Bayesian Networks

Reliability Engineering and System Safety ()
Contents lists available at ScienceDirect
Reliability Engineering and System Safety

journal homepage: www.elsevier.com/locate/ress
Methods for building Conditional Probability Tables of Bayesian Belief

Networks from limited judgment: An evaluation for Human
Reliability Application
L. Mkrtchyan, L. Podollini n, V.N. Dang
Paul Scherrer Institute, Switzerland
art ic l e i nf o
Keywords:
Bayesian Belief Networks
Human Reliability Analysis
Expert judgment
Conditional Probability Tables
a b s t r a c t
The present paper evaluates ve methods for building Conditional Probability Tables (CPTs) of Bayesian
Belief Networks (BBNs) from partial expert information: functional interpolation, the Elicitation BBN, the
Cain calculator, Fenton et al. and Red et al. methods. The evaluation considers application to a specic
eld of risk analysis, Human Reliability Analysis (HRA). The ve methods are particularly suited for HRA
models calculating the human error probability as a function of inuencing factor assessments. The
performance of the methods is evaluated on two simple examples, designed to test aspects relevant for
HRA (but not exclusively): the representation of strong factor inuences and interactions, the representation of uncertainty on the BBN relationships, and the method requirements as the BBN size
increases. The evaluation underscores modelling limitations related to the treatment of multi-factor
interdependencies and of different degrees of uncertainty in the factor relationships. The functional
interpolation method is the least susceptible to these limitations; however, its elicitation requirements
grow exponentially with the model size. Besides expert judgment, HRA applications of BBNs include the
use of empirical data, combination of data and judgment, information from existing HRA methods: the
building of the CPTs in these applications is outside the scope of the evaluation.
& 2016 Elsevier Ltd. All rights reserved.
1. Introduction
Bayesian Belief Networks (BBNs) are increasingly being used in
risk analysis applications to model the effect of multiple, diverse,
inter-related inuences on risk. Their ability to incorporate diverse
types of factors has allowed the construction of comprehensive
models for risk assessment including hardware, human, and
organizational failures, as well as diverse risk-conditioning events,
as in [13]. Risk analysis applications, dealing with rare events,
often have to cope with the scarcity of data available to understand the complex interactions leading to failure events: in these
applications, BBNs have proven useful to formalize, represent, and
quantify subjective knowledge on uncertain events. At the other
extreme, in data-rich applications (e.g. some medical diagnosis
and nancial applications), BBNs are typically used for data mining
to make sense of causal or inuencing relationships and build
predictive models learnt from data [4]. Other applications fall
between these categories and BBNs are generally developed by
combining available data and expert judgment.
n
Corresponding author. Tel.: 41 56 310 53 56.

E-mail address: luca.podollini@psi.ch (L. Podollini).
The focus of the present paper is on the development of BBNs

from expert judgment, for cases in which data is not available or
not adequate to determine the BBN relationships. While expert
judgment may be used in many phases of BBN development
(including the node and structure denition), the present paper
focuses on the quantication of the BBN relationships, i.e. the
Conditional Probability Distributions (CPDs). There is general
agreement that this is the most delicate part of the BBN development [5]; although, concerns that some applications may lack
transparency in the process of node and structure denition were
raised recently [6,7].
In general, the elicitation of judgments to assess CPDs has followed three approaches, often combined: direct assessment of the
probabilities by one or multiple experts; elicitation of probability
ranking on a qualitative scale (to avoid the shortcomings, e.g.
biases, of directly eliciting probabilities from experts); elicitation
of selected model relationships (or, more generally, of partial
model information) and derivation of the remaining relationships,
to complete or ll up the Conditional Probability Tables (CPTs), via
different methods. For the rst two approaches, the issues to be
addressed are typical of applications in which a large number of
probabilities are elicited, e.g. avoiding different types of biases and
ensuring consistency in the assessments (these issues are
http://dx.doi.org/10.1016/j.ress.2016.01.004
0951-8320/& 2016 Elsevier Ltd. All rights reserved.
Please cite this article as: Mkrtchyan L, et al. Methods for building Conditional Probability Tables of Bayesian Belief Networks from
limited judgment: An evaluation.... Reliability Engineering and System Safety (2016), http://dx.doi.org/10.1016/j.ress.2016.01.004i
L. Mkrtchyan et al. / Reliability Engineering and System Safety ()
presented in detail in [8,9]). Of course, when applying lling-up

methods, these issues may also need to be addressed. These
methods to populate CPTs on the basis of selected, elicited model
relationships are referred to in this paper as CPT building methods.
The most popular method to populate CPTs from partial
information is based on Noisy-OR gates [4,10,11]. The Noisy-OR
model entails assessing the effect on the outcome of the presence
of one factor at a time, with all other factors being absent. In their
typical implementation, Noisy-OR gates require binary BBN nodes
and model factor inuences as independent of the presence of the
other factors. A number of extensions of the Noisy-OR model have
been developed, generally addressing either dependent inuences
or multi-state nodes (see e.g., [12] for a brief summary). Alternative methods adopt interpolation algorithms [1315] that typically extract information on the factor inuences from selected,
specically elicited CPDs. Fenton et al. [16] base their method on a
further concept, according to which CPTs are derived from
weighted functions of the inuencing factors.
The available CPT building methods differ in their theoretical
basis, the base information on which they derive CPDs, the elicitation requirements, and the interpretation and extrapolation of
the factor inuences. The present paper analyses a selection of
these methods for use in Human Reliability Analysis (HRA)
applications. HRA is the area of risk analysis dealing with identifying risk-signicant human failure events, understanding and
modelling their causes and inuencing factors, and quantifying
their probability. Besides sharing many aspects with BBNs for
more general risk analysis applications, BBNs for HRA often
attempt to formally combine cognitive models, empirical data, and
expert judgment with the aim to enhance the empirical basis of
HRA methods, [7,17]. BBNs have a number of attractive features for
HRA and in general for elds with shortage of data and consequent
reliance on subjective judgments: their intuitive graphical representation, the possibility of combining diverse sources of information, the use the probabilistic framework to characterize
uncertainties. In terms of modelling capabilities, BBNs allow
modelling strong factor effects and interactions: this potentially
allows (provided that these effects can be quantied) to overcome
the assumption of some methods of independent factor effects
(e.g. SPAR-H [18], HEART [19]). Attractive features and research
gaps of the BBN modelling framework are discussed in a recent
review by the authors of the present paper [7]. Given the relevance
of the issue of CPT building and the variety of available methods, it
becomes important to clarify their strengths and limitations, to
evaluate their suitability for HRA, and to identify gaps to be
addressed by research.
The present analysis focuses on methods applicable to BBNs with
multi-state nodes; these are better suited for HRA applications than
those with binary nodes because the characterization of inuencing
factors in HRA methods generally involves multiple levels, e.g. SPARH [18], HEART [19] and CREAM [20]. The performance of ve
methods is examined: the functional interpolation method [14], the
Elicitation BBN (EBBN) method [13], the Cain calculator [15], Fenton
et al. [16] and Red et al. [2]. Two small BBNs representing very
simplied HRA models were used to benchmark the performance.
Small BBNs were selected to allow a comprehensive comparison of
the produced CPDs. The two examples are designed to test some
aspects relevant for HRA modelling: the representation of strong
factor inuences and of factor interactions, the representation of
uncertainty on the BBN relationships, and the method requirements
as the size of the BBN increases.
Although on-going data collection efforts [2123] aim at
reducing the need for expert judgment in HRA, the need to combine empirical data and expert judgment is likely to persist for at
least the middle term. This will be the case especially for application scopes for which data will continue to be difcult to obtain
(e.g., in nuclear PSA applications, HRA for accident mitigation

conditions and external initiating events) and also for industrial
sectors in which the collection of HRA data may be less advanced
than in the nuclear industry. Recent studies providing empirical
human error probability estimates focus on data usable for
quantication of rst generation HRA methods [2426]. Current
efforts on data collection for newer generation methods (emphasizing the role of the context and decision-making errors) are not
yet providing statistically solid gures [21,23,27]. This continued
need for expert inputs to HRA methods motivates this work's focus
on the use of expert data in the construction of BBNs for HRA. The
choice of BBNs with multi-state nodes is also related to this focus.
A motivation for the adoption of binary nodes in the recent efforts
to use empirical data to develop BBNs for HRA [17,28] is that it
makes data collection easier and statistically stronger. On the other
hand, binary BBN nodes are generally associated with models
based primarily on the presence or absence of an inuencing
factor. Reducing the number and complexity of the model relationships is an advantage but may make the expert elicitation
more difcult because factors may be dened less specically,
their states enveloping broader sets of conditions.
Note that in the literature, the application of BBNs to HRA
problems goes beyond models solely built on expert judgment,
addressed by the present paper. For example, the already mentioned [17,28] quantify the CPDs from databases of human failure
events (although in both [17,28] the available data is not sufcient
to a statistically solid determination of the CPDs and expert
judgment is used in combination). Other studies develop BBNs as
extensions of existing HRA methods, such as SPAR-H [29]: in these
cases the quantication of BBNs is based on the underling method
relationships. CPT populating methods have also been applied in
the HRA literature. A Noisy-Or lling up algorithm is used in [30],
therefore assuming independent parent nodes. Reference [31]
determines the CPDs by linear interpolation according to the
number of states of the parent nodes representing positive conditions. As discussed in [7], the interpolation does not differentiate
among congurations with different parent nodes in the same
number of positive states. For HRA applications, however factors
can be characterized by strong interactions so that their effect can
be strongly dependent on which factors are in their positive and
negative states. In [32], pre-dened triangular functions are
associated to different factor strength category. The functions for
each strength category differ by their variance: the stronger the
inuence, the smaller the variance. The functions for each parent
are then weighted to derive the child CPT. The method in [32]
shares similar characteristics with the Fenton et al. and Red et al.
methods (use of weights and inuence functions) and therefore is
not included in the present evaluation.
The rest of this paper is organized as follows. Section 2 briey
presents BBNs and introduces CPT building methods in general.
Section 3 presents the design of the evaluation study. This is based
on testing the methods on how they reproduce relevant modelling
features for HRA, implemented on specic evaluation criteria. Two
simple HRA models are introduced (BBN 1, BBN 2), so that the
differences in the CPTs produced by the methods can be easily
traced. Section 4 presents the application of the ve methods to
the two HRA models, on two analysis cases for each model (cases
NOUNC and UNC). Section 5 presents the evaluation of the
methods. Section 6 provides concluding remarks.
2. Bayesian Belief Networks: introduction

This section briey introduces BBNs. Part of the section is taken
from [7]. For extensive treatment of the BBN modelling framework
the reader should refer to [4] and [10], for example.
Fig. 1 shows a BBN example with two child nodes (D and E) and
three parent nodes (A, B, C). In their most typical representation,
BBN nodes are associated with mutually exclusive discrete states
(with reference to Fig. 1: a1, a2 are the states of node A; b1, b2 are
the states of node B; ; e1, e2 are the states of node E). Such
discrete nodes are generally preferred because the associated
calculation scheme is simpler compared to that for continuous
nodes; nevertheless, the adoption of continuous nodes is advisable
(or even necessary) for some applications [33]. Fig. 1 shows an
example of BBN with binary states, but multi-state nodes are also
frequently used. For BBNs with discrete state nodes, the BBN
relationships are in the form of Conditional Probability Distributions (CPDs); each child state requires a CPD for each possible
combination of its parents' states (Table 1 gives an example of the
CPDs for node E).
The BBN formalism allows visualizing and quantifying properties of the joint probability distribution of A, B, C, D, E. In particular,
for the most typical HRA applications, nodes A, B, C would model
observable PSFs; node D would model an error context or error
mechanism (possibly not observable); node E would model the
occurrence of the human error. The estimation of a probability of
human error, given a particular conguration of the factor states is
the primary interest in most HRA applications. Alternatively, the
BBN may be used to infer knowledge on the factor states given
that an error has been observed (i.e. to infer what is likely to have
inuenced the occurrence of the error). The former use of BBN is
often referred to as predictive, while the latter as diagnostic.
Data-rich applications such as medical diagnosis and nancial
applications typically rely on data to build CPTs by learning algorithms. Depending on the amount and quality of the available data,
the algorithms are based on the maximum likelihood estimator,
Bayesian estimator and more sophisticated procedures to cope
with realistic data sets with missing, incorrect, sparse entries
(refer to [4] and [5] for a comprehensive treatment). For rare-event
applications, BBNs are typically constructed based on input from
expert domain. In this phase, the typical tools are questionnaires,
interviews and panel discussions. The issues to deal with are
typical of applications in which a large number of probabilities are
elicited from experts, e.g., avoiding different types of biases,
ensure consistency in the assessments [8,9].
The development of lling-up methods, or more in general,
methods to determine CPDs from limited information is an
important subject of research for BBNs, motivated either because
of the difculty to collect statistically signicant data covering all
BBN relationships, or because the elicitation of an excessive
number of probabilities may become impractical and prone to
inconsistencies. The methods to build BBNs from limited data and
judgment information are fundamentally different. The former are
often based on Bayesian update, with prior knowledge either from
aggregated data or from expert judgment [5]. The latter are based
on eliciting specic information from experts, for example selected
Fig. 1. A simple example of a BBN.
Table 1
CPT for node E of the BBN in Fig. 1.
A
A a1
D d1
D d2
D d1
D d2
E e1
E e2
0.1
0.9
0.4
0.6
0.3
0.7
0.8
0.2
A a2
CPDs and/or factor importance weights and then build the CPTs.
The fundamental difference is that in general for the former case
one has to make the best use of the data available, with limited
possibility to drive the data to satisfy specic information needs.
On the other hand, the latter methods aim at focussing the elicitation on specic information, to efciently develop the
overall model.
The majority of lling-up methods have been developed for
BBNs with binary nodes. The general idea of these methods is to
associate operators linking the state of the parent and child nodes,
the mostly used being: Boolean operators (typically, OR, AND,
M-out-of-N) [10], weighted averages [10], and Noisy-OR gates
[4,10,11]. Boolean operators and weighted averages represent
deterministic relationships between parent and child nodes: the
uncertainty in the outcome (child state) depends on the uncertainty in the presence of the inuencing factor, but not in its effect.
This is a strong modelling constraint for many applications in
which factors have non-deterministic and/or uncertain relationships. The most widely used method is the Noisy-OR model, proposed in [11]. In the Noisy-OR model, the probability of the outcome is the product of the probabilities of the outcome in presence of one factor at a time, with all other factors being absent. In
this model, the factor inuence on the outcome is independent on
the presence of the other factors. A number of extensions of the
Noisy-Or model have been developed, generally addressing either
dependent inuences or multi-state nodes [12].
Another type of CPT building methods uses interpolation
methods [1315]. These methods are typically based on extracting
information on the factor effects from known relationships
(anchor CPDs) and extrapolating to the whole CPTs. In the functional interpolation method [14], CPDs elicited at the anchor
positions are approximated by functions (e.g., Normal functions),
described by parameters; the parameters of the missing CPDs are
obtained by interpolating those corresponding to the anchor ones.
The EBBN method [13] is based on piecewise linear functions
interpolating among the elicited CPDs, and on state inuencing
factors and importance weights. The Cain calculator uses interpolation factors derived from CPDs at the anchors to populate the
missing relationships in CPTs [15].
Another well-established method, not based on interpolation
of known CPDs is the one from Fenton et al. [16], implemented
within the AgenaRisk software [38]. According to this method, the
CPTs are derived from weighted functions of the inuencing factors. The functions (truncated normal) and the weights are
determined from qualitative statements (most often, but quantitative analysis may be of support) on the general tendency of the
inuencing factor effect. The method presented by Red et al. [2] is
also based on functional relationships between inuencing factors
and outcome nodes; here, the parameters of the function (exponential) are determined based on the elicitation of selected CPDs.
It is worth noting that, given the fast increase in the number of
CPDs with the size of the BBN, an effective way to reduce the CPD
quantication requirements (from data, expert judgment, and
combinations of both) is to act on the BBN structure. Indeed, for
data-rich applications, structure-learning algorithms generally
automatically search for the BBN structure that optimizes some
predened criteria, under some constraints [4] and [10]. These

algorithms are generally not usable for HRA because of their data
requirements. Indeed, in [17], an HRA application, the graphical
structure is not directly obtained by data, but it is informed by an
iterative process involving judgment and Correlation and Factor
Analyses. The combined and iterative use of data and judgment to
determine the BBN structure is typical of situations in which data
is available, but not enough to run structure-learning algorithms.
For BBNs totally built on judgment, the development of the
appropriate structures is also key for the subsequent quantication of the CPDs and the nal BBN use. Guidelines for the choice to
include variables, their denitions and presence of links can be
found, for example, in [10,33,34].
Finally, note that to quantify the full BBN, the values of the PSF
nodes have to be assessed as well. In typical HRA applications of
PSF-based methods, this entails following the PSF denition guidance provided by the various HRA methods to determine the
proper PSF rating (for example in the methods SPAR-H [18], HEART
[19]). If the HEP assessment is supported by a BBN, quantication
of the PSF node is not particularly different: the development of
the BBN would need to include the PSF rating scale to support the
analyst in the choice of the appropriate PSF state. The development of PSF denition and rating scales is outside the scope of the
present paper. Some relevant references for the subject including
recent advances and open issues are [35,36].
The present analysis considers methods applicable to BBNs
with multi-state nodes: the functional interpolation method [14],
the EBBN method [13], the Cain calculator [15], Fenton et al. [16]
and Red et al methods [2].
3. Design of the evaluation study

3.1. Overall design
The overall study design is shown in Fig. 2. The evaluation
study aims at testing whether the considered CPT building
methods are able to produce BBN models featuring desirable
aspects for HRA applications: strong factor inuences and interactions, different uncertainty in the factor relationships, method
requirements as the BBN size increases the motivation for these

aspects is discussed later in this section.
Two simple hypothetical HRA models are considered (BBN
1 and BBN 2, each with two analysis cases NOUNC and UNC, see
Section 3.2): these models implement a common expert judgement knowledge base which features the desirable aspects under
investigation. The idea underlying the introduction of the common
knowledge base is to mimic an individual expert, using his/her
mental model to build CPTs assisted by the different methods: the
methods are then tested on how they are able to reproduce the
mental model, over specic evaluation criteria.
The knowledge base is used by each method to produce the
whole CPT. This entails that method-specic input parameters are
derived for each method, in all four cases. Indeed, each method
has different requirement in terms of the information needed to
build the CPT (e.g. different anchoring CPDs, parent inuence
weights, parameters of CPD building functions). Challenges concerning the implementation of this step are discussed in Section
3.3. Finally, each method is evaluated with respect to evaluation
criteria specically tailored on the aspects implemented in the
knowledge base. The implementation of the criteria into the
knowledge base will be presented in Section 3.3.
The motivation for the criteria will now be introduced. The rst
criterion (strong factor inuences and interactions) represents the
need for HRA models to quantitatively represent complex factor
effects on the error probability. Negative conditions from single or
few factors can lead to large increases in the error probability
values compared to when the factors are characteristic of nominal
or positive conditions. Especially when multiple factors come into
play simultaneously, their effect on the error probability can be
much larger than when they act alone. This culminating effect into
error-forcing contexts is at the basis of the newer HRA methods
(e.g. ATHEANA, [37]) and it emerges from patterns of human error
databases [17,28] and simulator studies [36]. On the other hand,
factor effect also compensate each other: for example, very good
level of operating crew training and clear plant indications may
compensate for unspecic procedural guidance (again emerging
from experience [17,28]).
The second criterion relates to the fact that, realistically, the
model relationships may not all be known with the same condence: some relationships can be easier or more difcult to
Fig. 2. High level design of the evaluation study.
assess for experts; correspondingly, some relationships may have

larger uncertainty due to lower condence by one expert or to less
agreement across more than one expert. Typically, the condence
on very high error probability values may be strong when multiple
factors simultaneously concur to the error-forcing context. On the
other hand, larger uncertainty can be expected for assessments in
presence of counterbalancing effects, which are generally the most
difcult situations to evaluate. Also, if data is available (either from
operational experience or from simulated environment), the data
coverage is generally not homogeneous, as in the dataset utilized
in [17] and [28]. Indeed, also the presence of some data may
strengthen the condence on some relationships, as it was the
case in [17], with signicant difference in the amount of data
covering the different BBN relationships.
The third criterion relates to the practical usability of the
method. It evaluates the degree of elicitation burden in each
method as the size of BBN increases (in particular, the evaluation
criterion reects whether the increase is linear or exponential).
3.2. The two example BBNs
The rst BBN has two input nodes (HumanMachine Interface,
HMI; and Task Complexity, TC) and one output node (the Human
Error Probability, HEP), Fig. 3(1). The second BBN, Fig. 3(2),
includes a third input node, Time Pressure, TP. The size of the BBNs
has been kept small to trace in detail the differences in the CPTs
produced by the methods: the CPDs can be compared one-by-one
across all methods. For larger BBNs, with more CPDs, the comparison would have been unnecessarily much more complex.
Given the number of states of each parent node (Fig. 3 and
Table 2), the number of CPDs to be determined is 3 4 12 and
3 3 4 36 for BBN 1 and BBN 2, respectively.
The states of the input nodes (Table 2) describe the degree of
inuence of each input factor on the probability of human failure.
Note that the denition of the output node, modelling the HEP, is
rather different than in typical HRA applications of BBNs. In these,
the output node generally models whether the Human Failure
Event occurs: therefore the node is binary and the probability of
the failed state is the HEP. In the present application, a multi-state
node is used, with the states modelling different degrees of error
probability (Low, Medium, and High). This allows modelling the
uncertainty in the HEP assessment, by assigning different probabilities to each HEP state. Indeed, the precision of the assessment
is related to the probability interval covered by each state.
All methods for CPT building described in this paper are
applicable for BBNs with ordered states (from low to high values of
the underlying conceptual variable) and in which the parent state
effect on the child state has constant direction, with monotonic

and positive inuence. With reference to the BBNs in Fig. 3, when
one of the parent state increases (e.g. TC state from VL to
Nom), the value of the child node increases or stays constant
(and therefore the CPD shifts towards higher values). In probabilistic terms, larger values for the probabilities of higher parent
states make higher child states more likely. This property is generally referred to as positive (or negative, if in the opposite
direction) qualitative inuence [39,40] or monotonicity [41]. The
present paper deals with positive inuence nodes; yet, the
methods apply in similar way for negative inuence nodes. This
requirement on the qualitative inuence is not limiting for HRA
applications: as is the case for the simple models in Fig. 3, the
factors in HRA models are often dened with a constant direction
of inuence on the HEP.
Four out of ve methods discussed in this paper assume that
the BBN nodes have discrete labelled states. On the other hand, the
Fenton et al. method [16] uses the concept of ranked nodes that
represent abstractions of some underlying continuous quantities,
typically ranging between 0 and 1. The notion of ranked nodes is
further detailed in Section 4.4.1.
3.3. Knowledge base and evaluation criteria
The use of a common expert judgment knowledge base allows
investigating how each method uses the same information set to
build up the model. A challenge for the implementation of this
idea is that the information required by each method is generally
different in type (e.g. qualitative/quantitative, probabilistic/
numerical) and in the model characteristics to be elicited (e.g.
Table 2
The BBN nodes and their states.
Node
States
HumanMachine Interface
Strongly Success Forcing (SSF)

Nominal (Nom)
Less Than Adequate (LTA)
Error Forcing (EF)
Very Low (VL)
Nominal (Nom)
Very High (VH)
Very Low (VL)
Nominal (Nom)
Very High (VH)
Low
Medium
High
Task complexity
Time pressure
Human error probability
Fig. 3. The two example BBNs: BBN 1 (left) and BBN 2 (right).
general importance of factors or different subsets of the model

relationships).
The strategy used in the present paper to cope with the above
issue has been as follows. The expert judgment knowledge base is
presented both as qualitative statements for selected parent congurations and as CPDs associated to these congurations as
shown in Fig. 4 for BBN 1: this dual representation was decided to
satisfy most of the input requirements from the methods. The
selected parent congurations are those combining all possible
extreme states of the parents, i.e. for example, for BBN 1 all possible combinations of HMI SSF or HMI EF and TC VL or
TC VH, likewise for BBN 2. The positions of these combinations
within the CPTs are shown in Figs. 4 and 5 for BBN 1 and BBN 2,
respectively. These combinations provide the anchoring judgments for which the methods build the remaining CPDs. The 3D
plots of Figs. 4 and 5 only report the most likely state for each of
the anchor parent combinations. In general, indeed, all combinations (anchors and those generated by the method) are associated
with a CPD. Based on the anchor CPDs, the CPT building methods
populate the missing CPDs, lling up all the combinations of parent node states in Figs. 4 and 5.
The knowledge base will be introduced in what follows, with
emphasis on the implementation of the evaluation criteria
(Table 3). BBN 1, case NOUNC, has been primarily designed to test
single factor effects: Fig. 4 shows that the two factors HMI and TC
have different inuence on the HEP when they are one at a time in
their error-forcing state (anchors 2 and 3). The inuence of HMI
being in the Error forcing state while factor TC being in the Very
low state (anchor 2) is larger than that of factor TC being in the
Very high state while factor HMI being in the Strongly success
forcing state (anchor 3): indeed, the resulting HEP for the two
anchors 2 and 3 is High and Medium, respectively. The CPT
building methods are evaluated based on whether their CPT reect
this effect (Table 3, criterion 1, Difference importance of factors).
A related consideration from the assumed knowledge base is that
when HMI is in its Error forcing state, the HEP is High, independently on the state of TC (dominance of factor HMI when in its
error-forcing state, see Table 3).
Concerning the knowledge base for BBN 2, similarly, the HMI
node has higher inuence on the HEP than the TC and TP nodes
Fig. 4. Assumed expert judgment knowledge base: most likely child state at
anchors (BBN 1).
(Fig. 5). Again, the inuence of HMI being in the Error forcing state
while the other two factors TC and TP being in the Very low state
(anchor 2) is larger than that of either of TC or TP factors being in the
Very high state while factor HMI being in the Strongly success
forcing state (anchor 3 and 5): indeed, the resulting HEP for anchor
2 is High while for anchors 3 and 5 are Medium. In a similar way
as in BBN 1, when HMI is in its Error forcing state, the HEP is
High, independently on the state of the other two factors. The two
single-factor aspects Difference importance of factors and Dominance of one factor in one state in Table 3 are analysed for BBN 2 as
well. The knowledge base of BBN 2 has been dened such to model
another typical situation for HRA: the interaction of multiple factors
(two in this case), evaluation aspect Combined error-forcing effect
of Table 3. In particular, the HEP is assessed as High in case both
factors TC and TP are in their mostly error-forcing state, while the
HMI factor is strongly success-forcing (anchor 7). The two factors
interact in the sense that their combined effect is larger than when
they act alone (in which case the HEP state would be Medium, as
in anchors 3 and 5).
Two cases are considered for each example, each associated with
the same qualitative statements from Figs. 4 and 5, but with different degrees of uncertainty associated to the parent node inuence (evaluation criterion 2, Table 3). Fig. 6 shows the assumed
expert knowledge base for the two BBN 1 cases: no uncertainty,
NOUNC case on the left; uncertainty, UNC case on the right. The
knowledge base has been assumed such that anchor 3 has the largest uncertainty in the HEP state distribution (Fig. 6). The features of
the two cases for BBN 2 are similar: see the result gures for the
CPDs associated to the knowledge base Figs. 11 and 12 subplot (a),
dark blue bars, for NOUNC and UNC cases, respectively).
For evaluation criterion 3, the increase in the number of parameters to be elicited in BBN 2, respect to BBN 1 is considered. The
interest for this criterion is to extrapolate the trend (linear,
exponential) of the increasing requirement for the application the
methods.
Finally, as mentioned, each method requires a different set of
anchor relationships to build the BBN model. For example, the
EBBN method [13], the Cain calculator [15], and the functional
interpolation method [14] build the BBN models from elicited
CPDs of selected parent congurations (although the methods
differ in terms of which parent congurations are selected). In
contrast, Fenton et al. method [16] requires the denition of an
inuence function and the assessment of a number of shape
parameters. Tables 4 and 5 show the positions of the anchors
required by the methods for BBN 1 and BBN 2, respectively (Fenton
et al. method does not require a dened set of anchors).
In the denition of the anchor positions, it was decided to match
the requirement of [14], because the latter involves the lowest
number of anchors for BBN 1 (Table 4). In case one method requires
information beyond the basis of Figs. 3 and 5, the CPDs produced by
the method in [14] are used as the basis for this information. An
alternative approach could have been to introduce additional judgments to the base of Figs. 3 and 5 to satisfy each method requirements. The former approach was preferred to avoid that the
Fig. 5. Assumed expert judgment knowledge base: most likely child state at anchors (BBN 2).
4.1. The functional interpolation method
Two assumptions are required for the application of the method.

The rst concerns the approximating functions, the second concerns
the interpolation rule. As done in [14], in the present paper, Normal
functions are used (the parameters would be the mean and the
standard deviation), and the interpolation is linear on the factor states.
The parameters of the approximating functions are obtained by
minimizing the sum of the squared difference between each value
of the anchor CPDs and the corresponding value of the Normal
function (appropriately normalized so that the values sum to 1).
The Normal function is dened on an underlying continuous scale:
the scale ranges from -1 to 1 with the NEF, MEF, and EF state
corresponding to the values of 1, 2, 3, respectively (similarly, for
the other underlying factor scales). This underlying scale is only
used for the interpolation, and, given the linearity of the interpolation across the states, has no effect on the nal CPDs produced. The reader can refer to [14] for more details.
4.1.1. Summary description of the functional interpolation method

[14]
The method populates the missing CPDs in three steps. First, the
available CPDs (anchors assessed by expert judgment) are approximated by functions, described by parameters. Second, the parameters
of the full set of approximating functions are obtained by interpolating among the available CPDs. In the last step, the approximating
functions are then discretized back to obtain the full set of CPDs.
4.1.2. Application of the functional interpolation method

Tables 7 and 8 show the values of the parameters of the functions
approximating the anchor CPDs (grey shadow cells) and of the interpolated values (white cells) for BBN 1, NOUNC and UNC cases,
respectively. Note the values of the parameter are the same in the
two cases, while those of parameter are different. In the NOUNC case
(Table 7, right) all anchor points are associated with low uncertainty
( 0.2); this is in line with the knowledge base anchor CPDs of Fig. 6
differences in the produced BBNs would be due to different base

judgments, as opposed to differences in the methods.
4. Five methods for building CPTs

Table 6 summarizes the steps involved in the ve lling-up
methods used in the study. The applications of the methods on BBN
1 are presented in Sections 4.1 through 4.5. Each section is dedicated
to one method and is divided into two parts: description of the
method and application to BBN 1, cases UNC and NOUNC. The
application of all ve methods to BBN 2, UNC and NOUNC cases, is
presented in Section 4.6. Section 5 presents the method evaluation.
Table 3
CPT building methods evaluation criteria.
Description
Criteria 1 Factor inuences and
interactions
Criteria 2 Representation of
uncertainty
Criteria 3 Elicitation burden
Specic aspect investigated
Implementation in the knowledge base
Difference importance of factors
HEP takes different values if factors are one at a time in their EF state (e.g. BBN 1 case
NOUNC: HEP high if HMI EF and TC VL; HEP Medium if HMI SSF and TC VH)
Dominance of one factor in one
HEP is peaked on high if HMI EF, independently on the state of the other factors (BBN
state
1 and BBN 2 in NOUNC and UNC cases)
Combined error-forcing effect
In BBN 2 (NOUNC and UNC cases):
HEP High if TC VH, TP VH, HMI SSF
Different uncertainty in the
BBN 1, case UNC: uncertainty in HEP state for HMI SSF and TC VH larger than other
relationships
anchors
BBN 2, case UNC: uncertainty in HEP state for HMI SSF, TC VH, TP VL (anchor 3) and
HMI SSF, TC VH, TP VH (anchor 7) larger than other anchors
Increase in elicitation burden with BBN 2 larger than BBN 1 by one node
BBN size increase
Note: The implementation of the HRA aspects in the knowledge base aims at testing the methods: it does not represent any cognitive model or HRA method in particular.
Fig. 6. Assumed expert judgment knowledge base: CPDs for the parent congurations selected as anchors for BBN 1; NOUNC case, on the left; UNC case, on the right.

Table 4
Position of the elicited anchors, BBN 1, NOUNC and UNC cases.
Table 5
Position of the elicited anchors, BBN 2, NOUNC and UNC cases.
(1). In UNC case (Table 8, right), the values of the parameter reect
the different uncertainty in the anchor CPDs, with anchor 3
({HMISSF, TCVH}, bottom left) being associated the largest
uncertainty.
The full sets of CPDs for the two cases are plotted in Figs. 7 and 8
where the dark blue bars correspond to anchor CPDs and the light
blue bars are generated (Figs. 7 and 8 show the results of all lling-up
methods, to favour the comparison). In both NOUNC and UNC cases,
the produced CPDs reect the higher inuence on the HEP of factor
HMI than of factor TC, which was featured into the expert judgment
knowledge base in Fig. 6. Indeed, in both Figs. 7 and 8 subplot (a), if
HMI is in its EF state (last column), the CPD does not depend on the TC
state and replicates the anchor CPD. Concerning the rest of the CPDs,
for the NOUNC case, Fig. 7 subplot (a) shows that the CPDs are either
highly peaked on one HEP state, or, in practice, equally split on two
states. In this case of no uncertainty, the shape of the CPD results from
a narrow Normal function, determined by the values of the two
parameters and (the latter being constant and small, Table 7): if
is such that the Normal function is contained into one bar (e.g. in case
1.67 or 2.33), the highly peaked CPDs are produced; if is right inbetween two HEP states (e.g. in case 1.5 or 2.5) the equally split
CPDs are produced. Indeed, in the NOUNC case, with no uncertainty in
the anchor CPDs, the uncertainty in the derived CPDs reects the
degree of belief in which HEP state the output variable is associated
with, rather than on the value of the output variable itself (which, for
example, in case of HMI SSF and TCNom, is narrowly peaked
on 1.5).
Moving to consider the UNC case, Fig. 8 subplot (a) shows that the
uncertainty in the anchor CPDs results in larger uncertainty reected
in the produced CPDs. In Fig. 8 the CPD shapes are determined by the
progressive shift of both parameters and . In particular, concerning
, the CPDs progressively atten as they approach anchor 3. Indeed,
in this UNC case, the uncertainty in the derived CPDs reects both the
degree of belief in which HEP state the output variable is associated
with, as well as the value of the output variable itself.
4.2. EBBN method
4.2.1. Summary description of the EBBN method [13]
This method is based on piecewise linear functions interpolating among the CPDs elicited at the anchors. The functions
relate the probability of each child node state to an inuence
factor, which is used to characterize the strength of inuence of a
particular parent conguration. The CPDs for a specic parent
4. Calculate importance weights for each

parent node
5. Derive piecewise linear functions for each
child node state
6. For each CPD, calculate probability as
weighted sum of contributions from each
parent state
4. Discretize approximating functions to
CPDs
2. Approximate anchors with functions
1. Elicit CPDs for anchors (selection of anchors depends on method)
2. Assign inuence factors for parent node

states
3. Interpolate among available CPDs to
3. Calculate joint inuence factor for each
obtain full set of approximating functions parent conguration
2. Calculate interpolation factors for parent nodes

3. Calculate missing CPDs by
using interpolation factors
1. Elicit weighted function representing

central tendency of parent effect on child
2. Assign numerical values to child and parent nodes states 2. Elicit weights representing strength of
inuence of parent node on child node
3. Elicit variances (representing experts
3. Calculate weights for parent nodes, based on relative
change in child node for full range of parent node (other uncertainty)
parents at average or nominal value)
4. Elicit the R index
4. Derive CPDs from the approximation
to a TNormal distribution
5. Calculate distance between child node states and
weighted average parent states
6. Derive CPDs from distance measures
Red et al. method [2]

Cain calculator [15]
EBBN method [13]
Functional interpolation method [14]
Table 6
Summary of steps for each method (texts in italic identies the steps requiring expert elicitation; steps in normal font are performed by the algorithms underlying the methods).
Fenton et al. method [16]
conguration are determined based on the piecewise linear

functions, inuence factors, and parent importance weights.
Inuence factors are derived for each parent state (3 4 for BBN 1)
and parent congurations (3*4 for BBN 1); the latter is referred to as
joint inuence. The inuence factors convert the ranking of a state or
conguration of states on the numeric scale (0, 1). For example, for
BBN 1, the inuencing factors of HMI are 0, 0.33, 0.66, 1 for states SSF,
Nominal, LTA, EF, respectively. The joint inuence ranges from 0 for
combination {HMI SSF, TC VL} to 1 for {HMIEF, TC VH}.
Two groups of anchors are required (their position is shown in
Table 4). The rst group includes the parent congurations associated
with the largest probability value for each state of the child node: for
BBN 1, these are the congurations associated with the largest probability of being in each of the three states of the child node, Low,
Medium and High. For both BBN 1 NOUNC and UNC cases, for the
child states Low and High, these congurations are covered by the
knowledge base of Fig. 4 and correspond to anchors 1 and 4,
respectively. As discussed in Section 3.3, the CPD produced by the
functional interpolation method is taken (Section 4.1.2). From the CPTs
shown Figs. 7 and 8 subplot (a), the conguration with the largest
probability associated to the HEP state Medium is {HMINom,
TC Nom}.
The second group of anchors is used to derive the importance
weights for each parent node. The anchors correspond to the
combinations obtained with one of the parents is in its highest
state (i.e., in the state that mostly favours a high HEP, see Section
3.3), while the remaining parent nodes are in their lowest state
(i.e., in the states that mostly favours a low HEP). For example, for
BBN 1 the congurations to derive the weights of the HMI and TC
nodes would be {HMISSF, TC VH} and {HMI EF, TC VL}.
The next step is to derive the piecewise probability functions for
each state of the child node (Fig. 9). These functions relate the
probability of each child state with the inuence factor and are
derived based on the rst group of elicited CPDs. For each child
state, the functions interpolate the elicited probability values and
the joint inuence factors corresponding to the elicited conguration (i.e. anchor 1, anchor 2 for BBN 1) and {HMI Nom, TCNom}).
The missing CPDs are then derived as follows. Given a specic
parent state conguration, the probability of each child state is
calculated as a weighted sum of contributions from each factor.
The contribution from each factor to the state probability is
obtained as the average probability over the interval of inuence
identied by the joint inuence and the single factor inuence. As
an example, Fig. 10 shows the process for parent conguration
{HMIEF, TC Nom}, with the inuencing factor for state EF of
node HMI, IF(HMI EF), equal to 1, the inuencing factor for state
Nom of node TC, IF(TC Nom), equal to 0.5, and the joint inuencing factor for the two states, JIF(HMI EF, TC Nom), equal to
0.7. For this combination, the intervals of inuence for HMI and TC
range from 0.7 to 1 and 0.5 to 0.7, respectively. For each factor, the
interval of inuence ranges from the lower to the higher value
between the IF and the JIF (Fig. 10). Then, as an example, the
probability of the child node being in state High is calculated as:
PHEP HIGHj HMI EF; TC Nom
wHMI PHMI HEP HIGHj HMI EF; TC Nom
wTC PTC HEP HIGHj HMI EF; TC Nom
4.2.2. Application of the EBBN method
The CPTs resulting from the application of this method are
shown in Figs. 7 and 8 subplot (b) for BBN 1 NOUNC and UNC
cases, respectively. The derived factor weights are as follows:
wHMI 0.75; wTC 0.25; for BBN 1 NOUNC case
wHMI 0.68; wTC 0.32; for BBN 1 UNC case
10

Table 7
Parameters of the CPD approximating functions (mean and standard deviation): anchor assessment on shaded cells; interpolated values
on white cells (BBN 1, NOUNC case).
Table 8
Parameters of the CPD approximating functions (mean and standard deviation): anchor assessment on shaded cells; interpolated values
on white cells (BBN 1, UNC case).
The weights reect the higher inuence of node HMI on the

HEP, as represented in the knowledge base in Fig. 6. In both
NOUNC and UNC cases, the values of the weights for HMI are
larger than for TC because of the larger change in CPD when
moving from anchor 1 {HMISSF, TC VL} to anchor 2 {HMIEF,
TC VL}, compared to when moving to anchor 3 {HMI SSF,
TC VH} (see Fig. 6). However, although the weights reect the
assumed importance ranking, the CPDs corresponding to HMI EF
do not reect the statement in the knowledge base that the HEP is
high independently on the node TC. Indeed, in Figs. 7 and 8 subplot (b), the CPD for {HMI EF, TCNom} does not replicate the
complete belief on HEP High as those for anchors 2 and 3 do.
The reason for this is in the derivation of the missing CPDs as
weighted sums of the individual factor contributions. All factors
contribute to the weighted sum, therefore making it problematic
to model the dominance of one for some specic states (see the
example calculation of the CPD for {HMI EF, TC Nom} from the
previous Section 4.2.1).
As mentioned, the CPD distributions result from the combination of the average values of the linear functions within the
interval of inuence for each factor. Therefore, the behaviour of the
linear functions and the factor range of inuence also determine
the uncertainty reected in the CPDs. Comparing Fig. 9 left and
right, the effect of uncertainty in anchor 3 lays in the function
values for the central interpolation point at factor inuence 0.167;
the larger uncertainty for BBN 1 UNC case compared to BBN
1 NOUNC case is reected in a lower value for the probability of
HEP being Medium (0.714 vs. 1.0), and larger values for those of
being high and low (0.143 vs. 0.0). Comparing Figs. 7 and 8 subplot
(b), it can be seen that, although the CPDs in Fig. 8 are associated
with larger uncertainty than those in Fig. 7, this difference is
indeed quite small. This is due to the fact that the factors contribute to the CPDs over a range of inuence, which smooths the
differences in the relationships uncertainty, so that the result is
less sensitive to the change in the anchor 3 CPD (compared e.g. to

the previous method in Section 4.1).
4.3. The Cain calculator
4.3.1. Summary description of the Cain calculator [15]
The Cain calculator directly exploits the positive qualitative
inuence discussed earlier in Section 3.2. In particular, when one of
the parent state increases (e.g. TC state from VL to Nom): the
probability of the lowest child state (HEPLOW) either decreases or
remains unchanged; the probability of the highest child state
(HEP HIGH) either increases or remains unchanged. For each
increment in the parent states, the Cain calculator derives two
interpolation factors associated to the change in the probabilities of
the highest and lowest child states.
The required anchor CPDs for BBN 1 are shown in Table 4. The
method requires at rst to elicit the CPDs for the two parent congurations associated with the lowest and highest effect on the child
state: for BBN 1, these combinations are {HMI SSF, TCVL} and
{HMI EF, TCVH}, corresponding to anchors 1 and 4 in Fig. 4. Then,
the method requires the CPDs of the combinations obtained by progressively increasing the state of one parent, while keeping the other
at the lowest state: for example, for parent HMI, this corresponds to
the congurations {HMISSF, TCVL}, {HMINom, TC VL},
{HMI LTA, TC VL}, and {HMIEF, TCVL}. Thus, the number of
CPDs to elicit in this step is the same as the total number of parent
nodes states (34 for BBN 1). As shown by Table 4, anchors 1, 2 and
3 of Fig. 4 cover some of the required CPDs; the remaining anchor
CPDs are taken from those produced by [14] (Figs. 7 and 8, subplot (a)).
The two interpolation factors of a parent node are calculated for
each one-state increment in the parent state. The interpolation
factors are calculated as the proportion between the change in
probability of the child states due to the one-state increment and
the change due to the maximum increment. The interpolation factor
ranges from 0 (largest change in probability of the child node due to
11
Fig. 7. The CPDs for BBN 1, NOUNC case. (a) The functional interpolation method. (b) EBBN method. (c) Cain calculator. (d) Fenton et al. method. (e1) Red et al. method (all
CPDs generated by the Red et al. method). (e2) Red et al. method (anchors from the functional interpolation method). (For interpretation of the references to color in this
gure, the reader is referred to the web version of this article.)
the change in the parent state) to 1 (smallest change in probability

of the child node due to the change in the parent state). For example
VL-Nom
IFTC
HEP HIGH indicates the interpolation factor associated
to the change in the probability of the HEP state High, when the
state of node TC increases from VL to Nom. Interpolation factors
are required for all except one of the parent nodes, independently of
the structure of the BBN and on which parent is omitted. For the
present application, the interpolation factors for node HMI are
omitted. The factors are calculated for the two extreme states of the
child node Low and High, for the present case). Thus, in total,
BBN 1 requires four interpolation factors: two (one for each extreme
state of the child) for the two consecutive increments of the parent
node states (Low to Nom and Nom to High).
The idea for lling the CPT is to progressively determine the
CPDs of the parent congurations that differ by one state of one
parent node from one conguration with known CPD. Starting
12
Fig. 8. The CPDs for BBN 1, UNC case. (a) The functional interpolation method. (b) EBBN method. (c) Cain calculator. (d) Fenton et al. method. (e1) Red et al. method (all
CPDs generated by the Red et al. method). (e2) Red et al. method ( anchors from the functional interpolation). (For interpretation of the references to color in this gure,
the reader is referred to the web version of this article.)
from the anchor congurations, the set of known CPDs progressively increases until all CPDs are populated. To derive a missing
CPD from a known CPD (differing by one state of one parent node),
the interpolation factors are used to determine the proportion of
change in the probability of the child states.
4.3.2. Application of the Cain calculator

The interpolation factors for BBN 1 NOUNC and UNC cases are
as follows.
For BBN 1 NOUNC case for HEP HIGH, the interpolation factors for each increment of the TC node states are equal to one.
13
Fig. 9. The piecewise linear functions for CPDs for the BBN 1 case NOUNC (left) and UNC (right).
Therefore, the CPD values for HEP HIGH in each column of Fig. 7
subplot (c), which are derived with the method, are the same as
those values for corresponding anchors in the column. This means
column wise in Fig. 7 subplot (c) the CPDs for HEP HIGH state
will be the same. Consequently, when HMI is error forcing the
CPDs will be skewed towards high HEPs independently of complexity change, which reects the knowledge base of this example.
The interpolation factor for HEP LOW when complexity changes
from nominal to high is equal to 0, which means that independently of base congurations each time complexity is changing
from nominal to high, the CPD value for HEP LOW is the same as
the one for the worst case scenario (anchor 4 in Fig. 4) in which
normally the belief on HEP LOW is either 0 or close to 0.
Therefore, the belief on HEP LOW in CPDs of the last row Fig. 7
subplot (c) will replicate the values of low HEPs of anchor 4 in
Fig. 4.
For BBN 1 UNC case also the CPDs corresponding to HMI EF
reect the knowledge base in that the error forcing HMI leads to
high HEPs, independently of the task complexity (Fig. 8(c), last
column). This is again because the interpolation factor for switch
of TC node from very low to nominal when HEP HIGH is very
close to one (0.954). Therefore, in practice, the CPD of {HMIEF,
TC Nom} replicates that of anchor 2. The effect of uncertainty
results in the interpolation factors for TC node state increment
from NOMINAL to VH (Table 9: 1 vs. 0.734 for HEP HIGH and 0 vs.
0.271 for HEP LOW). Generally, as expected, the effect is the
larger uncertainties in CPDs of Fig. 8 compared to Fig. 7.
Table 9
Application of the Cain calculator to BBN 1: interpolation factors for NOUNC and
UNC cases, TC node.
IFTC
VL-Nom
(HEP HIGH)
NOUNC 1.000
UNC
0.954
Nom-VH
(HEP HIGH)
VL-Nom
(HEP LOW)
Nom-VH
(HEP LOW)
1.000
0.736
0.500
0.499
0.000
0.271
distribution. The mean is determined based on the general

tendency of the effect of the parent nodes on the child node, as a
weighted function of the parent node values (on the underlying
continuous scale). Four weighted functions are proposed in [16],
Mean Average, Minimum, Maximum, Mix of Minimum and Maximum. The variance 2 represents the degree of uncertainty on the
child node value and is elicited from experts. The method is
implemented as part of the AgenaRisk software [38].
To derive the CPTs, the method requires determining the
applicable weighted function, the weights, and the uncertainty
degree. Reference [16] provides qualitative guidance on how to
determine these elements. In the present work, the assessments of
Figs. 4 and 5 are used to understand the general effect of the
parent states and to infer the appropriate weighted function. The
corresponding function parameters (weights and variances) are
chosen such to reproduce the anchoring CPDs in Fig. 6. The choice
of the appropriate function and of its parameters is not based on
an automated procedure, but on expert judgment, consistent with
the guidelines from [10].
4.4. The Fenton et al. method

4.4.1. Summary description of the Fenton et al. method [16]
This method is different from the other methods described in
previous sections because of the elicited information and the
mathematical background (the method is not based on CPD
interpolations). The derivation of the CPT is based on weighted
functions which, for each child node, dene the central tendency
of the parent inuence on the child. Similarly as the method in
[14], the Fenton et al. method is applicable for BBNs with nodes
that represent abstractions of some underling continuous quantities, typically ranging between 0 and 1. This type of nodes is
referred to as ranked [16]. For example, for BBN 1 and BBN 2, the
three states of the output node HEP discretize the underlying
continuum of the human failure probability.
The child node's probabilities are derived by associating a
doubly truncated normal distribution (TNormal) to the continuous variable underling the factor labels and discretizing it on
the range associated to each label. For each parent conguration,
the probability density (on the underlying continuous scale) for
the child node is obtained from the approximation to a TNormal
distribution TNormal (, 2, 0, 1) where the parameters and 2
are the mean and the variance of the un-truncated Normal
4.4.2. Application of the Fenton et al. method

The Weighted Maximum function has been chosen for both
BBN 1 NOUNC and UNC cases. This function favours values of the
child node close to the maximum value of one of the parents. The
function is suitable to reproduce the dominating effect of factor
HMI as represented in the knowledge base.
The corresponding CPDs are plotted in Figs. 7 and 8 subplot (d).
The values of the weights that were found to give the best match
with the knowledge base were 5 and 1 for HMI and TC, respectively (for both NOUNC and UNC cases). Note these values indeed
reect the larger importance of factor HMI when at its strongly
error-forcing state. Concerning the variance, for BBN 1 NOUNC
case the lowest possible value allowed by the software AgenaRisk
was chosen (5.0E 4, corresponding to the value of the variance of
the Normal distribution from which the TNormal is derived). For
the UNC case, the value was chosen to reproduce the CPD for
anchor 3 (HMI SSF, TC VH) (the value of 5.0E 2 has been
chosen).
Fig. 7 subplot (d) shows that, for the NOUNC case, the Fenton
et al. method produces CPDs that well reproduce the information
in the knowledge base. The representation of the uncertainty is
similar to that of the functional interpolation method (Section 4.1).
14
Fig. 10. Example of CPD calculation with the EBBN method; HMI EF, TC Nominal, BBN 1 NOUNC case.
The shape of the CPDs results from a narrow truncated Normal

function: similar to the case in Section 4.1, CPDs split across two
child states are produced if the function is centred in-between the
states. The split between the states reects the degree of belief in
which state the output HEP variable is associated.
Considering the UNC case, the produced CPDs reect the larger
uncertainty in the relationships compared to NOUNC case,
resulting from the larger value of the variance parameter (Fig. 8
subplot (d)). However, in its basic application, the Fenton et al.
method applies the same value of variance to all parent congurations. The value of 2 has been determined to reproduce the
larger uncertainty in anchor 3. The result is that the CPDs do not
reect the high condence in the other anchors 1, 2, and 4 (Fig. 8).
Note that, in case of strong dependence of the CPD on the parent
conguration, such as the present case, reference [10] suggests
differentiating the functional relationships across the combinations: each parent conguration can be associated a different
weighted function, and different values of the weights and of . To
reproduce the knowledge base underlying UNC case, functional
relationships with lower value of could be used for anchors 1, 2,
and 4. This method may allow reproducing the anchor evaluations; however, it would not provide guidance on the functional
relationships in the rest of the CPDs (for the parent state congurations which are not anchors).
4.5. The Red et al. method
4.5.1. Summary description of the Red et al. method [2]
This method derives CPDs based on exponential functions of
the distance between child node states and a weighted average
parent states. The rst step to calculate the distance is to assign
numerical values to the states of all nodes. A natural choice is to
assign equally spaced values, unless some special factor features
advise against this: the values of (0, 0.5, 1) and of (0, 0.33, 0.66, 1)
were assigned for the three-state and four-state nodes, respectively (the same values as used for the EBBN method).
The second step involves the calculation of weights for each
parent node. As suggested in [2], this entails assessing (by expert
judgment) the relative change in the expectation value of the child
node, when one parent node changes from its lowest to its highest
state (e.g. HMI from SSF to EF), while the other parent is xed to its
average (or nominal, depending on the interpretation) state (e.g.
TC set to Nominal). For the present application, the expectation
values are calculated from the CPDs of the corresponding parent
congurations, shown in Table 4 as the anchors for this method.
The weights are then derived by normalizing the relative changes
such that they sum to 1. Note that, as none of anchors is part of the
knowledge base in Figs. 4 and 5, the CPDs are taken from those
produced with the functional interpolation method in Section 4.1
For a specic parent conguration, the distance Z j of child state

j from the weighted average parents' state is calculated as the
weighted sum of the differences between the values associated to
the child state j and those of the parent states. For example, the
distance measures of the child node states High, Medium and Low
from a parental conguration {HMISSF, TC VL} are 1, 0.5 and
0 respectively.
The probability distribution for the child node state j is then
derived as:
e RZ j
P j Pn
RZ j
j1e
where the numerator determines the probability mass between n
possible states of the child node, and the denominator is a normalisation factor. The R index determines the degree of the distribution of the probability mass between child node states. The
higher the R index, the lower the probability that the child node is
in a state distant from parent node states. The value of the R index
is elicited from experts; Ref. [2] gives some guidance for this.
4.5.2. Application of the Red et al. method
The CPTs resulting from the application of this method for BBN
1 are shown in Figs. 7 and 8 (e1 and e2) for NOUNC and UNC cases,
respectively.
Two plots are presented. Figs. 7 and 8 subplot (e1) shows all
CPDs as quantied by the Red et al. method; on Figs. 7 and 8
subplot (e2), the anchor CPDs used to derive the weights replace
the corresponding quantied CPDs. The parent weights are:
wHMI 0.98; wTC 0.02; for BBN 1 NOUNC case
wHMI 0.73; wTC 0.27; for BBN 1 UNC case
Similarly as for the previously described methods, the weight
associated to the HMI node is larger than that of the TC node, for
both NOUNC and UNC cases. Note for NOUNC case the dominance
of the HMI weight: this is due to the very small difference in the
CPDs for congurations {HMI Nom, TC VL} vs. {HMINom,
TCVH} (relevant for calculating wTC) compared to the difference
in the CPDs for congurations {HMI SSF, TC Nom} vs. {HMI EF,
TCNom} (relevant for calculating wHMI). The effect of the large
difference in the weights is clearly visible in Fig. 7, where the CPDs
are, in practice, independent on the TC for all states of the HMI
node. Note that if the anchor CPDs replace the corresponding
quantied CPDs (Fig. 7, subplot (e1)) the model would reproduce
the unphysical situation that the expected HEP does not monotonically increase with the TC factor (Fig. 7, subplot (e2)). For this
reason, for this method, only the CPTs with all CPDs generated by
the Red et al. method will be discussed.
Concerning the value of the R index, similar considerations as
made for the Fenton et al. method apply. The uncertainty
15
Fig. 11. The CPTs generated with all ve methods for BBN 2, NOUNC case. (a) The functional interpolation method. (b) EBBN method. (c) Cain calculator. (d) Fenton et al.
method. (e) Red et al. method (all generated CPDs). (For interpretation of the references to color in this gure, the reader is referred to the web version of this article.)
16
Fig. 12. The CPTs generated with all ve methods for BBN 2, UNC case. (a) The functional interpolation method. (b) EBBN method. (c) Cain calculator. (d) Fenton et al.
method. (e) Red et al. method (anchors from the functional interpolation method). (For interpretation of the references to color in this gure, the reader is referred to the
web version of this article.)
represented by the CPDs is reected in one parameter, applicable

to all CPDs. For NOUNC case, a high value for the R index (R 10)
was found to well represent the high condence in the relationships. Instead, as it might be expected, UNC case has been more
problematic, because the application of the same value of R across
all CPDs does not allow representing the different degrees of
uncertainty in the relationships of the knowledge base. The plots
in Fig. 8 are obtained with the value of R 6, which was set to
reproduce (by inspection) degrees of uncertainty comparable to
that for the anchor CPDs.
4.6. Application of the methods to BBN 2
Figs. 11 and 12 show the produced CPDs for the NOUNC and UNC
cases, respectively (anchor CPDs shown in dark blue). The functional
interpolation method [14] uses all eight anchors of the knowledge
base (Fig. 5). The values of the parameter for NOUNC and UNC
cases are the same, similarly to the results for BBN 1 (Section 4.1.2,
Tables 7 and 8). NOUNC case is associated with low uncertainty
degrees for all CPDs ( 0.2); instead, in UNC case, the CPDs are
associated different uncertainty degrees, anchor 3 {HMISSF,
TCVH, TPVL} and anchor 7 {HMISSF, TCVH, TPVH} having
the largest degrees ( 0.9) (note the position on the CPT and the
numbering of the knowledge base anchors are shown in Fig. 5).
For what concerns the EBBN method, the piecewise linear
functions are derived interpolating across anchors 1 and 8 of Fig. 5
and conguration {HMI Nominal, TCVL, TP Nominal}: these
are the parent congurations with the largest probabilities for the
HEP states Low, Medium, and High, respectively. The weights
of the HMI, TC and TP nodes are obtained from anchors 2, 3 and 5,
respectively, corresponding to parent congurations with one
parent at a time in its highest state. The obtained weights are:
wHMI 0.66, wTC wTP 0.17 for BBN 2, NOUNC case

wHMI 0.57, wTC 0.24, wTP 0.19 for BBN 2 UNC case.
Compared to BBN 1, populating BBN 2 with the EBBN method
requires only one additional anchor evaluation, to generate the
weight of the additional parent node TP. Indeed, in general, to
obtain the weights, the EBBN method requires as many assessments as the number of parents.
The application of the Cain calculator requires nine anchoring
CPDs, Figs. 11(c) and 12(c). Among these nine, anchors 1, 2, 3, 5 and
8 are part of the knowledge base. With respect to BBN 1, two
additional anchor evaluations are needed: {HMI SSF, TC VL,
TP Nominal} and {HMI SSF, TC VL, TP VH}, required to calculate the interpolation factors for the additional node TP. Table 10
shows the interpolated factors for BBN 2 TC and TP nodes (as
mentioned in Section 4.3 the interpolation factors of one of parents can be omitted) for NOUNC and UNC cases:
In general, when applying the Cain calculator, the number of
additional evaluations required grows with the number of additional states the new node has.
Table 10
The interpolation factors of the Cain calculator for BBN 2 NOUNC and UNC cases.
IFTC
NOUNC
UNC
IFTP
NOUNC
UNC
VL-Nom
(HEP High)
Nom-VH
(HEP High)
VL-Nom
(HEP Low)
Nom-VH
(HEP Low)
1.000
0.954
1.000
0.736
0.500
0.499
0.000
0.271
1.000
0.999
1.000
0.959
0.500
0.521
0.000
0.042
17
Moving to the Fenton et al. method, the Weighted Maximum

function was found to be the most appropriate function to reect
the BBN 2 knowledge base, based on similar consideration as for
BBN 1. The following values of the weights were chosen:
wHMI 5, wTC wTP 1 (normalizing: 0.72, 0.14, 0.14) for

NOUNC case
wHMI 5, wTC 2, wTP 1 (normalizing: 0.62, 0.25, 0.13) for

UNC case.
Concerning the values of the uncertainty parameter, 2, the
same values as for BBN 1 were chosen, i.e. 5e 4 and 5e 2, for
NOUNC and UNC cases, respectively.
Finally, regarding the Red et al. method, compared to BBN 1,
the development of BBN 2 requires the use of two additional
anchors, to determine the weight of the additional node. The
obtained weights are:
wHMI 0.34, wTC wTP 0.33 for BBN 2, NOUNC case,

wHMI 0.46, wTC 0.24, wTP 0.29 for UNC case.
Concerning the values of the R index, similar considerations as
for BBN 1 were made, resulting in the values of 10 and 6 for
NOUNC and UNC cases, respectively.
5. Evaluation of the performance of the methods

The evaluation summary is presented in Table 11. The rst evaluation aspect relates to whether the methods reect the different
inuence of the factors on the HEP when they are one at a time in
their error-forcing state. For both BBN 1 and BBN 2 in both NOUNC
and UNC cases, all methods are able to produce models reecting the
larger importance associated to the HMI node (Table 11). This is
evident from the CPDs for BBN 1 congurations {HMI EF, TC VL}
and {HMISSF, TCVH} (Figs. 7 and 8) and BBN 2 congurations
{HMIEF, TC VL, TP VL}, {HMI SSF, TC VH, TPVL} and
{HMISSF, TC VL, TPVH} (Figs. 11 and 12): in all cases the CPDs
for congurations with HMIEF are more shifted towards the high
state, compared to the rest of the CPDs with only one factor at a time
in its error-forcing state. As discussed in the previous chapters, this
reects also in larger values for weights associated to the HMI factor,
for the methods foreseeing the calculation of weights (EBBN, Fenton
et al., Red et al.).
The second aspect relates to whether the methods represent
the dominance effect of factor HMI is its error-forcing state: when
HMI is in this state, the HEP is High, independently on the state
of TC. The relevant plots of the result gures are those corresponding to HMI EF, i.e. for BBN 1, the last column for all subplots
in Figs. 7 and 8, for BBN 2, the last column for each of the three
subplot groups in Figs. 11 and 12. This aspect is well represented
by the CPDs produced by the functional interpolation method and
the Cain calculator in which all relevant CPDs (in practice) replicate. On the contrary, the EBBN, Fenton et al. and Red et al.
methods do not model this aspect. Indeed, as can be seen in Fig. 11
(b), (d) and (e) the generated CPDs when HMI is in error forcing
state depends on the task complexity and time pressure. Modelling the dominance of HMI state in EF state becomes a challenge
for methods that are based on linear combinations of the factor
effects. For the EBBN method, this issue has already been highlighted by BBN 1 (Section 4.2.2) as connected with the use of a
weighted average as the basis for the calculation of the CPDs.
Similar considerations apply to the Fenton et al. and Red et al.
methods. Concerning Fenton et al. although the use of the WMAX
function allows modelling some level of dominance of one factor,
the dominance of the HMI node in state EF is not represented
18
Table 11
Summary evaluation of the CPT building methods.
Criteria #
Description
Specic aspect investigated
Functional int.
EBBN
Cain
Fenton et al.
Red et al.
Factor inuences and interactions
Representation of uncertainty
Elicitation burden
Exponential
2
3
Difference importance of factors

Dominance of one factor in one state
Combined error-forcing effect
Different uncertainty in the relationships
Increase of elicitation burden with BBN size increase
Linear
Linear
when the other two nodes are in their SF states (see conguration
{HMIEF, TC VL, TP VL} in Fig. 11(d)). The deviation from the
knowledge base of the produced CPDs is even stronger when
adopting the Red et al. method (see for example the CPD for
conguration {HMIEF, TC VL, TP VL} in Fig. 11(e) has a maximum on HEP Low). In particular, as presented in Section 4.6 for
NOUNC case the three nodes are assigned very similar weights and
therefore do not differentiate in the factor importance. The value
of the weights is determined by the selected anchor CPDs, which
in the case of the Red et al. method result in very similar
expected values for the two congurations used to calculate the
weights (see the CPDs for {HMI SSF, TC Nom, TP Nom} vs.
{HMIEF, TC Nom, TP Nom}, {HMI Nom, TC VL, TP Nom}
vs. {HMINom, TC VH, TP Nom}, and {HMI Nom, TCNom,
TP VL} vs. {HMINom, TC Nom, TP VH}.
The third evaluation aspect relates to the interaction of the two
factors TC and TP when in their Very high state, while the HMI
factor is strongly success-forcing (Anchor 7 of the knowledge base,
i.e. conguration {HMISSF; TC VH; TP VH}). In this case, the
methods give remarkably different results (Figs. 11 and 12: relevant
plots are those corresponding to {HMISSF; TCVH; TP VH}). The
knowledge base of Fig. 5 represents the strong belief in the HEP node
state High, resulting from the combined effect of the two TC and TP
nodes being both in highest states. Only the functional approximation [14] and the Red et al. methods reproduce this belief, while all
other methods produce CPDs peaked on the Medium state. The
Red et al. method does not produce dominant weights, therefore
any of two parents being in the error forcing states will lead to HEPs
peaked on the high state (however, note that for the Red et al.
method this effect is not well reproduced in the BBN 2 UNC case
(Fig. 12). In general, anchor 7 is directly used only by the functional
approximation method. The EBBN method and Cain calculator build
on anchor CPDs obtained by varying the state of one node at a time.
The result is that the methods are challenged when the effect of
multiple nodes acting simultaneously is very different from when
they act alone. Concerning Fenton et al. method, the CPD for anchor
7 results from two opposing effects. On the one hand, the use of the
Weighted Maximum function favours high values of the child node
in presence of high values of one or more parent nodes. On the other
hand, the state Strongly success-forcing of the most important
node HMI tends to favour low values of the child state. Similar to
BBN 1 UNC case, different expressions could be used in the Fenton
et al. method to model different inuences of the parent states based
on their states. However, the method is easily applicable if the same
functional relationship is used for all parent combinations of a child
node. If different relationships are used to accommodate known
effects in specic combinations, then guidance is missing on which
relationships to use for the combinations where enough information
is not available. For modelling PSF relationships, the functional
approximation method appears to be the most exible. This may not
be surprising because, in its conceptual formulation, it uses information on single parameter effects and on all factor interactions (all
pairs, triplets, and so forth).
The second criterion deals with the treatment of uncertainty.
Fig. 12 shows the performance of the methods for BBN 2 UNC case,
characterized by larger uncertainty in the CPDs associated to
Linear
Linear
anchors 3 and 7 (Fig. 5). The gure conrms the considerations

made for BBN 1 concerning the different capability of reproducing
different degrees of uncertainty across the relationships. On the
one hand, the functional approximation method and the Cain
calculator produce CPDs reecting the different degrees featured
in the knowledge base, producing both peaked CPDs (e.g. see
congurations corresponding to {HMI EF and TP Nom}, last
column of the second group of subplots of Fig. 12) as well as more
uncertain ones. On the other hand, the EBBN, Fenton et al. and
Red et al. methods have difculties in differentiate. Similar
explanations as made for BBN 1 apply and are not repeated here.
Similar to the earlier discussion on PSF relationships, the functional interpolation method demonstrates the largest exibility to
represent varying degrees of uncertainty in the BBN relationships.
This results from the adoption of anchor points covering both
single and multiple factor effects and from the use of the uncertainty parameter for the determination of the interpolating
functions.
Concerning the last criterion, indeed, all CPT building methods
aim at reducing the number of CPDs to be determined to populate
the complete CPT. However, as the size of the BBN increases, the
corresponding increase in the information required is very different across the methods.
The amount of information required by the EBBN, Cain, Fenton
et al. and Red et al. methods increases linearly with the BBN size.
As shown by the BBN 1 and BBN 2 examples, the EBBN method
requires one additional CPD for each parent added to the BBN. The
Red et al. method requires elicitation of two additional CPDs for
each additional parent node which are necessary to derive the
weight of the additional node. Instead, for the Cain calculator, the
number of CPDs increases with the number of parent states. In the
Fenton et al. method, similar to the EBBN method, the information
requirement increases with the number of parents; however, differently from the EBBN method and Cain calculator, the information required is the value of the weight of each additional parent,
instead of CPDs (assuming that the introduction of the new parents would not change the functional relationship between parent
and child nodes).
With regard to the selected ve methods in this paper, the
functional interpolation, EBBN method, Cain calculator and Red
et al. method require elicitation of CPDs in selected anchors;
however the anchors differ depending on the requirements of the
method.
Concerning the number of elicited anchors, the Cain calculator
requires the largest number of elicitations (seven out of 12 for BBN
1 and nine out of 36 for BBN 2). The functional interpolation
method asks the CPDs of four and eight parent congurations
(linked to extreme scenarios) for BBN 1 and 2, respectively.
However, this method requires exponentially increasing anchor
evaluations when the number of parent nodes increases. For
example, with respect to BBN 1 for BBN 2 four additional assessments are required whereas for EBBN method only one additional
anchor evaluation is needed.
The EBBN method is the only method among the discussed four
methods for which the number of elicit anchors not only depends
on the number of parent nodes but also on the number of states of
19
the child node. For each additional state of the child node, the
EBBN method requires an additional anchor evaluation. For
example, if for BBN 1 the HEP child node would be associated with
a set of ve states {VERY LOW, LOW, MEDIUM, HIGH, VERY HIGH}
then with respect to BBN 1 two additional elicitations are necessary which correspond to congurations giving the largest probability to the additional states.
The analysis presented in this paper allows pinpointing to two

main aspects that CPD building methods need to be able to
address: strong factor inuences (single and multi-factor) and
proper uncertainty characterization. The functional interpolation
method is better suited to represent these, but it becomes too
costly as the BBN size increases. The latter limitation would be
worth further investigations.
6. Conclusions
References
This paper focused on methods to reduce the expert elicitation

requirements in populating the CPTs of a BBN. The performance of
ve methods of literature has been analysed, with reference to two
simple BBNs. The methods are based on extracting model information elicited from selected CPDs or specic inuences, and
extrapolating this to the whole CPT. The use of the small examples
allowed detailed comparison of the CPDs produced by the methods. The ve methods were applied on a common knowledge base,
hypothetically elicited from experts. The knowledge base was
designed to test the methods on some modelling aspects important for HRA: strong inuence of single and multiple factors, different degrees of uncertainty in the relationships, growth of elicitation burden with the BBN size. Indeed, these aspects may be
relevant beyond HRA, therefore the conclusions can be generalized
to other elds sharing similar characteristics. A challenge for the
present analysis has been that the information required by the
methods is generally different in type and model characteristics to
be elicited. The solution has been to present the knowledge base in
both qualitative and quantitative terms, such to satisfy different
information requirements while maintaining the same physical
properties behind the knowledge base. Also, the methods require
different anchor CPDs, which may not necessarily be covered by
the knowledge base. In this case, the CPDs produced by one
method (the functional interpolation method) were used as
additional anchor CPDs. This allows a fair comparison: the differences in the CPDs produced are due to the differences in the
building methods, as opposed to differences in elicited CPDs.
The analysed methods indeed allow reducing the elicitation
effort required to populate the CPTs. The effort grows linearly with
the BBN size for all methods except for the functional interpolation. In terms of modelling capabilities, all methods allow representing the different importance of the various inuencing factors;
yet, challenges arise when these inuences become very strong. A
major challenge is when representing multi-factor inuences.
Indeed, in HRA models, factors interact: their combined effect is
generally very different from when they act alone. The representation of these interactions is problematic for those methods
eliciting information on the inuence of factors taken one at a
time (the EBBN, the Cain calculation, Red et al.). Proper representation of uncertainty in the factor relationships is also a very
important requirement. In this respect, the methods based on
functional representation of the CPDs (the functional interpolation, Fenton's et al., Red et al.) are more traceable, because they
allow the explicit representation of this uncertainty. However, the
methods by Fenton's et al. and Red et al. have difculty to
represent the possibly different degrees of uncertainty in the
relationships. The method allowing the largest modelling exibility is the functional interpolation: this is perhaps not surprising
because the method requires the highest amount of information
(for larger BBNs). The method can be suitable for medium-sized,
BBNs with well-arranged inuence levels (such to reduce the
possible factor combinations) and multi-state nodes. For binary
BBNs, the method is no better than simply eliciting all factor
combinations.
[1] Francis RA, Guikema SD, Henneman L. Bayesian Belief Networks for predicting
drinking water distribution system pipe breaks. Reliab Eng Syst Saf
2014;130:111.
[2] Red W, Mosleh A, Vinnem JE, Aven T. On the use of the hybrid causal logic
method in offshore risk analysis. Reliab Eng Syst Saf 2009;94(2):44555.
[3] Groth K, Wang C, Mosleh A. Hybrid causal methodology and software platform
for probabilistic risk assessment and safety monitoring of socio-technical
systems. Reliab Eng Syst Saf 2010;95(12):127685.
[4] Jensen FV, Nielsen TD, Bayesian. Network and decision graphs. New York, NY,
USA: Springer science; 2007.
[5] Druzdzel MJ, van der Gaag L. Building Probabilistic Networks: where do the
numbers come from? IEEE Trans Knowl Data Eng 2000;12(4):4816.
[6] Fenton N, Neil M, Lagnado DA. A general structure for legal arguments about
evidence using Bayesian networks. Cognit Sci 2013;37(1):61102.
[7] Mkrtchyan L, Podollini L, Dang VN. Bayesian belief Networks for Human
reliability analysis: a review of applications and gaps. Reliab Eng Syst Saf
2015;139:116.
[8] Cooke RM. Experts in uncertainty. New York: Oxford university press; 1991.
[9] Pearl J. Fusion, propagation, and structuring in belief networks. Artif Intell
1986;29(3):24188.
[10] Fenton NE, Neil MD. Risk assessment and decision analysis with Bayesian
networks. Boca Raton, Florida, US: CRC Press; 2013.
[11] Pearl J. Probabilistic reasoning in intelligent systems: networks of plausible
inference. San Francisco, CA, US: Morgan Kaufmann; 1988.
[12] Xiang Y, Jia N. Modeling causal reinforcement and undermining for efcient
cpt elicitation. IEEE Trans Knowl Data Eng 2007;19(12):170818.
[13] Wisse BW, van Gosliga SP, van Elst NP, Barros AI. Relieving the elicitation
burden of bayesian belief networks. In: Proceedings of the sixth Bayesian
modelling applications workshop on UAI. Helsinki, Finland; 2008.
[14] Podollini L, Mkrtchyan L, Dang VN. Aggregating expert-elicited error probabilities to build HRA models. In: Proceedings of ESREL 2014. Sept. 1418,
Wroclaw, Poland; 2014. p. 108391.
[15] Cain J. Planning improvements in natural resource management. Guidelines
for using Bayesian networks to support the planning and management of
development programmes in the water sector and beyond. Wallingford, UK:
Centre for Ecology and Hydrology, Gromwarsh Gifford; 2001.
[16] Fenton NE, Neil M, Caballero JG. Using ranked nodes to model qualitative
judgments in Bayesian networks. IEEE Trans Knowl Data Eng 2007;
19(10):142032.
[17] Groth KM, Mosleh A. Deriving causal Bayesian networks from human reliability analysis data: a methodology and example mode. Proc Inst Mech Eng
Part O: J Risk Reliab 2012;226(4):36179.
[18] Gertman D, Blackman H, Byers J, Haney L, Smith C, Marble J. The SPAR-H
method. NUREG/CR-6883:US Nuclear Regulatory Commission; 2005.
[19] Williams JC. HEART a proposed method for assessing and reducing human
error. In: Proceedings of the 9th advances in reliability technology symposium.
University of Bradford; 1986.
[20] Hollnagel E. Cognitive reliability and error analysis method: CREAM. New
York: Elsevier; 1998.
[21] Chang YJ, Bley D, Criscione L, Kirwan B, Mosleh A, Madary T, Nowell R,
Richards R, Roth EM, Sieben S, Zoulis A. The SACADA database for human
reliability and human performance. Reliab Eng Syst Saf 2014;125:11733.
[22] Prvakova S, Dang VN. A review of the current status of HRA data. In: Proceedings of European safety and reliability conference (ESREL 2013). September 29October 2, Amsterdam, The Netherlands; 2013.
[23] Hallbert B, Morgan T, Hugo J, Oxstrand J, Persensky JJ. A formalized approach
for the collection of HRA data from nuclear power plant simulators. NUREG/
CR-7163, INL/EXT-12-26327 U.S. NRC Wash. D.C.; 2014.
[24] Shirley R, Smidts BC, Li M, Gupta A. Validating THERP: assessing the scope of a
full-scale validation of the technique for human error rate prediction. Ann
Nucl Energy 2015;77:194211.
[25] Preischl W, Hellmich M. Human error probabilities from operational experience of German nuclear power plants. Reliab Eng Syst Saf 2013;109:1509.
[26] Preischl W, Hellmich M. Human error probabilities from operational experience of German nuclear power plants, Part II. Reliab Eng Syst Saf
2016;148:4456.
[27] Liao H, Groth K, Stevens-Adams S. Challenges in leveraging existing human
performance data for quantifying the IDHEAS HRA method. Reliab Eng Syst Saf
2015;144:15969.
[28] Sundarmurthi R, Smidts C. Human reliability modelling for next generation
system code. Ann Nucl Energy 2013:13756.
20
[29] Groth KM, Swiler LP. Bridging the gap between HRA research and HRA
practice: a Bayesian network version of SPAR-H. Reliab Eng Syst Saf
2013;115:3342.
[30] Cai B, Liu Y, Zhang Y, Fan Q, Liu Z, Tian X. A dynamic Bayesian networks
modelling of human factors on offshore blowouts. J Loss Prev Process Ind
2013:63949.
[31] Martins MR, Maturana MC. Application of Bayesian Belief networks to the
human reliability analysis of an oil tanker operation focusing on collision
accidents. Reliab Eng Syst Saf 2013;110:89109.
[32] Vinnem JE, Bye R, Gran BA, Kongsvik T, Nyheim OM, Okstad EH, Seljelid J, Vatn
J. Risk modelling of maintenance work on major process equipment on offshore petroleum installations. J Loss Prev Process Ind 2012;25(2):27492.
[33] Langseth H, Portinale L. Bayesian networks in reliability. Reliab Eng Syst Saf
2007;92(1):92108.
[34] Neil M, Fenton N, Nielson L. Building large-scale Bayesian networks. Knowl
Eng Rev 2000;15(3):25784.
[35] Groth KM, Mosleh A. A data-informed PIF hierarchy for model-based Human
Reliability Analysis. Reliab Eng Syst Saf 2012;108:15474.
[36] Forester J, Dang VN, Bye A, Lois E, Massaiu S, Bromberg H, Braarud P, Boring
R, Mnnist I, Liao H, Julius J, Parry G, Nelson P. The international HRA
[37]
[38]
[39]
[40]
[41]
empirical study lessons learned from comparing HRA methods predictions to

HAMMLAB simulator data (NUREG-2127). Washington DC, USA: US Nuclear
Regulatory Commission; 2014.
Forester J, Kolaczkowski A, Cooper S, Bley D, Lois E. ATHEANA user's guide
(NUREG-1880). Washington DC, USA: US Nuclear Regulatory Commission;
2007.
Agena Ltd, AgenaRisk Software. www.agenarisk.com; 2007.
Wellman MP. Fundamental concepts of qualitative probabilistic networks.
Artif Intellig 1990;44(3):257303.
Helsper EM, van der Gaag LC, Groenendaal F. Designing a procedure for the
acquisition of probability constraints for Bayesian networks. Engineering
knowledge in the age of the semantic web. Berlin Heidelberg: Springer; 2004.
p. 28092.
Van Der Gaag LC, Tabachneck-Schijf HJ, Geenen PL. Verifying monotonicity of
Bayesian networks with domain experts. Int J Approx Reason 2009;50(3):
429436.

Bayesian Networks

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bayesian Networks

Uploaded by

Copyright:

Available Formats

Reliability Engineering and System Safety ()

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

Methods for building Conditional Probability Tables of Bayesian Belief

Corresponding author. Tel.: 41 56 310 53 56.

The focus of the present paper is on the development of BBNs

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

presented in detail in [8,9]). Of course, when applying lling-up

(e.g., in nuclear PSA applications, HRA for accident mitigation

2. Bayesian Belief Networks: introduction

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

Fig. 1. A simple example of a BBN.

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

predened criteria, under some constraints [4] and [10]. These

3. Design of the evaluation study

requirements as the BBN size increases the motivation for these

Fig. 2. High level design of the evaluation study.

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

assess for experts; correspondingly, some relationships may have

effect on the child state has constant direction, with monotonic

Strongly Success Forcing (SSF)

Human error probability

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

general importance of factors or different subsets of the model

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

4.1. The functional interpolation method

Two assumptions are required for the application of the method.

4.1.1. Summary description of the functional interpolation method

4.1.2. Application of the functional interpolation method

differences in the produced BBNs would be due to different base

4. Five methods for building CPTs

Criteria 3 Elicitation burden

Specic aspect investigated

Implementation in the knowledge base

Difference importance of factors

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

4. Calculate importance weights for each

2. Approximate anchors with functions

1. Elicit CPDs for anchors (selection of anchors depends on method)

2. Assign inuence factors for parent node

2. Calculate interpolation factors for parent nodes

1. Elicit weighted function representing

Red et al. method [2]

Fenton et al. method [16]

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

conguration are determined based on the piecewise linear

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

The weights reect the higher inuence of node HMI on the

less sensitive to the change in the anchor 3 CPD (compared e.g. to

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

the change in the parent state) to 1 (smallest change in probability

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

4.3.2. Application of the Cain calculator

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

distribution. The mean is determined based on the general

4.4. The Fenton et al. method

4.4.2. Application of the Fenton et al. method

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

The shape of the CPDs results from a narrow truncated Normal

For a specic parent conguration, the distance Z j of child state

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()

L. Mkrtchyan et al. / Reliability Engineering and System Safety ()