0% found this document useful (0 votes)
23 views8 pages

Continuous/Discrete Non Parametric Bayesian Belief Nets With Unicorn and Uninet

This document summarizes a continuous/discrete non-parametric Bayesian belief network system called UNINET that is used to model factors influencing air transport safety. UNINET contains 133 probabilistic nodes representing variables like pilot fatigue and 330 functional nodes capturing fault trees. It models how variables like pilot experience and training influence probabilities of human error. The network can be conditioned on specific variable values to observe how probabilities change. Conditional rank correlations are used to quantify influences between variables and elicitations from experts help infer these correlations.

Uploaded by

centscoup1
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
23 views8 pages

Continuous/Discrete Non Parametric Bayesian Belief Nets With Unicorn and Uninet

This document summarizes a continuous/discrete non-parametric Bayesian belief network system called UNINET that is used to model factors influencing air transport safety. UNINET contains 133 probabilistic nodes representing variables like pilot fatigue and 330 functional nodes capturing fault trees. It models how variables like pilot experience and training influence probabilities of human error. The network can be conditioned on specific variable values to observe how probabilities change. Conditional rank correlations are used to quantify influences between variables and elicitations from experts help infer these correlations.

Uploaded by

centscoup1
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8

Continuous/Discrete Non Parametric Bayesian Belief Nets with

UNICORN and UNINET

R.M. Cooke1, D. Kurowicka, B. Ale A. Roelen


A. M. Hanea, O. Morales, Department of Technical National Aeronautics
D. A. Ababei Administrative Science Laboratory
Department of Mathematics Delft University of Technology Amsterdam
Delft University of Technology The Netherlands The Netherlands
Mekelweg 4, 2628 CD Delft
The Netherlands
1
[email protected]

Abstract
Hanea et al. (2006) presented a method for quantifying and computing continuous/discrete non
parametric Bayesian Belief Nets (BBN). Influences are represented as conditional rank correlations, and
the joint normal copula enables rapid sampling and conditionalization. Further mathematical
background is in Kurowicka and Cooke (2007). This article sketches the current stage of development.
The driving application currently involves 133 continuous and discrete probabilistic nodes, and 330
functional nodes. Boolean functions enable fault trees to be fully represented as functional nodes in a
BBN. Repeated nodes are easily handled with the identity function. Current perspectives and challenges
conclude the paper.

1. Introduction
UNICORN is a standalone uncertainty analysis software package. The name of the package stands for
“UNcertainty analysis wIth CORrelatioNs” and its main focus is dependence modeling for high
dimensional distributions. Random variables can be coupled using a number of dependence structures.
These can be either:

• Empirical multivariate distribution,


• Markov Trees,
• Vines or,
• Bayesian Belief Nets (BBNs).

Markov Trees and Vines can be used with a user specified copula. BBNs can also be used with
arbitrary copulae (see Hanea et al. 2007), however the current implementation works only with the
normal copula.

UNINET is a continuous and discrete non parametric Bayesian belief net system, functioning as
module of UNICORN. It was developed to support a Causal Model of Air Transport Safety (CATS)
under contract with the Dutch Ministry of Transport and Waterworks (Ale et al 2005, 2006, 2007).
UNICORN is available free from http://dutiosc.twi.tudelft.nl/~risk, together with supporting scientific
documentation. This article briefly describes the driving application, and outlines the design
philosophy and main program features.
2. Driving Application
The application driving the development of continuous non parametric BBNs is a causal model for air
transport safety. The model so far covers the flight phases take-off, en route and landing. It involves
probabilistic nodes whose marginal distributions are, in most cases retrieved from field data. In a few
cases structured expert judgment is applied (Cooke, 1991). The influences between probabilistic nodes
are quantified by expert judgment (Morales et al appearing) In addition to probabilistic nodes, there
are functional nodes that capture fault tree modeling via Boolean functions. The model is still under
development, but a substantial portion has been completed and quantified. This portion involves 133
discrete and continuous probabilistic nodes and 330 functional nodes. The model is pictured below;
neither the graphic resolution, nor the purposes of this article permits a detailed picture of the
individual nodes. The probabilistic nodes are measurable variables which influence human error
probabilities. Expected values of these probabilities are fed into the fault trees, which are shown at the
upper layers of Figure 1:

Functional nodes

Figure 1: BBN for air safety, 330 functional nodes, 133 probabilistic nodes.

3. Design Philosophy
The basic design philosophy was introduced in Kurowicka and Cooke (2005) and Hanea et al. (2006).
Influence is represented as conditional rank correlation.

The conditional rank correlations between parent and child are shown in Figure 2. The order of
conditioning can be changed by the user. This representation of influence was chosen because
(Kurowicka and Cooke 2005):

• The numerical values of the conditional rank correlations are algebraically independent,
• Together with the conditional independence statements implied by the graph, the univarate
marginal distributions, and a copula realizing the correlations, they uniquely determine the
joint distribution. If zero correlation corresponds to the independent copula, then the
conditional probability statements implied by the graph are satisfied.
• Influences can be added without changing the values already chosen (unlike partial regression
coefficients),
• Conditioning must be performed by simulation or by the ‘Hybrid’method of Hanea et al 2006,
except in the case of the joint normal copula, where conditioning on probabilistic nodes can be
done analytically.
• Tested protocols for eliciting these values from experts exist, and do not depend on the
expert’s marginal distributions (see below).

Figure 2: Assignment of conditional rank correlations to nodes, for Crew Suitability En Route (Crew
suit ER)

Switching to the distribution view, we can observe the effects of conditionalization. Figure 3 shows
the unconditional distributions for Crew Suitability En Route. The horizontal histogram for FatigueER
indicates that this is a discrete distribution. Dealing with rank correlations for discrete variables is
discussed in Hanea and Kurowicka (2007). The expectations and standard deviations are shown below
each distribution. Crew suitability is influenced by the captain’s suitability and also that of the first
officer. The captain’s suitability is defined as the number of captains, out of 10,000, who would fail
the next proficiency check test. The negative correlation between the Captain’s experience in flight
hours (CapExp) and CapSuitER reflects the fact that those with more experience are less likely to fail
the “prof. check test”. Captain’s training (CapTraining) is the time since the last training course.
In Figure 4 we conditionalize on the value CapExp = 10,000 flight hours. The expected number of
captains, out of 10,000 failing the next prof check test, given that they have 10,000 hours flight
experience, drops from 234 to 171. The standard deviation drops from 235 to 166. Conditionalizing
the entire distribution in this way takes about 5 sec. on a fast PC.

Figure 3: Unconditional distributions for crewsuitER


Figure 4: Conditionalization on CapExp = 10,000 hrs.

According to this model an un-recovered loss of control from spatial disorientation in the en-route
flight phase might happen if a flight crew member is spatially disoriented AND the flight crew fails to
maintain control. Nodes in Figure 5 are Boolean functions of probabilistic nodes. The probability of
the “top” event in the unconditional joint distribution is 9.56e-8.

Figure 5: Un-recovered loss of control after spatial disorientation.

Figure 6: Un-recovered loss of control after spatial disorientation


Suppose that we condition on the expected number of captains, out of 10,000 failing the next prof
check test being equal to 1,000. After updating the joint distribution, the probability of an un-
recovered loss of control after spatial disorientation becomes approximately 4.7 times larger than in
the unconditional distribution. Observe that both the probability of a flight crew member being
spatially disoriented and the flight crew failing to maintain control have increased to 8.82e-7 and
0.325 respectively. The unconditional (gray) and conditional (black) distributions are both visible in
the FlgtcrewFail2maintctrl ER box (in Figure 4 the overlap is hardly visible).

4. Elicitation
Methods for eliciting conditional rank correlations from experts can also be applied to infer these
correlations from data, if such data are available. The method employed in this application is based on
the fact that the normal copula is used for all distributions. To assess the correlations between
CapsuitER and its parents, experts are first asked:

Suppose for a given captian, the experience is known to lie above the median value, what is now your
probability that CapsuitER is also above its median value?

If CapExp has no influence onCapsuitER, the answer will be ½; if there is a strong positive influence,
the answer will be near one, with strong negative influence it will be near zero.

After answering this question, the expert is asked:

Suppose for a given captian, the experience is known to lie above the median value AND ALSO the
time to last training is above its median value, what is now your probability that CapsuitER is also
above its median value?

The possible values are now constrained by the previous answer and by the choice of copula. An
elicitation tool was created to compute the feasible bounds for the current value given the preceding
information. A screen shot from this tool is shown below in Figure 7. The first conditional exceedance
probability, 0.7, led to a conditional rank correlation of 0.57. The second exceedance probability is
constrained to the interval [0.4, 1].

The experts in this case were commercial airline pilots, air traffic controllers, and air safety
professionals. After initial hesitation, they become quickly comfortable with these assessment tasks,
and learned to appreciate the meaning of the dependence relations. The elicitation tool proved very
helpful in this regard. Details of the elicitation procedure are described in Morales et al. (appearing).

5. Conclusion

The problems identified in Hanea et al. (2006) and Kurowicka and Cooke (2005) regarding continuous
and discrete BBNs have been largely overcome, along the lines indicated in those publications. The
examples in this paper show that the conditional rank correlation representation of influence, with the
joint normal copula, can handle large problems without excessive assessment burden, without
excessive computational burden, and without reverting to partial regression coefficients in ‘normal
units’ different from the physical units of the variables involved. Elicitation protocols are easy and
intuitive, and have been successfully applied with domain experts without special training in the
elicitation tasks. Combining BBN’s with fault trees, or indeed with any other functional nodes poses
no special problems. Of course we do not obtain a list of minimal cut sets. Repeated basic events in a
fault tree are handled simply by making them functionally dependent on one another.

Figure 7: Elicitation of conditional rank correlations via conditional exceedance probabilities

Challenges remain, however. The requisite computational speed can only be obtained with the joint
normal copula. It is hoped that other copula could be used in the future. Another challenge is posed by
conditioning on functional nodes or conditioning on intervals of probabilistic variables. In the current
framework, such conditionalization can only be preformed on samples generated by the BBN. More
sophisticated sampling and computational methods might offer other possibilities.

References
Cooke, R.M. (1991) Experts in Uncertainty, Osford University Press.

Hanea, A.M., Kurowicka, D. and Cooke, R.M. (2006). Hybrid Method for Quantifying and Analyzing
Bayesian Belief Nets, Quality and Reliability Engineering International 22, pp 709-729.

Kurowicka, D. and Cooke, R.M. (2005). Distribution - Free Continuous Bayesian Belief Nets. In A.
Wilson, N. Limnios, S. Keller-McNulty and Y. Armijo, (Eds.), Modern Statistical and Mathematical
Mathematical Methods in Reliability. pp 309-323

Morales, O., Kurowicka. D. and Roelen, A. (appearing). Eliciting Conditional and Unconditional Rank
Correlations from Conditional Probabilities, Reliability engineering and System Safety.

Hanea, A. and Kurowicka, D. (2007) Mixed Non-Parametric Continuous and Discrete Bayesian
Belief Nets , MMR, Glasgow.
Kurowicka, D. and Cooke, R.M. (2007). Sampling algorithms for generating joint uniform
distributions using the vine-copula method. Computational Statistics & Data Analysis, 51(6), pp
2889-2906 .

Ale, B.J.M., L.J. Bellamy, R.M. Cooke, L.H.J. Goossens, A.R. Hale, D. Kurowicka, A.L.C. Roelen,
E.Smith (2005). Development of a causal model for air transport safety. In Kolowrocki (Eds)S,
Advances in Safety and Reliability (ESREL), Taylor and Francis Group, London, ISBN, 0415 383404.

Ale,B.J.M., L.J. Bellamy, R.M. Cooke, L.H.J.Goossens, A.R. Hale, A.L.C.Roelen, E. Smith, (2006).
Towards a causal model for air transport safety – an ongoing research project, SAFETY SCIENCE,
44(8), pp 657–673.

B.J.M. Ale , L.J. Bellamy, R. van der Boom J., R.M. Cooke, L.H.J. Goossens, .R. Hale, D. Kurowicka
and O. Morales, A.L.C. Roelen , J. Spouge (2007). Further development of a Causal model for Air
Transport Safety (CATS); building the mathematical heart Reliability and Societal Safety – Aven &
Vinnem (eds) Taylor & Francis Group, London, ISBN 978-0-415-44786-7

You might also like