climatological AES (AES-2) Surface hydrological AES (AES-3) Ecological AES (AES-4)
Seismicity Precipitation River flow and lake level changes Plant uptake/evolution
Fault activation Glaciation Site flooding Animal uptake/evolution
Fracturing Deep weathering Recharge to groundwater Uptake by deep rooting species
Regional Tectonics Transient greenhouse-gas induced
warming
Freshwater sediment transport and
deposition
Microbial interactions
(Human intrusion) River, stream, channel erosion
(downcutting)
Effects of terrestrial ecological
development on hydrology
Natural thermal effects
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 525
Furthermore, stochastic effects and data uncertainty can be
considered altogether in a set of the distributions. Taking
this consideration along with the epistemic uncertainty of
the parameters, we can outline a domain variable with ‘true
but unknown’ distributions.
All upper and lower bounds for the parameters of the
domain variables given in Table 2 are assumed in the
illustrative application, partly based on available infor-
mation addressed in the followings. In particular, in case of
handling the domain variables with both epistemic and
aleatory uncertainties, we assume that a variance of the
epistemic distribution closes to a mode of specific
representative, which intends solving an issue that how we
can provide the parameters of the domain variables when we
have a lack of critical data set for them. By giving high
confidence on the variance of the epistemic distribution like
this, we can focus on the problems desired to be solved.
† Delay time. In this paper, the time between initial burial
and the beginning time of concrete structure failure is
defined as the delay time (T
delay
) for release. Waste form
hold-up and container lifetime cause a natural delay and
then a reduction in the vault release terms. The delay
time of the release varies depending on adsorption of
waste and concrete. It is noted that while a probable
failure mode of concrete may be a gradual failure, there
is more or less a possibility of a rapid failure. It is
assumed that a geological AES can directly affect the
integrity of concrete structure. Minimum T
delay
is
determined to be around 10 years. Maximum T
delay
is
determined by the maximum time of the simulation.
† Retardation coefficient. The delay in release can be
described by a variety of analytic models. For example,
leaching by dissolution or adsorption by buffer material
is one of the models. The delay mechanism and its
relevant time scales are the most critical data. It is usual
to assume that groundwater transport processes are
dominated by various phenomena such as advection,
diffusion, and sorption, etc. All the heterogeneous
phenomena can be lumped together into a model for
retardation coefficient. Partition coefficients of radio-
nuclides need calculating the retardation coefficient of
soil layer. As a domain variable, the soil partition
coefficient (SPC) can influence mass transfer between
soil and an aquifer, and consequently the radionuclide
concentrations in drinking water consumed by human
beings are affected by this. Uniform distribution for this
epistemic parameter uncertainty is assumed with lower
and upper bounds for each specific radionuclide.
† Infiltration rate. The source terms can be provided by
estimating a flux of groundwater near the repository.
Estimating the flux directly consists in finding the
amount of water that infiltrates into soils, redistributes,
and enters the disposal unit as a function of time. It is
noted that the infiltration rate (IR) is very site specific
[51] and depends on a natural stochastic process. It is also
well known that soil hydraulic conductivity is a critical
factor for deriving the infiltration rate and Darcy
infiltration velocity can be limited by the specified
hydraulic conductivity of the vault roof [43]. Beyeler et
al. [50] have a specific model that the infiltration rate is
the product of application rate (AR) and fraction of the
applied water that will percolate deeply beneath a root
zone and become infiltration. They estimated a minimum
irrigation water requirement, which uses an empirical
relationship between soil permeability and the proportion
of water developed by the US Bureau of Reclamation.
Based on this consideration, we guess that the scenarios
such as climatological AES or ecological AES may
greatly affect the infiltration rate.
† Dilution volume. The contribution to an ingestion dose
from the use of groundwater depends on many domain
variables. One of the major domain variables, dilution
volume (DV) directly relates with a transport modelling
of several contaminants through an underground hydro-
logical system. All the radionuclides released from the
buried waste are assumed to be diluted in the volume of
aquifer defined as equivalent tube stream area times
transverse length of the aquifer to a well. If a time period
long enough to change the groundwater resources is
provided, it is accepted to consider a lot of natural
stochastic processes directly or indirectly affecting the
volume of aquifer. Reference data from Beyeler et al.
[50] is adopted again for choosing a nominal value in this
paper.
4.2. Evaluations associated with information gained from
the probabilistic inference of Bayesian network
This section devotes our concerns to combine the
probabilistic inference of Bayesian network with
the estimation of parameters. Likelihood estimation of
Table 2
Subjective probability distributions representing the state of knowledge about fixed but uncertain parameters in using default values for Th-230 in the model
Fixed quantity (units) T
delay
(yr) SPC (m
3
/kg) IR (mean) (m/yr) IR (s.d.) (m/yr) DV (mean) (m
3
) DV (s.d.) (m
3
)
Epistemic distribution LU U U LU T U
Min 10 3.0!10
3
0.1 1!10
K3
6.0!10
4
1.5!10
4
Mode – – – – 9.1!10
4

Max 1000 3.4!10
3
0.3 2.4!10
K2
1.22!10
5
2.5!10
4
T
delay
, delay time; SPC, soil partition coefficient; IR, infiltration rate; DV, dilution volume; s.d., standard deviation; LU, loguniform; U, uniform; T, triangular.
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 526
the AESs is a starting point in the analysis. As a practical
assumption, an AES stochastically occurs in space and in
time so that the underlying measure of interest (e.g. the
occurrences of an AES in a specified time interval) is faced
with a stochastic process. Therefore, a frequency estimation
of AES occurrence is related to the realization of aleatory
characteristics of the parameters. The occurrences of AESs
for non-overlapping intervals are independent of each other.
As summarized by Siu [52], the Poisson distribution is
commonly used in parameter estimation problems to model
all processes where events are generated over all the time.
Considering this perspective, the probability of observing N
events in a time period T is given by
PlN events in time Tíl(t)¦ Z
ll(t)$T¦
N
N!
e
Kl(t)$T
; (11)
where l(t) is an expected occurrence rate of an AES. Given
the jth AES occurrence rate, l
j
(t), we can estimate the
probability of occurrence of a first time event within a
specified time interval t:
p(nR1) Z1 KexplKl
j
(t)$t¦: (12)
Thus, if the expected occurrence rate is provided,
Eq. (12) easily calculates a scenario probability. A best-
estimate value may be taken to each occurrence rate of
AESs. However, it is noted that the occurrence rate of an
AES should be estimated with the support of empirical
database and/or expert judgment. Historical data collection
and relevant methodology development are then necessary
in order to establish a reliable data base about stochastic
information of AES, which is beyond the scope of this
study. Instead of choosing insufficient empirical data, the
frequency about AES likelihood is just estimated based on
the subjective engineering judgment which uses a concept
of categorized selection criterion, as denoted in Table 3. For
preparing each likelihood category of Table 3, we set each
occurrence interval and median occurrence frequency. As
explained in the previous section, by referring to the survey
of relevant information, we made a decision to choose the
likelihood category for each AES, which resulted in suitable
designation so as category 3 (‘intermediate’ likelihood) for
AES-1 and AES-3, and category 2 (‘likely’ likelihood) for
AES-2 and AES-4, respectively.
As recognized by Frank [8], owing to environmental
circumstances by many phenomena in the groundwater
during the release of radioactive materials, the
consequences such as an exposure dose will involve both
epistemic and aleatory uncertainties. Also, it is already
identified that the subjective probability distributions
representing parameter uncertainties may undergo changes
over the time owing to the scenarios causing perturbations
[53]. We are sure that, in the consideration of environmental
transfer models, there are strong dependencies between
environmental scenarios and the characteristics of the
geosphere’s and biosphere’s domain variables. Therefore,
we assume that the subjective parameters of the domain
variables, which are critical for estimating the end point’s
measure of ground release of radionuclides, can vary
according to the occurrence of various AESs.
As explained earlier, it is reasonable to assume that the
occurrence of AESs directly influences estimating uncertain
parameters of the domain variables. If that is the case, the
dependency between scenarios and variables can obviously
affect all the input types of a model. However, no precisely-
defined methodology exists in identifying adequate input
types and standardizing the inputs. Very often, while there is
an incomplete database, the analyst tends to resort to
subjective inputs by some experts who make full use of their
experience or the literatures published.
The informal approach for performing uncertainty
analyses involves a technique varying one parameter or
set of parameters at first and then observing the deviation
from a base-case prediction [22]. This kind of approach was
considered in the study so as to obtain reasonable
uncertainty estimates based on the dependency between
scenarios and domain variables.
In order to realize this consideration, we adopted a
so-named ‘mixture prior’ concept. This concept refers to a
special approach using correction factors for representing
the parameter uncertainty. By introducing the ‘mixture
prior’ we can articulate the effects of uncertain parameters
under the occurrence of AESs. The term ‘contamination’ is
addressed in a ‘mixture prior’. The ‘contamination’ means a
generation of ‘contaminated prior’. The ‘contaminated
prior’ represents a variety of input parameters which reflect
the previously-mentioned dependency. Therefore, the
‘mixture prior’ shows a mixing status of a normal prior
(as termed for representing ordinary uncertainty distribution
in this paper) and the ‘contaminated prior’.
Under those representations, first, we define an arbitrary
prior set G to elicit a linear estimator of an uncertain
parameter q. Next, the priors p(q) ‘close’ to a single normal
prior p
0
(q) can be realized with a class of possible
‘contamination’ [54]:
G Zlp : p(q) Z(1 K4)p
0
(q) C4q(q); q2J¦; (13)
where J is a class of possible ‘contamination’, and 4 is an
adjustable correction factor with 0%4!1 which reflects
how p closes to p
0
. Next, if we attempt to give the prior
class q(q) with a maximum likelihood estimator (MLE) for
reflecting a possible ‘contamination’, then Eq. (13) becomes
Table 3
Categorization of AES likelihood in the frequency estimation
Degree of likelihood Estimated occurrence
interval
Occurrence
frequency (median)
(1) Highly likely 0–10
2
y post-closure 2.0!10
K2
/yr
(2) Likely 10
2
–10
3
y post-closure 2.0!10
K3
/yr
(3) Intermediate 10
3
–10
4
y post-closure 2.0!10
K4
/yr
(4) Unlikely 10
4
–10
5
y post-closure 2.0!10
K5
/yr
(5) Highly unlikely 10
5
–10
6
y post-closure 2.0!10
K6
/yr
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 527
a similar form of a convex combination of the MLE of the
‘contaminated prior’ and a best estimate of the normal prior
[55]. Furthermore, the factor 4 will give a quantitative
measure for showing a possibility of MLE of the
‘contaminated prior’.
As modelled in Section 2, the information gained from
the probabilistic inference of Bayesian network, i.e.
probabilities on the query random variables obtained from
evidential reasoning can be used to approximately predict
the factor 4. However, in fact, the prediction of MLE of the
‘contaminated prior’ requires engineering/scientific knowl-
edge specific to the problem, as well as probabilistic
modelling expertises.
4.3. Results of example application and discussion
The previous section indicates that we can connect the
results of probabilistic inference from Bayesian network
with the consequence evaluation model addressed. In a
practical matter, we had to provide some supporting
information to get a variety of quantitative expectations
about the relationship between domain variables and AESs.
As a previous step for estimating the risk from future
evolution, supporting information is directly used in the
program input of Bayesian inference algorithm.
Preparing specific information on the CPTs used in the
inference program seems not to be an easy task. None-
theless, based on current knowledge of the relationship
between domain variables and scenarios, which are usually
gathered from relevant experts, we can do that. This kind of
solution can be explained from the other case. When we
consider a traditional level 2 PRA for nuclear power plants
particularly in the determination of containment progression
event branch, we believe that it is unavoidable to give a
single representative estimate of the qualitative probability
based on user’s degree-of-belief on corresponding severe
accident phenomena [56]. Even though the PRA considers
interval probabilities covering the uncertainty in the
expressions of branch possibility, only nominal point
estimate values primarily based on engineering judgments
are used in the quantification. This is similar to the approach
adopted for getting the CPTs in this paper.
As a general assumption, each individual AES must last
for sufficient period so as to cause significant characteristic
changes to the domain variables. Multiple dependencies
may occur by the simultaneous AESs. However, since its
consideration needs doing complex graphical represen-
tations in Bayesian network, it is not treated for simplifying
the problem.
4.3.1. Conditional probabilities
We had primarily considered the dependencies between
4 domain variables and 4 AESs with the information
explained in Section 4.1. While keeping this consideration
in mind, conditional probabilities for representing these
relationships were determined using single representative
estimate. Furthermore, we had already addressed an
approximate inference program with a bounded-variance
likelihood weighting algorithm in order to calculate the
conditional probability of query nodes of Bayesian network.
Before practically using the program, we decided to apply a
popular network structure with four simple nodes—the
so-called ‘WetGrass’ problem [37], for the purpose of doing
verification and showing actualization of applied inference
algorithm. We found a very fast convergence during the
verification process, where the convergence was checked
with the following stopping rule provided by the developers
of the algorithm [38,57]:
S
T
R
4l(1 C3)
3
2
ln
2
d
(14)
where T is a number of samples; 3 is a relative error; d is a
failure probability; and l is a defined constant. We obtained
values for the query nodes that were close to the exact
probability within an error of G0.005, with 3Z0.1 and
dZ0.05.
Next, we let the program apply to actual simulations
when the network structure was given. In preparing a
simulation input for CPTs, we defined a dependency matrix
with an ad hoc basis, where the dependency relationships
with four grades—high, medium, low, and zero—were
assumed. All the random variables were simulated con-
currently with each AES. Fig. 4 provides two running
outlines from the simulation of all the AESs, where the
convergence in the iteration of Monte Carlo sampling is
desired. The simulation easily reached a stable state except
initial transitions. Also, in most cases, the simulation made a
progression quickly and ended within 30,000 times’
iteration. With the acceptance of this simulation on
approximate probabilistic inference, therefore, we can
provide the solutions of the query random variables
depending on each AES, as denoted in Fig. 5.
Fig. 5 provides some remarkable insights, acquired in the
simulation results, on the relationship between random
variables used in Bayesian network and various AESs. First,
the random variable corresponding to a disturbance of T
delay
seems to have a great causal influence to the occurrence of
AES-1 (geological AES); similar insights are followed with
that of IR to the AES-2 (climatological AES) and with that
of DV to the AES-3 (surface hydrological AES), respect-
ively. Second, even though we let any strong dependency of
random variables to the AES-4 (ecological AES) provide,
which seems to be usual in case of the random variable for
SPC, significant results are hardly shown in representing
relative importance of the random variables against the
AES-4. Last, compared with the others, the random variable
for SPC reveals relatively-small conditional probabilities; it
seems that low confidence of the degree-of-belief to the
disturbance of this variable is reflected in its likelihood
against all the AESs.
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 528
4.3.2. Prediction examples of the end point’s measures
By making a mixed prior directly by taking the
information given in the form of Eq. (13), we can move
on a first step to estimate the end point’s measures of the
environmental transfer model. Since the contaminated prior
is provided in the evaluation process following after
occurrences of AESs, the effects of AESs must be reflected
in the parameters of random variables. The parameters of
the random variables were prepared by approximating
values referred to a maximum or minimum bound of normal
parameters depending on their probability distributions. For
instance, an upper bound of delay time would be extremely
reduced according to the existence of AES-1, provided the
dependency between geological phenomena and degra-
dation of engineered barrier of repository is considered. On
the other hand, an upper bound of the location parameter in
a domain variable may be near to the best estimate of a
normal prior depending on specific situation. Similar
considerations were given for the other domain variables,
where a variable-specific treatment was done.
Historically, a form of complementary cumulative
distribution functions (CCDFs) has been used as a favourite
form for portraying uncertainty results of risk. CCDF is an
easily-plotted risk curve, where an exceeding probability
against an end point’s measure can be expressed. In this
study, the CCDF construction process with a Monte Carlo
sampling technique was adopted from Ref. [25].
Before making the simulations, we must identify the
endpoint’s measure of this paper with a screening factor
(SF) as it was suggested by the NCRP [42] for representing
the consequence (e.g. exposure dose). The screening factor
has the dimension of committed effective dose (Sv) for a
unit concentration of radioactivity (Bq) in the media, finally
has a unit of Sv/Bq as provided in the NCRP model. Only
Th-230 was selected for the purpose of illustrative
evaluation. We also tried to usually get 10,000 times’
simulation with all the combinations of AESs. Next, the
comparative results between normal (i.e. non-AES) case
and abnormal (i.e. occurrence of AES) cases were
illustratively provided. Even though the frequencies of the
occurrence of AESs in Table 3 were reasonably assumed, it
is necessary to confirm that the adopted values do not cause
any problem in predicting the end point’s measures. We
performed a sensitivity study to extensively survey the
effects of frequency variation. This sensitivity study can be
potentially used to check the importance of the AES
frequency. Fig. 6 indicates that the variation of occurrence
frequency of AES-2 has no strong influence on the results
such as mean values of the end point’s measure. In other
words, although an extremely-frequent occurrence of AES-
2 is assumed, the deviation against the case of non-AES is
only 29% approximately. For the purpose of giving our
confidence to this insight, however, it may further need
determining the most limiting case in that the scenario
frequency influences the uncertainty propagation.
Various CCDF fractiles (i.e. fractions of percentile) for
exceeding probabilities of the end point’s measure are
shown in Fig. 7, which gives us good insights in the
uncertainty propagation for non-AES case. It is noted that,
Fig. 4. Convergence representation of Bayesian inference program for the cases of AES-1 and AES-4.
Fig. 5. Comparison chart in the estimation of conditional probability of
random variables from the Bayesian inference program, classified with
each AES. (Note: variable 1Ztime delay, variable 2ZSPC, variable 3ZIR,
variable 4ZDV, respectively).
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 529
in terms of estimating an end point’s measure, there is no
big difference between the uncertainty output (e.g. mean of
mean CCDF) and the deterministically-calculated value by
the NCRP. In the case of a subjective confidence interval,
however, Fig. 7 reveals that there is a great possibility for
showing a wide band. For instance, the end point’s measure
stands at about 2.5 times’ difference in the interval between
95 and 5% (i.e. 90% confidence interval) at a credible
exceeding possibility limit of 0.05.
Since we are interested in the ranking for the importance
of AESs to the impacts of the end point’ measure, a focused-
down analysis on each AES has been conducted. The mean
CCDF curves of the AESs resulted from the uncertainty
propagation are depicted in Fig. 8. The results are made with
nominal frequencies given in Table 3. Compared with
the non-AES case, we can easily identify the relative
importance of AESs at any subjective confidence level.
In particular, the occurrence of AES-2 has a strong impact
on the outcomes of the end point’ measure because high
dependency of AES-2 on the domain variables such as IR
and DV has been already anticipated as shown in Fig. 5. On
the other hand, the results for AES-1 and AES-3 curves are
almost the same; this is partly because there were no big
differences of conditional probability between AES-1 and
AES-3 except only for the variable of ‘delay time’, as shown
in Fig. 5, and partly because we had already let same
likelihood category for both AES-1 and AES-3, as
previously explained.
To investigate the importance of the parameters of the
delay time, we have performed another sensitivity study.
Just after defining two cases with fixed inputs for the delay
time, i.e. one for 10 years and the other for 300 years,
we made a separate simulation and ultimately compared
with a standard case having previously-given distribution.
In this case, randomly occurrence of AESs was assumed
with the other domain variables remained unchanged. Fig. 9
shows an outstanding profile for the case of Cs-137.
Fig. 6. Dose in unit of screening factor and occurrence rate of AES
according to a variation of the AES-2 frequency. Arithmetic means of 5%,
mean, and 95% fractiles are shown. (Occurrence rate of AES means the
AES occurrence counts per total simulation times. SF denotes the screening
factor).
Fig. 7. Comparison of uncertainty propagation results for non-occurrence
AES case.
Fig. 8. Mean values according to the AESs, including non-occurrence of
AES.
Fig. 9. Mean values according to random AES occurrences, showing
sensitivity results for the delay time (Radionuclide: Cs-137).
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 530
Wide band of the end point’s measure is derived from the
input variation of delay time. According to this result, we
can assert that it is more desirable for the uncertainty
analysis of relatively-short lived radionuclides to prepare a
best-estimate value of dominant domain variable (i.e. the
delay time in this case).
5. Summary and conclusions
Since Bayesian networks can contribute to identify the
causality of an uncertain system’s behaviour, we introduced
its inference process quantitatively estimating the depen-
dency between stochastic scenarios and affected domain
variables of the system. A general approach integrating
the Bayesian network concept to the nuclear risk assessment
was proposed and illustratively demonstrated. After simpli-
fying the network corresponding with a problem-specific
structure, we developed and verified an approximate
probabilistic inference program using bounded-variance
likelihood weighting algorithm, finally shown to be
adequate in the verification test for clarifying the depen-
dency relationships under different problem queries. In
addition, in estimating the consequence of example pathway
problem, i.e. environmental transfer of radionuclides from
radioactive waste disposal, a Monte Carlo simulation
process accounting for epistemic and aleatory uncertainties
was provided. Ultimately, specific models, including a
model for propagating uncertainty of relevant parameters,
were developed with a comparison of variable-specific
effects due to the occurrence of diverse altered evolution
scenarios (AESs). After defining and identifying the terms
for specific AES, we classified them into four major
categories.
While analysing the outcomes of the illustrative
evaluation of the environmental transfer pathway, we
could identify exceeding probabilities of the end point’s
measure and importance ranking on each AES. The domain
variables important in the consideration of causality were
found, and also dominant contributors to the uncertainty
propagation might be identified.
Not a few remarkable features appeared in applying the
proposed approach; the knowledge base for establishing
dependency relationships was prepared and enhanced
during the simulation; it was worthwhile to establish
definite rationale on the dependency information for getting
general insights through probabilistic inference modelling
in the prediction of future uncertain state of a system.
Therefore, further researches need enhancing the knowl-
edge base of various modelling areas, especially in the
prediction of the occurrence frequency of AESs and in
the assessment of causality between domain variables and
stochastic scenarios. We have also an understanding that
more appropriate data for domain variables should be
provided to refine the results. A systematic technique for
expert knowledge survey and utilization, as well as the use
of experimental results, can support to enhancing the
predictability. However, if the epistemic uncertainty in the
distribution of event occurrence frequencies will be
addressed in this study, a more complicated simulation
process may be required.
While taking a modelling approach proposed in this
paper, we made a causal inference process including
uncertainty estimates, also got a number of results to
improve the current knowledge base for prioritizing future
risk-significant variables in an actual site. The identified
specific results presented will probably be useful in making
the prediction of future system states. Furthermore, the
proposed approach will probably give a general and flexible
structure for doing further extension studies and appli-
cations. For instance, this study can be extensively applied
to the evaluation of all the exposure pathway models,
provided the relationships between AESs and domain
variables are searched in detail. The evaluation can be
also intensively done for any specific scenarios such as a
specific phenomenon categorized in this paper.
References
[1] Kaplan S, Garrick BJ. On the quantitative definition of risk. Risk Anal
1981;1(1).
[2] Helton JC. Risk, uncertainty in risk, and the EPA release limits for
radioactive waste disposal. Nucl Technol 1993;101.
[3] Slaper H, Blaauboer R. A probabilistic risk assessment for accidental
releases from nuclear power plants in Europe. J Hazard Mater 1998;
61:209–15.
[4] Protection from Potential Exposure: A Conceptual Framework. ICRP
64; 1993.
[5] Modarres M. What every engineer should know about reliability and
risk analysis. New York: Marcel Dekker Inc; 1993.
[6] Garrick BJ, Chistie RF. Probabilistic risk assessment practices in the
USA for nuclear power plants. Safety Sci 2002;40:177–201.
[7] Solomon K, Giesy J, Jones P. Probabilistic risk assessment of
agrochemicals in the environment. Crop Prot 2000;19:649–55.
[8] Frank MV. Treatment of uncertainty in space nuclear risk assessment
with examples from Cassini mission applications. Reliab Eng Syst Saf
2000;66:203–21.
[9] Wilmot RD. The treatment of climate-driven environmental change
and associated uncertainty in post-closure assessments. Reliab Eng
Syst Saf 1993;42:181–200.
[10] Rowe WD. Understanding uncertainty. Risk Anal 1994;14:743–50.
[11] Pate-Cornell ME. Uncertainties in risk analysis: six levels of
treatment. Reliab Eng Syst Saf 1996;54:95–111.
[12] Caruso MA, Cheok MC, Cunningham MA, Holahan GM, King TL,
Parry GW, Ramey-Smith AM, et al. An approach for using risk
assessment in risk-informed decisions on plant-specific changes to the
licensing basis. Reliab Eng Syst Saf 1999;63:231–42.
[13] Oberkampf WL, DeLand SM, Rutherford BM, Diegert KV, Alvin KF,
et al. Error and uncertainty in modeling and simulation. Reliab Eng
Syst Saf 2002;75:333–57.
[14] Hoffman FO, Hammonds JS. Propagation of uncertainty in risk
assessments: the need to distinguish between uncertainty due to lack
of knowledge and uncertainty due to variability. Risk Anal 1994;14:
707–12.
[15] Hofer E. When to separate uncertainties and when not to separate.
Reliab Eng Syst Saf 1996;54:113–8.
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 531
[16] Siu N, Malik S, Bessette D, Woods H. Treating aleatory and epistemic
uncertainties in analyses of pressurized thermal shock. PSAM 5 2000.
[17] Helton JC, Martell M, Tierney MS. Characterization of subjective
uncertainty in the 1996 performance assessment for the waste
isolation pilot plant. Reliab Eng Syst Saf 2000;69:191–204.
[18] Preyssl C. Safety risk assessment and management—the ESA
approach. Reliab Eng Syst Saf 1995;49:303–9.
[19] Parry GW. The characterization of uncertainty in probabilistic risk
assessments of complex systems. Reliab Eng Syst Saf 1996;54:
119–26.
[20] May LM, Gaborek B, Pitrat T, Peters L. Derivation of risk based wipe
surface screening levels for industrial scenarios. Sci Total Environ
2002;288:65–80.
[21] Hattis D, Burmaster DE. Assessment of variability and uncertainty
distributions for practical risk analyses. Risk Anal 1994;14:713–30.
[22] Helton JC. Uncertainty and sensitivity analysis techniques for use in
performance assessment for radioactive waste disposal. Reliab Eng
Syst Saf 1993;42:327–67.
[23] Mckone TE. Uncertainty and variability in human exposures to soil
contaminants through home-grown food: a Monte Carlo assessment.
Risk Anal 1994;14(4):449–63.
[24] Bogen KT. Methods to approximate joint uncertainty and variability
in risk. Risk Anal 1995;15(3):411–9.
[25] Helton JC, Davis FJ, Johnson JD. Characterization of stochastic
uncertainty in the 1996 performance assessment for the waste
isolation pilot plant. Reliab Eng Syst Saf 2000;69:167–89.
[26] Liu C, Bennett DH, Kastenberg WE, McKone TE, Browne D. A
multimedia, multiple pathway exposure assessment of atrazine: fate,
transport and uncertainty analysis. Reliab Eng Syst Saf 1999;63:
169–84.
[27] Charniak E. Bayesian network without tears. AI Magazine 1991;
12(4):50–63.
[28] Henrion M, Breese JS, Horvitz EJ. Decision analysis and expert
systems. AI Magazine 1991;12(4):64–91.
[29] Pearl J. Russel J. Bayesian networks. UCLA cognitive systems
laboratory, technical report (R-277); November 2000.
[30] Heckerman D, Mamdani A, Wellman MP. Real-world applications of
Bayesian networks. Commun ACM 1995;38(3):24–68.
[31] Heckerman D. Causal independence for probability assessment and
inference using Bayesian networks. Technical report MSR-TR-94-08.
Microsoft Research; March 1994.
[32] Diez FJ. Parameter adjustment in Bayes networks. The generalized
noisy OR-gate. In: Proceedings of the 9th conference on uncertainty in
artificial intelligence, Washington DC, 1993. p. 99–105.
[33] Pearl J. Graphical models for probabilistic and causal reasoning. In:
The computer science and engineering handbook. Boca Raton: CRC
Press Inc; 1997. p. 697–714.
[34] Buntine W. A guide to the literature on learning probabilistic
networks from data. IEEE Trans Knowl Data Eng 1996;8:195–210.
[35] Murphy KP. Dynamic Bayesian networks: representation, inference
and learning. PhD Thesis. UC Berkeley, Computer Science Division;
July 2002.
[36] Jordan MI, Weiss Y. Graphical models: probabilistic inference. 2nd
ed. The handbook of brain theory and neural networks. Cambridge,
MA: MIT Press; 2002.
[37] Guo H. Hsu WH. A survey of algorithms for real-time Bayesian
network inference. AAAI/KDD/UAI-2002 joint workshop on real-
time decision support and diagnosis systems. Edmonton; July 2002.
[38] Dagum P, Luby M. An optimal approximation algorithm for Bayesian
inference. Artif Intell 1997;93:1–27.
[39] Poole D. Average case of a search algorithm for estimating prior and
posterior probabilities in Bayesian networks with extreme probabil-
ities. Proceedings of the 13th IJCAI; 1993. p. 606–12.
[40] Jordan MI, Ghahramani Z, Jaakkola TS, Saul LK. An introduction to
variational methods for graphical models. Machine Learn 1999;37:
183–233.
[41] Koplik CM. Status report on risk assessment for nuclear waste
disposal. NP-1197, interim report EPRI; October 1979.
[42] National Council on Radiation Protection and Measurements.
Screening models for releases of radionuclides to atmosphere, surface
water, and ground, NCRP report No. 123 I; 1996.
[43] Safety assessment of near surface radioactive waste disposal facilities:
model intercomparison using simple hypothetical data (test case 1).
Final report of NSARS, IAEA-TECDOC-846; November 1995.
[44] Galson DA, Swift PN, Anderson DR, Bennett DG, Crawford MB,
Hicks TW, Wilmot RD, et al. Scenario development for the waste
isolation pilot plant compliance certification application. Reliab Eng
Syst Saf 2000;69(1–3):129–49.
[45] Thorne MC. The use of expert opinion in formulating conceptual
models of underground disposal systems and the treatment of
associated bias. Reliab Eng Syst Saf 1993;42:161–80.
[46] Thompson BGJ, Sagar B. The development and application of
integrated procedures for post-closure assessment, based upon Monte
Carlo simulation: the probabilistic systems assessment (PSA)
approach. Reliab Eng Syst Saf 1993;42:125–60.
[47] McGuire RK, Bullen DB, Cook N, Coppersmith KJ, Kemeny J, Long
A, Pearson FJ, et al. Demonstration of risk-based approach to high-
level waste repository evaluation. EPRI NP-7057, Risk Engineering,
Inc.; October 1990.
[48] Kozak MW, Olague NE, Rao RR, McCord JT. Evaluation of a
performance assessment methodology for low-level radioactive waste
disposal facilities—evaluation of modeling approaches. NUREG/CR-
5927, SAND91-2802. vol. 1. Albuquerque: Sandia National Labora-
tories; 1993.
[49] Sheppard ML, Thibault DH. Default soil solid/liquid partition
coefficients, Kds, for four major soil types: a compendium. Health
Phys 1990;59(4):471–8.
[50] Beyeler WE, Hareland WA, Duran FA, Brown TJ, Kalinina E,
Gallegos DP, Davis PA. Residual radioactive contamination from
decommissioning—parameter analysis, draft report for comment.
NUREG/CR-5512, SAND99-2148. vol. 3. Albuquerque: Sandia
National Laboratories; 1999.
[51] Kozak MW, Olague NE, Rao RR, McCord JT. Evaluation of a
performance assessment methodology for low-level radioactive waste
disposal facilities—validation needs. NUREG/CR-5927, SAND91-
2802. vol. 2. Albuquerque: Sandia National Laboratories; 1995.
[52] Siu NO, Kelly DL. Bayesian parameter estimation in probabilistic risk
assessment. Reliab Eng Syst Saf 1998;62:89–116.
[53] Evaluating the reliability of predictions made using environmental
transfer models, Safety Series No. 100, IAEA; 1989.
[54] Berger J. Statistical decision theory and Bayesian analysis. 2nd ed.
Berlin: Springer; 1985.
[55] Carlin P, Louis T. Bayes and empirical Bayes methods for data
analysis. 2nd ed. London: Chapman and Hall/CRC; 2000.
[56] Severe accident risks: an assessment for five US nuclear power plants,
US Nuclear Regulatory Commission, NUREG-1150; 1990.
[57] Pradhan M. Dagum P. Optimal Monte Carlo estimation of belief
network inference. Proceedings of the 12th conference of uncertainty
in AI, Portland OR; August 1996.
C.-J. Lee, K.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 532

This approach has been widely introduced by Hoffman [14]. K. environmental forecasting. a number of factors affect a variation of the risk results. data. Lee. generally known as a probabilistic risk assessment (PRA). the risk from scenarios can be systematically predicted. Since the prediction for such an evolution is extremely difficult. the risk under any of the scenarios should be estimated. Unfortunately. The scenarios will be formed by combining both initiating events and mitigating capabilities. Hofer [15]. and so on. some remarkable studies have been conducted. The future risk from radioactive waste disposal facilities will be categorized into this case. Evolution examples coming from stochastic processes in the environment involve potential changes of weather. we can classify states of systems with two—active and passive. has been used to provide a probability distribution of the end point’s measures in various fields such as nuclear reactor operation. etc. through a lot of operating experience. which includes an adverse impact to the natural environment. a waste disposal facility has to be isolated from natural circumstances. and ecological condition. we are not sure that the performance criteria produce best-estimate values under incomplete analysts’ knowledge bases. The performance criteria used to facilitate comparison with specific conservative regulatory criteria. and experience with similar courses of actions in the past [1]. In addition. Next. The depth of uncertainty depends upon our total state of knowledge. since the release of risk-dominant radionuclides from a disposal site is expected to occur after hundreds or even thousands years. However. or the number of casualty. Therefore. therefore. as denoted by Oberkampf [13]. The results of a deterministic approach are usually judged in view of specific measures known as performance criteria. Therefore. Especially. The subjective probability distribution represents an uncertainty against the impact of such a lack of knowledge or stochastic processes. on the other hand. a knowledge base might not be easily provided. as well as by taking the support of standardized and enhanced evaluation techniques. ‘What can go wrong?’ If we imagine an initial stage of scenario. These risk prediction would require a systematic process for the scenarios or accident sequences. They can be used to interpret the various contributors to the risk with which are compared and ranked [5]. a hazard source itself can be a trigger for specific end point’s measures. Siu [16]. hydraulic profile. exposure dose of humans. the state of system is called ‘active’ if an event can be directly controlled according to its happening simultaneously. geological form. with natural hazards being considered. [6–8]. and known as so-called ‘divide and conquer’ concept. and Helton [17] etc. In terms of treating scenarios.J. Furthermore.516 C. With a view to preventing exposure or unacceptable risk to human health. the state of system is called ‘passive’ if any event cannot be directly controlled according to its happening simultaneously. a probabilistic approach in quantifying the risk. space accident prediction.-J. it is just necessary to demonstrate that waste disposal treatment is reliable and that both man and environment can be adequately protected. a temporal evolution by change in both external and internal environmental conditions becomes a fundamental factor in estimating the risk. Various hazard sources exist in diverse natural or man-made systems. the scenario in the risk analysis can be defined as a propagating feature of specific initiating event which can go to a wide range of undesirable consequences [18]. usually there are no direct methods of assuring the credibility of the evolution. Since the system performance in the passive state is very sensitive depending on a variety of scenarios. The operating risk from nuclear power plants is categorized into this case. A lot of risk analyses have been performed to actually estimate a risk profile from both uncertain future states of hazard sources and undesirable scenarios. which focussed on uncertainty propagation with emphasis on the separation principle for both uncertainty types in every variables. which answers to the question. Diverse approaches for categorizing uncertainty have set forth last two decades. these measures with limiting (or regulatory) indicators which affect about humans or environment. the variables in a model domain (called here as ‘domain variables’). Lee / Reliability Engineering and System Safety 91 (2006) 515–532 frequency. a comprehensive risk prediction needs a complete listing of scenarios and estimation of their frequencies. we had to consider a system-specific feature originating from a very broad range of aleatory and epistemic uncertainties. The issue of numerical solution error has been generally ignored in risk analysis. In this case. long periods of evolution in natural environment facing diverse stochastic processes have influence on the uncertainty [9]. On the other hand. whose risk has been easily assessed using a traditional PRA technique. The risk assessment of a disposal site includes specific models for addressing the major pathways of radionuclides ill-affected human beings. Due to both lack of the knowledge and stochastic features. We can identify. as well as rankings of occurrence probability of the scenarios. The risk profile coming from combination of likelihood and consequences of future transport and exposure depends . the potential risks to future generations may be subjected to complex uncertainty propagation assessment via PRA techniques. Most risk analysts [10–12] agree that all the variables used in the risk analysis models contain both epistemic and aleatory uncertainty because of their unique nature. upon all the evidence.. besides an issue of numerical solution error due to the mathematical modelling approximations and temporal discretization in simulations. In estimating. which is also beyond our concerns. The uncertainty must be practically treated in the estimation of the end point’s measures. whose risk then scarcely has assessed via the traditional PRA technique. In this case. First. Recently.

It is then obvious that current methodology of uncertainty analysis of a waste disposal site should be revisited with these diverse considerations.g. a system analyst may want knowing whether or not it is worthwhile to reinforce engineering protection features in order to defence a random hazard (e. including medical diagnosis (e. forecasting. they have been popularly utilized [28]. language understanding.C. Subjective probability distributions on the numerical results are given by uncertainty analysis and are used to represent the confidence level where a risk criterion is being met. then that reduces the likelihood of causes of the others. In order to consider the effects owing to the evolution of environmental conditions of waste disposal facilities.g. Furthermore. These findings notice us not to lose sight of how dominating domain variables contribute to the uncertainty of the risk in relation with diverse complex phenomena and/or scenarios. This section provides an overview of scenario categorization. heuristic search (e.e. 2. The uncertainties arisen from different evolutions should be adequately propagated throughout the analysis.1. † Estimating the probability of the query variables by any inference algorithm. Interpreting the significance of results of a PRA in view of the uncertainties is also important if the PRA results are applied to make meaningful risk-informed decisions. so called it as ‘explaining away’. Microsoft troubleshooter). As these reasoning processes with useful evidence are made. While it is important to characterize the overall uncertainty. After a survey of relevant studies [9. the consequence. . map learning.g. For example. The reasoning processes can be operated by propagating probabilistic evidence in any directions. Overview Bayesian network also known as belief networks. an initiating event). this paper proposes a quantitative assessment framework combining an inference process of Bayesian network with a traditional risk analysis. were not practically linked to the PRA. The last term of ‘sets of triples’. where network means a graphical model.e. A deterministic model of radioactive waste disposal. † Learning the structure or defining parameters of the query variables.e. the first and the second terms used in ‘sets of triples’ of risk definition are delineated in detail in this paper. may be easily identified using deterministic evaluation models. abduction. hydrological.e. The remainder of this paper consists of four sections. Bayesian networks have been applied to diverse technical or societal problems. It is required to explicitly consider distribution functions of random variables that define both aleatory and epistemic uncertainty associated with each variable in a scenario evaluation.e. In order to build overall reasoning processes for Bayesian networks. as to the evolution modelling for stochastic domain variables under adversely-environmental conditions in waste repositories.-J. K. The last section summarizes and concludes the study. answers to a variety of queries are required. and ‘explaining away’ are made: for example. Therefore. if a forward direction of reasoning from a cause (i. as well as characteristics of the domain variables. it is equally worth understanding well what factors or variables mainly propagate the uncertainty [19]. a descendant) is occurred. The probabilistic inference algorithm employed is explained in Section 3 in detail. The use of Bayesian networks helps to answer such the queries even when no experiment about the effects of reinforced protection is available. and causal networks. Lee. only epistemic uncertainty may be treated [20–24]. In some variables. including an explanation of the dependency between scenarios and domain variables is presented in Section 4. QMR [Quick Medical Reference]). in terms of providing decision analysis and expert systems using artificial intelligence (AI). Since Bayesian network constitutes a model of reasoning process in a system. we found most scenarios developed for a disposal site. In addition. In addition. In utilizing Bayesian networks. PathFinder. and ecological aspects of the site with all the scenarios.20. where various forms of reasoning [29] such as prediction. Markov random fields. sources of the uncertainty at future states can be reduced by pertinently setting up dependency relationships interrelating geological. pattern classification. without considerations on above-mentioned dependency treatment required. scenario analyses up to the present have not given a specific and realistic approach. a predecessor) to an effect (i. i. This section also provides some illustrative results and discussion.J. is a new concept for reasoning complex uncertain problems. a descendant) to a cause (i. as well as epistemic uncertainty. Framework development for the risk assessment using Bayesian networks 2.25. it is general to communicate qualitative assumptions about cause-effect relationships and to derive causal inferences from a combination of diverse assumptions. if a backward direction of reasoning from an effect (i. then prediction is formed. A major portion of Section 4 is devoted to the development of the evaluation models in relation with the information given in Section 3. on the other hand. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 517 on aleatory uncertainty.26]. Bayesian network can provide a basis to perform uncertainty propagation of query variables. following three steps are generally engaged: † Structuring the graphical models for causality representation. if we have more than two causes and one of those causes is found. [27]. then abduction is formed. etc. probabilistic networks.18. Section 2 provides the methodology framework for using Bayesian network in the study. a predecessor) is occurred. real-time monitoring.

Moreover. Fig. after assigning a probability (or its distribution) to a node conditional on that node’s predecessors. (d) module 2: a detailed DAG between V and R with n-elements of X. Xk is an actual system performance indicator corresponding to a domain variable k under the dependency of initiating events). where M denotes an inhibitory indicator (or mechanism) that prevents any cause E from giving an effect R and thus has two states— true (on) and false (off). the joint probability (or its distribution) over the whole network is determined [26]. we can consider a possible interaction to hinder a direct impact from insignificant causes. construction of Bayesian networks helps to capture engineer’s intuitive understanding of complex systems or problems. their effects). 1(b). and arcs.e. they are far from popular in the nuclear PRA. Accordingly. which will be explained in the following sections. since a practical implementation of Bayesian networks requires complex modelling of the inter-relationships between various initiating events and engineered mitigating features which diversify the risk profile. A random variable V denotes a trigger (a kind of filtering interaction) reflecting the fact that the input E cannot always produce the output R. In the PRA adopting Bayesian networks. As a last step for figuring G. arcs are drawn from a cause to an effect.e. a whole reasoning process is deemed to be a complex task in itself. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 As explained by Heckerman [30]. it seems that we Fig. (c) module 1: a detailed DAG between E and V using a concept of M. Incomplete DAGs encode expert knowledge in the form of missing arrows. subsequent computations are pursued by symbolic manipulation of probabilistic expressions. However. then we have to make a further expansion in G. K. not in a full reasoning process of Bayesian networks.e. A structure of Bayesian network is generally represented by an annotated directed acyclic graph (DAG) G. linking lines. it is kept in mind that an initiating event simply means a conditional initiation for any occurrence of the scenarios. If we want resolving the causality (i. we propose a conceptual framework to apply Bayesian networks to the nuclear PRA. The causality represented in four decomposition steps: (a) a basic DAG simply representing a cause and an effect. 1. With this desire.J. i. Unfortunately. the structure in G is divided into two modules providing with decomposing paths—‘from E to V’ and ‘from V to R’. Lee. Fig. for describing the performance of a system). As a first step for representing G. realize a probabilistic and causal dependence between random variables that they connect. as addressed in the above. E denotes an occurrence of any initiating event while R denotes a response of given problem. Bayesian networks have been explicitly applied as a popular graphical modelling language to represent the causality under uncertain relational situations in a real world. Presented with a step-by-step decomposition of the network structure for solving a problem (e. In general. the DAGs are regarded as a substantial language for communicating conditional independence assumptions. Representation of the cause-effect dependency In this section. 1(b). a trigger) in the decomposition of Fig. cannot provide any practical reasoning for the problem only with this structure. after considering the causality from the initiating events to the engineered features. we can figure a problem to a simple structure as outlined in Fig. and for representing qualitative causal inferences [25]. 1(b). relationship of cause and effect) in detail. we . shown with dotted boxes in Fig. (Note: All the symbols in a node represent random variables. In other words. we suggest a special technique for applying the general reasoning concept of Bayesian networks to the PRA of a nuclear facility. causes) to consequences (i.e. V denotes a trigger or a filtering interaction which reflects the fact that the cause E cannot always produce the effect R.g. we are mainly interested in the direct representation of the real circumstance.e. Usually. we can represent this idea by expanding the upper module in Fig. M denotes an inhibitory indicator or mechanism that prevents any cause E from giving an effect R. 1(a). 1 shows an overall process for making DAG G which finally gives a causality between random variables of E denoting a cause (input like an occurrence of any initiating event) and R denoting an effect (output like a response of given problem). However. Once the network is configured. Therefore. Since our approach mainly focuses on connecting scenarios (i.518 C. 2. where some nodes denote variables of interest. Since the DAGs represent high dimensional uncertainty distributions. (b) more extended DAG consisting of 2 modules. The trigger V remains in a hidden state. 1(c) shows a result using AND gate and NOT operator. by introducing the upper module (i.2.-J.

a multiply-connected network with four query nodes is outlined in Fig. Vs. i. if this assumption is accepted. PjZP(XkjVj only). to denote that variable x is in state k. for example. In the general graphical representation it is noted that the circle nodes denote continuous random variables. Ei in a root node represents the happening of an initiating event i which affects physical domain variables of given system. Direct inference of the network structure shown in Fig. K.3. we may take up the concept of a noisy-OR gate [28. For instance. state of a random variable. we can usually configure the network structure. . furthermore. 1(d) is given only that input is present. According to this consideration. a clear node indicates a hidden state and a shaded represents an observed.e. Bayesian networks take advantage of expressing a graphical description of the dependencies between random variables.C. Bayesian network needs a corresponding database and a set of explicit assumptions about our prior probabilistic knowledge of a domain. It is also noted that when we make a network structure. in some variables there is only one connection. 2. i. and Xs are configured as nodes. For the purpose of showing an illustration. Modelling considerations for building a network structure In a graphical structure adopted in Bayesian networks causality plays a great role. squares denote discrete. The logic of the structure provided in Fig. we can easily assume that the causal mechanism in the connection path from Vj to Xk is mutually independent. i. To do this. In describing Bayesian networks. 2 is not easy to solve and requires numerous probability distributions because a causal complexity exists between both ancestors and descendants. when identifying direct physical connections we should consult a graphical structure. As denoted in Fig.-J. where the expression P(XkjVj only) means a conditional probability of kth X node given only Vj node. ancestors before descendants. but in others. Using this variable ordering and a concept of conditional independencies. we also introduce some random variables which can give a system response R following after the instantiation (it commonly means that a random variable becomes a true state) of triggering. 1(c). to express the causality. By way of interfacing structure model to data. We also write xZk. With this modelling concept. therefore. It is noted that almost all the observed things can be represented with the evidence.31]. if an individual trigger (Vj following Ei) occurs together with its corresponding inhibitory indicator M. the instantiation of their descendant (Xk) in Fig.J. we can fully represent the causal relationships from the occurrence of initiating events to the system response. four Es. 2 shows a fully-specified Bayesian network corresponding with an actual inference problem. The network structure has a characteristic to encode conditional independencies among random variables. As far as the response R from an instantiation of Xk is concerned. We always want making a graphical structure predict responses and infer causes from evidence. Therefore. for example.e. Fig. Lee. In addition. we call it a state of X. 1(d). The states of unshaded nodes in the figures. therefore. Three R nodes are also provided for the states of the end point’s measure). 2. the nodes must always be numbered in topological order.e. 2 is explained in detail. For the purpose of simplifying the inference after structuring of the model. In the figure. because the process which prevents the input signal from being sufficient is independent of the others. 3 or 4 connections are possible. Practically. Let’s take an example for expressing this case. Absence of a direct link between any E and R. we may define it as an adverse impact on the system or any other consequential measure of the system. When we observe every state of a variable in any set X. we use a lower-case letter to represent any single state of a random variable and an upper-case letter as a set of Fig. available information about the states of nodes. Therefore. the noisy-OR gate model can be selected. all kinds of evidences can be propagated within the structure. a missing edge explicitly means that there is an independent relationship between random variables. we attempted to represent a more general structure model for application of a nuclear PRA. diverse connections from any jth node of V (Vj) to any kth node of X (Xk) exist depending on their causal characteristics. reveals that there is no direct influence of initiating events on the final system status. As a descendant node is affected by different trigger nodes. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 519 have to fully represent the lower module of Fig. With the above-mentioned decomposition steps. in the selected graphical models. let a random variable Xk an actual system performance indicator corresponding to a domain variable k under the dependency of initiating events. specified by a specific function incorporating independencies between nodes relevant to the converging path. 1(b) as shown in Fig. cannot be observed in an actual circumstance. respectively. 2. A Bayesian network with fully treating causal relationships in PRA problems (As an illustrative purpose.

Thus. It is well known that the principle of inclusion–exclusion can compensate the combined belief of outcome by the degree of belief that is common to a different set of causes.. Vn Þ Z 1 K ð1 K pV1 Þð1 K pV2 Þ. they can have two simple states—true and false. Actually.ð1 K pVn Þ (1) where the left expression specifies the probability distribution of Xk conditional on its parents of V1. the notion of probability is known for the best way to actualize uncertain beliefs. we often encounter situations where an effect has many potential causes. . since it may be reasonable that multiplyoccurring initiating events are relatively rare under the time span considered. 3 might be a Poissondistributed probability with parameters depending on their occurrence frequencies.n. Suppose n parents X1. with adopting assumption about mutually exclusive situation for preventing dual initiating events.. only one end effect is also proposed for the purpose of illustrative application. pVj means the input probability of the parents Vj to Xk given an arbitrary number j. Vn. However. According to the characteristic of Bayesian networks. Given any combination of the input. K. we can simplify the structure of Bayesian networks with one root node.. As a result. X2.J. in complex networks. and allows evidence propagation in time proportional to the number of parents of each variable [32].. each input variable has a probability of being sufficient to cause an output variable. a joint probability distribution over the set can be represented by the product of the individual probability distributions over each Yi conditional on its parents. however. 3. if we practically consider the actual occurrence of initiating events. while it is exponentially increased (2n with binary state) in the application of other general gate types. we must specify the probability distribution of R conditional on every state of its parents. parents’ inputs should be carefully treated so as not to induce an impractical problem.1. the probability of a descendant is expressed with a form of minimal cut upper bound (using a principle of inclusion–exclusion) as PrðXk jV1 . In other words.2. Given the response R shown in Fig. a node can be gone to diverse states representing continuous values. . (1) to an equation with first order probability. it is usual that a lot of variables such as a decision variable are inherently discrete. this is not fit for a general case. the noisy-OR gate simplifies knowledge acquisition. and its role in Bayesian network is to provide a belief basis about qualitative encoding of a network structure. Let’s take another example for continuous random variable. V2 . Approaches for a probabilistic inference in Bayesian network As denoted by Henrion [28].. 3. the causal independence relationship in a graphical model informs us how a global high-dimensional probability distribution over all the variables can be factored into a product of local probability distributions over lower dimensional spaces. Furthermore.. In an ideal case. the above reduced equation can be further reduced to a single term but disjunctive element on the possibility of each initiating event. is quite a demanding task.n and given a specified DAG G in a Bayesian network. we have to specify mn probability distributions for representing this causality. Xn bearing on a descendant R can have m states. However. A simplified DAG of Bayesian network for considering non-multiple initiating events in PRA (For the purpose of showing an illustration. jZ1. Representation of variable states Generally. Y2. iZ1. As discussed by Heckerman [31]. If a set YZ{Y1. the value of root nodes given in Fig.520 C. respectively. In solving various real problems. random variables have been regarded as a matter of state. Fig. 3. saves memory space. The main advantage of the noisy-OR gate model is that the number of parameters in the variables of interest may be reduced proportionally to the number of causes (n). n nodes for random variable Xs and one node for a state of the endpoint’s measure are provided). Lee / Reliability Engineering and System Safety 91 (2006) 515–532 In the noisy-OR gate model.. most values of the probability would be obtained by engineering judgment rather than by insufficient empirical data. . V2. Lee. Even though a variable representing time-dependent situations may be inherently continuous.-J. we can reduce Eq. Overlooking based on our available knowledge base.. As a consequence.. it can be also treated as discrete in the simulation for the convenience of computation. expressing our beliefs for the random variables’ states in a term of probability. it is well acknowledged that.. In Bayesian networks.Yn} is composed of n nodes of discrete random variable Yi. 3.2. Any random variable is represented by either a continuous scalar or a set of discrete values. n denotes a maximum number of V node. When modelling a real world.

Therefore.x. Paimi g where mi is the number of parents of Yi. Therefore. Paimi Z fimi .C. we need specifying a complete joint probability distribution (JPD) over all the random variables. if nothing is known about R. x. since we let a fixed structure in the study.qv Þ x Prðrjx. Lee. we have the conditional probability as P Prðe.2.-J. If we expand the Eq. which implicitly provides a well-known definition of ‘d-separation (directional separation)’ [33] in Bayesian network model. coming into existence of any states of random variables is termed an instantiation. and assign their values with a corresponding probability set fi Z ffi1 . . which is a process for structure learning. The expression of Eq. fimi g. observing the effect(s) of a given model and trying to estimate the causes is termed a diagnostic reasoning.qxk Þ k (4) As explained previously.qr Þ Prðxk jvj .qv Þ k Prðxk jvj . . In the case of Fig. v. a parameter fitting problem is exposed and it can be resolved with available reasoning process considered [34]. On the other hand. rÞ In the above equation. Conditional probability estimation In Bayesian networks. we can obtain an explicit form of the conditional probability: PrðxjrÞ P P Q Prðrjx.v. 3. Bayesian network defines a probability distribution over the possible assignments of values to all the variables as PrðYjGÞ Z n Y iZ1 because the dependence between Xi and Xj is conditional on the certainty of V.qr Þ k Prðxk jvj . the JPD can be expressed. The probability of each instantiation can be simply calculated as the product of the conditional probabilities of its predecessors.. 3. if we will apply Fig.v. the state values of the descendant node are wholly determined by the values of the parent nodes. x.v : (3) PrðrÞ e. Full JPD over a given domain can be provided with a combination of conditional independencies in network structure and local probability models. There are two kinds of probabilistic inferences. if n-element Xs are identified in network. Under the circumstances of diverging connections in Fig. in case of converging connections where any of the paths from parent nodes Xs to descendant node R also remain unchanged. where isj. However. v.x. lowercase letters are commonly used here to represent particular instantiations of variables. and r determined from the overall network structure. (4) denotes the parameter of each query node...rÞ ZPrðei jqe ÞPrðvj jei . it is desirable to efficiently utilize an alternative approach with the consideration of conditional independence assumption. 3 as an example. set of x. (4). (3) by applying Eq. where any of the paths from one parent node V to descendant nodes Xs remain unchanged. in order to obtain an adequate solution on the reasoning with parameter fitting. observing the root(s) of a given model and trying to forecast the effect is termed a predictive reasoning. In a learning process with fixed graphical structure. then R’s parents (Xs) are being interindependent so that no information can be transmitted between them.vj . (2) implies that.r) is the joint probability distribution of a set of random variables e.qv ÞPrðrjx. From the rules of probability. in Bayesian networks it is possible to use a learning process supported by data on the structure and parameters of variables. including specific risk attributes. based on the chain rule of probability: Y Prðei . Pr(e.3.x Prðe.J. (3) can be expanded in a way that reflects the structure of Bayesian network itself [29]. In reality. A bold lowercase letter x represents a set of instantiations of X elements.qxk Þ P P Q ZP e Prðei jqe Þ v Prðvj jei . Each q shown in right side terms of Eq. The probabilistic inference focuses on estimating the posterior distributions of random variables when relevant evidences are observed. it is necessary to consider an appropriate probabilistic inference algorithm [35]. 3. rÞ Prðx.qr Þ e Prðei jqe Þ v Prðvj jei . Lee / Reliability Engineering and System Safety 91 (2006) 515–532 521 If we denote the parents’ set of any ith variable Yi in G by Pai Z fPai1 . rÞ PrðxjrÞ Z Z P e. The probabilistic inference should give PrðYi Z yi jPai1 Z fi1 . which is generally referred to as a probabilistic inference. 3. methodology adoption for structure learning is out of scope. On the other hand. for example. v. Simulation algorithm for a probabilistic inference In performing a traditional PRA. Many structure identification methods or learning algorithms are also used to optimize a structure. given the parents’ sets and G.qxk Þ (5) In functional and logical determinations. each variable Yi is conditionally independent of the other parents. It is also noted that all the classical and logical operators can be brought when defining unknown parameters in Bayesian networks. K. information can be transmitted from any Xk node to the other Xk0 node through the parent V unless V is instantiated. A key point of practical reasoning is to take the evidence propagation by computing the answer probabilistically for particular queries about the domain. . GÞ: (2) Eq. a variety of sensitivity analysis are generally used to examine the importance of various parameters and assumptions for risk modelling. since this distribution may not be feasible to estimate. We can state that the variables of Xk are conditionally independent of V . Both forms of reasoning can be selected in Bayesian networks.

e) and a weight distribution u(y. it weights each sample by the product of the conditional probabilities of the evidence variables. and Markov Chain Monte Carlo algorithm (sometimes Gibbs sampling) are popularly used. in this case Y to y. eÞ Z Y Ei2E Pr½Ei jPaðEi ފYZy. This updated algorithm is adopted in the study. They have realized utilizing stochastic simulation-based. given its parents. To provide an inference probability Pr(YZyjEZe) which is an extended expression of similar term as denoted in Eq. Furthermore. It is remarkable that a probabilistic inference problem can be easily solved if we have a JPD of the random variables involved. respectively. likelihood weighting algorithm. have appeared as an alternative to the overall algorithms in Bayesian networks with extreme probabilities [39]. use a special technique. (7) where the expressions are conditioned. approximate probabilistic inferences have been widely surveyed by many researchers. it fixes the values on evidence variables and samples only query variables left. which exploits the law of large numbers to approximate large sums of random variables by their means [40]. also. i. The accuracy of stochastic simulation-based algorithms generally depends on the size of samples. For applying to Bayesian networks with extreme conditional probabilities. The weights in a network make local assertions about the relationship between neighbouring nodes. . in the likelihood weighting. Variational-based inference methods. in order to use the algorithm. However. That is. EZe) into a path probability distribution p(y. rejection sampling. Based on the consideration for characteristics of the study. and E to e. meanwhile. It has been frequently provided to query problems within the specified bounds. and so on. Disadvantages of the Monte Carlo approach are that it can be slow to converge and be hard to diagnose its convergence. (6) and uðy. as they converge very slowly if there are extreme probabilities in the conditional probability tables (CPTs) [37]. since general marginalization using joint probabilities requires exponential time in the probabilistic inference. The following explicit advantages have given to the algorithm: † It can be applicable to any type of nodes.EZe) and Pr(EZe). it is noted that the algorithm displays much faster convergence time than a simple stochastic simulation. In the following discussions.EZe .522 C. As a special case. discrete or continuous. Next. and a final ratio is used to compute the probability. although they depend on the characteristics given a problem. these algorithms have a little disadvantage. let E denote a set of observed node. Also.e) that is equal to ‘true’ if YZy instantiates the node X to x. irrespective of the structure of Bayesian network. One of the categories in approximate inference algorithms. A wide variety of approximate inference algorithms for Bayesian networks share many common traits. K.EZe .J. Finding a sufficient and effective inference algorithm in static Bayesian networks is desirable with a computing efficiency of simulation considered. we finally selected the best out of a lot of inference algorithms. to denote an instantiation of their arguments. Instead of randomly choosing a value. Y the nodes not contained in E. Yi\E denotes ith arbitrary node which is contained in a set Y but not included in the evidence subset E. eÞ Z Y Yi2Y Pr½Yi jPaðYi nEފYZy. the likelihood weighting algorithm utilizes a technique to decompose the full joint probability Pr(YZy. stochastic simulation-based algorithms such as direct sampling. However. (3). Various exact inference algorithms have been also developed for the probabilistic inference [37]. However. likelihood weighting algorithm. with a relative error 3. In order to make these approximations. each sample has a weight. Lee. we must generate the relative approximation of two probabilities Pr(YZy. and variationalbased techniques [38]. for example Q Pr½Ei jPaðEi ފYZy. but is still slow under extremely small conditional probability values. which search for high probability configurations through a space of possible values. Advantage of the Monte Carlo algorithms includes their simplicity of implementation and theoretical guarantees of convergence. it may be reasonable to apply a simple simulation-based algorithm. since an exact inference solution in real-world networks seems to be impractical. let’s define a new binary random variable c(y. searchbased approximation algorithms.e. For instance.e) presented by pðy. Inference algorithms turn these local assertions into global assertions about the relationships between nodes. A lot of efficient methodology or algorithms for evidence propagation in Bayesian networks have been proposed [36]. † Each event generated is consistent with an evidence.EZe .-J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 an adequate answer to a query in a reasonable amount of time. † It avoids the inefficiency of rejection sampling by generating only events that are consistent with the evidence. we had to consider more efficient methods. the algorithm takes values with conditional probability as a likelihood weight. Dagum [38] proposed a bounded-variance likelihood weighting algorithm to provide efficient probabilistic inference in polynomial time. To meet the above demands. search-based. emerged as a variation of stochastic logic simulation for the purpose of giving an algorithm speedup.

However. Also. eÞ Z False. release pathway models should be assembled so that they can represent radionuclide transport from the repository to the environment and ultimately to the human being. A basic mechanism governing the release in underground media is the mass transfer in a porous medium where pore water is a main agent for radionuclide transfer. From a geochemical standpoint. Near surface disposal is one of the preferred options for comparatively large volumes of low and intermediate-level wastes.C.eÞjcðy. and also for wastes arising from radionuclide applications in hospitals and research establishments. if X Z x in y.e) becomes X E½c$uŠ Z cðy. it may be unable to identify the key exposure pathway because both high likelihood and a great consequence cannot generally come together. In general. In other words. Radionuclide release characteristics A simple. which can ultimately go to the biosphere. particularly in association with a radiological waste disposal facility. 4. if we sample the distribution p(y. The simulation of the algorithm uses a random number generator to uniformly select values for given subset of evidence nodes and subsequently propagates them to pick values for the other query nodes. hypothetic site with an engineered vault design is adopted to the environmental transfer model. The engineered barriers. Model inputs Release pathways strongly depend on both a location and design of a repository. E Z eŠ.e)u(y.e) in sufficient counts. generally not below 50 m in depth at near surface disposal repository. which arise during nuclear power plant operation. the rate how much radionuclides are released into the media where they are transported. Key exposure pathways can be defined with the consideration of both a likelihood and significance. these considerations are important in the determination of a delay time of release. Lee.e. and backfills within the vault. a normal delay of the source terms will occur because a delay from breach of waste container and a relatively-slow leaching are usually expected. geosphere) such as dispersion. retention by geologic media. with respect to the risk concept.J. as well as waste containers and surrounding materials.eÞpðy. Wastes are normally disposed at or near the ground surface. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 523 and is equal to ‘false’ otherwise: ( True. cðy.eÞpðy.eÞZFalse ZPr½E ZeŠ: determined. and local retardation of nuclide movement on container debris. It is noted that the likelihood of a peak dose can be also represented as an end point’s measure resulting from the scenario evaluation. if X sx in y: Then the expectation of the random variable c(y.e) and u(y. Also the expectation of weight distribution is represented by X X E½uŠ Z pðy. Statistics are then maintained by the corresponding values that the nodes take out. and inventory uses by a critical human group are considered in the model of exposure pathways. Once the repository has been . in spite of the engineered features.eÞZTrue y2U X y2U y2U C cðy. and this process gives a desired answer. The end point’s measure of interest can be simply defined as a peak dose to a person drinking from a biosphere well. After an Therefore. Various processes in the underground environment (i.1. eÞuðy.e)u(y. a transport of radionuclides through groundwater in the geological formation. In obtaining the simulation output.1. Therefore. both natural and engineered systems form a series of multiple and potentially independent barriers against the release of waste [41]. K.e) can converge to a desired inference probability. Source terms describe the release of radionuclides from the repository to the pathways through the geosphere and far field. Especially. can also contribute to the repository performance. and holdup in sub-layer aquifer usually relate to the release amount of radioactive materials to accessible environment (i.1.eÞuðy. 4.eÞuðy. backfill materials. eÞpðy. the release timing is greatly affected by the assumptions concerning occurrence time of breach of the engineered barriers. An example application to an environmental transfer model of the nuclear facility The application involves an example for estimating the risk in a nuclear facility. the disposal system is a good domain example for this study because its performance potentially depends on a variety of scenarios.-J. Altogether. a key feature seems to make a practical estimation of the query variable whose parameters are an unknown function even under the simulation inputs. biosphere).eÞ Z cðy. 4. A normal exposure pathway in the post-closure repository is generally defined by the non-accidental leaching of radionuclides from the repository to human beings.eÞuðy. there are several considerations in defining the source terms. or other engineered barriers etc. the limiting ratio of c(y. Several key factors such as delay times from different engineered barriers. eÞ Z Pr½Y Z y.e. y2U where a sample space U for Y contains all the instantiations of YZy. the rate and extent of container degradation which determines when and how much waste is exposed to the leaching and the other release mechanisms.eÞjcðy. unavoidable uncertainties remain unsolved in the time when waste packages are breached and in the way how the radionuclides are released.

Any scenarios leading to similar consequences can be grouped into the same category of AES.i $lL T0 where initial source term inventory (Bq) maximum leaching period (y) leach rate of the parent nuclide (yK1) consumption of drinking water. we can calculate the screening dose rate Di. taken to be 800 L yK1 V dilution volume. AESs are defined as unusual happenings that can cause significant alteration or momentary changes of underground geochemical. Only a groundwater ingestion pathway is considered in this paper. inhalation. all the soil layers are composed of saturated zone.e. flow of the groundwater system.524 C. and soil and vegetable ingestion pathways are not addressed because any given pathway model has no problem for fulfilling the objective of the study. Sometimes aggregation of events and corresponding special situations can go into scenarios. we get a lot of system information. annual committed effective dose of radionuclide i as: Di Z M0. (9) are the infiltration rate and the retardation coefficients for each radionuclide of interest.i ingestion dose factor for radionuclide i (Sv BqK1) Xi decay correction factor (dimensionless) Basic site data are adopted from the study conducted with the IAEA research program on the safety assessment of near surface radioactive waste disposal facilities [43]. Therefore. According to the NCRP model. the groundwater system becomes an important part of pathway for radionuclide transfer. the length of the facility becomes perpendicular to the direction of groundwater flow.J. Scenario classification The combination of events. The model gives a screening technique that can be employed to demonstrate compliance with environmental standards or other administratively-set reference levels against releases of radionuclides to the biosphere. and processes causing diverse natural phenomena can be taken as a scenario. which was prepared by the National Council on Radiation Protection and Measurements (NCRP) [42].e. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 elapse of a delay period. significant initiating events and concurrent responses resulting in adverse disturbance of the systems concerned can go to the major AESs. the repository is located in a stable geological formation with an approximate depth of 3 m.i V iZ0 i (10) (8) where R I H n retardation coefficient (dimensionless) groundwater infiltration rate (m yK1) thickness of the layer of buried soil (m) soil porosity (dimensionless). The degree of saturation of medium. the average yearly fraction of the initial inventory of the parent radionuclide X0 left from time tdelay to tdelayCTav is derived as X0 Z where lL leach rate (yK1) lr radiological decay constant of parent (yK1) 0 Tav averaging time tdelay delay time for release (y) The leach rate defined in the NCRP model is lL Z I RHn (9) ½1 K eKðlLCl0 ÞTav ŠeKtdelay l0 Tav ðlL C lr Þ 0 r r the parent nuclide over whole lifetime is also provided in Ref.1. and chemical properties of radionuclides may be key uncertain elements in determining the release rate. it may be enough to classify the scenarios with four major AES M0. i. Although this adoption would intentionally limit the work scope. 4. the geosphere is also characterized as an unconfined sand aquifer with defined properties. [42]. Thus. With consideration of the screening factor for the groundwater ingestion pathway. 4. hydrological. K. features. Since the main objective of the NCRP model is to give a simple means in predicting the results for screening analyses. Referring to previous scenario studies [44.3. average fractions of each progeny appeared from .45]. total volume of water in the aquifer (L yK1) DFing. disturbance means an abnormal state departed from the normal state.-J. which implicitly means the properties of unsaturated soil layer such as partition coefficient or bulk density are the same as those of saturated soil layer. For nuclides with significant progeny. The considerable domain variables in Eq. We introduce the concept of ‘altered evolution scenarios (AESs)’ here as one of extreme and special types of the scenarios. yet there is sufficient realism to identify the feasibility of modelling approach suggested in this paper. with only one time burial assumed.i T0 lL UDW N UDW X X $DFing. The others such as direct irradiation. and/or mechanical properties which are directly linked with the system performance.2. In this paper. i. Deterministic pathway model We adopted a deterministic model in the pathway analysis. For convenience’ sake. conservative approaches and parameters with effective dose factors for 826 radionuclides were proposed in many environmental transport pathways. Lee. The screening dose assessment gives a defensible basis for making regulatory decisions because resulting dose is likely to be overestimated rather than underestimated.1.

causing an increase in the release quantity of radionuclides. It is noted that large faults causing a direct release of radionuclides are expected to be improbable because of the plastic behaviour and self-sealing capabilities of clay. The primary hazard considered is the rupture of waste canisters induced by the faults. also a water budget can be changed.e. which can alter the amount of water available for infiltration.e. A lot of phenomena such as freshwater sediment transport and deposition. natural thermal effects. ultimately the soil partition coefficients. Observing many historical surface fault ruptures. a biotic intrusion is considerably increased. 4. the AESs has a stochastic behaviour.1. ‘Fixed but unknown’ distributions represent the epistemic uncertainty about characteristics at a particular site unless site-specific information is available for the corresponding parameters. The phenomena are combined with the normal evolution pathway creating a lumped scenario. Adverse effects of this case result in a considerable decrease in the dilution volume. Lee. strongly influence two variables of the system. even though its likelihood depends on root depths. stream. We can include a human intrusion scenario in this AES. i. reported available data must be implicitly interpreted. (2) Climatological AES. The soil/liquid partition coefficients of radionuclides can be derived from soil-to-plant concentration ratios for specific plants and types of soil. The extreme water consumption by floras can also cause to reduce the width of a water table or an aquifer. Adopted values for the parameters depend on assigned subjective distributions.4. as given in Table 2. which becomes a considerable increase or decrease in the aquifer characteristic. and processes for the classification of AESs Geological AES (AES-1) Seismicity Fault activation Fracturing Regional Tectonics (Human intrusion) Climatological AES (AES-2) Precipitation Glaciation Deep weathering Transient greenhouse-gas induced warming Surface hydrological AES (AES-3) River flow and lake level changes Site flooding Recharge to groundwater Freshwater sediment transport and deposition River. faults) have been regarded to be critical in the geological stochastic analysis. and may depend on a time-dependent characteristic [46]. as well as water table changes. we need determining how much the infiltration rate or an aquifer flow can change from a nominal value due to the AES. Progression toward trees. This AES may greatly influence the groundwater recharge and discharge. Another possible consequence will be the variation of aquifer flow. for instance.-J.50]. channel erosion (downcutting) Natural thermal effects Ecological AES (AES-4) 525 Plant uptake/evolution Animal uptake/evolution Uptake by deep rooting species Microbial interactions Effects of terrestrial ecological development on hydrology categories. The quantity changes of a water table and/or an aquifer can also occur resulting from the disturbances of local surface hydrological behaviour due to this AES. In general. He also notes that changes in ground stress following an earthquake can induce pore pressure changes and groundwater flow.C. Therefore. It is well acknowledged that the infiltration rate is highly dependent upon the weather. and site flooding have the potential to change regional hydrological properties over time scales of hundreds years. Earthquake or tectonic displacements (i. McGuire [47] indicates that often the mapview width of the fault zone during an earthquake is not restricted to a narrow zone along the primary fault but a fault zone that is wide up to kilometers. The major consequences in this case will probably be a reduction of infiltration rate or aquifer flow. therefore. The AES representing a climatic variation can cause medium level perturbations in the geosphere. as outlined in Table 1. It is assumed that the geological AES can. Uncertainty in a variable is also treated as ‘aleatory’ when the distribution is determined from a random process. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 Table 1 Example compilation of events. we assume that the AES indirectly influence the change of plant-to-soil concentration ratios. Progression may tend to either increase or decrease the infiltration [48]. (1) Geological AES. (3) Surface hydrological AES. so result in altering a normal evolution of the system. . may lead to larger canopy hold-up. delay time for release and dilution volume. Consulting this correlation. ultimately changing the release quantity of radionuclides. spatial and temporal variations (soil erosion) in precipitation have the greatest effect on the infiltration rate. Vegetative progression is expected to occur near surface disposal after a loss of waste control. Sheppard and Thibault [49] defined a partition coefficient correlation of leafy vegetables based on plant-to-soil concentration ratios.J. surface hydrological AES. (4) Ecological AES. Consideration of domain variables Nominal parameters of the domain variables are mainly adopted from available references [43. features. This causes two effects. Nevertheless such the faults with low or intermediate level may be responsible for local disturbances in the properties of geological barrier. for applying to the waste disposal repository. To develop a subjective distribution of the parameters in the model. K.

infiltration rate. standard deviation. Uniform distribution for this epistemic parameter uncertainty is assumed with lower and upper bounds for each specific radionuclide. The delay in release can be described by a variety of analytic models. SPC. One of the major domain variables. s. They estimated a minimum irrigation water requirement. † Dilution volume. Based on this consideration. LU. in case of handling the domain variables with both epistemic and aleatory uncertainties. Taking this consideration along with the epistemic uncertainty of the parameters. Furthermore.5!104 – 2. [50] is adopted again for choosing a nominal value in this paper. The delay mechanism and its relevant time scales are the most critical data.0!10 – 3. If a time period long enough to change the groundwater resources is provided. Maximum Tdelay is determined by the maximum time of the simulation. It is assumed that a geological AES can directly affect the integrity of concrete structure. etc.) (m/yr) LU 1!10 – 2. it is accepted to consider a lot of natural stochastic processes directly or indirectly affecting the volume of aquifer. we can outline a domain variable with ‘true but unknown’ distributions. soil partition coefficient. † Infiltration rate. partly based on available information addressed in the followings. † Delay time. The delay time of the release varies depending on adsorption of waste and concrete. For example.5!104 Tdelay. U. which intends solving an issue that how we can provide the parameters of the domain variables when we have a lack of critical data set for them. we can focus on the problems desired to be solved. which uses an empirical relationship between soil permeability and the proportion of water developed by the US Bureau of Reclamation. triangular. All the heterogeneous phenomena can be lumped together into a model for retardation coefficient. [50] have a specific model that the infiltration rate is the product of application rate (AR) and fraction of the applied water that will percolate deeply beneath a root zone and become infiltration. we guess that the scenarios such as climatological AES or ecological AES may greatly affect the infiltration rate. the soil partition coefficient (SPC) can influence mass transfer between soil and an aquifer. Waste form hold-up and container lifetime cause a natural delay and then a reduction in the vault release terms. It is noted that the infiltration rate (IR) is very site specific [51] and depends on a natural stochastic process. The contribution to an ingestion dose from the use of groundwater depends on many domain variables. Partition coefficients of radionuclides need calculating the retardation coefficient of soil layer. leaching by dissolution or adsorption by buffer material is one of the models. redistributes. As a domain variable.1 – 0.d. † Retardation coefficient.526 C.2. loguniform.-J. It is usual to assume that groundwater transport processes are dominated by various phenomena such as advection. It is noted that while a probable failure mode of concrete may be a gradual failure. Estimating the flux directly consists in finding the amount of water that infiltrates into soils. Reference data from Beyeler et al. DV. Lee.1!104 1. Likelihood estimation of . In particular. T. All the radionuclides released from the buried waste are assumed to be diluted in the volume of aquifer defined as equivalent tube stream area times transverse length of the aquifer to a well.4!10K2 K3 DV (mean) (m3) T 6.. and consequently the radionuclide concentrations in drinking water consumed by human beings are affected by this. Evaluations associated with information gained from the probabilistic inference of Bayesian network This section devotes our concerns to combine the probabilistic inference of Bayesian network with the estimation of parameters. the time between initial burial and the beginning time of concrete structure failure is defined as the delay time (Tdelay) for release. Beyeler et al. and enters the disposal unit as a function of time. IR. dilution volume. and sorption. Minimum Tdelay is determined to be around 10 years. we assume that a variance of the epistemic distribution closes to a mode of specific representative. uniform.0!10 9. 4.J.4!103 3 IR (mean) (m/yr) U 0. there is more or less a possibility of a rapid failure. It is also well known that soil hydraulic conductivity is a critical factor for deriving the infiltration rate and Darcy infiltration velocity can be limited by the specified hydraulic conductivity of the vault roof [43]. dilution volume (DV) directly relates with a transport modelling of several contaminants through an underground hydrological system. stochastic effects and data uncertainty can be considered altogether in a set of the distributions. All upper and lower bounds for the parameters of the domain variables given in Table 2 are assumed in the illustrative application. By giving high confidence on the variance of the epistemic distribution like this. delay time.22!105 4 DV (s. In this paper. diffusion. K. The source terms can be provided by estimating a flux of groundwater near the repository.d.3 IR (s.) (m3) U 1. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 Table 2 Subjective probability distributions representing the state of knowledge about fixed but uncertain parameters in using default values for Th-230 in the model Fixed quantity (units) Epistemic distribution Min Mode Max Tdelay (yr) LU 10 – 1000 SPC (m3/kg) U 3.d.

For preparing each likelihood category of Table 3.-J. A bestestimate value may be taken to each occurrence rate of AESs. Considering this perspective. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 527 the AESs is a starting point in the analysis. This kind of approach was considered in the study so as to obtain reasonable uncertainty estimates based on the dependency between scenarios and domain variables.0!10K2/yr 2. we can estimate the probability of occurrence of a first time event within a specified time interval t: pðnR 1Þ Z 1 K expfKlj ðtÞ$tg: (12) Thus. (12) easily calculates a scenario probability. while there is an incomplete database. it is already identified that the subjective probability distributions representing parameter uncertainties may undergo changes over the time owing to the scenarios causing perturbations [53]. the Table 3 Categorization of AES likelihood in the frequency estimation Degree of likelihood (1) Highly likely (2) Likely (3) Intermediate (4) Unlikely (5) Highly unlikely Estimated occurrence interval 0–102 y post-closure 102–103 y post-closure 103–104 y post-closure 104–105 y post-closure 105–106 y post-closure Occurrence frequency (median) 2. q 2Jg. the Poisson distribution is commonly used in parameter estimation problems to model all processes where events are generated over all the time. which is beyond the scope of this study. In order to realize this consideration.0!10K6/yr consequences such as an exposure dose will involve both epistemic and aleatory uncertainties. The ‘contamination’ means a generation of ‘contaminated prior’. Lee. As recognized by Frank [8]. the probability of observing N events in a time period T is given by PfN events in time TjlðtÞg Z flðtÞ$TgN KlðtÞ$T e . as denoted in Table 3. Therefore. Historical data collection and relevant methodology development are then necessary in order to establish a reliable data base about stochastic information of AES. which resulted in suitable designation so as category 3 (‘intermediate’ likelihood) for AES-1 and AES-3. (13) becomes . Under those representations. it is noted that the occurrence rate of an AES should be estimated with the support of empirical database and/or expert judgment. respectively. However. it is reasonable to assume that the occurrence of AESs directly influences estimating uncertain parameters of the domain variables. If that is the case. and 4 is an adjustable correction factor with 0%4!1 which reflects how p closes to p0.C. can vary according to the occurrence of various AESs. if the expected occurrence rate is provided. the analyst tends to resort to subjective inputs by some experts who make full use of their experience or the literatures published. by referring to the survey of relevant information. no preciselydefined methodology exists in identifying adequate input types and standardizing the inputs. K. and category 2 (‘likely’ likelihood) for AES-2 and AES-4. if we attempt to give the prior class q(q) with a maximum likelihood estimator (MLE) for reflecting a possible ‘contamination’. The ‘contaminated prior’ represents a variety of input parameters which reflect the previously-mentioned dependency. Given the jth AES occurrence rate.0!10K3/yr 2. The informal approach for performing uncertainty analyses involves a technique varying one parameter or set of parameters at first and then observing the deviation from a base-case prediction [22]. an AES stochastically occurs in space and in time so that the underlying measure of interest (e. By introducing the ‘mixture prior’ we can articulate the effects of uncertain parameters under the occurrence of AESs. which are critical for estimating the end point’s measure of ground release of radionuclides. Next. we made a decision to choose the likelihood category for each AES. Very often. the ‘mixture prior’ shows a mixing status of a normal prior (as termed for representing ordinary uncertainty distribution in this paper) and the ‘contaminated prior’. Also. we set each occurrence interval and median occurrence frequency. We are sure that. The occurrences of AESs for non-overlapping intervals are independent of each other. This concept refers to a special approach using correction factors for representing the parameter uncertainty. However. lj(t). As a practical assumption. then Eq. Next.0!10K5/yr 2. Therefore. we define an arbitrary prior set G to elicit a linear estimator of an uncertain parameter q. we assume that the subjective parameters of the domain variables. first. the dependency between scenarios and variables can obviously affect all the input types of a model.g. the frequency about AES likelihood is just estimated based on the subjective engineering judgment which uses a concept of categorized selection criterion. the occurrences of an AES in a specified time interval) is faced with a stochastic process. in the consideration of environmental transfer models. Instead of choosing insufficient empirical data. a frequency estimation of AES occurrence is related to the realization of aleatory characteristics of the parameters. we adopted a so-named ‘mixture prior’ concept. Therefore.0!10K4/yr 2. (13) where J is a class of possible ‘contamination’. the priors p(q) ‘close’ to a single normal prior p0(q) can be realized with a class of possible ‘contamination’ [54]: G Z fp : pðqÞ Z ð1 K 4Þp0 ðqÞ C 4qðqÞ. As explained earlier. N! (11) where l(t) is an expected occurrence rate of an AES. Eq.J. As summarized by Siu [52]. owing to environmental circumstances by many phenomena in the groundwater during the release of radioactive materials. The term ‘contamination’ is addressed in a ‘mixture prior’. there are strong dependencies between environmental scenarios and the characteristics of the geosphere’s and biosphere’s domain variables. As explained in the previous section.

Nonetheless. Preparing specific information on the CPTs used in the inference program seems not to be an easy task. However. it is not treated for simplifying the problem. Lee. the simulation made a progression quickly and ended within 30. In a practical matter. We found a very fast convergence during the verification process. 4 provides two running outlines from the simulation of all the AESs. the prediction of MLE of the ‘contaminated prior’ requires engineering/scientific knowledge specific to the problem. which seems to be usual in case of the random variable for SPC. When we consider a traditional level 2 PRA for nuclear power plants particularly in the determination of containment progression event branch. probabilities on the query random variables obtained from evidential reasoning can be used to approximately predict the factor 4. it seems that low confidence of the degree-of-belief to the disturbance of this variable is reflected in its likelihood against all the AESs. Before practically using the program. As modelled in Section 2. While keeping this consideration in mind.1. we defined a dependency matrix with an ad hoc basis.005.1 and dZ0. we can do that. This is similar to the approach adopted for getting the CPTs in this paper. As a previous step for estimating the risk from future evolution. In preparing a simulation input for CPTs.J. as denoted in Fig. Furthermore. in most cases. acquired in the simulation results. and zero—were assumed. on the relationship between random variables used in Bayesian network and various AESs. 5 provides some remarkable insights.528 C. This kind of solution can be explained from the other case. based on current knowledge of the relationship between domain variables and scenarios.000 times’ iteration. low. All the random variables were simulated concurrently with each AES. Next. the information gained from the probabilistic inference of Bayesian network. we can provide the solutions of the query random variables depending on each AES. only nominal point estimate values primarily based on engineering judgments are used in the quantification.e. we let the program apply to actual simulations when the network structure was given. as well as probabilistic modelling expertises. compared with the others. With the acceptance of this simulation on approximate probabilistic inference. each individual AES must last for sufficient period so as to cause significant characteristic changes to the domain variables.05. 4. 4. we had to provide some supporting information to get a variety of quantitative expectations about the relationship between domain variables and AESs. the factor 4 will give a quantitative measure for showing a possibility of MLE of the ‘contaminated prior’. where the convergence in the iteration of Monte Carlo sampling is desired. we had already addressed an approximate inference program with a bounded-variance likelihood weighting algorithm in order to calculate the conditional probability of query nodes of Bayesian network.57]: 4lð1 C 3Þ 2 ln d 32 ST R (14) where T is a number of samples. the random variable corresponding to a disturbance of Tdelay seems to have a great causal influence to the occurrence of AES-1 (geological AES).3. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 a similar form of a convex combination of the MLE of the ‘contaminated prior’ and a best estimate of the normal prior [55].3. . Fig. we believe that it is unavoidable to give a single representative estimate of the qualitative probability based on user’s degree-of-belief on corresponding severe accident phenomena [56]. Furthermore. similar insights are followed with that of IR to the AES-2 (climatological AES) and with that of DV to the AES-3 (surface hydrological AES). since its consideration needs doing complex graphical representations in Bayesian network. Even though the PRA considers interval probabilities covering the uncertainty in the expressions of branch possibility. 3 is a relative error. First. respectively. K. However. with 3Z0. medium. even though we let any strong dependency of random variables to the AES-4 (ecological AES) provide.-J. therefore. where the convergence was checked with the following stopping rule provided by the developers of the algorithm [38. d is a failure probability. The simulation easily reached a stable state except initial transitions. Second. i.1. 5. Also. conditional probabilities for representing these relationships were determined using single representative estimate. We obtained values for the query nodes that were close to the exact probability within an error of G0. supporting information is directly used in the program input of Bayesian inference algorithm. for the purpose of doing verification and showing actualization of applied inference algorithm. and l is a defined constant. in fact. we decided to apply a popular network structure with four simple nodes—the so-called ‘WetGrass’ problem [37]. Multiple dependencies may occur by the simultaneous AESs. the random variable for SPC reveals relatively-small conditional probabilities. Conditional probabilities We had primarily considered the dependencies between 4 domain variables and 4 AESs with the information explained in Section 4. where the dependency relationships with four grades—high. As a general assumption. significant results are hardly shown in representing relative importance of the random variables against the AES-4. Fig. which are usually gathered from relevant experts. Last. Results of example application and discussion The previous section indicates that we can connect the results of probabilistic inference from Bayesian network with the consequence evaluation model addressed.

Historically. Next. Convergence representation of Bayesian inference program for the cases of AES-1 and AES-4. On the other hand. For the purpose of giving our confidence to this insight. provided the dependency between geological phenomena and degradation of engineered barrier of repository is considered. Before making the simulations. (13). It is noted that.-J. it is necessary to confirm that the adopted values do not cause any problem in predicting the end point’s measures. The parameters of the random variables were prepared by approximating values referred to a maximum or minimum bound of normal parameters depending on their probability distributions. We also tried to usually get 10. 4. we can move on a first step to estimate the end point’s measures of the environmental transfer model. Even though the frequencies of the occurrence of AESs in Table 3 were reasonably assumed. .2. where an exceeding probability against an end point’s measure can be expressed. an upper bound of the location parameter in a domain variable may be near to the best estimate of a normal prior depending on specific situation. variable 4ZDV. (Note: variable 1Ztime delay. where a variable-specific treatment was done. although an extremely-frequent occurrence of AES2 is assumed. exposure dose). Prediction examples of the end point’s measures By making a mixed prior directly by taking the information given in the form of Eq. classified with each AES. Fig. We performed a sensitivity study to extensively survey the effects of frequency variation. variable 3ZIR.g. the deviation against the case of non-AES is only 29% approximately. Similar considerations were given for the other domain variables. In other words. 4.e. occurrence of AES) cases were illustratively provided. Since the contaminated prior is provided in the evaluation process following after occurrences of AESs. fractions of percentile) for exceeding probabilities of the end point’s measure are shown in Fig. CCDF is an easily-plotted risk curve. 7. a form of complementary cumulative distribution functions (CCDFs) has been used as a favourite form for portraying uncertainty results of risk. finally has a unit of Sv/Bq as provided in the NCRP model. For instance. Only Th-230 was selected for the purpose of illustrative evaluation. Various CCDF fractiles (i. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 529 Fig. K.000 times’ simulation with all the combinations of AESs.J. Lee.C. The screening factor has the dimension of committed effective dose (Sv) for a unit concentration of radioactivity (Bq) in the media. 6 indicates that the variation of occurrence frequency of AES-2 has no strong influence on the results such as mean values of the end point’s measure. the CCDF construction process with a Monte Carlo sampling technique was adopted from Ref. it may further need determining the most limiting case in that the scenario frequency influences the uncertainty propagation.e. This sensitivity study can be potentially used to check the importance of the AES frequency.3. 5. the comparative results between normal (i. variable 2ZSPC. we must identify the endpoint’s measure of this paper with a screening factor (SF) as it was suggested by the NCRP [42] for representing the consequence (e. which gives us good insights in the uncertainty propagation for non-AES case. however. an upper bound of delay time would be extremely reduced according to the existence of AES-1. Comparison chart in the estimation of conditional probability of random variables from the Bayesian inference program. Fig. respectively). [25].e. the effects of AESs must be reflected in the parameters of random variables. non-AES) case and abnormal (i. In this study.

as previously explained. we can easily identify the relative importance of AESs at any subjective confidence level.5 times’ difference in the interval between 95 and 5% (i. a focuseddown analysis on each AES has been conducted. randomly occurrence of AESs was assumed with the other domain variables remained unchanged. 7 reveals that there is a great possibility for showing a wide band. Fig. On the other hand. For instance. 90% confidence interval) at a credible exceeding possibility limit of 0. The mean CCDF curves of the AESs resulted from the uncertainty propagation are depicted in Fig.530 C. 8. however. and partly because we had already let same likelihood category for both AES-1 and AES-3. Dose in unit of screening factor and occurrence rate of AES according to a variation of the AES-2 frequency. Fig. we have performed another sensitivity study. In particular.e. Arithmetic means of 5%. this is partly because there were no big differences of conditional probability between AES-1 and AES-3 except only for the variable of ‘delay time’. 9 shows an outstanding profile for the case of Cs-137. 5. To investigate the importance of the parameters of the delay time. 7. there is no big difference between the uncertainty output (e. showing sensitivity results for the delay time (Radionuclide: Cs-137). The results are made with nominal frequencies given in Table 3.J. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 Fig. Lee. (Occurrence rate of AES means the AES occurrence counts per total simulation times. . SF denotes the screening factor). Mean values according to random AES occurrences. 5. i. the end point’s measure stands at about 2. Fig. In the case of a subjective confidence interval.e. one for 10 years and the other for 300 years. Just after defining two cases with fixed inputs for the delay time. K.g. Fig. mean. Fig. 8. in terms of estimating an end point’s measure. including non-occurrence of AES. Compared with the non-AES case. Mean values according to the AESs. we made a separate simulation and ultimately compared with a standard case having previously-given distribution. In this case. 9. and 95% fractiles are shown. 6. as shown in Fig.-J. Since we are interested in the ranking for the importance of AESs to the impacts of the end point’ measure. the occurrence of AES-2 has a strong impact on the outcomes of the end point’ measure because high dependency of AES-2 on the domain variables such as IR and DV has been already anticipated as shown in Fig. the results for AES-1 and AES-3 curves are almost the same.05. mean of mean CCDF) and the deterministically-calculated value by the NCRP. Comparison of uncertainty propagation results for non-occurrence AES case.

Diegert KV. Reliab Eng Syst Saf 2000. a more complicated simulation process may be required. we could identify exceeding probabilities of the end point’s measure and importance ranking on each AES. Alvin KF. Cheok MC.1(1). if the epistemic uncertainty in the distribution of event occurrence frequencies will be addressed in this study. environmental transfer of radionuclides from radioactive waste disposal.54:95–111. we can assert that it is more desirable for the uncertainty analysis of relatively-short lived radionuclides to prepare a best-estimate value of dominant domain variable (i. Probabilistic risk assessment of agrochemicals in the environment. Treatment of uncertainty in space nuclear risk assessment with examples from Cassini mission applications. finally shown to be adequate in the verification test for clarifying the dependency relationships under different problem queries. Furthermore. When to separate uncertainties and when not to separate. Chistie RF. However. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 531 Wide band of the end point’s measure is derived from the input variation of delay time. can support to enhancing the predictability. The treatment of climate-driven environmental change and associated uncertainty in post-closure assessments. [15] Hofer E. A probabilistic risk assessment for accidental releases from nuclear power plants in Europe. Reliab Eng Syst Saf 2002.-J. we classified them into four major categories.e. the knowledge base for establishing dependency relationships was prepared and enhanced during the simulation. as well as the use of experimental results. et al. and also dominant contributors to the uncertainty propagation might be identified. [13] Oberkampf WL. it was worthwhile to establish definite rationale on the dependency information for getting general insights through probabilistic inference modelling in the prediction of future uncertain state of a system. According to this result. Uncertainties in risk analysis: six levels of treatment. After defining and identifying the terms for specific AES. References [1] Kaplan S. On the quantitative definition of risk.66:203–21. Blaauboer R. Error and uncertainty in modeling and simulation.75:333–57.J. and the EPA release limits for radioactive waste disposal. [8] Frank MV. Cunningham MA. Nucl Technol 1993. Garrick BJ. a Monte Carlo simulation process accounting for epistemic and aleatory uncertainties was provided. Crop Prot 2000. [14] Hoffman FO. Risk Anal 1994. i.54:113–8. While analysing the outcomes of the illustrative evaluation of the environmental transfer pathway. Rutherford BM. After simplifying the network corresponding with a problem-specific structure. specific models. [4] Protection from Potential Exposure: A Conceptual Framework.e.19:649–55. A systematic technique for expert knowledge survey and utilization. Therefore. Probabilistic risk assessment practices in the USA for nuclear power plants. Reliab Eng Syst Saf 1996. Reliab Eng Syst Saf 1996. An approach for using risk assessment in risk-informed decisions on plant-specific changes to the licensing basis. Understanding uncertainty. New York: Marcel Dekker Inc. we developed and verified an approximate probabilistic inference program using bounded-variance likelihood weighting algorithm.101. Ramey-Smith AM. Ultimately. . we introduced its inference process quantitatively estimating the dependency between stochastic scenarios and affected domain variables of the system.14: 707–12. Risk. the proposed approach will probably give a general and flexible structure for doing further extension studies and applications. Not a few remarkable features appeared in applying the proposed approach. [7] Solomon K. K. A general approach integrating the Bayesian network concept to the nuclear risk assessment was proposed and illustratively demonstrated. including a model for propagating uncertainty of relevant parameters. Giesy J. Summary and conclusions Since Bayesian networks can contribute to identify the causality of an uncertain system’s behaviour. also got a number of results to improve the current knowledge base for prioritizing future risk-significant variables in an actual site. In addition. 1993. Risk Anal 1981. 61:209–15. We have also an understanding that more appropriate data for domain variables should be provided to refine the results. [6] Garrick BJ. [12] Caruso MA. Holahan GM. this study can be extensively applied to the evaluation of all the exposure pathway models. especially in the prediction of the occurrence frequency of AESs and in the assessment of causality between domain variables and stochastic scenarios. J Hazard Mater 1998.42:181–200. [10] Rowe WD. Lee. [5] Modarres M. were developed with a comparison of variable-specific effects due to the occurrence of diverse altered evolution scenarios (AESs). Risk Anal 1994.C. Parry GW. What every engineer should know about reliability and risk analysis. Propagation of uncertainty in risk assessments: the need to distinguish between uncertainty due to lack of knowledge and uncertainty due to variability. the delay time in this case).40:177–201. Jones P.63:231–42. uncertainty in risk. DeLand SM. Hammonds JS. in estimating the consequence of example pathway problem. The identified specific results presented will probably be useful in making the prediction of future system states. further researches need enhancing the knowledge base of various modelling areas. The domain variables important in the consideration of causality were found. Reliab Eng Syst Saf 1993. [9] Wilmot RD. et al. ICRP 64. 1993. [3] Slaper H. For instance. [11] Pate-Cornell ME. The evaluation can be also intensively done for any specific scenarios such as a specific phenomenon categorized in this paper.14:743–50. King TL. we made a causal inference process including uncertainty estimates. Safety Sci 2002. [2] Helton JC. Reliab Eng Syst Saf 1999. 5. provided the relationships between AESs and domain variables are searched in detail. While taking a modelling approach proposed in this paper.

2nd ed. In: Proceedings of the 9th conference on uncertainty in artificial intelligence. [46] Thompson BGJ. Risk Anal 1995. Russel J. Uncertainty and variability in human exposures to soil contaminants through home-grown food: a Monte Carlo assessment. [29] Pearl J.54: 119–26. Albuquerque: Sandia National Laboratories. July 2002. Evaluation of a performance assessment methodology for low-level radioactive waste disposal facilities—evaluation of modeling approaches. [22] Helton JC. Dagum P. Washington DC. [36] Jordan MI. A survey of algorithms for real-time Bayesian network inference. AI Magazine 1991. Davis PA. and ground. Demonstration of risk-based approach to highlevel waste repository evaluation. US Nuclear Regulatory Commission. 123 I.93:1–27. Status report on risk assessment for nuclear waste disposal. [50] Beyeler WE. [21] Hattis D. [57] Pradhan M. [49] Sheppard ML. Brown TJ. 1999. [40] Jordan MI.42:161–80. McCord JT. Coppersmith KJ. Hsu WH.63: 169–84. Causal independence for probability assessment and inference using Bayesian networks. Residual radioactive contamination from decommissioning—parameter analysis. 99–105. 1997. August 1996. IAEA. [56] Severe accident risks: an assessment for five US nuclear power plants. NUREG-1150. [47] McGuire RK. Kelly DL. 1995. Peters L. The handbook of brain theory and neural networks. 1993. p. 2nd ed. Burmaster DE. vol. Jaakkola TS. Reliab Eng Syst Saf 1996. 606–12. Saul LK. Swift PN. Bennett DG.38(3):24–68. 100. vol. [19] Parry GW.42:327–67. 1989. SAND91-2802. [24] Bogen KT. Cook N. McCord JT. Reliab Eng Syst Saf 1995. 2. Statistical decision theory and Bayesian analysis. The characterization of uncertainty in probabilistic risk assessments of complex systems. Safety risk assessment and management—the ESA approach. for four major soil types: a compendium. Scenario development for the waste isolation pilot plant compliance certification application. Olague NE. Machine Learn 1999. Cambridge.14(4):449–63. Decision analysis and expert systems. [30] Heckerman D. Bullen DB. 12(4):50–63. Ghahramani Z. Davis FJ. [25] Helton JC.49:303–9. McKone TE. Derivation of risk based wipe surface screening levels for industrial scenarios. [18] Preyssl C. Rao RR. SAND99-2148. Sci Total Environ 2002. [33] Pearl J. Martell M.8:195–210. Wellman MP. p. Proceedings of the 12th conference of uncertainty in AI. Real-world applications of Bayesian networks. 1985. Treating aleatory and epistemic uncertainties in analyses of pressurized thermal shock. Artif Intell 1997. NUREG/CR-5927. November 2000. [16] Siu N. Woods H. [26] Liu C. Pearson FJ. Thibault DH. Reliab Eng Syst Saf 1999. Pitrat T. [17] Helton JC. [28] Henrion M. Berlin: Springer. Albuquerque: Sandia National Laboratories. et al. Johnson JD. transport and uncertainty analysis. March 1994. Sagar B. Bessette D. IEEE Trans Knowl Data Eng 1996. Health Phys 1990. Risk Anal 1994. et al. AI Magazine 1991.69:191–204. [20] May LM. Default soil solid/liquid partition coefficients. [48] Kozak MW. Weiss Y. 3.59(4):471–8. Breese JS. Average case of a search algorithm for estimating prior and posterior probabilities in Bayesian networks with extreme probabilities. Evaluation of a performance assessment methodology for low-level radioactive waste disposal facilities—validation needs. Bennett DH. [27] Charniak E. Assessment of variability and uncertainty distributions for practical risk analyses. Commun ACM 1995. A multimedia. Bayesian network without tears. SAND912802. Characterization of stochastic uncertainty in the 1996 performance assessment for the waste isolation pilot plant.42:125–60. Microsoft Research. Hareland WA. UCLA cognitive systems laboratory. Boca Raton: CRC Press Inc. London: Chapman and Hall/CRC. An introduction to variational methods for graphical models. NP-1197. 2000. multiple pathway exposure assessment of atrazine: fate. Risk Engineering. Proceedings of the 13th IJCAI. [32] Diez FJ.62:89–116.14:713–30. Reliab Eng Syst Saf 1998.69:167–89. October 1990. . Hicks TW. Mamdani A. Final report of NSARS. Bayesian networks. Characterization of subjective uncertainty in the 1996 performance assessment for the waste isolation pilot plant. NUREG/CR-5512. Anderson DR. [43] Safety assessment of near surface radioactive waste disposal facilities: model intercomparison using simple hypothetical data (test case 1).15(3):411–9. Reliab Eng Syst Saf 1993. Malik S. Reliab Eng Syst Saf 2000. Lee. [41] Koplik CM. Risk Anal 1994. PSAM 5 2000. draft report for comment. Reliab Eng Syst Saf 1993. Dynamic Bayesian networks: representation.. Technical report MSR-TR-94-08. 1993.J. An optimal approximation algorithm for Bayesian inference. Computer Science Division. [38] Dagum P. The generalized noisy OR-gate. Reliab Eng Syst Saf 2000. [35] Murphy KP. EPRI NP-7057. 1990. In: The computer science and engineering handbook. Portland OR. K. Tierney MS. Uncertainty and sensitivity analysis techniques for use in performance assessment for radioactive waste disposal. Horvitz EJ. inference and learning. Albuquerque: Sandia National Laboratories. July 2002. Reliab Eng Syst Saf 1993. Graphical models for probabilistic and causal reasoning. UC Berkeley. Luby M.-J. IAEA-TECDOC-846. Graphical models: probabilistic inference.37: 183–233. [45] Thorne MC. NCRP report No. Kastenberg WE. vol. Parameter adjustment in Bayes networks. Rao RR. Long A. Methods to approximate joint uncertainty and variability in risk. technical report (R-277). [34] Buntine W. Reliab Eng Syst Saf 2000. Bayesian parameter estimation in probabilistic risk assessment. Lee / Reliability Engineering and System Safety 91 (2006) 515–532 [39] Poole D. p. Kds. The development and application of integrated procedures for post-closure assessment. AAAI/KDD/UAI-2002 joint workshop on realtime decision support and diagnosis systems. Optimal Monte Carlo estimation of belief network inference. Gallegos DP. Wilmot RD. Crawford MB. Browne D. interim report EPRI. NUREG/CR5927. [51] Kozak MW. 1993. [54] Berger J. [55] Carlin P. [42] National Council on Radiation Protection and Measurements. 2nd ed. Bayes and empirical Bayes methods for data analysis. Duran FA. Safety Series No. [44] Galson DA. The use of expert opinion in formulating conceptual models of underground disposal systems and the treatment of associated bias. based upon Monte Carlo simulation: the probabilistic systems assessment (PSA) approach. 697–714.532 C. 1996. Olague NE. Inc. Screening models for releases of radionuclides to atmosphere.12(4):64–91. 1. Louis T. [52] Siu NO. Edmonton. MA: MIT Press. [31] Heckerman D. Kalinina E. November 1995. surface water. [53] Evaluating the reliability of predictions made using environmental transfer models. A guide to the literature on learning probabilistic networks from data. Gaborek B. 2002. [37] Guo H. PhD Thesis. October 1979.69(1–3):129–49. [23] Mckone TE.288:65–80. Kemeny J.

Sign up to vote on this title
UsefulNot useful