Copyright by Xiaomin You 2010
Transcript of Copyright by Xiaomin You 2010
Copyright
by
Xiaomin You
2010
The Dissertation Committee for Xiaomin You Certifies that this is the approved
version of the following dissertation:
Risk Analysis in Tunneling with Imprecise Probabilities
Committee:
Fulvio Tonon, Supervisor
Robert B. Gilbert
Lance Manuel
Timothy P. Smirnoff
Ellen M. Rathje
Risk Analysis in Tunneling with Imprecise Probabilities
by
Xiaomin You, B.E.; M.E.
Dissertation
Presented to the Faculty of the Graduate School of
The University of Texas at Austin
in Partial Fulfillment
of the Requirements
for the Degree of
Doctor of Philosophy
The University of Texas at Austin
August 2010
Dedication
This thesis is dedicated to my parents.
v
Acknowledgements
First, I would like to thank my advisor, Dr. Fulvio Tonon, for giving me this
opportunity and supporting me during the entire process. I appreciate his guidance,
patience, and insightful discussion. Without him this thesis would not have been a reality.
Sincere thanks are extended to Drs. Robert Gilbert, Lance Manuel, Timothy
Smirnoff, and Ellen Rathje for being on my committee, and for valuable comments and
advice. Special thanks are due to Dr. Timothy Smirnoff for his continued support and
traveling thousands of miles to attend my comprehensive exam and the final defense.
I want to thank my labmates: Ran, Sang Yeon, Seung Han, Yuannian, Pooyan,
and Heejung, for the collaboration and inspiring discussion on lab tests and coursework.
I also thank my friends: Songcheng, Jiabei, Rui, Manxiang, Tommy, Lu, Jianli,
and Weihong. Thank you all for making my life at Austin more fun.
I especially thank my parents for their continuous support and unconditional love
throughout my life. I would also like to thank my parents-in-law for their love and care.
Finally, very special thanks to my husband, Wen, for his enthusiasm and support. Thank
you for loving me since I was just a girl.
This work was supported by the International Tunneling Consortium (ITC).
Thanks are due to all ITC members.
vi
Risk Analysis in Tunneling with Imprecise Probabilities
Publication No._____________
Xiaomin You, Ph.D
The University of Texas at Austin, 2010
Supervisor: Fulvio Tonon
Due to the inherent uncertainties in ground and groundwater conditions, tunnel
projects often have to face potential risks of cost overrun or schedule delay. Risk analysis
has become a required tool (by insurers, Federal Transit Administration, etc.) to identify
and quantify risk, as well as visualize causes and effects, and the course (chain) of events.
Various efforts have been made to risk assessment and analysis by using conventional
methodologies with precise probabilities. However, because of limited information or
experience in similar tunnel projects, available evidence in risk assessment and analysis
usually relies on judgments from experienced engineers and experts. As a result,
imprecision is involved in probability evaluations. The intention of this study is to
explore the use of the theory of imprecise probability as applied to risk analysis in
tunneling. The goal of the methodologies proposed in this study is to deal with imprecise
information without forcing the experts to commit to assessments that they do not feel
comfortable with or the analyst to pick a single distribution when the available data does
not warrant such precision.
vii
After a brief introduction to the theory of imprecise probability, different types of
interaction between variables are studied, including unknown interaction, different types
of independence, and correlated variables. Various algorithms aiming at achieving upper
and lower bounds on previsions and conditional probabilities with assumed interaction
type are proposed. Then, methodologies have been developed for risk registers, event
trees, fault trees, and decision trees, i.e. the standard tools in risk assessment for
underground projects. Corresponding algorithms are developed and illustrated by
examples. Finally, several case histories of risk analysis in tunneling are revisited by
using the methodologies developed in this study. All results obtained based on imprecise
probabilities are compared with the results from precise probabilities.
viii
Table of Contents
List of Tables ......................................................................................................... xi
List of Figures ....................................................................................................... xv
Chapter 1 Introduction.................................................................................. 1
1.1 Research Motivation ............................................................................. 1
1.2 Literature Review.................................................................................. 5
1.2.1 Risk Analysis in Tunneling.......................................................... 5
1.2.2 Imprecise Probabilities................................................................. 7
1.3 Research Objectives and Plans ............................................................. 8
1.4 Dissertation Outline ............................................................................ 10
Chapter 2 A Short Introduction to Imprecise Probabilities ........................ 12
2.1 Basic Concepts.................................................................................... 13
2.1.1 Gamble....................................................................................... 13
2.1.2 Lower and Upper Previsions...................................................... 14
2.1.3 Geometrical Interpretation of Previsions ................................... 16
2.1.4 Lower and Upper Probabilities .................................................. 19
2.2 Two Properties of Rational previsions................................................ 20
2.2.1 Avoiding Sure Loss.................................................................... 20
2.2.2 Coherence .................................................................................. 22
2.3 Set Ψ Constructed on the basis of Specified Previsions..................... 24
2.4 Natural Extension................................................................................ 28
2.5 Decision making ................................................................................. 29
2.6 Why Use Imprecise Probabilities?...................................................... 32
Chapter 3 Different types of interaction between variables ....................... 37
3.1 Constraints on marginals..................................................................... 37
3.2 Problem Formulation .......................................................................... 39
3.2.1 Unknown interaction.................................................................. 39
3.2.2 Analysis with independent variables ......................................... 40
ix
3.2.2.1 Epistemic Irrelevance/Independence........................... 41
3.2.2.2 Conditional Epistemic Irrelevance/Independence....... 46
3.2.2.3 Strong Independence ................................................... 47
3.2.2.4 Conditional strong Independence................................ 48
3.2.3 Analysis with uncertain correlation ........................................... 49
3.3 Algorithms .......................................................................................... 50
3.3.1 Previsions and conditional probability in joint distributions ..... 51
3.3.2 Unknown Interaction ................................................................. 53
3.3.3 Independence ............................................................................. 59
3.3.3.1 Epistemic Irrelevance/Independence........................... 59
3.3.3.2 Conditional Epistemic Irrelevance/Independence....... 77
3.3.3.3 Strong Independence ................................................... 81
3.3.3.4 Conditional strong Independence................................ 86
3.3.4 Uncertain Correlation................................................................. 88
3.4 Summary ............................................................................................. 90
Chapter 4 Failure and Decision Analysis ................................................... 94
4.1 Introduction......................................................................................... 94
4.2 Failure analysis with Imprecise Probability........................................ 98
4.2.1 Event Tree Analysis (ETA) ....................................................... 98
4.2.1.1 ETA with conditional probabilities ............................. 98
4.2.1.2 ETA with total probabilities ...................................... 107
4.2.1.3 ETA with Combination of conditional probabilities and total probabilities ............................................................ 115
4.2.2 Fault Tree Analysis .................................................................. 117
4.2.3 Combination of Event Tree Analysis and Fault Tree Analysis ...... ......................................................................................... 126
4.3 Decision Analysis with Imprecise Probabilities ............................... 127
4.3.1 Standard form of decision tree................................................. 128
4.3.2 Algorithm of decision analysis with imprecise probabilities... 130
4.3.3 Decision analysis with uncertain new information.................. 136
4.3.3.1 Input Data .................................................................. 136
x
4.3.3.2 Algorithm to calculate ( )†|LOW iE X Y s− ................ 138
4.3.3.3 Discussion ................................................................. 149
4.3.4 Lower and upper values of information................................... 155
Chapter 5 Case Histories .......................................................................... 169
5.1 ETA applied to the design of a underwater tunnel ........................... 169
5.2 FTA applied to the Stockholm Ring road Tunnels ........................... 175
5.3 Decision analysis: the optimal exploration plan for the Sucheon Tunnel........................................................................................................... 184
5.4 Risk register for the East Side CSO project...................................... 203
Chapter 6 Summary and Future Work...................................................... 206
6.1 Summary ........................................................................................... 206
6.1.1 Algorithms for different types of interaction in imprecise probabilities.............................................................................. 206
6.1.2 Application to the standard tools in risk analysis .................... 207
6.1.2.1 Event tree analysis..................................................... 207
6.1.2.2 Fault tree analysis...................................................... 207
6.1.2.3 Decision tree analysis ................................................ 208
6.1.2.4 Risk register............................................................... 209
6.2 Future Work...................................................................................... 209
6.2.1 Elicitation and Assessment with Imprecise Probabilities ........ 209
6.2.2 Improvement on algorithms for different types of interaction. 210
6.2.3 Cost/Contingency and Schedule Estimation............................ 210
Appendix A Explicit Formulation for Optimization Problems in Section 4.2.1.2 .................................................................................................. 211
Appendix B Input Data and Results of Optimal Exploration Plan........... 213
Appendix C Risk register for East Side CSO Project, Portland, Oregon. 218
Reference ............................................................................................................ 229
Vita .................................................................................................................... 233
xi
List of Tables
Table 1-1 The Intervals for the frequency of occurrence in ITA Guidelines ...... 3
Table 1-2 An Example of Risk Matrix (Pennington et al., 2006)........................ 6
Table 3-1a Example 3-1: Solutions of the linear programming problems (3.45) for
the lower and upper probabilities for T............................................ 56
Table 3-2 Example 3-1: Extreme distribution of Ψ for the linear programming
problems (3.45) ................................................................................ 57
Table 3-3a Example 3-2: Solutions of the optimization problems (3.17) for lower
and upper probabilities for T............................................................ 65
Table 3-4a Example 3-2: Solutions of the optimization problems (3.17) for upper
and lower conditional probabilities that the bolt is Type B given the type
of the nut. ......................................................................................... 67
Table 3-5 Example 3-2: Solutions of the optimization problems (3.63) for lower and
upper probabilities for T................................................................... 68
Table 3-6 Example 3-2: Solutions of the optimization problems (3.63) for upper and
lower conditional probabilities that the bolt is Type B given the type of
the nut............................................................................................... 68
Table 3-7 Example 3-3: Solutions of the optimization problems (3.64) for lower and
upper probabilities for T. (same solutions for (3.65))...................... 71
Table 3-8a Example 3-3: Solutions of the optimization problems (3.64) for upper
and lower conditional probabilities that the bolt is Type B given the type
of the nut. ......................................................................................... 72
Table 3-9 Example 3-3: Solutions of the optimization problems (3.66) for lower and
upper probabilities for T................................................................... 74
xii
Table 3-10 Example 3-3: Solutions of the optimization problems (3.66) for upper
and lower conditional probabilities that the bolt is Type B given the type
of the nut. ......................................................................................... 74
Table 3-11 Example 3-3: Extreme distribution of Ψ for the linear programming
problem (3.66).................................................................................. 75
Table 3-12 Example: Solutions of the optimization problems (3.71) for upper and
lower conditional probabilities P(1,2,3)+P(1, 2 , 3 )+P( 1 ,2, 3 ). ........ 81
Table 3-13 Example 3-5: Lower and upper probabilities for T and lower and upper
conditional probability ( )1|2 1 2|P S B S B= = on all extreme distributions
of Ψ (optimal solutions are highlighted.) ......................................... 85
Table 3-14 Example: Solutions of the optimization problems (3.79) for upper and
lower conditional probabilities P(1,2,3)+P(1, 2 , 3 )+P( 1 ,2, 3 ). ........ 87
Table 3-15 Example: Solutions of the optimization problems (3.81) for upper and
lower probabilities for event T = (A, A), (B, B), (C, C). ............. 89
Table 3-16 Example: Solutions of the optimization problems (3.81) for upper and
lower probabilities for P(S1=B|S2=B). ............................................. 90
Table 3-17 Example: Solutions of the optimization problems (3.81) for upper and
lower probabilities for P(S2=B|S1=B). ............................................. 90
Table 3-18 Summary of all algorithms for different types of independence........ 93
Table 4-1: Solutions for the optimization problems (19) for the upper and lower
probabilities of failure.................................................................... 113
Table 4-2: Extreme points for the set of joint distributions of E1 and E2. ....... 114
Table 4-3: Extreme points for the set Ψ of joint distributions for Scomb and S3. combP
........................................................................................................ 114
xiii
Table 4-4: Solutions for the optimization problems (4.32) for the upper and lower
probabilities of Event E1,1. ............................................................. 124
Table 4-5: Solutions for the optimization problems (4.33) for the upper and lower
probabilities of Event E1,2. ............................................................. 124
Table 4-6: Solutions for the optimization problems (4.34) for the upper and lower
probabilities of Event E1. ............................................................... 125
Table 4-7: Extreme joint distributions of E2,1 and E2,2. ................................... 126
Table 4-8: Extreme joint distributions of E1 and E2. ....................................... 126
Table 4-9: Construction Cost Matrix............................................................... 132
Table 4-10: Exploration Reliability Matrix....................................................... 143
Table 4-11: Set of extreme joint probabilities XEXT . ....................................... 145
Table 4-12: Extreme probabilities conditional to the exploration results. ........ 145
Table 4-13: Exploration Reliability Matrix....................................................... 148
Table 4-14: Set of extreme joint probabilities, XEXT . ...................................... 149
Table 4-15: Extreme probabilities conditional to the exploration results. ........ 149
Table 4-16: Set of extreme joint probabilities XEXT . ....................................... 164
Table 4-17: Extreme probabilities conditional to the exploration results. ........ 165
Table 4-18: Extreme marginals on exploration results and real geological states ...
........................................................................................................ 167
Table 4-19: Unique optimal construction strategies and value of information . 168
Table 5-1 Success probabilities of safety measures......................................... 170
Table 5-2 Probabilities of criticality and occurrence of accident .................... 173
Table 5-3 Probabilities of accident at different risk levels .............................. 173
Table 5-4 Success probabilities of safety measures in imprecise probabilities ......
........................................................................................................ 174
xiv
Table 5-5 Probabilities of criticality and occurrence of accident .................... 175
Table 5-6 Probabilities of accident at different risk levels .............................. 175
Table 5-7 Occurrence probabilities of events at the bottom of fault-trees ...... 178
Table 5-8 Calculated occurrence probabilities of events................................. 183
Table 5-9 Description of Geologic States........................................................ 185
Table 5-10 Description of Construction Strategies.......................................... 185
Table 5-11 Construction Cost (per meter) ....................................................... 186
Table 5-12 Tunnel Section and Precise Prior Probabilities ........................... 186
Table 5-13 Exploration Reliability Matrix ...................................................... 187
Table 5-14 Imprecise Prior Probabilities ......................................................... 188
Table 5-15 Imprecise Exploration Reliability Matrix...................................... 189
Table 5-16 Optimal Construction Strategies Obtained.................................... 192
Table 5-17 Optimal Exploration Plans and the Corresponding Savings. ........ 203
Table 5-18 Description of occurrence probability in the East Side CSO project. ..
........................................................................................................ 204
Table B-1 Extreme Prior Probabilities for each tunnel section ....................... 213
Table B-2 Extreme conditional probabilities of exploration result given real
geological states ............................................................................. 215
Table B-3 Value of information (with relaxed constraints)............................. 215
Table B-4 Value of information (with strict constraints)................................. 216
Table B-5 Value of information (with precise probabilities)........................... 217
xv
List of Figures
Figure 2-1 Gamble X, the mapping from Ω to R............................................ 13
Figure 2-2 (a) 3-D space of probability of singletons; (b) probability simplex in a
3-D space ......................................................................................... 17
Figure 2-3 Geometrical Interpretation of Previsions ......................................... 18
Figure 2-4 Three lower previsions which are incurring sure loss...................... 21
Figure 2-5 Three coherent lower previsions ...................................................... 23
Figure 2-6 Three incoherent lower previsions ................................................... 24
Figure 2-7 Probability simplex .......................................................................... 27
Figure 2-8 The natural extension lower prevision of the new gamble 1 2X X+ ....
.......................................................................................................... 29
Figure 2-9 Relaxed constraint............................................................................ 31
Figure 2-10 Strict constraint .............................................................................. 32
Figure 3-1 Example 3-1: marginal sets Ψi in the 3-dimensional spaces (n1 = n2 = 3)
.......................................................................................................... 55
Figure 3-2 Set of extreme points for the case of epistemic independence......... 63
Figure 3-3 An example of Credal network ........................................................ 78
Figure 3-4 Non-convex set of joint distribution, ΨS, under strong independence82
Figure 4-1: Example of Event Tree. ................................................................... 95
Figure 4-2: Example of Fault Tree. .................................................................... 96
Figure 4-3: Example of Decision Tree. .............................................................. 97
Figure 4-4: Event-tree with N levels................................................................... 99
Figure 4-5: Event-tree in Example 4-1. ............................................................ 106
xvi
Figure 4-6: Example 4-1: sets Ψ1,0 and Ψ1,2 in the 3-dimensional spaces of the
probability of the singletons........................................................... 106
Figure 4-7: Event Tree when total probabilities are assigned to all events. ..... 111
Figure 4-8: Event Tree with mixed information consisting of conditional
probabilities and total probabilities................................................ 116
Figure 4-9: Event Tree equivalent to the tree in Figure 4-8 that contains only
probabilities conditional to the upper level events. ....................... 117
Figure 4-10: Sub-tree with OR-gate. .................................................................. 118
Figure 4-11: Sub-tree with AND-gate. ............................................................... 119
Figure 4-12: Fault tree analysis for the failure probability of sub-sea tunnel project
with imprecise probabilities........................................................... 122
Figure 4-13: (a) Decision tree, (b) its standard form, and (c) reduced decision tree
with only the optimal choices. ....................................................... 129
Figure 4-14: Decision tree for the tunnel............................................................ 132
Figure 4-15: Decision tree for the tunnel with exploration. ............................... 143
Figure 4-16: Effect of imprecision...................................................................... 152
Figure 4-17: Effect of reliability......................................................................... 152
Figure 4-18: Case (1): Intersection point in set Ψ ............................................ 154
Figure 4-19: Case (2): No intersection point in set Ψ ...................................... 155
Figure 4-20: Decision tree for the tunnel exploration......................................... 160
Figure 5-1 Construction site plan .................................................................... 170
Figure 5-2 Event tree for initiating event of poor ground conditions .............. 171
Figure 5-3 Stockholm Ring Road project plan in 1992 ...................................... 176
Figure 5-4 Fault tree for damage to lime trees due to tunneling activities ...... 179
Figure 5-5 Fault tree for Branch A .................................................................. 180
xvii
Figure 5-6 Fault tree for Branch B................................................................... 181
Figure 5-7 Fault tree for Branch C................................................................... 182
Figure 5-8 Geological profile and layout of the Sucheon Tunnel ................... 184
Figure 5-9 Decision tree for Section i of the Sucheon Tunnel without additional
exploration ..................................................................................... 190
Figure 5-10 Decision tree for Section i of the Sucheon Tunnel with imperfect
additional exploration .................................................................... 191
Figure 5-11 Decision tree for determining the value of information for Section 1 of
the Sucheon Tunnel: No additional exploration branch ................ 194
Figure 5-12 Decision tree for determining the value of information for Section 1 of
the Sucheon Tunnel: Imperfect additional exploration branch...... 195
Figure 5-13 Decision tree for determining the value of information for Section 1 of
the Sucheon Tunnel: Perfect additional exploration branch .......... 196
Figure 5-14 Value of imperfect exploration (with relaxed constraints) .......... 198
Figure 5-15 Value of imperfect exploration (with strict constraints) .............. 198
Figure 5-16 Value of perfect exploration (with relaxed constraints)............... 199
Figure 5-17 Value of perfect exploration (with strict constraints) .................. 199
Figure 5-18 Value of updating to perfect exploration (with relaxed constraints)...
........................................................................................................ 200
Figure 5-19 Value of updating to perfect exploration (with strict constraints) ......
........................................................................................................ 200
1
Chapter 1 Introduction
1.1 RESEARCH MOTIVATION
The inherent uncertainties in ground and groundwater conditions make tunnel
projects facing potential risks in cost overrun or schedule delay. Between 1994 and 2005,
catastrophic accidents in tunneling projects incurred more than $600 millions loss
(Wannick, 2006). In March 2009, the subway tunnel collapse in Cologne incurred at least
€ 400 million loss in addition to two fatalities (TunnelTalk, 2009). Risk analysis, an
important tool to identify risk, quantify risk, visualize causes and effects, and the course
(chain) of events, is becoming more and more important.
Because of the large amount of loss, tunneling insurance has become “notoriously
unprofitable” (Wannick, 2006), and insurers have not been willing to enter the tunnel
insurance market. Given this background, the International Tunnel Insurance Group
(2006) published The Code of Practice for Risk Management of Tunnel Works (ITIG
Code) to improve the insurability of tunnel projects. The ITIG code advocates a
systematic risk management approach in all phases of tunnel construction.
The International Tunnelling Association has published guidelines for tunneling
risk management (Eskesen et al., 2004). Referred to as “ITA Guidelines,” they aim to
improve present risk management significantly . Risk management is an overall term that
includes risk identification, risk assessment, risk analysis, risk elimination, and risk
mitigation and control. The ITA Guidelines stress the importance of risk analysis and
state that risk analysis should be an integral part of most underground construction
projects. In ITA guidelines, risk assessment and analysis are required as important
techniques to quantify risks and help to make decisions.
2
In conventional risk assessment and analysis, uncertainties are measured by
precise probabilities. (An example can be found in Whitman, 1984.) However,
uncertainties in the real world include two parts: stochastic uncertainties and epistemic
uncertainties. The first arise from the variance of the event itself and the second from
ignorance about the subject matter. Stochastic uncertainties are the inherent physical
property of a specific event, and we are able to perform a quantitative assessment of
variance by using classic theories of precise probabilities, such as the estimation of wind
load by a probabilistic model (Simiu et al, 2001). Epistemic uncertainties about the
subject matter can be reduced if the amount of information is expanded. Because of
limited information, assigning a precise value as to the probability of an event may not be
practical. Probability is often evaluated imprecisely. Available evidence in risk
assessment and analysis usually relies on judgments from experienced engineers and
experts. As a result, imprecision is involved in probability evaluations. For instance, the
frequency of occurrence should be evaluated by 5 predefined intervals according to the
ITA Guidelines, as shown in Table 1-1. Given the 5 predefined intervals, experts’
evaluations are estimates or imprecise judgments, for example, ‘likely to happen.’ When
using precise probabilities, the central value ‘0.1’ is used in the later analysis.
Consequently, the question of why the single value ‘0.1’ is used to represent the whole
interval ‘0.03 – 0.3’ arises. This simplification applied in the risk analysis is not
defensible.
3
Table 1-1 The Intervals for the frequency of occurrence in ITA Guidelines (Eskesen et al., 2004).
The author has participated in several risk analysis workshops for tunneling
projects. In these workshops, at the very beginning almost all probability evaluations
were given by implicit judgment such as “event A is of very high probability for
occurrence.” Then a crisp number such as “90%” was proposed to describe the “very
high probability.” If an expert expressed discomfort with this number, then it was
decreased to 80% or lower. Such experience illustrates the subjectivity involved when
evaluating probabilities. Experts also cannot distinguish between 82% and 85%
according to their experience. Subjectivity is a consequence of ignorance of the subject
matter itself and lack of information. Precision may not be necessary in such situations.
Various efforts have been made to consider the imprecision in probabilities.
Resulting theories include random sets (Matheron, 1975), fuzzy sets (Zadeh, 1965), and
probability bounds—also referred to as p-boxes (Ferson et al., 2002). These approaches
have been applied in risk analysis (Tonon et al. 2000, Huang et al. 2001). In fuzzy sets, a
membership function must be pre-defined; however, this term may not be familiar in the
tunneling industry. Normalized fuzzy sets (interpreted in a probabilistic sense) are
consonant random sets; also, equivalent random sets can be derived from p-boxes
(Bernardini and Tonon 2010, Pages 51-52, and 88-89). Indeed, random sets are the
4
special cases of imprecise probabilities. Imprecise probabilities completely reflect the
imprecision within upper and lower previsions (or expectations) of gambles (or bounded
real functions), as will be introduced in Chapter 2. The fundamental principles of this
theory are: avoiding sure loss, coherence, and natural extension. As Walley (1991, Page
2) explained:
“a probability model avoids sure loss if it cannot lead to behavior that is certain to be harmful. This is a basic principle of rationality. Coherence is a stronger principle which characterized a type of self-consistency. Coherent models can be constructed from any set of probability assessments that avoid sure loss, through a mathematical procedure of natural extension which effectively calculates the behavioral implication of the assessments.”
For example, imprecise probabilities admit and allow imprecision in probability
evaluation and assessment. Implicit judgments, like ‘A is more probable than B’, can be
applied to construct sets of probability measures. Any probability distribution compatible
with given judgments should be considered in the analysis. On the other hand,
“a central assumption of the Bayesian theory is that uncertainty should always be measured by a single probability measure, and values should always be measured by a precise utility function.” (Walley, 1991, Page 3)
A thorough motivation for imprecise probability in tunneling will be given in Section 2.6.
The intention of this study is to explore the use of the theory of imprecise
probability as applied to risk analysis in tunneling. The goal of the methodologies
proposed in this study is to deal with imprecise information without forcing the experts to
commit to assessments that they do not feel comfortable with or the analyst to pick a
single distribution when the available data does not warrant such precision. It allows
considering different types of interaction between events, including unknown interaction,
independence, and uncertain correlation. Likewise, it allows, for example, contractors to
distinguish between the maximum buying price for additional information at bid stage
5
and the minimum selling price of that information. Methodologies have been developed
for risk registers, event trees, fault trees, and decision trees, i.e. the standard tools in risk
assessment for tunnels.
1.2 LITERATURE REVIEW
1.2.1 Risk Analysis in Tunneling
Although ITIG Code states the significance of risk analysis, it does not
recommend a specific risk analysis method. Both qualitative and quantitative risk
analysis methods are recommended in ITA Guidelines. In a qualitative risk analysis, the
frequency and consequence of a hazard are rated based on predefined intervals (e.g.,
Table 1-1). According to the ratings, the risk level of a considered hazard is determined.
This qualitative method is called the ‘risk matrix method’. For example, Table 1-2 shows
the risk matrix for the East Side CSO project, Portland, Oregon (Pennington et al., 2006),
in which risks are divided into three levels: tolerable, as low as reasonably practicable
(ALARP), and intolerable. Risk register methodology is frequently used in risk
management for tunneling projects. It records all identified hazards with risk description,
probability rating, consequence rating, risk level determined by a risk matrix, and
suggested mitigations for high level risks.
While the qualitative method can provide a big picture of the risks, thereby
enabling the engineers to prepare pro-active mitigation measures, it is considered too
coarse to provide reliable quantitative risk estimates. Thus, the Monte-Carlo simulation is
often applied as the quantitative risk analysis method, but a probability distribution has to
be provided, even in the case of limited information. Some standard tools such as the
event tree, fault tree, and decision tree are suggested for risk analysis in ITA Guidelines.
6
These tools are widely used outside the tunneling industry. However, they can be and
have been used for risk analysis in tunneling without major adjustments. The event tree
analysis in the design of a TBM tunnel (Hong et al., 2009), the application of fault tree
analysis to the Stockholm Ring Road Tunnels (Sturk et al., 1996), and the optimal
exploration plan for the Sucheon Tunnel by decision tree analysis (Karam et al., 2007),
all adopt precise probabilities.
Table 1-2 An Example of Risk Matrix (Pennington et al., 2006)
Besides ITA Guidelines and ITIG Code, some guidelines developed in other civil
engineering industries can be applied in tunneling projects as well. For example, the
Association for the Advancement of Cost Engineering published a Recommended
Practice (AACE International, 2008) to provide guidelines for risk analysis by using
range estimating. In the Recommended Practice, a “double-triangular” probability
distribution must be predefined to describe the uncertainties. The U.S. Federal Highway
7
Administration also published Risk Assessment and Allocation for Highway
Construction Management (FHWA, 2006) and recommends the risk matrix method and
Monte-Carlo simulation method, which is more frequently used in cost estimation.
Federal Transit Administration also requires risk management for all federally-funded
projects (FTA, 2003). Similarly, the Washington State Department of Transportation
(WADOT, 2010) and California Department of Transportation (Caltrans, 2007) have
issued guidelines for risk management. Both recommend risk analysis methods similar to
those identified by the FHWA.
In summary, the qualitative method considers imprecise inputs but the analysis
results are coarse, while the quantitative method requires precise inputs to proceed with
the risk analysis. Considering the uncertainty inherent in tunneling projects, forcing
experts to commit to precise probability evaluation is not practical. Much needs to be
done in risk analysis in tunneling with imprecise inputs.
1.2.2 Imprecise Probabilities
With imprecise inputs, the question of how to generate a reasonable evaluation in
probability measures is critical in the risk analysis and decision-making process. One
solution is imprecise probabilities. Imprecise Probability, as proposed by Walley (1991),
is used as a generic term to cover all mathematical models which measure chance or
uncertainty without sharp numerical probabilities. All analysis conducted in this study is
by means of imprecise probabilities.
In the theory of imprecise probability, a basic problem in the risk analysis is how
to consider the interaction between events or variables. Usually, interactions include
unknown interaction, independence, and correlation. Because a set of probability
measures is considered in the theory of imprecise probabilities, the interaction between
8
events or variables in imprecise probabilities is more complex than that in precise
probabilities. Couso, Moral, and Walley (2000) firstly proposed and explicitly
distinguished the concepts of different types of independence in imprecise probabilities;
however they did not develop systematic algorithms to deal with independence. Cano and
Moral (2000) proposed algorithms for imprecise probabilities to consider different types
of independence, but the concepts of independence in this reference is different from the
definitions proposed by Couso, Moral, and Walley (2000), which are more generally
accepted in the field of imprecise probabilities. Campos and Cozman (2007) investigated
the computation of lower and upper previsions under epistemic independence (one type
of independence, as will be discussed in Chapter 3), and algorithms were developed to be
applied in a graphical model—Credal networks. For the interaction considering uncertain
correlation, the author is not aware of any reference describing a general algorithm
dealing with the uncertain correlation in imprecise probabilities, though Ferson et al.
(2004) and Berleant and Zhang (2004) considered the case that dependence between
variables are partially determined in probability bound analysis. Therefore, the challenge
remains in the area of algorithms in imprecise probability for dealing with different types
of interaction.
1.3 RESEARCH OBJECTIVES AND PLANS
We have identified three research objectives that will constitute novel work in the
field of risk analysis:
(1) Our first research objective comes from considering different types of
interactions between variables under imprecise probabilities. The objective is the creation
of efficient algorithms for unknown interaction, concepts of independence at different
levels, and uncertain correlation.
9
(2) The second research objective is the advancement of failure analysis and
decision-making with imprecise probabilities. Efficient algorithms are proposed for
failure analysis and decision analysis, respectively.
(3) The final research objective investigates the application of the proposed
methodologies to risk analysis in tunnel projects.
Our research plan consists of the sequence of actions that must be completed in
order to fulfill the three research objectives:
(1) For each type of interaction, corresponding algorithms will be developed to
calculate both prevision bounds and conditional probability bounds, subject to constraints
on marginal distributions over finite joint spaces. Two different types of constraints on
the marginals will be considered: previsions bounds and extreme distributions.
Algorithms written in terms of joint distributions or marginal distributions will be
discussed and compared. All algorithms will be justified and illustrated by simple
examples, and results will verify the influences of different interactions and the
equivalence between various algorithms.
(2) The tools adopted to improve failure analysis and decision-making are event-
tree analysis, fault-tree analysis and decision-tree analysis. In event-tree analysis,
different types of given information on probabilities will be considered; in fault-tree
analysis, various interactions between events will be taken into account; in decision-tree
analysis, we will study making decisions with imprecise probabilities and present how to
assess the value of the information. Different types of given information on probabilities,
interpretations of interactions will be discussed and compared in the study.
10
(3) The application of the proposed methodologies on risk analysis in tunnel
projects will include risk assessment with a risk register, failure analysis, and decision-
making in tunnel engineering.
1.4 DISSERTATION OUTLINE
Including this introductory chapter, the dissertation is organized as follows.
Chapter 2 presents the basic concepts in the theory of imprecise probabilities. The basic
concepts include gambles, previsions, and two important properties in imprecise
probabilities: avoiding sure loss and coherence. At the end of this chapter, the idea of
decision-making with imprecise probabilities is introduced. All concepts and properties
are illustrated by simple examples.
Chapter 3 studies different types of interaction between variables. First, we
formulate the available information on marginals, and then introduce concepts of
interaction, including unknown interaction, different types of independence, and
correlated variables. For each type of interaction, systematic algorithms are proposed,
justified, and illustrated by examples. The algorithms aim at achieving upper and lower
bounds on previsions and conditional probabilities on joint finite spaces subject to the
constraints on marginals and the assumed interaction type.
Chapter 4 presents algorithms for failure analysis (i.e. Event-Tree Analysis and
Fault-Tree Analysis) and decision analysis with imprecise probabilities. Available
information is used to construct a convex set of probability distributions that are then
considered during failure analysis and decision making. In the failure analysis, our aim is
to determine the upper and lower bounds of a prevision (expectation of a real function) or
of the probability of failure; in the decision analysis, our objective is to determine the
optimal action(s). Corresponding algorithms are developed and illustrated by examples.
11
Chapter 5 revisits several case histories of risk analysis in tunneling by using the
methodologies developed in previous chapters. Section 5.1 applies event-tree analysis
with imprecise probabilities to obtain the bounds on the occurrence probability of
accidents during the construction of an underwater tunnel. Section 5.2 deals with the
probability of the environmental damage occurring due to the construction activities of
the Stockholm Ring Road Project. Section 5.3 revisits the Sucheon Tunnel, where the
imprecision of probabilities are considered and finally the optimal exploration plan is
determined. Section 5.4 introduces the application of risk register methodology in the
East Side CSO Project in Portland, Oregon. All results obtained based on imprecise
probabilities are compared with the results from precise probabilities.
Chapter 6 presents a summary of the conclusions drawn from the dissertation and
provides recommendations for future work.
12
Chapter 2 A Short Introduction to Imprecise Probabilities
Uncertainties in the real world arise from both the variance of the event itself and
the ignorance about the subject matters. The variance is a physical property of a specific
event, and we are able to perform a quantitative assessment of variance with the classic
theory of probability (Durrett, 1996), such as the estimation of wind load by a
probabilistic model (Simiu et al, 2001). Compared with the variance independent of
human efforts, ignorance about the subject matter can be reduced if the amount of
information is expanded. Because of the limited information, assigning a precise value as
the probability of an event is not practical. In this case, the evaluation of probability is
often presented implicitly; for example, ‘A is more probable than B’. How to combine
such implicit expressions and generate a rational evaluation is critical in a decision-
making process. One solution for this problem is Imprecise Probabilities (Walley, 1991)
which measure chance or uncertainty without sharp numerical probabilities. In the theory
of imprecise probabilities, the considered probability is not unique. All probabilities
which are compatible with available information should be taken into consideration. Such
probabilities form a convex set Ψ . The objective of this chapter is to introduce the
theory of Imprecise Probabilities. The starting point of this chapter is introducing the
concept of a gamble, followed by the upper and lower previsions of a gamble. Two
important properties of the upper and lower previsions are presented. Finally, the basic
concepts in decision-making with imprecise probabilities are introduced, where two types
of constraints: relaxed constraints and strict constraints are considered and explained.
13
2.1 BASIC CONCEPTS
2.1.1 Gamble
A gamble X is a bounded real-valued function on the set of possible states of
affairs Ω (Walley, 1991). In other words, a gamble in nature is a mapping from Ω to
a bounded section of real axis, as illustrated in Fig. 1, where iω denotes a possible state
of affairs and ia is the corresponding real number in the real axis, R.
Figure 2-1 Gamble X, the mapping from Ω to R
To illustrate, consider a tunnel boring machine (TBM) advancing in soft ground
as an example, often facing such problem: whether there are any existing obstructions or
not. Let 1 be the reward for no obstructions, 0 for small obstructions, and -1 for existing
obstructions. In this case, the set of possible states (also called Possibility Space) is Ω =
no obstructions ahead, small obstructions ahead, big obstructions ahead, and the
gamble X is the function as follows:
ω1
ω2
ω3
ω4
X
X
X
X
a1 a2 a3 a4
14
1
2
3
1, 0,( ) -1,
No obstructions aheadSmall obstructions aheadXBig obstructions ahead
ωωωω
=⎧⎪ == ⎨⎪ =⎩
(2.1)
A particular gamble that will be frequently used in the following is a 0-1 valued
gamble. Consider a set A⊆ Ω and a gamble X such that:
if Aω∈ , then ( )X ω =1; otherwise ( )X ω =0 (2.2)
then we call this gamble X a 0-1 Valued Gamble. Since A is an event in the possibility
space Ω, a 0-1 valued gamble identifies an event in Ω because it is a characteristic
function for that event. On the other hand, a generic gamble is a generalization of an
event, and this makes imprecise probabilities very powerful as will be explained in
Section 2.1.4 and 2.3. The 0-1 valued gamble is widely used in risk analysis, because we
are often interested whether the event will happen or not.
2.1.2 Lower and Upper Previsions
In precise probabilities, only one probability distribution P is defined on Ω, thus
the expectation of a gamble X can be obtained precisely. However, when a convex set of
probability distributions, Ψ, is considered in the theory of imprecise probabilities, for a
given gamble X there is not a unique expectation (also called prevision): there will be as
many expectation values as distributions in Ψ. For a given gamble X, let P be a distribution in Ψ and let ( )P XE be the
expectation of X calculated with distribution P. The lower prevision LOWE is defined as
follows:
( ) ( )infLOW PPE X X
∈Ψ= E (2.3)
and the upper prevision UPPE is defined as follows:
15
( ) ( ) ( ) ( )inf supUPP LOW P PP PE X E X X X
∈Ψ ∈Ψ= − − = − − =E E (2.4)
Actually, the Separation Lemma (Walley, 1991, Theorem 3.3.2, Page 133) assures that, if
Ψ is convex, then
( ) ( )minLOW PPE X X
∈Ψ= E and ( ) ( )minUPP PP
E X X∈Ψ
= − −E ( )max PP
X∈Ψ
= E (2.5)
i.e. the infimum and supremum are achieved by some P∈Ψ.
A different point of departure in the theory of imprecise probabilities is taken
when, instead of specifying Ψ, a lower prevision or a set of lower previsions is specified.
Then, the question is whether these previsions satisfy any rationality requirement and
what type of set Ψ they define. This topic is dealt with in Section 2.2.
In Imprecise Probability, the behavioral interpretation of the theory of the lower
prevision LOWE of gamble X is used to model the decision maker’s attitudes about the
gamble; namely, the value of the lower prevision LOWE reflects the decision maker’s
behavioral disposition as the maximum buying price for the gamble X; the upper
prevision, UPPE , is interpreted as the minimum selling price for the gamble X (Walley,
1991). There is no need for UPPE to be the same as LOWE ; for example, bookmakers
and insurers make their living on the difference UPPE - LOWE . Bayesians impose that
UPPE = LOWE in all circumstances—it seems highly unlikely that the highest buying
price be the same as the lowest selling price—and if so, then bookmakers and insurers
would be broke!
When the available information is limited, judgments can be given in terms of
lower previsions and upper previsions of some gambles. In the previous example of TBM
advancing in the soft ground (gamble X), if 0.1 is assigned as its lower prevision and 0.4
as the upper prevision, the two pieces of information can be written as
16
( ) 0.1LOWE X =
( ) 0.4UPPE X = (2.6)
In this example, the highest price for the decision maker to buy the gamble X is 0.1, and
the minimum selling price of the gamble X is 0.4.
Generally, the selling price of a gamble is higher than the buying price. Based on
this fact, we are able to earn a sure gain. In line with it, the upper prevision ( )UPPE X is
usually higher than the lower prevision ( )LOWE X .
Sometimes, a decision maker has to face a gamble about which he/she has no
information. In this situation, the most conservative solution in the theory of imprecise
probabilities is to set ( ) infLOWE X X= , and consequently ( )UPPE X is sup X, which are
called vacuous previsions. Bayesians deal with such situations by using a unique
probability distribution (usually choose the uniform distribution), which is improper
because (1) it cannot model the case of complete ignorance; (2) it actually has strong
behavioral implication by insisting that ( ) ( )LOW UPPE X E X= : the maximum buying price
for X is always equal to its minimum selling price.
2.1.3 Geometrical Interpretation of Previsions
Consider an n-dimensional possibility space Ω = ω1,…, ωn, and an n-
dimensional space (P(ω1),…, P(ωn)) (the space of probabilities of singletons), where the
i-th component P(ωi) is the probability of the i-th possible state. Notice that any
probability distribution is constrained by P(ωi)≥0 and ∑P(ωi)=1, which make the space
of probabilities of singletons bounded and closed. Figure 2-2(a) is an example in a 3-
dimensional space, where the space of singletons’ probabilities is the intersection of
plane ( ) ( ) ( )1 2 3 1P P Pω ω ω+ + = and three half-spaces P: P(ωi)≥0. Figure 2-2(b)
17
shows the corresponding probability simplex, which is the view of the 3-dimentional
space of the singleton’s probabilities (Figure 2-2(a)) down the line connecting point (1, 1,
1) through origin, and therefore the probability simplex is an equilateral triangle. The
three vertices of the triangle represent the occurrence of the possible states. Every point
Q inside the triangle represents a probability distribution and the probability of iω is
measured by the perpendicular distance of the point Q to the side opposite to vertex
P( iω )=1. For example, the point Q = (1/3, 1/3, 1/3) in Figure 2-2(b) indicates that 1ω
through 3ω have the same probability of 1/3.
(a) (b)
Figure 2-2 (a) 3-D space of probability of singletons; (b) probability simplex in a 3-D space
The expectation of gamble Xi: ( )ij jaω → (Figure 2-1) defines a hyper-plane in
the space of probabilities of singletons as follows: ( ) ( ) ( ) ( ) ( ) ( )1 1( ) ... ...i i i
P i j j n nX a P a P a Pω ω ω= + + + +E (2.7)
0
( )2P ω
Plane: ( ) ( ) ( )1 2 3 1P P Pω ω ω+ + =
Point(1/3, 1/3, 1/3)
Parallel to the line from (1, 1, 1) through origin
( )1P ω
( )3P ω
( )1P ω
( )2P ω ( )3P ω
1
1 1
18
and therefore a lower prevision ELOW on a gamble Xi defines a half-space Ψ(i): ( ) ( ) ( ) ( ) ( ) ( ) ( ) 1 1: ( ) ... ... ( )i i i i
P i j j n n LOW iP X a P a P a P E Xω ω ωΨ = = + + + + ≥E (2.8)
Consider again the example in Figure 2-2. Assume gamble X1 such that X1 = 0 for 1ω ω= , X1 = 1 for 2ω ω= , and X1 = -1 for 3ω ω= , and that the lower prevision for
gamble X1 is 0. Then, Ψ(1) is P: P(ω2) – P(ω3) > 0, as shown in Figure 2-3.
(a) (b)
Figure 2-3 Geometrical Interpretation of Previsions
If a lower prevision ELOW is assigned on a set of gambles K, it defines a set Ψ as
intersection of half-spaces Ψ(i): ( )iΨ = Ψ∩ , Xi ∈ K (2.9)
Since each Ψ(i) is convex, Ψ is convex as well.
If set Ψ is convex, it can also be defined by its extreme points (vertices), which
are the points that cannot be expressed as a convex combination of any other points in set
Ψ (Rockafellar 1991, Page 162). Properties of Ψ and the algorithm for calculating
extreme points are discussed in Section 2.3.
0
( )2P ω
( )1P ω
( )3P ω
( )1P ω
( )2P ω ( )3P ω
Ψ(1): P: P(ω2) – P(ω3) > 0 Ψ(1): P: P(ω2) – P(ω3) > 0 1
1
1
19
2.1.4 Lower and Upper Probabilities
Compared with probabilities, previsions are more general and can capture the
available information without losing any available information; therefore, previsions are
more fundamental than probabilities (Walley 1991). Here is why.
As noted in Section 2.1.1, 0-1 valued gambles are the characteristic functions of
events. Since the expectation of an event is equal to its probability, upper and lower
previsions of 0-1 valued gambles are upper and lower probabilities of events. Assume n
possible states in Ω: 1ω , … , nω . The expectation of a 0-1 valued gamble X defined in
Eq.(2.2) can be written as follows:
( ) ( ) ( )1 1( ) ... ...P i i n nX a P a P a Pω ω ω= + + + +E (2.10)
where 1ia = if i Aω ∈ , and 0ia = if i Aω ∉ .
Upper and lower probabilities give hyper-planes in the space of probabilities of
singletons:
( ) ( ) ( )1 1( ) ...P n n LOWX a P a P E Xω ω= + + =E (2.11)
and
( ) ( ) ( )1 1( ) ...P n n UPPX a P a P E Xω ω= + + =E (2.12)
These hyper-planes’ normal components are either 0 or 1, whereas within previsions
normals of hyper-planes can assume any values (Bernardini and Tonon 2010, Page 68).
Upper and lower probabilities cannot represent the assessments like ‘the probability of A
is at least twice the probability of B’, i.e. P(A) > 2P(B) P(A)–2P(B) > 0, because in
this case one normal component is equal to –2. Indeed, they cannot even capture the
assessment” “event A is more probable than B”, i.e. P(A) > P(B) P(A)–P(B) > 0,
because here the normal component is equal to –1. Therefore, lower and upper
probabilities are only a special case of lower and upper previsions, and are not elaborate
20
enough to define a general convex set in the space of the singleton’s probabilities. In this
sense, prevision can capture the available information without losing any available
information, as explained further in Section 2.3 (weak* compactness theorem).
2.2 TWO PROPERTIES OF RATIONAL PREVISIONS
When the available information is conveyed as lower and upper previsions on a
set of gambles, K, only the previsions that are consistent and not conflicting with others
should be used. A rational prevision should avoid sure loss and be coherent as defined
below.
2.2.1 Avoiding Sure Loss
In loose terms, a lower prevision which incur sure loss indicates irrational
judgments, because:
(1) we are willing to pay more for X∈K than the supremum that we can get back;
or
(2) the prevision of X∈K contradicts the prevision of another gamble Y∈K or of a
finite linear combination of gambles in K,
and the set Ψ of probability distributions is empty.
In mathematical terms, the lower prevision ELOW avoids sure loss if ( )1
sup 0nj LOW jj
X E X=⎡ ⎤− ≥⎣ ⎦∑ (2.13)
whenever 1n ≥ , and jX K∈ (Wally 1991, Page 68).
Take TBM advancing in soft ground as an example. Consider three gambles 1X ,
2X , and 3X defined as three pieces of information that the contractor can buy and will
give the following rewards ( in units of utility):
21
1
21
3
1, 0,( ) 0,
No obstructions aheadSmall obstructions aheadXBig obstructions ahead
ωωωω
=⎧⎪ == ⎨⎪ =⎩
(2.14)
1
22
3
0, 1,( ) 0,
No obstructions aheadSmall obstructions aheadXBig obstructions ahead
ωωωω
=⎧⎪ == ⎨⎪ =⎩
(2.15)
1
23
3
0, 0,( ) 1,
No obstructions aheadSmall obstructions aheadXBig obstructions ahead
ωωωω
=⎧⎪ == ⎨⎪ =⎩
(2.16)
Suppose that the contractor is willing to pay 0.4 units of utility for each gamble (i.e. the
lower prevision of each gamble is 0.4). As a result, the total amount paid for three
gambles is 0.4+0.4+0.4 = 1.2, while the reward obtained by the contractor from the
gambles is 1. As a result, the lower prevision has incurred a sure loss (-0.2). As shown in
Figure 2-4, every assessment ( ( ) 0.4LOW iE X = ) is represented by a half space (see
Section 2.1.3), and the three half spaces have no intersection (Ψ = ∅), thus, sure loss is
incurred by the three previsions.
Figure 2-4 Three lower previsions which incur sure loss
( )2 0.4LOWE X =
( )1P ω
( )2P ω ( )3P ω
( )3 0.4LOWE X =
( )1 0.4LOWE X =
22
2.2.2 Coherence
Once it is ensured that a lower prevision on K avoids sure loss, the next step is to
work out the full implications for buying prices of the initial assessment, i.e. for the lower
previsions on K so that the prevision for each X∈K actually bounds (is an active
constraint for) Ψ. In a mathematic term, a general definition of coherence is given by
Walley (1991, Page 73): a lower prevision ELOW is coherent for a set of gambles K if the
following condition is satisfied.
( ) ( ) 0 01sup 0n
j LOW j LOWjX E X m X E X
=⎡ ⎤− − − ≥⎡ ⎤⎣ ⎦⎣ ⎦∑ (2.17)
whenever m and n are non-negative integers and Xo and Xj are in K. If K is a linear space,
then ELOW is coherent when the following three axioms (Walley 1991, Page 63) are
satisfied:
( ) infLOWE X X≥ when X∈K (2.18)
( ) ( )LOW LOWE X E Xλ λ= when X∈K and 0λ > (2.19)
( ) ( ) ( )LOW LOW LOWE X Y E X E Y+ ≥ + when X∈K and Y∈K (2.20)
An equivalent form of Eq.(2.20) in the conjugated upper prevision EUPP is
( ) ( ) ( )UPP UPP UPPE X Y E X E Y+ ≤ + when X∈K and Y∈K (2.21)
In general, previsions are coherent when each gamble X∈K is accepted by a
decision maker, and what he/she pays for a finite combination of some gambles X∈K is
lower than the lower prevision of some gamble X0∈K.
Consider again the example of TBM advancing in soft ground introduced in
Section 2.2.1. If now the lower previsions for gamble X1 (X2) is 0.3 (0.4), and the lower
prevision for gamble (X1+X2) is assigned to be 0.8, then the contractor is willing to pay
0.3 to buy gamble X1 and 0.4 for gamble X2, i.e., he is expected to accept the price of 0.7
(i.e., 0.3+0.4) for buying both gamble X1 and gamble X2. Since 0.7 is lower than the
23
highest price that the contractor is willing to pay for (X1+X2), i.e., 0.8, then the assigned
lower prevision on the three gambles holds the property of coherence. As shown in
Figure 2-5, all three lines (that represent the previsions of X1, X 2, and X1+X2) intersect
each other. If the lower prevision of gamble (X1+X2) is 0.6 (see Figure 2-6), then the line
( )1 2 0.6LOWE X X+ = does not intersect the area constructed by ( )1 0.3LOWE X = and
( )2 0.4LOWE X = and thus the prevision is not coherent. The lower prevision
( )1 2 0.6LOWE X X+ = does not contribute in defining Ψ, i.e. it is an inactive constraint.
In order for it to intersect and be consistent with the lower prevision on X1 and X2, we
must set ( )1 2 0.7LOWE X X+ = : this is the natural extension of (X1+X2) that will be
introduced in Section 2.4.
Figure 2-5 Three coherent lower previsions
( )2 0.4LOWE X =
( )1 0.3LOWE X =
( )1P ω
( )2P ω ( )3P ω
( ) ( )1 2 0.7LOW LOWE X E X+ =
Ψ
( )1 2 0.8LOWE X X+ =
24
Figure 2-6 Three incoherent lower previsions
2.3 SET Ψ CONSTRUCTED ON THE BASIS OF SPECIFIED PREVISIONS
From Section 2.1.3, assessments specified as lower and upper previsions on a set
of gambles K, i.e. EP(Xj) ≥ ELOW(Xj) and Xj∈K, can be written in terms of linear
inequalities of probabilities of the singletons P(ωi) and thus each set P:
EP(Xj)≥ELOW(Xj); Xj∈K is convex. Suppose the specified lower prevision ELOW avoids
sure loss, i.e. set Ψ is not empty. Ψ is the intersection over all convex and closed hence
compact sets P: EP(Xj)≥ELOW(Xj); Xj∈K , and therefore Ψ is convex and compact.
The weak*-compactness theorem (Walley 1991, Pages 145 – 146) states that,
there is a one-to-one correspondence exist between lower previsions ELOW and compact
convex sets Ψ. The proof of the theorem—
“relies on the fact that any closed convex set in a locally convex topological space is the intersection of the closed half-spaces containing it.” (Walley 1991, Page 146)
( )2 0.4LOWE X =
( )1 0.3LOWE X =
( ) ( )1 2 0.7LOW LOWE X E X+ =
( )1P ω
( )2P ω ( )3P ω
( )1 2 0.6LOWE X X+ =
Ψ
25
Because set Ψ is convex and is the intersection of the closed half spaces P:
EP(Xj)≥ELOW(Xj); Xj∈K , each gamble Xj supports convex set Ψ at some point. Notice
that the one-to-one correspondence between lower previsions and convex sets is not valid
for probabilities because probability hyper-planes’ normal components are restricted to 0s
and 1s (Section 2.1.3). Given a generic convex set Ψ, any bounding plane may have any
normal vectors.
The following extreme point theorem (Walley 1991, Pages 146 – 147) i shows
that each gamble supports convex set Ψ at some extreme point:
Assume that the lower prevision ELOW avoids sure loss, and let EXT denote the set of all
extreme points of Ψ. The extreme point theorem states that:
(a) EXT is non-empty.
(b) Ψ is the smallest convex set containing EXT.
(c) If the lower prevision ELOW is coherent, for every gamble Xj∈K there is a
extreme point P in set EXT such that EP(Xj) = ELOW(Xj).
Set Ψ may be constructed by assessments (lower prevision on a set of gambles K)
and defined by its extreme points. Let |Ω| = n, and |K| = q, the general algorithm for
determining all extreme points (or vertices) is as follows:
1) In the |n|-dimensional space of the singleton’s probabilities, Ψ is bounded by q
linear inequalities, n non-negative constraints on singletons ωi, ( ) 0iP ω ≥ , and the
constraint that ∑P(ωi)=1. Consider (n-1) inequalities at a time in addition to the
constraint that ∑P(ωi)=1.
2) Write the (n-1) inequalities as equalities: an n×n system of linear equations is
obtained. Compute the unique solution P.
26
3) If P satisfies the remaining (n+q+1) – n = q + 1 constraints, P is an extreme point
of the joint distribution set Ψ, otherwise it is not.
4) Repeat 1) to 3) until all combinations of inequalities are considered.
Let us consider again the example of a TBM advancing in soft ground and
construct set Ψ by the following available information:
①. The upper prevision of gamble X1 is 0.5:
( )1 0.5UPPE X = , i.e., P( 1ω )≤0.5 (2.22)
②. No obstructions is more probable than small obstructions:
P( 1ω )≥P( 2ω ) (2.23)
③. No obstructions is more probable than big obstructions:
P( 1ω )≥P( 3ω ) (2.24)
Considering that the sum of the probabilities is equal to 1, i.e.,
P( 1ω )+P( 2ω )+P( 3ω ) =1, by using the algorithm stated previously, all effective
combinations to determine the extreme points of set Ψ are as follows
① & ②( )( ) ( )
( ) ( ) ( )
1
1 2
1 2 3
1/ 2
1
P
P P
P P P
ω
ω ω
ω ω ω
=⎧⎪
=⎨⎪ + + =⎩
(2.25)
① & ③( )( ) ( )
( ) ( ) ( )
1
1 3
1 2 3
1/ 2
1
P
P P
P P P
ω
ω ω
ω ω ω
=⎧⎪
=⎨⎪ + + =⎩
(2.26)
② & ③( ) ( )( ) ( )
( ) ( ) ( )
1 2
1 3
1 2 3 1
P P
P P
P P P
ω ω
ω ω
ω ω ω
=⎧⎪
=⎨⎪ + + =⎩
(2.27)
Solve the equation sets and obtain the three extreme probability distributions:
27
( ) ( ) ( )( )1 3 3, ,P P Pω ω ω = (0.5, 0.5, 0); it satisfies ③ (2.28)
( ) ( ) ( )( )1 3 3, ,P P Pω ω ω = (0.5, 0, 0.5); it satisfies ① (2.29)
( ) ( ) ( )( )1 3 3, ,P P Pω ω ω = (0.33, 0.33, 0.33); it satisfies ② (2.30)
Since each of the three solutions satisfies the remaining inequality, all three solutions are
extreme points.
Figure 2-7 Probability simplex
From Section 2.1.3, recall that the probability simplex is the view of Ψ down the
line connecting (1,1,1) through the origin. The three lines shown in Figure 2-7, three lines
represent the three linear inequalities in Eqs. (2.22) through (2.24) when written as
equalities; the associated half spaces indicated by arrows represent the inequalities.
Finally, an intersection is obtained, which is the set of all probability distributions
reflecting the available information. The coordinates of the three extreme points
(vertices) of the intersection Ψ are the same as the ones in Eqs. (2.28) through (2.30).
( )3P ω
( )2P ω ( )1P ω
②
③ ①
Ψ
28
2.4 NATURAL EXTENSION
Avoiding sure loss and coherence are two important properties that ensure that a
previsions on a set of gambles is rational. Given a coherent prevision on a set of
gambles K, when assessing the lower and upper previsions for a new gamble X, it is
heavy work to verify whether the prevision on X avoids sure loss and is coherent. The
solution is natural extension. Natural extension enables us to construct previsions for
new gambles from the given previsions for the specified gambles.
As stated by Walley (1991, Page 122), for a new gamble X, its lower prevision
( )LOWE X is the maximum buying price for X that is constructed from the specified
buying prices LOWE (Xj), Xj ∈ K through linear combination. This process is natural
extension. The lower prevision for the new gamble X can be obtained as follows (Walley
1991, Page 122):
( ) ( )1
sup :n
LOW j j LOW jj
E X X X E Xα α λ=
⎧ ⎫⎡ ⎤= − ≥ −⎨ ⎬⎣ ⎦⎩ ⎭∑ for some 0n ≥ ,
jX K∈ , 0jλ ≥ , α ∈
(2.31)
If the lower prevision is coherent on K, it is coherent on K∪X.
The natural extension theorem (Walley 1991, Page 136) states that the natural extension
of X∉K can be obtained as ( ) ( )( )minLOW PP
E X X∈Ψ
= E (2.32)
where set Ψ is defined in Eqs. (2.8) and (2.9).
In the example of a TBM advancing in soft ground of Section 2.2.2, let P =
(P(ω1), P(ω2), P(ω3)) be a probability distribution over the possibility space Ω. Given K = X1, X2, ( )1 0.3LOWE X = and ( )2 0.4LOWE X = , the natural extension for X = X1 + X2 is
obtained as 0.7 by applying Eq.(2.32): ( ) ( ) ( )( )1 2minLOW PE X P Pω ω
∈Ψ= + , where Ψ =
( ) ( ) 1 2: 0.3, 0.4P P Pω ω≥ ≥ . As illustrated in Figure 2-8, graphically, the natural
29
extension of the lower prevision to the new gamble 1 2X X+ is achieved at the point of
intersection for the lower prevision on the two specified gambles X1 and X2, and thus
( )1 2LOWE X X+ = 0.3+0.4 = 0.7.
Figure 2-8 The natural extension (lower prevision) of the new gamble 1 2X X X= +
2.5 DECISION MAKING
This section introduces two concepts optimal (maximal) and preference, which
are the basic concepts of decision-making within imprecise probabilities. Consider a set of gambles K, gamble X∈K is optimal (or maximal) in set K when ( ) 0UPPE X Y− ≥ for
all X∈K (Walley 1991, page 161). Here the condition of ( ) 0UPPE X Y− ≥ is equivalent
to ( ) 0LOWE Y X− ≤ , which is not a very strong condition. As long as any gamble
satisfies this condition, it is optimal. In other words, we may have multiple optimal
gambles in set K.
According to Walley (1991, page 155-156), preference is interpreted as follows:
the decision maker is disposed to choose one gamble rather than the other one, i.e. a
( )2 0.4LOWE X =
( )1 0.3LOWE X =
( )1 2 0.7LOWE X X+ =
( )1P ω
( )2P ω ( )3P ω
Ψ
30
preferred gamble in set K is always unique. Consider two gambles X1 and X2. Given event
A, gamble X1 is preferred to X2 if the following inequalities are satisfied (Walley 1991,
Page 156, coherence theorem):
( )1 2 | 0LOWE X X A− ≥ (2.33)
On the other hand, X2 is preferred to X1 if the following conditions are satisfied
( )2 1 | 0LOWE X X A− ≥ (2.34)
Otherwise, the preference (or choice) between X1 and X2 is indeterminate. A general method to calculate the lower expectation ( )|LOW i jE X X A− is by
using the natural extension theorem (Eq. (2.32)):
( ) ( )( )| min |LOW i j i jE X X A X X A∈Ψ
− = −P
E (2.35) where P is a conditional probability measure for gamble ( )i jX X− given event A,
and Ψ is the set of conditional probability measures P . Since the expectation operator
E is linear, one may rewrite Eq. (2.35) as
( ) ( )( )( ) ( )( )
,
| min |
min | |i i j j
LOW i j i j
i j
E X X A X X A
X A X A∈Ψ
∈Ψ ∈Ψ
− = −
= −
P
P P
E
E E (2.36)
where iP and jP are conditional probability measures for gambles iX and jX
given event A, respectively, and iΨ and jΨ are the sets of iP and jP , respectively.
Now ( )|LOW i jE X X A− may be obtained by solving the optimization problem
(2.36) where, besides the constraints ( i i∈ΨP , j j∈ΨP ) on the sets of probability
measures written in terms of vertices or upper and lower previsions, there are additional
constraints (called relaxed constraint and strict constraint), which are applied only if two
gambles do not affect the uncertainty on Ω. Here a simple example illustrates the
situation.
Consider a decision situation in which the construction strategy must be decided
based on the rock mass condition quantified by the Q-system of rock mass classification
31
(Barton, 1974). For example, the construction strategies consist of gamble X1 (full face
excavation with nominal support) or gamble X2 (full face excavation with extensive
support). The uncertain geologic states are: G1 (Q value > 40) or G2 (Q value < 40),
which are the only two elements in possibility space Ω. No matter which construction
strategy is selected, the uncertainty in the geologic states is the same. In imprecise
probabilities, the same uncertainty could be interpreted in two ways:
(1) Relaxed constraint: Sets of probability measures for geologic states should be the
same no matter whether X1 or X2 is chosen, i.e.
1 2Ψ = Ψ (2.37)
As shown in Figure 4-19, the relaxed constraint requires sets Ψ1 and Ψ2 be the same but
two different probability distributions P1 and P2 can be selected to carry out the analysis.
Figure 2-9 Relaxed constraint
(2) Strict constraint: the probability measures for geologic states should be the same no
matter whether X1 or X2 is chosen, i.e.
1 2P P= , 1 1P ∈Ψ , 1 2P ∈Ψ (2.38)
Ψ1
P1
Ψ2
P2
32
As illustrated by Figure 2-10, two probability distributions P1 and P2 are enforced to be
the same, which is a stronger constraint than the relaxed constraint illustrated in Figure
4-19, where the two distributions P1 and P2 are not necessarily the same.
Figure 2-10 Strict constraint
The choice between Eqs.(2.37) and (2.38) could be determined based on the context or by
the risk analyst.
However, it is worth noting that in some cases the uncertainty on possible states
could be affected by the decision making. For example, the failure probability of the
tunnel will depend on the construction strategy chosen. In this case, Eqs.(2.37) or (2.38)
should not be applied, and the only constraints are the ones that define sets 1Ψ and 2Ψ .
Details about the methodology of decision-making with imprecise probabilities
will be discussed in Chapter 4.
2.6 WHY USE IMPRECISE PROBABILITIES?
Now that the basic knowledge about the theory of imprecise probabilities has
been outlined, it is important to discuss the rationale for utilizing imprecise probabilities
in this study. Arguments for imprecise probabilities are summarized in Walley (1991,
Ψ1
P1
Ψ2
P2
33
pages 3-6). Here the author explain several of those statements in the perspective for
tunneling projects.
(1) Lack of information
Insufficient information is the most important inducement for choosing imprecise
probabilities. Imprecise probabilities can be applied to problems when information is
limited, either because it is impossible to collect more or because it is not practical or
economically feasible to do so. For example, we will never drill boreholes every meter
along the tunnel axis for a perfect geological investigation report. Imprecise probabilities
allow us to rationally reason with the available information, which may still be enough to
make decisions. If it is not enough to make decisions, the decision maker rests assured
that no unwanted information is unwittingly added to force him to make a decision
despite the lack of information (an example will be given in Section 4.3.3.3.1).
Indeed, the methods of imprecise probabilities yield a range of results according
to the amount of information available for analysis. Suppose minimal previous
experience is available from tunneling projects similar to the one proposed. The
difference between upper and lower probabilities is used to reflect that lack of
information. An extreme case is “complete ignorance”, for example, construction of a
tunnel with an entirely new construction method, where the vacuous previsions model
appropriately takes account of the maximal imprecision situation. No single precise
probability can do this. As experience is gained and information accumulates, upper and
lower probabilities become close to one another and imprecision decreases. The opposite
extreme case is the project in which a large amount of information is available. Then as a
special case of imprecise probabilities, precise probabilities are proper to use. For
34
instance, a negative exponential distribution can be used to describe the length of intact
rocks in borehole (Harrison and Hudson, 2000, Page 93).
(2) Descriptive realism
As a result of lack of information, beliefs about the subject maters are often
indeterminate. Recall the risk analysis workshops for tunneling projects introduced in
Chapter 1: experts cannot really differentiate between probabilities of 82% and 85%
based on their experience. Imprecise probabilities can model such indeterminacy in
beliefs.
(3) Conflicting assessments
Conflict or disagreement on probability and utility assessment may arise when
they are obtained from different sources even if each source is precise. For example,
consider the case in which two or more experts give different assessments on the
probability of the event “leakage in tunnel lining”. Imprecise probabilities can reflect the
extent of disagreement; for example, a simple technique is to use of intervals.
Generally, distinctly different opinions may result when experts use different
evidence or different assessment strategies. The conflict or disagreement can be
eliminated or reduced by sharing the experts’ information, or by using a more elaborate
strategy to improve the original assessments.
(4) Elicitation
The imprecise probability model enables us to express our beliefs by probabilistic
judgments like “A is more probable than B.” Although such judgments are imprecise,
they are easier to elicit than precise models. Through the process of elicitation, a set of
35
rational probability measures can be constructed to properly reflect available evidences
(see the example in Section 2.4).
(5) Natural extension
In the theory of imprecise probability, a decision maker can always determine the
bounds on the prevision or probability of a new gamble based on the specified previsions.
Remember the example in Section 2.4: the lower prevision of new gamble 1 2X X+ is
determined given ( )1 0.3LOWE X = and ( )2 0.4LOWE X = .
(6) Statistical inference
Imprecision in probability assessment decreases when statistical data
accumulates. For example, before performing any test, we may roughly guess the in-situ
stresses based on the overburden, the topography and the tectonic history of a site. After a
large amount of in-situ stress measurements, a more precise estimation of in-situ stresses
can be provided.
(7) Robust results
Robust results are not affected by a small variation in the probability measures.
The theory of imprecise probability is inherently robust, i.e., the results are
“automatically robust, because they do not rely on arbitrary or doubtful assumptions”
(Wally, 1991, Page 5).
(8) Indeterminacy
A strong argument in favor of precise probabilities is that the decision maker must
choose a unique, optimal option; however imprecise probabilities often result in
36
indeterminacy. It is difficult to assess the probabilities precisely when the information
required is not available. In such cases, if precise probabilities are imposed and a ‘best’
choice is selected, the choice may not be defensible.
Analytical techniques of imprecise probabilities may fail to yield the “best”
option. Instead, they generate multiple optimal options with indeterminate preferences.
To determine a unique optimal option by this method, two major strategies can be
adopted: (a) undertake further analysis or search for more information to determine the
“best” option; for example, drill additional boreholes to collect geological data; or (b)
arbitrarily select one among the optimal options. (Walley, 1991, Page 239) i.e., the
decision maker has the “freedom of choice,” as explained by Walley (1991, Page 241)
“Perhaps we should simply accept this degree of arbitrariness in choice, and call it ‘freedom of choice’. Rather than follow rules to eliminate it. Reasoning cannot always determine a uniquely reasonable course of action, especially when there is little information to reason it.”
37
Chapter 3 Different types of interaction between variables
This chapter studies different types of interaction between discrete variables
within imprecise probabilities. First, we formulate the available information on
marginals, and then introduce concepts of interaction, including unknown interaction,
different types of independence, and correlated variables. For each type of interaction,
systematic algorithms are proposed, justified, and illustrated by simple examples. The
algorithms aim at achieving upper and lower bounds on previsions and conditional
probabilities on joint finite spaces subject to the constraints on marginals and the
assumed interaction. All theorems presented in this chapter are developed and proved by
the author.
3.1 CONSTRAINTS ON MARGINALS
The usual situation we are facing in risk analysis is to search the upper and lower
bounds for a specific real function (or gamble) of interest with constraints on marginals.
There are two ways to present the constraints on marginals. One is to define upper and
lower previsions on marginals; the other is by specifying extreme distributions of
marginals.
First, let us consider the prevision bounds. We use bold letters for column vectors
or matrices, and corresponding non-bold letter as the components of a vector or matrix.
An n–column vector pi is a probability measure on the finite space Si= jis : j = 1,…, ni,
and its j-th component is Pi( jis ). Let fi
k : Si → , k= 1,…,ki be a set of bounded functions
(gambles) on Si. fik( j
is ) is the j-th component of the n–column vector fik. The expectation
(prevision) is:
( ) ( ) ( )1
in Tk k j j ki i i i i i i
jE f f s P s
=
⎡ ⎤ = =⎣ ⎦ ∑ f p (3.1)
38
If Ψ, a set of distributions p is given, upper and lower previsions of a new gamble
f in Ψ are calculated as follows (Eq. (2.6) in Chapter 2) [ ] [ ] [ ] [ ]max ; minUPP LOWE f E f E f E f
∈Ψ∈Ψ= =
pp (3.2)
On the other hand, if bounds on previsions on marginals, kLOW iE f⎡ ⎤
⎣ ⎦ and k
UPP iE f⎡ ⎤⎣ ⎦ , are provided, the set of distributions pi, Ψi that is compatible with the given
information, is
( ) ( ) : ; 1,..., ; 1; 0i
Tk k k T ji i LOW i i i UPP i i i inE f E f k k p⎡ ⎤ ⎡ ⎤Ψ = ≤ ≤ = = ≥⎣ ⎦ ⎣ ⎦p f p 1 p (3.3)
Alternatively, constraints may be given in terms of extreme distributions of Ψ
(Section 2.3). Let ETXi indicate the set of extreme distributions (vertices) of Ψi, and
iEXTξp be the ξ -th extreme distribution of variable Si, or the ξ -th vertex of Ψi. Thus, Ψi
is the set of convex combinations of iEXT
ξp :
1 1: , 1, 0
j j
ii i i i i iEXTc c cξ ξ
ξ ξ ξ ξ
ξ ξ= =
⎧ ⎫⎪ ⎪Ψ = = = ≥⎨ ⎬⎪ ⎪⎩ ⎭
∑ ∑p p p (3.4)
Consider a probability distribution P on a two-dimensional joint space S = ( )1 2, jis s : i = 1,…, n1; j = 1,…, n2, where the i,j-th entry of the n1×n2 matrix P is the
probability mass on a joint elementary event ( )1 2, jis s . Then the marginal pi can be written
as pi = P ( )jn1 . Here ( )jn1 is an nj-column vector of unit components. We can rewrite the
constraints in Eqs.(3.3) and (3.4) as follows, respectively:
( ) ( ) ; 1,..., ; 1,2, 2,1j
Tk k kLOW i i UPP i inE f E f k k i j⎡ ⎤ ⎡ ⎤≤ ⋅ ≤ = = =⎣ ⎦ ⎣ ⎦f P 1
( ) ( )1 2
,1; 0T i jn n p⋅ ⋅ = ≥1 P 1
(3.5)
( )1 1
, 1, 0; 1,2j j
i ii i in EXTc c c iξ ξ
ξ ξ ξ ξ
ξ ξ= == = ≥ =∑ ∑P 1 p (3.6)
The set of joint distributions Ψ is composed by all joint probabilities P compatible
with the constraints in Eq. (3.5) or (3.6) and the assumed interaction between variables.
39
3.2 PROBLEM FORMULATION
This section introduces three types of interaction between variables: unknown
interaction, independent variables, and correlated variables in Imprecise Probability.
Concepts of unknown interaction and notions of independence are introduced in Couso et
al. (1999 and 2000) and Vicig (1999). Tonon et al. (2010) and Bernardini and Tonon
(2010) further explicitly express corresponding constraints in mathematical terms.
3.2.1 Unknown interaction
Unknown interaction is applied to the situation that no information is available on
the interaction between S1 and S2. ΨU, the set of probability measures on the joint space
S, is composed of all joint probability measures that respect the marginal rules, i.e.
marginals are in Ψi.:
( )2 1P S⋅× ∈Ψ ; ( )1 2P S ×⋅ ∈Ψ (3.7)
Unknown interaction has no more constraints than Eq.(3.7). Consider a joint
distribution P with (i, j)-th entry Pi,j. By expressing the previsions as a linear function of the probability mass, 1 2; , ,
1; 1
i n j n i j i ji j
a P= =
= =∑ , the problem under unknown interaction reads as
follows (Eq.(3.5)):
Minimize (Maximize) 1 2; , ,1; 1
i n j n i j i ji j
a P= =
= =∑
Subject to ( ) ( )
( ) ( )1 2
,1 2
; 1,..., ; 1,2, 2,1
1
0; 1,..., ; 1,...,
j
Tk k kLOW i i UPP i in
Tn n
i j
E f E f k k i j
p i n j n
⎡ ⎤ ⎡ ⎤≤ ⋅ ≤ = = =⎣ ⎦ ⎣ ⎦
⋅ ⋅ =
≥ = =
f P 1
1 P 1 (3.8)
When extreme distributions are given on the marginals, the problem of (3.8) can
be written as follows:
40
Minimize (Maximize) 1 2; , ,1; 1
i n j n i j i ji j
a P= =
= =∑ Subject to
( )
( )
1
2 1
2
1 2
1 11
2 21
1
1, 0; 1,2i
n EXT
Tn EXT
i i
c
c
c c i
ξξ ξ
ξ
ξξ ξ
ξ
ξξ ξ
ξ
=
=
=
= ⋅ =
= ⋅ =
= ≥ =
∑
∑
∑
p P 1 p
p P 1 p (3.9)
Constraints in Eqs.(3.8) and (3.9) are linear, thus the set of joint distributions
under unknown interaction, ΨU, is convex. If the conditional probability ( ) ( )1 1 1
1 2 2 2, |P s s S s= is of interest, replace the
objective functions in (3.8) and (3.9) by 11,1 ,11/ i n i
iP P==∑ ; likewise for other conditional
probabilities. Although Ψ remains convex, the objective function is no longer linear.
3.2.2 Analysis with independent variables
The simplest way to consider the interaction between variables is to assume
independence. Let Pi be the probability measure on the variable Si. The probability
measure on the joint measurable space (S1, S2) for independent variables in precise
probabilities is defined as the product measure 1 2 1 2: : [0,1]i iP P P U U U S= ⊗ = × ∈ →C (3.10)
where ( ) ( ) ( )1 2 1 2 1 1 2 2P P U U P U P U⊗ × = . The conditional probability measures for
independent variables yield the marginal probability measures ( ) ( ) ( )2 1 2 1 2 2 2| : 0;P S S s P s P s⋅× × = ⋅ ∀ >
( ) ( ) ( )1 1 2 2 1 1 1| : 0P S s S P s P s×⋅ × = ⋅ ∀ > (3.11)
This means that: if we learn that the actual value of S2 is s2, then our knowledge about the
probability measure for S1 does not change. Likewise for S1. The concept of independence
is equivalent to the following two properties: 1) the probability of S1 conditional to S2 is
41
equal to the marginal of S1, and vice versa; 2) the joint probability is equal to the product
of the marginals. The equivalence exists only when each variable is subject to a single
probability distribution.
Within imprecise probabilities, the two aforementioned properties and definitions
of independence are not always equivalent. As a result, notions of independence on joint
spaces are defined at different levels, including epistemic irrelevance/independence and
strong independence (see Campos et al. 1995, Couso et al. 1999 and 2000).
3.2.2.1 Epistemic Irrelevance/Independence
Consider the finite joint space S = S1×S2. Under epistemic irrelevance of the
second variable with respect to the first variable, the probability measure over S1
conditional to S2 is always in the set of Ψ1, regardless of the value of S2; however, the
probability measure for S1 could be different for different values of S2. The definition of
epistemic irrelevance indeed uses the concept of conditional probability. The set of joint
probability measures, 2|sEΨ , is just the largest set of joint measures that are extensions to
Eq.(3.11) (Bernardini and Tonon 2010): ( ) ( ) 2|
2 1 2 1 2 2 2 2 2: . | ; : : 0sE P S S s s P P sΨ = × × ∈Ψ ∀ ∃ ∈Ψ >P (3.12)
where 2|sEΨ is called the “irrelevant natural extension (Couso et al. 1999) of the two
marginals when the second experiment is epistemically irrelevant to the first” (i.e. the set
of acceptable gambles concerning the first experiment does not change when we learn the
outcome of the second experiment). Likewise for 1|sEΨ . It is important to notice that s1
may be selected by a different procedure for different values of s2. Therefore, in
Imprecise Probabilities, irrelevance of one experiment with respect to another is a
directional or asymmetric relation. This asymmetry disappears in precise probabilities
because both Ψ1 and Ψ2 contain only one probability distribution each.
42
Let matrix P1|2 be probability distributions over S1 conditional to the values of S2,
i.e. the i-th column of P1|2 is the conditional probability distribution over S1 given that S2
is equal to its i-th value. Following Eq.(3.12), a joint probability P in 2|sEΨ is
( ) ( ) ( )1 2 1 2 1 2 2 2|P U s P U S S s P s× = × × ; P = P1|2 Diag(p2) (3.13)
and thus 2|sEΨ in Eq.(3.12) can be written as (Bernardini and Tonon 2010)
( )( )
2 2
( )2 1 1
| ( )(1) (2)1|2 2 2 2 1|2 1 1 1: ; ...
i
s nE
n
Diag∈Ψ
⎧ ⎫⎛ ⎞⎪ ⎪⎜ ⎟⎪ ⎪Ψ = = ∈Ψ = ⎜ ⎟⎨ ⎬
⎜ ⎟⎪ ⎪⎪ ⎪⎝ ⎠⎩ ⎭p
P P p p P p p p (3.14)
where ( )iDiag p is a diagonal matrix with ip as its diagonal entries.
When the first variable is epistemically irrelevant to the second variable, a joint
probability P in 1|sEΨ is
( ) ( ) ( )1 2 1 1 1 2 1 2|P s U P s P S U s S× = × × ; P = Diag(p1) P2|1 (3.15)
Thus the set 1|sEΨ is as follows (Bernardini and Tonon 2010).
( )
( )( )
( )
1
1
(1)2
(2)2
| ( )1 2|1 1 1 2|1 2 2
( )2
.: : ; ;
.
.
T
T
s iE
Tn
Diag
⎧ ⎫⎛ ⎞⎪ ⎪⎜ ⎟⎪ ⎪⎜ ⎟⎪ ⎪⎜ ⎟⎪ ⎪⎜ ⎟⎪ ⎪⎜ ⎟Ψ = = ∈Ψ = ∈Ψ⎨ ⎬⎜ ⎟⎪ ⎪⎜ ⎟⎪ ⎪⎜ ⎟⎪ ⎪⎜ ⎟⎪ ⎪⎜ ⎟⎜ ⎟⎪ ⎪⎝ ⎠⎩ ⎭
p
p
P p P p P p
p
(3.16)
Consequently, the constraints for problems involving epistemic irrelevance are
constraints on the marginals , i.e. Eq.(3.7), and on the definition of epistemic irrelevance,
i.e. (3.14) or (3.16). For example, when the second variable is epistemically irrelevant to
the first variable, the complete optimization problems read as follows (Bernardini and
Tonon 2010):
43
Minimize(Maximize) 1 2; , ,1; 1
i n j n i j i ji j
a P= =
= =∑ Subject to
( ) ( )2( )(1) (2)1 1 1 2... n Diag=P p p p p
( )( )
1 2
( )1 1 1 1 1 2
2 2 2 2 2
( )( ) 1 2 ( ) 2
( )1 2 2
; 1,..., ; 1,..., 1;
; 1,..., ;
1 1,..., 1; 1
0 1,..., 1; 0
Tk k j kLOW UPP
Tk k kLOW UPP
T j Tn n
j
E f E f k k j n
E f E f k k
j n
j n
⎡ ⎤ ⎡ ⎤≤ ≤ = = +⎣ ⎦ ⎣ ⎦
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⋅ = = + ⋅ =
≥ = + ≥
f p
f p
1 p 1 p
p p
(3.17)
When marginals are assigned through their extreme distributions, the optimization
problems is (Bernardini and Tonon 2010):
Minimize(Maximize) 1 2; , ,1; 1
i n j n i j i ji j
a P= =
= =∑ Subject to
1 1 22
1 1 2
1 2
,,11 1 2
1 1 1
,21 2
1 1
,2 2 11 2
...
1 1,..., ; 1
0 1,..., ; 1,..., ; 0 ; 1,...,
nEXT EXT EXT
j
j
c c Diag c
c j n c
c j n c
ξ ξ ξξξ ξ ξ ξ ξ
ξ ξ ξ
ξ ξξ ξ
ξ ξ
ξ ξξ ξ ξ ξ
= = =
= =
⎛ ⎞ ⎛ ⎞= ⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
= = =
≥ = = ≥ =
∑ ∑ ∑
∑ ∑
P p p p
(3.18)
When each variable is epistemically irrelevant to the other, that is, irrelevance is
applied to both directions, it is called “epistemic independence” (Campos et al. 1995,
Couso et al. 1999 and 2000). Epistemic independence is symmetric. It is the appropriate
model when we are given information on two marginal sets of probability measures and
under the assumption that the uncertainty about either variable does not change when
some information about the other variable is available. However, this assumption is not
equivalent to the assumption that variables are stochastically independent, which actually
is the concept of strong independence defined in the next sub-section.
Following of Bernardini and Tonon (2010), let P be a n1×n2 matrix whose i,j-th entry is the probability mass on a joint elementary event ( )1 2, jis s , ( )1 2, jiP s s . If T1 = U1 ×
44
S2, T2 = S1 × s2, by noticing that T1∩T2 = U1×s2, two variables are epistemically
independent if ∀(s1, s2)∈ S1 × S2,
( ) ( ) ( )1 2 1 2 1 2 2 2|P U s P U S S s P s× = × × ; P = P1|2 Diag(p2) (3.19)
Likewise,
( ) ( ) ( )1 2 1 2 1 2 1 1|P s U P S U s S P s× = × × ; P = Diag(p1) P2|1. (3.20)
Thus, two variables are epistemically independent if ∀(s1, s2)∈ S1 × S2: ( ) ( ) ( )
( ) ( ) ( )2 2
1 1
| |1 1 1 2 1 1 2 2
| |2 2 1 2 2 2 1 1
:
:
s s
s s
P P U s P U P s AND
P P s U P U P s
∃ ∈Ψ × =
∃ ∈Ψ × = (3.21)
The set of joint probability distributions, ΨE, is (Bernardini and Tonon 2010):
2 1| |s sE E EΨ = ∈Ψ ∩ΨP (3.22)
Here 1|sEΨ and 2|s
EΨ are given by (3.14) and (3.16), respectively.
In the optimization problem under epistemic independence, the constraints are
given by the marginal rules, i.e. Eq.(3.7), and by the definition of epistemic
independence, i.e. (3.21). If information on marginals is given as bounded previsions, the
optimization problems are (Bernardini and Tonon 2010): Minimize(Maximize) 1 2; , ,
1; 1
i n j n i j i ji j
a P= =
= =∑ Subject to
( ) ( )
( )
( )( )
( )
2 1
2
1
( ) ( 1)(1) (2)1 1 1 2
(1)2
(2)( 1) 2
1
( )2
... n n
T
Tn
Tn
Diag
Diag
+
+
=
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟
= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
P p p p p
p
pP p
p
( )( )
1 2
( )1 1 1 1 1 2
( )2 2 2 2 2 1
( ) ( )( ) 1 2 ( ) 2 1
( ) ( )1 2 2
; 1,..., ; 1,..., 1;
; 1,..., ; 1,..., 1;
1 1,..., 1; 1 1,..., 1
0 1,..., 1; 0
Tk k j kLOW UPP
Tk k j kLOW UPP
T j T jn n
j j
E f E f k k j n
E f E f k k j n
j n j n
j n j
⎡ ⎤ ⎡ ⎤≤ ≤ = = +⎣ ⎦ ⎣ ⎦
⎡ ⎤ ⎡ ⎤≤ ≤ = = +⎣ ⎦ ⎣ ⎦
⋅ = = + ⋅ = = +
≥ = + ≥ =
f p
f p
1 p 1 p
p p 11,..., 1n +
(3.23)
45
When marginals are assigned through their extreme distributions, the optimization
problems become are (Bernardini and Tonon 2010): Minimize(Maximize) 1 2; , ,
1; 1
i n j n i j i ji j
a P= =
= =∑ Subject to
1 1 22 1
1 1 2
2
2
12
1
21
2
, , 1,11 1 2
1 1 1
,12
1
, 11
1
,2
1
,1
...
.
.
.
n nEXT EXT EXT
T
EXT
nEXT
Tn
EXT
c c Diag c
c
Diag c
c
c
ξ ξ ξξ ξξ ξ ξ ξ
ξ ξ ξ
ξξ ξ
ξ
ξξ ξ
ξ
ξξ ξ
ξ
ξ
+
= = =
=
+
=
=
⎛ ⎞ ⎛ ⎞= ⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
⎛ ⎞⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎜ ⎟⎜ ⎟
⎛ ⎞⎜ ⎟= ⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎜ ⎟
⎜ ⎟⎜ ⎟⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠
∑ ∑ ∑
∑
∑
∑
P p p p
p
P p
p
1 2,
2 121 1
, ,2 2 1 11 2
1 1,..., ; 1 1,...,
0 1,..., ; 1,..., ; 0 1,..., ; 1,...,
j j
j j
j n c j n
c j n c j n
ξ ξξ
ξ ξ
ξ ξξ ξ ξ ξ
= == = = =
≥ = = ≥ = =
∑ ∑
(3.24)
By observing Eqs.(3.17), (3.18), (3.23), and (3.24), we find that the joint
probability P is the product of two matrices, which makes the constraints non-convex.
Optimization problems involving non-convex constraints are known to be NP-hard (Horst
et al. 2000), i.e. there is no fully polynomial-time approximation scheme to solve the
problem. To reduce the computational difficulties, equivalent forms of the optimization
problems and algorithms will be discussed in Section 4.3.3. If the conditional probability ( ) ( )1 1 1
1 2 2 2, |P s s S s= is of interest, then replace the
objective functions in (3.17), (3.18), (3.23), and (3.24) by 11,1 ,11/ i n i
iP P==∑ ; likewise for
other conditional probabilities. Again, the objective function is no longer linear.
46
3.2.2.2 Conditional Epistemic Irrelevance/Independence
In last sub-section, we have discussed epistemic irrelevance and epistemic
independence between two variables. When a third variable is involved, then conditional
irrelevance/independence (Campos and Cozman 2007) may need to be considered.
Similar to the definition of epistemic irrelevance in Eq. (3.12), S1 is conditional
epistemic irrelevant to S2 for a given *3s if
( ) ( )* *2 1 3 2 3| |P S s s S s× ∈Ψ (3.25)
In matrix form, set ΨE over (S1, S2) given *3s is
( )
( )( )
( )
( )
( )1 1
*1 2 3
1 * (1) *1 3 2 3
*1 2 3
* ( ) *31 2 3
, | :
| 0 |
, |
0 | |
E
T
Tn n
S S s
P s s S s
S S s
P s s S s
Ψ =
⎫⎡ ⎤⎛ ⎞⎪⎢ ⎥⎜ ⎟ ⎪⎢ ⎥⎜ ⎟= ⎬⎢ ⎥⎜ ⎟ ⎪⎢ ⎥⎜ ⎟ ⎪⎝ ⎠ ⎢ ⎥⎣ ⎦⎭
P
P
P
P
(3.26)
where vector ( )( ) *2 3|i S sP is a conditional probability measure over S2 given *
3s .
Accordingly, S2 is conditional epistemic irrelevant to S1 given *3s if
( ) ( )* *1 2 3 1 3| |P S s s S s× ∈Ψ (3.27)
and set ΨE over (S1, S2) given *3s is
( )
( ) ( ) ( )( )
( )2
2
*1 2 3
1 *2 3
( )* (1) * *1 2 3 1 3 1 3
*32
, | :
| 0
, | | ... |
0 |
E
n
n
S S s
P s s
S S s S s S s
P s s
Ψ =
⎫⎛ ⎞⎪⎜ ⎟⎪⎡ ⎤ ⎜ ⎟= ⎬⎣ ⎦ ⎜ ⎟⎪⎜ ⎟⎪⎝ ⎠⎭
P
P P P (3.28)
It is easy to see that conditional epistemic irrelevance is an asymmetric and directional
property as well. If both Eqs. (3.25) and (3.27) are satisfied, S1 is conditional epistemic
independent to S2 given S3.
47
3.2.2.3 Strong Independence
In Imprecise Probability, strong independence is equivalent to stochastic
independence in precise probability. As a result, the set of probability measures, ΨS, is
composed of all product measures: 1 2 1 1 2 2: ,S P P P P PΨ = = ⊗ ∈Ψ ∈Ψ (3.29)
where ( ) ( ) ( )1 2 1 2 1 1 2 2P P s s P s P s⊗ × = . Bernardini and Tonon (2010) pointed out that if
(3.21) and additional constraints below are satisfied, then epistemic independence
becomes strong independence. 2|
2 2 1 1: ss S P P∀ ∈ = and 1|1 1 2 2: ss S P P∀ ∈ = (3.30)
Strong independence is an appropriate model when the random experiments are
stochastically independent. The constraints in the optimization problem under strong
independence are constraints on marginals, i.e., Eq.(3.7), and the definition of strong
independence in Eq.(3.29). When the marginals are bounded by upper and lower
previsions, the complete optimization problems are (Bernardini and Tonon 2010):
Minimize(Maximize) 1 2; , ,
1; 1
i n j n i j i ji j
a P= =
= =∑ Subject to
1 2T=P p p
( )( )
1 2
1 1 1 1 1
2 2 2 2 2
( ) 1 ( ) 2
1 2
; 1,..., ;
; 1,..., ;
1 1
0 ; 0
Tk k kLOW UPP
Tk k kLOW UPP
T Tn n
E f E f k k
E f E f k k
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⋅ = ⋅ =
≥ ≥
f p
f p
1 p 1 p
p p
(3.31)
If the extreme points are given on the convex sets of marginal probability distributions,
the optimization problems under strong independence become (Bernardini and Tonon
2010):
48
Minimize(Maximize) 1 2; , ,1; 1
i n j n i j i ji j
a P= =
= =∑ Subject to
1
1
2
2
1 2
1 2
1 11
2 21
1 21 1
1 21 2
1 ; 1
0; 1,..., ; 0 ; 1,...,
T
EXT
EXT
c
c
c c
c c
ξξ ξ
ξ
ξξ ξ
ξ
ξ ξξ ξ
ξ ξ
ξ ξξ ξ ξ ξ
=
=
= =
=
=
=
= =
≥ = ≥ =
∑
∑
∑ ∑
P p p
p p
p p (3.32)
Similar to the case of epistemic irrelevance/independence, constraints in (3.31) and (3.32)
under strong independence are NP-hard as well. Detailed algorithms to solve this
problem will be explained in Section 4.3.3. For the conditional probability ( ) ( )1 1 1
1 2 2 2, |P s s S s= , the objective functions in
(3.31) and (3.32) are replaced by 11,1 ,11/ i n i
iP P==∑ ; likewise for other conditional
probabilities.
3.2.2.4 Conditional strong Independence
Consider now strong independence conditional to a third variable. S1 is
Conditional strong independent of S2 given s3 if ( ) ( ) ( )1 2 3 1 3 2 3| | |P S S s P S s P S s× = × (3.33)
where ( )1 3|P S s and ( )2 3|P S s are given and equal to marginal probabilities of
( )1 2 3|P S S s× .
In matrix form, set Ψ over (S1, S2) given s3 is ( ) ( ) ( ) ( )1 2 3 1 2 3 1 3 2 3, | : | | |TS S s S S s S s S sΨ = × =P P P P (3.34)
where ( )1 3|S sP and ( )2 3|S sP are conditional probability distributions over S1
and S2 given s3, respectively.
49
Or, equivalently,
( ) ( ) ( ) ( )
1 3 2 31 2 3
3
P S s P S sP S S s
P s
× × ×× × = , for ( )3P s > 0 (3.35)
where ( )1 3P S s× and ( )2 3P S s× are given and equal to marginal probabilities of
( )1 2 3P S S s× × . The definition in Eq. (3.35) is the same as ‘weak conditional
independence’ defined in (Cano and Moral 2000), where ‘strong conditional
independence’ is defined for another situation, which we do not discuss in this study.
In the form of matrix, set ΨS over (S1, S2, s3) is
( ) ( ) ( ) ( )( ) ( )1 3 2 3
1 2 3 1 2 3 33
, ,, , : , , , 0
T
S
S s S sS S s S S s P s
P s
⎫⎪Ψ = = > ⎬⎪⎭
P PP P (3.36)
Compared with the constraints in Eq.(3.26) or (3.28), constraints in Eq.(3.36) are stricter.
Similar to the case of epistemic irrelevance and strong independence, we could expect
that set ΨS under conditional strong independence is a subset of set ΨE under conditional epistemic irrelevance: ( ) ( )1 2 3 1 2 3, , , ,S ES S s S S sΨ ⊆ Ψ .
3.2.3 Analysis with uncertain correlation
The correlation coefficient ρ is often used as a measure of linear dependence
between two variables S1 and S2: ( ) ( ) ( ) ( )
1 2 1 2
1 2 1 2 1 2,
S S S S
COV S S E S S E S E SD D D D
ρ−
= = (3.37)
where ( ) ( )22 , 1,2iS i iD E S E S i= − = , ( )E S is the expected value of variable S.
In this sub-section, we discuss the situation when the correlation coefficient ρ is
given imprecisely, i.e., bounded by an interval:
ρ ρ ρ≤ ≤ (3.38)
To eliminate the possibility of division-by-zero errors, Eq. (3.37) is re-written as
follows:
50
( ) ( ) ( )1 21 2 1 2 S SE S S E S E S D Dρ= + (3.39)
Substitute ρ in Eq. (3.39) with the bounds on ρ in Eq. (3.38), then one obtains:
( ) ( ) ( ) ( ) ( )1 2 1 21 2 1 2 1 2S S S SE S E S D D E S S E S E S D Dρ ρ+ ≤ ≤ + (3.40)
Therefore, the set of joint probability measures over ( )1 2,S S , Ψc, is obtained as follows ( ) ( ) ( ) ( ) ( )
( ) ( )
1 2 1 2
1 2
1 2 1 2 1 2
1 2
: ;
;
S S S SC T
n n
E S E S D D E S S E S E S D Dρ ρ⎧ ⎫+ ≤ ≤ +⎪ ⎪Ψ = ⎨ ⎬∈Ψ ∈Ψ⎪ ⎪⎩ ⎭
P
P1 1 P (3.41)
where Ψi, i = 1, 2, is the given convex set of marginals of Si.
It should be noted that in precise probabilities, the covariance matrix must be
positive semi-definite. Here we do not include this constraint in Eq. (3.41), because all
constraints in (3.41) are written in terms of joint probability distributions. The constraint
that the covariance matrix should be positive semi-definite will be satisfied automatically
as long as set Ψc is not empty.
3.3 ALGORITHMS
Section 3.2 formulated the problems under assumptions of unknown interaction,
different types of independence, and uncertain correlation respectively. This section will
focus on the algorithms to search the global optimizer where the upper and lower bounds
of objective functions are achieved.
Generally, there are two options to find the extreme values of functions on the
joint distribution:
Option (1) Using global optimization to find the point(s) at which the minimum or
maximum value of the objective function is achieved.
Option (2) First finding all extreme points for Ψ as described in Section 2.3 and then
restricting the search to these extreme points.
51
Option (1) directly focuses on the optimal solution, which is apparently more
efficient if only one maximum or minimum must be calculated. Option (2) could be used
only if the objective function and constraints are linear (e.g., probabilities or
expectations). However, the conditional probability over joint finite spaces, which is non-
linear, is an exception. In Section 3.3.1, we will prove that the minimum (or maximum)
value of a conditional probability is still achieved at the extreme points of the set of the
joint distributions Ψ.
3.3.1 Previsions and conditional probability in joint distributions
The following two theorems make calculations on joint spaces much easier:
Theorem 3-1 The minimum and maximum values of a prevision in the joint distribution
are achieved at some extreme points of the set of the joint distributions Ψ .
Proof: Any prevision in the joint distribution is a convex function.
Case 1: Ψ is convex. The global optimizer of a convex function relative to a convex set Ψ
occurs at some extreme point of Ψ (Rockafellar 1997, Page 342)
Case 2: Ψ is not convex. Let Ψ* be the convex hull of Ψ. Thus, Ψ* is the smallest convex
set which contains Ψ. Any extreme point of Ψ* is an extreme point of Ψ, and vice versa.
According to Case 1, the prevision will achieve its extreme values at some extreme points
of Ψ*, and thus some extreme points of Ψ. ◊
Theorem 3-2 The minimum and maximum values of a conditional probability in the joint
distribution are achieved at the extreme points of the set of the joint distributions Ψ. Proof: By inserting the marginal expression for ( )2 2
jp s into the expression for the
conditional probability ( )1|2 1 2| jip s s , one obtains:
52
( ) ( )1
, ,1 1|2 12 2
1, : | /
nj ji i i j i j
is s S p s s p p
=∀ ∈ = ∑ (3.42)
Let *P ∈Ψ and **P ∈Ψ be two extreme points of the set Ψ. The conditional
probability ( )1|2 1 2|i jp s s on each joint distribution is ( )1
, ,1|2,* 1 2 * *
1| /
ni j i j i j
ip s s p p
== ∑ , and
( )1
, ,1|2,** 1 2 ** **
1| /
ni j i j i j
ip s s p p
=
= ∑ , respectively. Assume ( )1|2,* 1 2|i jp s s ≥ ( )1|2,** 1 2|i jp s s , then
any point ( )newP between *P and **P may be written as ( )* **1λ λ+ −P P , 0 1λ≤ ≤ ,
i.e., ( )( ) * **
, , ,1new
i j i j i jp p pλ λ= + − . Consequently, the conditional probability based on the
new joint distribution ( )newP is:
( ) ( )
( )
( )
( )1 1 1
, , ,* **
1|2, 1 2, , ,
* **1 1 1
1|
1
i j i j i jnewji
new n n ni j i j i jnew
i i i
p p pp s s
p p p
λ λ
λ λ= = =
+ −= =
+ −∑ ∑ ∑ (3.43)
By subtracting Eq. (3.43) from (3.42), one obtains:
( ) ( ) ( )
( )1 1 1
, , ,* ** *
1|2, 1 1|2,* 12 2, , ,
* ** *1 1 1
1| |
1
i j i j i jj ji i
new n n ni j i j i j
i i i
p p pp s s p s s
p p p
λ λ
λ λ= = =
+ −− = −
+ −∑ ∑ ∑
( ) ( )
( )
1 1 1
1 1 1
, , , , , ,* ** * * ** *
1 1 1
, , ,* ** *
1 1 1
1 1
1
n n ni j i j i j i j i j i j
i i in n n
i j i j i j
i i i
p p p p p p
p p p
λ λ λ λ
λ λ
= = =
= = =
⎡ ⎤⎡ ⎤+ − − + −⎢ ⎥⎣ ⎦ ⎢ ⎥⎣ ⎦=
⎡ ⎤+ −⎢ ⎥
⎢ ⎥⎣ ⎦
∑ ∑ ∑
∑ ∑ ∑
( )
( )
1 1
1 1 1
, , , ,** * * **
1 1
, , ,* ** *
1 1 1
1
1
n ni j i j i j i j
i in n n
i j i j i j
i i i
p p p p
p p p
λ
λ λ
= =
= = =
⎛ ⎞− −⎜ ⎟⎜ ⎟
⎝ ⎠=⎡ ⎤
+ −⎢ ⎥⎢ ⎥⎣ ⎦
∑ ∑
∑ ∑ ∑
( )
( )
1 1 1 1
1 1 1 1 1
, , , , , ,** * * ** * **
1 1 1 1
, , , , ,* ** * * **
1 1 1 1 1
1 /
1 /
n n n ni j i j i j i j i j i j
i i i in n n n n
i j i j i j i j i j
i i i i i
p p p p p p
p p p p p
λ
λ λ
= = = =
= = = = =
⎛ ⎞ ⎛ ⎞− − ⋅⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠=⎡ ⎤ ⎛ ⎞
+ − ⋅⎜ ⎟⎢ ⎥ ⎜ ⎟⎢ ⎥⎣ ⎦ ⎝ ⎠
∑ ∑ ∑ ∑
∑ ∑ ∑ ∑ ∑
(3.44)
53
( )
( ) ( )
1 1
1 1 1
, , , ,** ** * *
1 1, , ,* ** **
1 1 1
1|2,** 1 1|2,* 12 2
1 / /1 /
| | ,
n ni j i j i j i j
n n ni ii j i j i j
i i i
j ji i
p p p pp p p
A p s s p s s
λ
λ λ = =
= = =
⎛ ⎞−= −⎜ ⎟⎜ ⎟⎡ ⎤ ⎝ ⎠+ −⎢ ⎥⎢ ⎥⎣ ⎦
⎡ ⎤= ⋅ −⎣ ⎦
∑ ∑∑ ∑ ∑
where ( )
( )
1
1 1
,**
1
, ,* **
1 1
10
1
ni j
in n
i j i j
i i
pA
p p
λ
λ λ
=
= =
−= ≥
+ −
∑
∑ ∑.
Since ( ) ( )1|2,* 1 1|2,** 12 2| |j ji ip s s p s s≥ , then one obtains ( )1|2, 1 2| jinewp s s - ( )1|2,* 1 2| 0jip s s ≤ ,
i.e. p1|2,*(s1i|s2
j) ≥ p1|2,new(s1i|s2
j). Likewise, p1|2,new(s1i|s2
j).≥ p1|2,**(s1i|s2
j)
In conclusion, given two extreme points on the set of joint distribution, *P and
**P , and p1|2,*(s1i|s2
j) ≥ p1|2,**(s1i|s2
j), any point ( )newP between them satisfies the
inequality p1|2,*(s1i|s2
j) ≥ p1|2,new(s1i|s2
j) ≥ p1|2,**(s1i|s2
j). Therefore, the minimum and
maximum values of conditional probability are achieved at the extreme points of the set
of joint distributions.
Therefore, regardless of the type of independence introduced next, the conditional
upper and lower probabilities reach the minimum or maximum values at an extreme point
of Ψ. ◊
3.3.2 Unknown Interaction
Because the constraints for an unknown interaction problem are linear, the set of
joint distributions, ΨU, is convex. According to Theorem 3-1 and Theorem 3-2, whether
the objective function is a prevision or a conditional probability on joint spaces, the
bounds are achieved at extreme points of ΨU. Both Option (1) and (2) are applicable to
the unknown interaction problem.
54
Example 3-1. Consider the case in which a bolt and its corresponding nut have to be applied in a
process. Box 1 contains bolts of types A, B, and C; Box 2 contains nuts of types A, B, and C.
Unfortunately, the manufacturer mixed up the labels, and the following information is available about Box
1 (i.e., S1): P(A) ≤ 2P(B); 2P(A) ≥ P(C); P(B) ≤ P(C), whose set Ψ1 is depicted in Figure 3-1a, and has
three vertices identified by vectors: 1
1EXTp = (0.29, 0.14, 0.57)T,
1
2EXTp = (0.5, 0.25, 0.25)T, and
1
3EXTp =
(0.2, 0.4, 0.4)T. Since bounds on expectations of general functions are given on S1, planes bounding Ψ1 are
not parallel to any coordinate plane (Figure 3-1a), and Ψ1 cannot be generated by bounds on probabilities
of events (Section 2.3). Available information about Box 2 (i.e., S2) is that 10% of nuts are either A or B,
80% are B, and the remaining 10% is indeterminate, then set Ψ2 is depicted in Figure 3-1b, and has four
vertices identified by vectors 2
1EXTp = (0, 0.9, 0.1)T,
2
2EXTp = (0, 1, 0)T,
2
3EXTp = (0.1, 0.8, 0.1)T, and
2
4EXTp = (0.2, 0.8, 0)T. A worker in the field takes one bolt from Box 1 and then one nut from Box 2; we
are interested in the joint probability of the picked bolt and nut. Since we are just given the marginal
probabilities, all we know is that the joint probability measure must satisfy Eq.(3.7). A pair of bolt and nut
is selected from each of the boxes, but we do not assume stochastic independence, and it is possible that a
correlated joint procedure is used to select the bolt and nut.
Let Si = A, B, C, and S = (A, A), (A, B), (A, C), (B, A), (B, B), (B, C), (C, A), (C, B), (C, C).
Under the unknown interaction assumption, ΨU contains all joint probabilities of elementary events on S
whose marginals are in Ψ1 and Ψ2. Consider the case in which the same type of bolt and nut is selected: T =
(A, A), (B, B), (C, C). Objective function and Constraints (3.8) are thus equal to:
55
(a)
(b)
Figure 3-1 Example 3-1: marginal sets Ψi in the 3-dimensional spaces (n1 = n2 = 3).
Minimize (Maximize) 1,1 2,2 3,3p p p+ + Subject to ( ) ( )( ) ( )( ) ( )
1,1 1,2 1,3 2,1 2,2 2,3
1,1 1,2 1,3 3,1 3,2 3,3
2,1 2,2 2,3 3,1 3,2 3,3
1,1 2,1 3,1
1,2 2,2 3,2
1,3 2,3 3,3
1,1 1,2 1,3 2,1 2,2 2,3 3,1
2 0
2 0
0
0.0 0.2
0.8 1.0
0.0 0.1
p p p p p p
p p p p p p
p p p p p p
p p p
p p p
p p p
p p p p p p p
− + + + ⋅ + + ≥
⋅ + + − + + ≥
− + + + + + ≥
≤ + + ≤
≤ + + ≤
≤ + + ≤
+ + + + + + + 3,2 3,3
,
1
0i j
p p
p
+ =
≥
(3.45)
With all extreme points of sets Ψ1 and Ψ2, the constraints in problems (3.9) now read as follows:
1,1 1,2 1,3 1 2 31 1 1
2,1 2,2 2,3 1 2 31 1 1
3,1 3,2 3,3 1 2 31 1 1
1,1 2,1 3,1 1 2 3 42 2 2 2
1,2 2,2 3,2 1 22 2
0.29 0.5 0.2 0
0.14 0.25 0.4 0
0.57 0.25 0.4 0
0.0 0.0 0.1 0.2 0
0.9 1.0 0
subject to
p p p c c c
p p p c c c
p p p c c c
p p p c c c c
p p p c c
+ + − − − =
+ + − − − =
+ + − − − =
+ + − − − − =
+ + − − − 3 42 2
1,3 2,3 3,3 1 2 3 42 2 2 2
.8 0.8 0
0.1 0.0 0.1 0.0 0
c c
p p p c c c c
− =
+ + − − − − =
1 2 31 1 11 2 3 42 2 2 2
,
1
1
0
0ii j
c c c
c c c c
c
p
ξ
+ + =
+ + + =
≥
≥
(3.46)
1
1EXTp
1
2EXTp
1
3EXTp
( )32P s
( )22P s( )1
2P s
( )32P s
( )22P s( )1
2P s2
1EXTp
2
2EXTp
2
3EXTp
2
4EXTp
56
Under the unknown interaction assumption, the lower and upper probabilities for the event T are
equal to 0 and 0.6, respectively; solutions are detailed in Table 3-1a and Table 3-1b.
Another option is to use the general algorithm of Option (2) to find all extreme distributions on the
joint space, and then calculate the objective function on all the extreme distributions to find the extreme
values of the objective function. Constraints in Eq.(3.45) generate 60 extreme distributions, which are
given in Table 3-2 together with the values of the objective functions p 1,1+ p 2,2+ p3,3. One can check that
these values are the same as the maximum and minimum in Table 3-1a and Table 3-1b.
It should be noted that the optimal solutions in unknown interaction problems are not necessary
achieved at extreme points of Ψ1 or Ψ2. The concepts of independence will provide narrower probability
intervals.
Table 3-1a Example 3-1: Solutions of the linear programming problems (3.45) for the lower and upper probabilities for T.
Solution for Joint P Marginal on S1 Marginal on S2
Min 0.00 0.29 0.000.09 0.00 0.060.00 0.57 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29, 0.14, 0.57)
(0.09, 0.84, 0.06)
Max 0.1 0.1 00 0.4 00 0.3 0.1
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
3EXTp =
(0.2, 0.4, 0.4) 2
3EXTp =
(0.1, 0.8, 0.1)
Table 3-1b Example 3-1: Solutions of the linear programming problems (3.46) for the lower and upper probabilities for T.
Solution for Joint P (c11, c1
2, c13) Marginal on S1 (c2
1, c22, c2
3, c24) Marginal on S2
Min 0 0.4 00.2 0 00 0.4 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0.47, 0.53, 0) (0.4, 0.2, 0.4) (0, 0, 0, 1)
2
4EXTp =
(0.2, 0.8, 0)
Max 0.2 0 00 0.4 00 0.4 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0, 0, 1)
1
3EXTp =
(0.2, 0.4, 0.4)
(0, 0, 0, 1) 2
4EXTp =
(0.2, 0.8, 0)
57
Table 3-2 Example 3-1: Extreme distribution of Ψ for the linear programming problems (3.45)
Extreme point of Ψ p1,1 p1,2 p1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,2
1,1 2,2 3,3p p p+ + p2,2/( p1,2+ p2,2+ p3,2)
1 0.00 0.40 0.00 0.20 0.00 0.00 0.00 0.40 0.00 0.00 0.00
2 0.20 0.00 0.00 0.00 0.40 0.00 0.00 0.40 0.00 0.60 0.50
3 0.00 0.27 0.00 0.20 0.00 0.00 0.00 0.53 0.00 0.00 0.00
4 0.00 0.29 0.00 0.00 0.14 0.00 0.00 0.57 0.00 0.14 0.14
5 0.00 0.29 0.00 0.14 0.00 0.00 0.00 0.57 0.00 0.00 0.00
6 0.00 0.50 0.00 0.00 0.25 0.00 0.00 0.25 0.00 0.25 0.25
7 0.00 0.20 0.00 0.00 0.40 0.00 0.00 0.40 0.00 0.40 0.40
8 0.00 0.40 0.00 0.10 0.00 0.10 0.00 0.40 0.00 0.00 0.00
9 0.00 0.27 0.00 0.10 0.00 0.10 0.00 0.53 0.00 0.00 0.00
10 0.10 0.00 0.10 0.00 0.40 0.00 0.00 0.40 0.00 0.50 0.50
11 0.00 0.29 0.00 0.14 0.00 0.00 0.00 0.51 0.06 0.06 0.00
12 0.00 0.29 0.00 0.00 0.14 0.00 0.20 0.37 0.00 0.14 0.18
13 0.00 0.29 0.00 0.14 0.00 0.00 0.06 0.51 0.00 0.00 0.00
14 0.20 0.09 0.00 0.00 0.14 0.00 0.00 0.57 0.00 0.34 0.18
15 0.00 0.23 0.06 0.14 0.00 0.00 0.00 0.57 0.00 0.00 0.00
16 0.06 0.23 0.00 0.14 0.00 0.00 0.00 0.57 0.00 0.06 0.00
17 0.00 0.29 0.00 0.00 0.14 0.00 0.00 0.47 0.10 0.24 0.16
18 0.00 0.29 0.00 0.00 0.04 0.10 0.00 0.57 0.00 0.04 0.05
19 0.00 0.29 0.00 0.04 0.00 0.10 0.00 0.57 0.00 0.00 0.00
20 0.00 0.19 0.10 0.00 0.14 0.00 0.00 0.57 0.00 0.14 0.16
21 0.00 0.50 0.00 0.00 0.25 0.00 0.20 0.05 0.00 0.25 0.31
22 0.00 0.50 0.00 0.20 0.05 0.00 0.00 0.25 0.00 0.05 0.06
23 0.20 0.30 0.00 0.00 0.25 0.00 0.00 0.25 0.00 0.45 0.31
24 0.00 0.50 0.00 0.00 0.25 0.00 0.00 0.15 0.10 0.35 0.28
25 0.00 0.50 0.00 0.00 0.15 0.10 0.00 0.25 0.00 0.15 0.17
26 0.00 0.40 0.10 0.00 0.25 0.00 0.00 0.25 0.00 0.25 0.28
27 0.00 0.20 0.00 0.00 0.40 0.00 0.20 0.20 0.00 0.40 0.50
28 0.00 0.20 0.00 0.20 0.20 0.00 0.00 0.40 0.00 0.20 0.25
29 0.00 0.20 0.00 0.00 0.40 0.00 0.00 0.30 0.10 0.50 0.44
30 0.00 0.20 0.00 0.00 0.30 0.10 0.00 0.40 0.00 0.30 0.33
31 0.00 0.10 0.10 0.00 0.40 0.00 0.00 0.40 0.00 0.40 0.44
32 0.00 0.29 0.00 0.00 0.14 0.00 0.10 0.37 0.10 0.24 0.18
33 0.00 0.29 0.00 0.10 0.00 0.04 0.00 0.51 0.06 0.06 0.00
34 0.00 0.29 0.00 0.10 0.04 0.00 0.00 0.47 0.10 0.14 0.05
35 0.10 0.19 0.00 0.00 0.14 0.00 0.00 0.47 0.10 0.34 0.18
58
Extreme point of Ψ p1,1 p1,2 p1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,2
1,1 2,2 3,3p p p+ + p2,2/( p1,2+ p2,2+ p3,2)
36 0.00 0.29 0.00 0.00 0.04 0.10 0.10 0.47 0.00 0.04 0.05
37 0.00 0.29 0.00 0.04 0.00 0.10 0.06 0.51 0.00 0.00 0.00
38 0.00 0.19 0.10 0.00 0.14 0.00 0.10 0.47 0.00 0.14 0.18
39 0.10 0.19 0.00 0.00 0.04 0.10 0.00 0.57 0.00 0.14 0.05
40 0.00 0.23 0.06 0.10 0.00 0.04 0.00 0.57 0.00 0.00 0.00
41 0.06 0.23 0.00 0.04 0.00 0.10 0.00 0.57 0.00 0.06 0.00
42 0.00 0.19 0.10 0.10 0.04 0.00 0.00 0.57 0.00 0.04 0.05
43 0.10 0.09 0.10 0.00 0.14 0.00 0.00 0.57 0.00 0.24 0.18
44 0.00 0.50 0.00 0.00 0.25 0.00 0.10 0.05 0.10 0.35 0.31
45 0.00 0.50 0.00 0.10 0.15 0.00 0.00 0.15 0.10 0.25 0.19
46 0.10 0.40 0.00 0.00 0.25 0.00 0.00 0.15 0.10 0.45 0.31
47 0.00 0.50 0.00 0.00 0.15 0.10 0.10 0.15 0.00 0.15 0.19
48 0.00 0.40 0.10 0.00 0.25 0.00 0.10 0.15 0.00 0.25 0.31
49 0.00 0.50 0.00 0.10 0.05 0.10 0.00 0.25 0.00 0.05 0.06
50 0.10 0.40 0.00 0.00 0.15 0.10 0.00 0.25 0.00 0.25 0.19
51 0.00 0.40 0.10 0.10 0.15 0.00 0.00 0.25 0.00 0.15 0.19
52 0.10 0.30 0.10 0.00 0.25 0.00 0.00 0.25 0.00 0.35 0.31
53 0.00 0.20 0.00 0.00 0.40 0.00 0.10 0.20 0.10 0.50 0.50
54 0.00 0.20 0.00 0.10 0.30 0.00 0.00 0.30 0.10 0.40 0.38
55 0.10 0.10 0.00 0.00 0.40 0.00 0.00 0.30 0.10 0.60 0.50
56 0.00 0.20 0.00 0.00 0.30 0.10 0.10 0.30 0.00 0.30 0.38
57 0.00 0.10 0.10 0.00 0.40 0.00 0.10 0.30 0.00 0.40 0.50
58 0.00 0.20 0.00 0.10 0.20 0.10 0.00 0.40 0.00 0.20 0.25
59 0.10 0.10 0.00 0.00 0.30 0.10 0.00 0.40 0.00 0.40 0.38
60 0.00 0.10 0.10 0.10 0.30 0.00 0.00 0.40 0.00 0.30 0.38
Min 0.00 0.00
Max 0.60 0.50
By replacing the objective function in (3.45) or (3.46) by the conditional probability that the bolt
is Type B given the type of nut, i.e. p2,2/( p1,2+ p2,2+ p3,2) , the bounds of the conditional probability are 0
and 0.5, which are larger than the interval [0.14, 0.4], i.e., the range of P1(B) in Ψ1.
In Table 3-2, the function p2,2/( p1,2+ p2,2+ p3,2) has been calculated at the 60 extreme points of Ψ.
One can check that the objective function achieves its maximum and minimum values at extreme points of
Ψ.
59
3.3.3 Independence
As stated in Section 3.2.2, the multiplication of two matrices makes the problems
of epistemic irrelevance/independence and strong independence non-convex and
consequently NP-hard. The following sub-sections will present the solutions to avoid
such computational difficulties.
3.3.3.1 Epistemic Irrelevance/Independence
When marginals’ expectations are bounded as in Eq.(3.3), these non-convex
optimization problems could be turned into linear ones by rewriting the problems in
terms of the joint distribution matrix P. The derivation is presented as follows.
We start from a problem of epistemic irrelevance. Given P = Diag(p1) P2|1, the marginal on S1, ( )2n⋅P 1 , may be written as
( )( ) ( )
( ) ( )( ) ( ) ( )
22
2 2
1 2|1
1 2|1 1 1
n n
n n
Diag
Diag Diag
⋅ = ⋅
= ⋅ = ⋅ =
P 1 p P 1
p P 1 p 1 p
(3.47)
i.e., the marginal on S1 is the same as p1. Therefore, the constraints on p1 (i.e.,
( )1 1 1 1 Tk k k
LOW UPPE f E f⎡ ⎤ ⎡ ⎤≤ ≤⎣ ⎦ ⎣ ⎦f p ), by substituting ( )2n⋅P 1 for p1, can be rewritten as
( ) ( )21 1 1
n
Tk k kLOW UPPE f E f⎡ ⎤ ⎡ ⎤≤ ⋅ ≤⎣ ⎦ ⎣ ⎦f P 1 (3.48)
By substituting ( )2n⋅P 1 for p1 in P = Diag(p1) P2|1, one obtains
( ) ( )( )21 2|1 2|1nDiag Diag= ⋅ = ⋅ ⋅P p P P 1 P (3.49)
Thus, the jth row of matrix P, ,j ⋅P , may be written as
( )( ) ( )2
2
, , , , ( )2|1 2 1
1; 1,...,
n Tj j j j m jn
mp j n⋅ ⋅ ⋅
=
⎛ ⎞= ⋅ = =⎜ ⎟
⎝ ⎠∑P P 1 P p (3.50)
where ,2|1j ⋅P is the jth row of matrix 2|1P .
60
As for constraints on p2, i.e.: ( ) ( )2 2 2 2
Tk k j kLOW UPPE f E f⎡ ⎤ ⎡ ⎤≤ ≤⎣ ⎦ ⎣ ⎦f p , by multiplying
it by 2
,
1
nj m
mp
=∑ (as
2,
10
nj m
mp
=≥∑ ), one obtains
( )2 2 2
, ( ) , ,2 2 2 2
1 1 1
n n nTk j m k j j m k j mLOW UPP
m m mE f p p E f p
= = =
⎡ ⎤ ⎡ ⎤⋅ ≤ ≤ ⋅⎣ ⎦ ⎣ ⎦∑ ∑ ∑f p (3.51)
By noticing Eq.(3.50) and substituting ( ), Tj ⋅P for 2
( ) ,2
1
nj j m
mp
=
⎛ ⎞⋅ ⎜ ⎟⎜ ⎟⎝ ⎠∑p in Eq. (3.51),
constraints on p2 are rewritten as 2 2
, , ,2 2 2
1 1
n nk j m j k k j m
LOW UPPm m
E f p E f p⋅
= =
⎛ ⎞ ⎛ ⎞⎡ ⎤ ⎡ ⎤⋅ ≤ ⋅ ≤ ⋅⎜ ⎟ ⎜ ⎟⎣ ⎦ ⎣ ⎦⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠∑ ∑P f (3.52)
Therefore, given the epistemic irrelevance of the first experiment with respect to
the second experiment, P = Diag(p1) P2|1, the optimization problem may be written in the
linear form:
Minimize 1 2; , ,1; 1
i n j n i j i ji j
a p= =
= =∑ 1 2; , ,1; 1
i n j n i j i ji j
a p= =
= =⎛ ⎞−⎜ ⎟⎝ ⎠∑
Subject to ( ) ( )2
2 2
1 1 1 1
, , ,2 2 2 2 1
1 1
; 1,..., ;
; 1,..., ; 1,..., ;
1; 0
n
Tk k kLOW UPP
n nk j m j k k j m
LOW UPPm m
T
E f E f k k
E f p E f p k k j n⋅
= =
⎡ ⎤ ⎡ ⎤≤ ⋅ ≤ =⎣ ⎦ ⎣ ⎦
⎛ ⎞ ⎛ ⎞⎡ ⎤ ⎡ ⎤⋅ ≤ ⋅ ≤ ⋅ = =⎜ ⎟ ⎜ ⎟⎣ ⎦ ⎣ ⎦⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠
⋅ ⋅ = ≥
∑ ∑
f P 1
P f
1 P 1 P
(3.53)
Likewise, when the second experiment is epistemically irrelevant to the first, P = P1|2Diag(p2) and ( )1 2
Tn =1 P p , i.e., the marginal on S2 is equal to p2. The linear
optimization problem is
Minimize(Maximize) 1 2; , ,1; 1
i n j n i j i ji j
a P= =
= =∑
Subject to
( )1 2
1
, , ,1 1 1 1 2
1 1
2 ( ) 2 2 2
; 1,..., ; 1,..., ;
; 1,..., ;
1; 0
n nTk m j k j k m jLOW UPP
m m
k T k kLOW n UPP
T
E f p E f p k k j n
E f E f k k
⋅
= =
⎛ ⎞ ⎛ ⎞⎡ ⎤ ⎡ ⎤⋅ ≤ ≤ ⋅ = =⎜ ⎟ ⎜ ⎟⎣ ⎦ ⎣ ⎦⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎡ ⎤ ⎡ ⎤≤ ⋅ ≤ =⎣ ⎦ ⎣ ⎦
⋅ ⋅ = ≥
∑ ∑f P
1 P f
1 P 1 P
(3.54)
61
where , j⋅P is the jth column of matrix P. Constraints in (3.54) are equivalent to the
constraints in (3.17), but much more the computational effort is substantially reduced
with respect to (3.17).
For the case of epistemic independence, the optimization problem (3.23) can be
reformulated in terms of linear constraints:
Minimize 1 2; , ,1; 1
i n j n i j i ji j
a p= =
= =∑ 1 2; , ,1; 1
i n j n i j i ji j
a p= =
= =⎛ ⎞−⎜ ⎟⎝ ⎠∑
Subject to
( )1 2
2 2
, , ,1 1 1 1 2
1 1
, , ,2 2 2 2 1
1 1
; 1,..., ; 1,..., ;
; 1,..., ; 1,..., ;
1; 0
n nTk m j k j k m jLOW UPP
m m
n nk j m j k k j m
LOW UPPm m
T
E f p E f p k k j n
E f p E f p k k j n
⋅
= =
⋅
= =
⎛ ⎞ ⎛ ⎞⎡ ⎤ ⎡ ⎤⋅ ≤ ≤ ⋅ = =⎜ ⎟ ⎜ ⎟⎣ ⎦ ⎣ ⎦⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎛ ⎞ ⎛ ⎞
⎡ ⎤ ⎡ ⎤⋅ ≤ ⋅ ≤ ⋅ = =⎜ ⎟ ⎜ ⎟⎣ ⎦ ⎣ ⎦⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
⋅ ⋅ = ≥
∑ ∑
∑ ∑
f P
P f
1 P 1 P
(3.55)
An efficient algorithm for calculating the extreme joint distributions for epistemic
irrelevance is given by the following two theorems:
Theorem 3-3 Let the extreme points of the convex sets of probability distributions on S1 and S2 be
1 1, 1,...EXTξ ξ ξ=p , and
2 2, 1,...EXTξ ξ ξ=p , respectively. If the first experiment is
epistemically irrelevant to the second, the set of extreme points of the joint distributions, 1|s
EΨ , is:
( ) ( )
1 2|1 1 1 2|1 2 1= : , ; 1,...i
EXT Diag EXT EXT i n= ⋅ ∈ ∈ =P p P p P , i.e.
( )( )
( )
12
1 1 2
12
1 2
1
: , ;=
1,...
i
n
T
EXTm m
EXT EXT EXT EXTT
EXT
Diag EXT EXTEXT
i n
η
η
η
⎧ ⎫⎛ ⎞⎪ ⎪⎜ ⎟⎪ ⎪⎜ ⎟
= ∈ ∈⎪ ⎪⎜ ⎟⎨ ⎬⎜ ⎟⎪ ⎪⎜ ⎟⎪ ⎪⎝ ⎠⎪ ⎪=⎩ ⎭
p
P p p p
p
(3.56)
Proof: Any p1∈Ψ1 and p2∈Ψ2 can be written as a linear combination of extreme points
in Ψ1 and Ψ2, respectively:
( )( )1 11 1 1
1 11 1 1 1... ... ... ...
TEXT EXT EXT
ξ ξξ ξλ λ λ=p p p p (3.57)
62
( )
( )
22
1 2 12
2
1,11,12 2
,2|1 2
1, ,2 2
...T
EXTi
n n T
EXT
ξ
ξ
ξ ξ
λ λ
λ
λ λ
⎛ ⎞⎡ ⎤⎜ ⎟⎢ ⎥ ⎜ ⎟= ⎢ ⎥⎜ ⎟⎢ ⎥ ⎜ ⎟⎢ ⎥ ⎜ ⎟⎣ ⎦ ⎝ ⎠
p
P
p
1
2
11 11
, ,2 12 2
1
0 1, 1,..., ; 1
0 1, 1,..., , 1,..., ; 1i ii n
ξξ ξ
ξ
ξξ ξ
ξ
λ ξ ξ λ
λ ξ ξ λ
=
=
≤ ≤ = =
≤ ≤ = = =
∑
∑
(3.58)
By inserting Eqs.(3.57) and (3.58) into ( )1 2|1Diag= ⋅P p P , one obtains:
( )
( )
( )
( )
2
2
1 21 1 1
1 2 1
12
2
111
,11,12 2
1 ,12
1, ,2 2
1
...
... ...
TEXT
Ti EXT
EXT EXT EXTn n
T
EXT
Diag
ξξ ξ
ξξ ξ
ξξ
ξ
λ
λ λλ
λ
λ λλ
⎛ ⎞⎜ ⎟⎛ ⎞⎛ ⎞
⎜ ⎟ ⎜ ⎟⎜ ⎟⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎡ ⎤⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎢ ⎥⎜ ⎟ ⎜ ⎟⎜ ⎟= ⎢ ⎥⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎢ ⎥⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎢ ⎥⎣ ⎦⎜ ⎟ ⎜ ⎟⎜ ⎟⎜ ⎟ ⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠ ⎜ ⎟⎝ ⎠ ⎜ ⎟
⎝ ⎠
p
pP p p p
p
(3.59)
Extreme points of P are achieved if and only if 11, 0,
mm
ξ ξλ
ξ=⎧
= ⎨ ≠⎩ and ,
21, 0,
ii
i
ξ ξ ηλ
ξ η=⎧
= ⎨ ≠⎩,
11,...,m ξ= , 21,...,iη ξ=
Therefore, ( )( )
( )
12
1
12
n
T
EXTm
EXT EXTT
EXT
Diag
η
η
⎛ ⎞⎜ ⎟⎜ ⎟
= ⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
p
P p
p
◊
Theorem 3-4 If the second experiment is epistemically irrelevant to the first, the set of
extreme points of the joint distributions is ( ) ( )
1|2 2 1|2 1 2 2 2= : , ; 1,...i
EXT Diag EXT EXT i n= ⋅ ∈ ∈ =P P p P p , i.e.
( ) ( )1 221 1
21 1 1 1
P :=
, ; 1,...
n
i
mEXT EXTEXT EXT
mEXTEXT
DiagEXT
EXT EXT i n
ηη
η
⎧ ⎫=⎪ ⎪⎨ ⎬⎪ ⎪∈ ∈ =⎩ ⎭
p p p
p p
… (3.60)
63
Theorem 3-3 and Theorem 3-4 enable us to efficiently find the extreme joint
distributions given the extreme distributions on the marginals. The upper limit for the
number of extreme joint distribution is 11 2
nξ ξ× when the first experiment is
epistemically irrelevant to the second experiment, and 21 2nξ ξ× when the second
experiment is epistemically irrelevant to the first one. However, the algorithms in
Theorem 3-3 and Theorem 3-4 cannot be used in the case of epistemic independence
because, under epistemic independence, the set of joint distributions is the intersection of
the two convex sets for the epistemic irrelevance cases, i.e., ΨE= 1|SEΨ ∩ 2|S
EΨ . As
illustrated in Figure 3-2, the extreme points of ΨE may not be extreme points of 1|SEΨ or
2|SEΨ . The only way to determine the extreme points for ΨE is to use the linear constraints
in (3.55) and apply the general algorithm for Option (2).
Figure 3-2 Set of extreme points for the case of epistemic independence
Example 3-2. Consider again the situation and knowledge available in Example 3-1. Suppose that now
two additional boxes (Boxes 3 and 4) of nuts are delivered to the construction site. The content is the same
as Box 2 in Example 3-1. This time, the worker in the field picks a bolt from box 1 (Experiment 1), and
then proceeds with Experiment 2, i.e.,
If it is Type A, then he picks a nut from Box 2.
If it is Type B, then he picks a nut from Box 3
2|SEΨ
1|SEΨ
EΨ
64
Otherwise, he picks a nut from Box 4.
We want to write down the optimization problems for finding upper and lower expectations on the
joint space and then calculate the upper and lower probabilities for the case in which the same type of bolt
and nut is selected, i.e. event T=(A, A), (B, B), (C, C). Finally, calculate upper and lower conditional
probabilities that the bolt is Type B given the type of nut, and contrast to upper and lower conditional
probabilities that the nut is Type B given the type of the bolt.
In this example, the first experiment is epistemically irrelevant to the second experiment because
1) The set of acceptable gambles concerning the second experiment does not change when we
learn the outcome of the first experiment.
2) s1 is selected according to some marginal distribution in Ψ1, and then s2 is selected according
to a distribution from Ψ2 that depends on s1.
3) s2 is selected by a different procedure for different values of s1
Let us first solve the problem with the quadratic constrains in (3.17) and (3.18), respectively. As a
consequence, quadratic constraints in Eq. (3.17) become:
Subject to
( )( )( )( )
(1)2
(1) (2)1 2
(3)2
T
T
T
Diag
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟=⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
p
P p p
p
( )( )( )
( )( )( )
(1)1
(1)1
(1)1
( )2
( )2
( )2
(1) (1)1 1
(1) (2) (3)2 2 2
(1) (2)2 2
1,2,0 0
2,0, 1 0
0, 1,1 0
0.0 1,0,0 0.2, 1, 2,3
0.8 0,1,0 1.0, 1, 2,3
0.0 0,0,1 0.1, 1,2,3
1; 0
1; 1; 1
0; 0
T
T
T
T j
T j
T j
T
T T T
j
j
j
− ≥
− ≥
− ≥
≤ ≤ =
≤ ≤ =
≤ ≤ =
⋅ = ≥
⋅ = ⋅ = ⋅ =
≥ ≥
p
p
p
p
p
p
1 p p
1 p 1 p 1 p
p p (3)2; 0≥p
(3.61)
65
Based on the extreme distributions given in Example 3-1, Eq. (3.18) gives the following quadratic
constraints:
Subject to
1,1 2,1 3,11 1 1
1,1 2,1 3,1 4,12 2 2 2
0.29 0.5 0.20.14 0.25 0.40.57 0.25 0.4
0.0 0.0 0.1 0.20.9 1.0 0.8 0.80.1 0.0 0.1 0.0
Diag c c c
c c c c
⎛ ⎞⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟= + +⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎝ ⎠
⎛ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟+ + +⎜ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎝
⋅
P
1,2 2,2 3,2 4,22 2 2 2
1,3 2,3 3,3 4,32 2 2 2
0.0 0.0 0.1 0.20.9 1.0 0.8 0.80.1 0.0 0.1 0.0
0.0 0.0 0.1 0.20.9 1.0 0.8 00.1 0.0 0.1
T
T
c c c c
c c c c
⎞⎟⎟⎟⎠
⎛ ⎞⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟+ + +⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎝ ⎠
⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟ ⎜ ⎟ ⎜ ⎟+ + +⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠
4,1,1 2,1 3,1
1 1 1 21
, ,1 2
.80.0
1; 1; 1, 2,3
0; 0
T
i j
ii j i j
c c c c j
c c=
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎛ ⎞⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠⎝ ⎠
+ + = = =
≥ ≥
∑
(3.62)
The upper and lower probabilities for the event T = (A, A), (B, B), (C, C) are equal to 0.11 and
0.48, respectively. The solutions summarized in Table 3-3a and Table 3-3b indicate that p2(1) ≠ p2
(2) ≠ p2(3)
at the minimizing solution, i.e. they do not satisfy stochastic independence. Likewise for the maximizing
solution.
Table 3-3a Example 3-2: Solutions of the optimization problems (3.17) for lower and upper probabilities for T.
Solution for Joint P ( )1
ip Marginal on S1 ( )
2jp Marginal on S2
Min
0.00 0.27 0.010.02 0.11 0.010.06 0.51 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29,0.14,0.57) 1
1EXTp = (0.29, 0.14, 0.57)
(0.00, 0.95, 0.05); (0.14, 0.79, 0.07); (0.11, 0.89, 0.00)
(0.07, 0.90, 0.03)
Max 0.04 0.16 0.000.00 0.40 0.000.03 0.33 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
3EXTp = (0.2, 0.4, 0.4) 1
3EXTp =
(0.2, 0.4, 0.4) 2
4EXTp =(0.2,0.8,0);
2
2EXTp = (0, 1, 0); (0.07, 0.83, 0.10)
(0.08, 0.88, 0.04)
66
Table 3-3b Example 3-2: Solutions of the optimization problems (3.18) for lower and upper probabilities for T.
Solution for Joint P ( )1
ip Marginal on S1 ( )
2jp Marginal on S2
Min 0.00 0.26 0.030.02 0.11 0.010.03 0.54 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29,0.14,0.57) 1
1EXTp =
(0.29,0.14,0.57)
(0, 0.91, 0.09); (0.15, 0.80, 0.05); (0.05, 0.95, 0)
(0.05,0.92,0.03)
Max 0.04 0.16 0.000.00 0.40 0.000.02 0.34 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
3EXTp =
(0.2, 0.4, 0.4) 1
3EXTp =
(0.2, 0.4, 0.4) 2
4EXTp =(0.2,0.8,0);
2
2EXTp = (0, 1, 0);
(0.02, 0.88, 0.10)
(0.06,0.90,0.04)
We started from noticing, and we based our solution on the observation that the first experiment is
epistemically irrelevant to the second experiment. As a consequence, upper and lower conditional
probabilities that the nut is Type B given the type of the bolt are equal to the marginal ones, i.e. 1 and 0.8,
respectively. In order to check that epistemic irrelevance is a directional, asymmetric property, let us
calculate upper and lower conditional probabilities that the bolt is Type B given the type of the nut. The
(non-linear) function to minimize and maximize is p2,2/( p1,2+ p2,2+ p3,2), where p1,2+ p2,2+ p3,2 is the second
component of the marginal on S2. The constraints are still given by Eqs. (3.17) or (3.18). Upper and lower
conditional probabilities are equal to 0.45 and 0.12, which are larger bounds than the marginal bounds, i.e.
0.4 and 0.14. Since the conditional probability bounds are different from the marginal probability bounds,
the second experiment is epistemically relevant to the first one. Results summarized in Table 3-4a and
Table 3-4b indicate that p2(1) ≠ p2
(2) ≠ p2(3), i.e., the minimizing and maximizing solutions again violate
stochastic independence.
67
Table 3-4a Example 3-2: Solutions of the optimization problems (3.17) for upper and lower conditional probabilities that the bolt is Type B given the type of the nut.
Solution for Joint P ( )
1ip Marginal on S1
( )2
jp Marginal on S2
Min
0.00 0.29 0.000.01 0.11 0.010.00 0.57 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29,0.14,0.57) 1
1EXTp =
(0.29,0.14,0.57) 2
2EXTp = (0, 1, 0);
(0.09, 0.81, 0.09);
2
2EXTp = (0, 1, 0);
(0.01, 0.97, 0.01)
Max
0.02 0.16 0.020.00 0.40 0.000.04 0.32 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
3EXTp =
(0.2, 0.4, 0.4) 1
3EXTp =
(0.2, 0.4, 0.4) 2
3EXTp =(0.1,0.8,0.1);
2
2EXTp = (0, 1, 0);
2
3EXTp =(0.1,0.8,0.1);
(0.06, 0.88, 0.06)
Table 3-4b Example 3-2: Solutions of the optimization problems (3.18) for upper and lower conditional probabilities that the bolt is Type B given the type of the nut.
Solution for Joint P ( )1
ip Marginal on S1 ( )
2jp Marginal on S2
Min 0.00 0.29 0.000.02 0.11 0.010.00 0.57 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29,0.14,0.57) 1
1EXTp =
(0.29,0.14,0.57) 2
2EXTp = (0, 1, 0);
(0.15,0.80, 0.05);
2
2EXTp = (0, 1, 0);
(0.02,0.97,0.01)
Max 0.04 0.16 0.000.00 0.40 0.000.04 0.32 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
3EXTp =
(0.2, 0.4, 0.4) 1
3EXTp =
(0.2, 0.4, 0.4) 2
4EXTp =(0.2,0.8,0);
2
2EXTp = (0, 1, 0);
2
3EXTp =(0.1,0.8,0.1);
(0.08,0.88,0.04)
Now we can rework Example 3-2 with the new linear algorithm (Eq.(3.53)), in which constraints
are written in terms of the joint distribution. Since the problem is linear, one may solve the problem with
two different methods: a linear optimization problem (Option (1)), and Option (2). When constraints are
given in terms of extreme points of the marginals, Theorem 3-3 will be used in Example 3-2.
Let us redo Example 3-2 by rewriting the constraints as in Eq. (3.53): Subject to (3.63)
68
( )( )( )
, ,1 ,
, ,2 ,
, ,3 ,
1,2,0 0
2,0, 1 0
0, 1,1 0
0.0 0.2 , 1,2,3
0.8 1.0 , 1,2,3
0.0 0.1 , 1,2,3
T
T
T
i i i
i i i
i i i
p i
p i
p i
⋅ ⋅
⋅ ⋅
⋅ ⋅
− ⋅ ≥
− ⋅ ≥
− ⋅ ≥
⋅ ≤ ≤ ⋅ =
⋅ ≤ ≤ ⋅ =
⋅ ≤ ≤ ⋅ =
P 1
P 1
P 1
P 1 P 1
P 1 P 1
P 1 P 1
1;0
T ⋅ ⋅ =≥
1 P 1P
Option (1): a linear optimization problem
Upper and lower probabilities of T are 0.48 and 0.11, and upper and lower conditional
probabilities are 0.45 and 0.12, the same as the results in previous calculation in Example 3-2, and may be
achieved at different solutions, listed in Table 3-5 and Table 3-6. The computational effort was greatly
reduced: for the linear optimization problem (3.63), it takes around 8 iterations to achieve the optimal
solutions; while for the optimization problem (3.61) with quadratic constraints, the solutions are achieved
after about 11 iterations. Thus, the computational effort was reduced.
Table 3-5 Example 3-2: Solutions of the optimization problems (3.63) for lower and upper probabilities for T.
Solution for Joint P Marginal on S1 Marginal on S2 Min
0.00 0.26 0.030.03 0.11 0.000.11 0.46 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29,0.14,0.57)
(0.14, 0.83, 0.03)
Max 0.04 0.16 0.000.00 0.40 0.000.04 0.32 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
3EXTp =
(0.2, 0.4, 0.4)
(0.08, 0.88, 0.04)
Table 3-6 Example 3-2: Solutions of the optimization problems (3.63) for upper and lower conditional probabilities that the bolt is Type B given the type of the nut.
Solution for Joint P Marginal on S1 Marginal on S2
Min
0.00 0.29 0.000.01 0.11 0.010.00 0.57 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29,0.14,0.57) (0.01, 0.97, 0.01)
69
Max
0.02 0.16 0.020.00 0.40 0.000.08 0.32 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
3EXTp =
(0.2, 0.4, 0.4) (0.10, 0.88, 0.02)
Option (2):
Use Eq. (3.56) or (3.60) to calculate extreme points of the convex set of joint probability
distributions. All 192 extreme points are found by this algorithm. One can check all values of objective
functions and find upper and lower probabilities of T are 0.48 and 0.11, and upper and lower conditional
probabilities are 0.45 and 0.12, the same as previous results.
Theorem 3-3 can also be used to solve the problem. One may first use Eq.(3.63) to generate all
3×43=192 extreme points, check and find that none is duplicated. Therefore, Theorem 3-3 also generates
192 extreme points, which are the same as the extreme points in Option (2).
Example 3-3. Consider again the situation and knowledge available in Example 3-1. In addition to the
knowledge available in Example 3-1, all we now know about the stochastic mechanism for picking the bolt
and nut, i.e. the joint probability measure P is that: (a) whatever the type of the bolt, the conditional
probability that the nut is A (B or C) lies in Ψ2; and (b) whatever the type of the nut, the conditional
probability that the bolt is A (B or C) lies in the convex set of Ψ1.
We want to write down the optimization problems for finding upper and lower expectations on the
joint space and then calculate the upper and lower probabilities for the case in which the same type of the
bolt and nut is selected, i.e. event T=(A, A), (B, B), (C, C). Finally, calculate the conditional upper an
lower probabilities that the bolt is Type B given the type of the nut, and contrast these values with the
conditional upper and lower probabilities that the nut is Type B given the type of the bolt.
This is a case of epistemic independence, where each experiment is epistemically irrelevant to the
other. As a consequence, this example is the symmetric counterpart of Example 3-2 above, where only the
70
first experiment was epistemically irrelevant to the other. Quadratic constraints in Eqs. (3.23) and (3.24)
become: Subject to
( )( )( )( )
( ) ( )
( ) ( ) ( )( )( )
(1)2
(1) (2) (2) (3) (4) (4)1 2 1 1 1 2
(3)2
( ) ( ) ( )1 1 1
( )2
( )2
;
1,2,0 0; 2,0, 1 0; 0, 1,1 0
0.0 1,0,0 0.2
0.8 0,1,0 1.0
T
T
T
T T Tj j j
T j
T j
Diag Diag
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟= =⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
− ≥ − ≥ − ≥
≤ ≤
≤ ≤
p
P p p P p p p p
p
p p p
p
p
( ) ( )2
( ) ( )1 1
( ) ( )2 2
0.0 0,0,1 0.1
1; 0
1; 01,2,3, 4
T j
T j j
T j j
j
≤ ≤
⋅ = ≥
⋅ = ≥=
p
1 p p
1 p p
(3.64)
Based on the extreme distributions given in Example 3-1, Eq. (3.23) gives the following
quadratic constraints:
Subject to
1,1 2,1 3,11 1 1
1,1 2,1 3,1 4,12 2 2 2
0.29 0.5 0.20.14 0.25 0.40.57 0.25 0.4
0.0 0.0 0.1 0.20.9 1.0 0.8 0.80.1 0.0 0.1 0.0
Diag c c c
c c c c
⎛ ⎞⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟= + +⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎝ ⎠
⎛ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟+ + +⎜ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎝
⋅
P
1,2 2,2 3,2 4,22 2 2 2
1,3 2,3 3,3 4,32 2 2 2
0.0 0.0 0.1 0.20.9 1.0 0.8 0.80.1 0.0 0.1 0.0
0.0 0.0 0.1 0.20.9 1.0 0.8 00.1 0.0 0.1
T
T
c c c c
c c c c
⎞⎟⎟⎟⎠
⎛ ⎞⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟+ + +⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎝ ⎠
⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟ ⎜ ⎟ ⎜ ⎟+ + +⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠
.80.0
T
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎛ ⎞⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠⎝ ⎠
(3.65)
71
1,2 2,2 3,21 1 1
1,3 2,3 3,31 1 1
1,4 2,41 1
0.29 0.5 0.20.14 0.25 0.40.57 0.25 0.4
0.29 0.5 0.20.14 0.25 0.40.57 0.25 0.4
0.29 0.50.14 0.250.57
c c c
c c c
c c
⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟ ⎜ ⎟ ⎜ ⎟+ +⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟ ⎜ ⎟ ⎜ ⎟= + +⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎛ ⎞⎜ ⎟ +⎜ ⎟⎜ ⎟⎝ ⎠
P
3,41
1,4 2,4 3,4 4,42 2 2 2
,1
0.20.4
0.25 0.4
0.0 0.0 0.1 0.20.9 1.0 0.8 0.80.1 0.0 0.1 0.0
T
i
c
Diag c c c c
c
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎛ ⎞ ⎛ ⎞⎜ ⎜ ⎟ ⎜ ⎟⎟+⎜ ⎜ ⎟ ⎜ ⎟⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟⎝ ⎠ ⎝ ⎠⎝ ⎠⎛ ⎞⎛ ⎞⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟⋅ + + +⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎝ ⎠⎝ ⎠
3
14
,2
1, ,
1 2
1; 1, 2,3, 4;
1; 1, 2,3, 4;
0; 0
j
i
i j
ii j i j
j
c j
c c
=
=
= =
= =
≥ ≥
∑
∑
The upper and lower probabilities for the event T=(A, A), (B, B), (C, C) are equal to 0.42 and
0.14, respectively. These bounds are tighter than those in Example 3-2 because now additional constraints
were added on P reflecting the fact that the second experiment is epistemically irrelevant to the first one.
The solutions are summarized in Table 3-7. They do not necessary satisfy stochastic
independence. In epistemic independence, if we learn that the actual value of s2 is s2*, then the probability
measure for s1 is again one of the probability measures in Ψ1, but in general not always the same for
different values s2*; and vice versa. Strong independence (dealt with in the next sub-section) imposes that
the probability measures be the same.
Table 3-7 Example 3-3: Solutions of the optimization problems (3.64) for lower and upper probabilities for T. (same solutions for (3.65))
Solution for Joint P ( )
1ip Marginal on
S1
( )2
jp Marginal on S2
Min 0.00 0.29 0.000.00 0.14 0.000.00 0.57 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
( )
1ip =
1
1EXTp =
(0.29,0.14,0.57) i=1,…,4
1
1EXTp =
(0.29,0.14,0.57)
( )2
jp =2
2EXTp =
(0, 1, 0); j=1,…,4
2
4EXTp =
(0, 1, 0)
72
Max 0.02 0.18 0.020.01 0.35 0.010.01 0.35 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0.22,0.37,0.41);
1
2EXTp =(0.5,0.25,0.25);
1
3EXTp =(0.20,0.40,0.40);
1
1EXTp =(0.29,0.14,0.57)
(0.22,0.37,0.41) (0.11, 0.80, 0.09); (0.03, 0.94, 0.03); (0.03, 0.87, 0.10); (0.05,0.88,0.07)
(0.05,0.88, 0.07)
Upper and lower conditional probabilities that the first (second) selection is Type B given the type
of the second (first) selection are obtained by minimizing and maximizing p2,2/( p1,2+ p2,2+ p3,2) (p2,2/( p2,1+
p2,2+ p2,3)). The constraints are still given by Eq. (3.64) or (3.65). Results are summarized in Table 3-8a-d.
The solutions show that upper and lower conditional probabilities that the first (second) selection is Type B
given the type of the second (first) selection are equal to the marginal ones, i.e. 0.4 and 0.14 (1 and 0.8),
respectively, which are the same as the provided bounds on the marginals. It is consistent with the
definition of epistemic independent that each experiment is epistemically irrelevant to the other one.
Table 3-8a Example 3-3: Solutions of the optimization problems (3.64) for upper and lower conditional probabilities that the bolt is Type B given the type of the nut.
Solution for Joint P ( )
1ip Marginal on
S1
( )2
jp Marginal on S2
Min 0.03 0.24 0.030.01 0.12 0.010.04 0.47 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =(0.29,0.14,0.57);
(0.38,0.11,0.52);
1
1EXTp =(0.29,0.14,0.57);
(0.37,0.12,0.51);
1
1EXTp =
(0.29,0.14,0.57) 2
2EXTp =(0.1,0.8,0.1)
(0.07,0.86, 0.07);
(0.07,0.86, 0.07);
(0.07,0.84, 0.10)
(0.07,0.84, 0.10)
Max 0 0.2 00 0.4 00 0.4 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
( )
1ip =
1
3EXTp =
(0.2,0.4,0.4)
i=1,…,4
1
3EXTp =
(0.2, 0.4, 0.4)
( )2
jp =2
2EXTp =
(0, 1, 0); j=1,…,4
2
2EXTp =
(0, 1, 0);
Table 3-8b Example 3-3: Solutions of the optimization problems (3.64) for upper and lower conditional probabilities that the nut is Type B given the type of the bolt.
Solution for Joint P ( )
1ip Marginal on
S1
( )2
jp Marginal on S2
Min 0.04 0.23 0.010.04 0.18 0.000.05 0.44 0.01
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0.28,0.23,0.49); (0.33,0.31,0.35);
(0.27,0.21,0.51); (0.32,0.16,0.53);
(0.28,0.23,0.49) (0.16,0.82, 0.02);
(0.18,0.80, 0.01);
(0.10,0.89, 0.02);
(0.13,0.85, 0.02)
(0.13,0.85, 0.02)
73
Max 0 0.2 00 0.4 00 0.4 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
( )
1ip =
1
3EXTp =
(0.2,0.4,0.4)
i=1,…,4
1
3EXTp =
(0.2, 0.4, 0.4)
( )2
jp =2
2EXTp =
(0, 1, 0); j=1,…,4
2
2EXTp =
(0, 1, 0);
Table 3-8c Example 3-3: Solutions of the optimization problems (3.65) for upper and lower conditional probabilities that the bolt is Type B given the type of the nut.
Solution for Joint P ( )
1ip Marginal on
S1
( )2
jp Marginal on S2
Min 0.04 0.24 0.020.02 0.12 0.010.02 0.48 0.05
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0.30,0.15,0.55);
1
2EXTp =(0.5,0.25,0.25);
1
1EXTp =(0.29,0.14,0.57);
1
1EXTp =(0.29,0.14,0.57);
(0.30,0.15,0.55) (0.12, 0.8, 0.08);
(0.12, 0.8, 0.08);
(0.04,0.88, 0.08);
(0.08,0.84, 0.10)
(0.08,0.84, 0.10)
Max 0.02 0.17 0.020.01 0.35 0.020.02 0.35 0.03
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0.22,0.38,0.40);
(0.45,0.26,0.28);
1
3EXTp =(0.2,0.4,0.4);
(0.27,0.27,0.47);
(0.22,0.38,0.40) (0.11,0.8, 0.09);
(0.04,0.91,0.05);
(0.04,0.88, 0.08); (0.05,0.87,0.07)
(0.05,0.87,0.07)
Table 3-8d Example 3-3: Solutions of the optimization problems (3.65) for upper and lower conditional probabilities that the nut is Type B given the type of the bolt.
Solution for Joint P ( )
1ip Marginal on
S1
( )2
jp Marginal on S2
Min 0.06 0.32 0.010.05 0.21 0.000.05 0.27 0.02
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0.39,0.27,0.34);
(0.40,0.30,0.30);
(0.40,0.26,0.34);
(0.29,0.16,0.55);
(0.39,0.27,0.34) (0.16, 0.82, 0.02);
(0.18, 0.80, 0.02);
(0.15,0.81, 0.04);
(0.16,0.81, 0.03)
(0.16,0.81, 0.03)
Max 0 0.2 00 0.4 00 0.4 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
( )
1ip =
1
3EXTp =
(0.2,0.4,0.4)
i=1,…,4
1
3EXTp =
(0.2, 0.4, 0.4)
( )2
jp =2
2EXTp =
(0, 1, 0); j=1,…,4
2
2EXTp =
(0, 1, 0);
Let us now redo Example 3-3 under epistemic independence by using the linear optimization
algorithm in Eq.(3.55). Rewrite the constraints as:
74
Subject to ( )( )( )
, ,
, ,
, ,
, ,1 ,
, ,2 ,
, ,3 ,
1,2,0 , 1,2,3
2,0, 1 , 1,2,3
0, 1,1 , 1,2,3
0.0 0.2 , 1,2,3
0.8 1.0 , 1,2,3
0.0 0.1 , 1,2,
T j T j
T j T j
T j T j
i i i
i i i
i i i
j
j
j
p i
p i
p i
⋅ ⋅
⋅ ⋅
⋅ ⋅
⋅ ⋅
⋅ ⋅
⋅ ⋅
− ≥ ⋅ =
− ≥ ⋅ =
− ≥ ⋅ =
⋅ ≤ ≤ ⋅ =
⋅ ≤ ≤ ⋅ =
⋅ ≤ ≤ ⋅ =
P 0 1 P
P 0 1 P
P 0 1 P
P 1 P 1
P 1 P 1
P 1 P 1 3
1;0
T ⋅ ⋅ =≥
1 P 1P
(3.66)
Option (1) consists of finding the solutions in the linear optimization problem (3.66). Solutions are
detailed in Table 3-9 and Table 3-10. Here the upper and lower probabilities of T are 0.42 and 0.14, and
upper and lower conditional probabilities are 0.4 and 0.14, still the same as the previous results in Example
3-3.
Table 3-9 Example 3-3: Solutions of the optimization problems (3.66) for lower and upper probabilities for T.
Solution for Joint P Marginal on S1 Marginal on S2
Min
0.00 0.29 0.000.00 0.14 0.000.00 0.57 0.00
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29,0.14,0.57) 2
4EXTp =
(0, 1, 0)
Max
0.02 0.18 0.020.01 0.35 0.010.01 0.35 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0.22,0.37,0.41) (0.05,0.88, 0.07)
Table 3-10 Example 3-3: Solutions of the optimization problems (3.66) for upper and lower conditional probabilities that the bolt is Type B given the type of the nut.
Solution for Joint P Marginal on S1 Marginal on S2
Min
0.00 0.26 0.030.00 0.13 0.010.00 0.52 0.05
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
1
1EXTp =
(0.29,0.14,0.57) (0, 0.92, 0.08)
75
Max
0.02 0.18 0.020.01 0.35 0.010.01 0.35 0.04
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
(0.22,0.37,0.41) (0.05,0.88, 0.07)
Option (2) consists of finding the extreme points of ψ by the linear constraints in Eq. (3.66) and
then calculating the objective functions on these extreme points. A total of 80 extreme points are found and
listed in Table 3-11. By calculating the values of the objective functions, one can find the same extreme
values as in the previous results, i.e. the upper and lower probabilities of T are 0.42 and 0.14, respectively,
and upper and lower conditional probabilities are 0.40 and 0.14, respectively.
As explained before, in the epistemically independent case, we cannot find all extreme
distributions through Theorem 3-3 or Theorem 3-4.
Table 3-11 Example 3-3: Extreme distribution of Ψ for the linear programming problem (3.66)
Extreme point of Ψ p1,1 p1,2 p1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,2
1,1 2,2 3,3p p p+ + p2,2/( p1,2+ p2,2+ p3,2)
1 0.00 0.29 0.00 0.00 0.14 0.00 0.00 0.57 0.00 0.14 0.14 2 0.00 0.50 0.00 0.00 0.25 0.00 0.00 0.25 0.00 0.25 0.25 3 0.00 0.20 0.00 0.00 0.40 0.00 0.00 0.40 0.00 0.40 0.40 4 0.06 0.23 0.00 0.03 0.11 0.00 0.11 0.46 0.00 0.17 0.14 5 0.00 0.26 0.03 0.00 0.13 0.01 0.00 0.51 0.06 0.19 0.14 6 0.10 0.40 0.00 0.05 0.20 0.00 0.05 0.20 0.00 0.30 0.25 7 0.04 0.16 0.00 0.08 0.32 0.00 0.08 0.32 0.00 0.36 0.40 8 0.03 0.43 0.00 0.05 0.22 0.00 0.05 0.22 0.00 0.24 0.25 9 0.06 0.25 0.00 0.03 0.13 0.00 0.03 0.50 0.00 0.19 0.14 10 0.02 0.26 0.00 0.03 0.13 0.00 0.03 0.53 0.00 0.15 0.14 11 0.00 0.45 0.05 0.00 0.23 0.03 0.00 0.23 0.03 0.25 0.25 12 0.04 0.17 0.00 0.02 0.34 0.00 0.09 0.34 0.00 0.38 0.40
76
Extreme point of Ψ p1,1 p1,2 p1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,2
1,1 2,2 3,3p p p+ + p2,2/( p1,2+ p2,2+ p3,2)
13 0.03 0.45 0.00 0.01 0.23 0.00 0.06 0.23 0.00 0.25 0.25 14 0.00 0.18 0.02 0.00 0.36 0.04 0.00 0.36 0.04 0.40 0.40 15 0.00 0.19 0.02 0.00 0.37 0.01 0.00 0.37 0.04 0.41 0.40 16 0.00 0.47 0.01 0.00 0.23 0.03 0.00 0.23 0.03 0.26 0.25 17 0.00 0.48 0.01 0.00 0.24 0.01 0.00 0.24 0.03 0.26 0.25 18 0.05 0.18 0.00 0.02 0.36 0.00 0.02 0.36 0.00 0.41 0.40 19 0.00 0.27 0.03 0.00 0.13 0.02 0.00 0.54 0.02 0.15 0.14 20 0.00 0.28 0.01 0.00 0.14 0.02 0.00 0.55 0.02 0.15 0.14 21 0.00 0.19 0.02 0.00 0.38 0.01 0.00 0.38 0.01 0.39 0.40 22 0.03 0.23 0.03 0.01 0.11 0.01 0.06 0.46 0.06 0.20 0.14 23 0.02 0.19 0.02 0.03 0.24 0.03 0.05 0.37 0.05 0.31 0.30 24 0.05 0.40 0.05 0.03 0.20 0.03 0.03 0.20 0.03 0.28 0.25 25 0.06 0.37 0.03 0.03 0.18 0.02 0.03 0.25 0.03 0.28 0.23 26 0.05 0.37 0.05 0.02 0.18 0.02 0.02 0.26 0.03 0.26 0.23 27 0.02 0.16 0.02 0.04 0.32 0.04 0.04 0.32 0.04 0.38 0.40 28 0.01 0.42 0.05 0.03 0.21 0.03 0.03 0.21 0.03 0.25 0.25 29 0.02 0.19 0.02 0.05 0.24 0.01 0.05 0.38 0.05 0.31 0.29 30 0.01 0.41 0.05 0.03 0.20 0.02 0.03 0.22 0.03 0.25 0.24 31 0.03 0.24 0.03 0.02 0.12 0.02 0.02 0.48 0.06 0.21 0.14 32 0.01 0.25 0.03 0.02 0.12 0.02 0.02 0.49 0.06 0.19 0.14 33 0.01 0.25 0.03 0.02 0.12 0.01 0.02 0.49 0.06 0.19 0.14 34 0.06 0.31 0.02 0.03 0.25 0.03 0.03 0.25 0.03 0.34 0.31 35 0.07 0.35 0.02 0.04 0.17 0.01 0.04 0.28 0.04 0.28 0.22 36 0.05 0.42 0.01 0.03 0.21 0.03 0.03 0.21 0.03 0.29 0.25 37 0.01 0.43 0.01 0.03 0.22 0.03 0.03 0.22 0.03 0.26 0.25 38 0.02 0.37 0.02 0.04 0.18 0.01 0.04 0.29 0.04 0.24 0.22 39 0.03 0.24 0.03 0.02 0.13 0.02 0.02 0.47 0.05 0.21 0.15 40 0.03 0.24 0.03 0.02 0.12 0.01 0.02 0.48 0.06 0.21 0.14 41 0.03 0.24 0.03 0.02 0.12 0.02 0.02 0.48 0.06 0.21 0.14 42 0.02 0.17 0.02 0.01 0.33 0.04 0.04 0.33 0.04 0.39 0.40 43 0.01 0.43 0.05 0.01 0.21 0.02 0.03 0.21 0.03 0.25 0.25 44 0.02 0.17 0.02 0.01 0.34 0.01 0.04 0.34 0.04 0.40 0.40 45 0.01 0.44 0.01 0.01 0.22 0.03 0.03 0.22 0.03 0.26 0.25 46 0.01 0.45 0.01 0.01 0.23 0.01 0.03 0.23 0.03 0.27 0.25 47 0.03 0.23 0.03 0.06 0.29 0.01 0.06 0.29 0.01 0.33 0.36 48 0.02 0.17 0.02 0.04 0.33 0.01 0.04 0.33 0.04 0.39 0.40 49 0.01 0.42 0.05 0.03 0.21 0.02 0.03 0.21 0.02 0.25 0.25 50 0.01 0.42 0.05 0.03 0.21 0.02 0.03 0.21 0.03 0.25 0.25
77
Extreme point of Ψ p1,1 p1,2 p1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,2
1,1 2,2 3,3p p p+ + p2,2/( p1,2+ p2,2+ p3,2)
51 0.02 0.17 0.02 0.01 0.34 0.04 0.01 0.34 0.04 0.40 0.40 52 0.02 0.18 0.02 0.01 0.35 0.01 0.01 0.35 0.04 0.42 0.40 53 0.06 0.32 0.02 0.03 0.25 0.01 0.03 0.25 0.03 0.35 0.31 54 0.05 0.42 0.01 0.03 0.21 0.01 0.03 0.21 0.03 0.29 0.25 55 0.01 0.44 0.01 0.03 0.22 0.01 0.03 0.22 0.03 0.26 0.25 56 0.02 0.17 0.02 0.01 0.34 0.04 0.01 0.34 0.04 0.40 0.40 57 0.02 0.18 0.02 0.01 0.35 0.01 0.01 0.35 0.04 0.42 0.40 58 0.04 0.29 0.04 0.02 0.15 0.02 0.07 0.36 0.02 0.20 0.18 59 0.04 0.21 0.01 0.02 0.17 0.02 0.08 0.42 0.02 0.23 0.21 60 0.03 0.24 0.03 0.02 0.12 0.02 0.06 0.48 0.02 0.16 0.14
61 0.04 0.30 0.01 0.02 0.15 0.02 0.08 0.37 0.02 0.21 0.18
62 0.04 0.20 0.01 0.02 0.20 0.03 0.08 0.40 0.03 0.27 0.25
63 0.03 0.24 0.01 0.02 0.12 0.02 0.06 0.49 0.02 0.17 0.14
64 0.03 0.25 0.03 0.02 0.13 0.02 0.02 0.50 0.02 0.17 0.14
65 0.01 0.26 0.03 0.02 0.13 0.02 0.02 0.51 0.02 0.15 0.14
66 0.02 0.20 0.02 0.05 0.24 0.01 0.05 0.39 0.01 0.28 0.29
67 0.01 0.26 0.03 0.02 0.13 0.02 0.02 0.51 0.02 0.15 0.14
68 0.05 0.23 0.01 0.02 0.18 0.02 0.02 0.45 0.02 0.25 0.21
69 0.03 0.26 0.01 0.02 0.13 0.02 0.02 0.51 0.02 0.18 0.14 70 0.01 0.26 0.01 0.02 0.13 0.02 0.02 0.53 0.02 0.16 0.14 71 0.04 0.29 0.04 0.03 0.26 0.03 0.03 0.26 0.03 0.33 0.32
72 0.02 0.18 0.02 0.01 0.33 0.04 0.04 0.33 0.04 0.38 0.39
73 0.01 0.43 0.05 0.01 0.21 0.02 0.03 0.21 0.02 0.25 0.25
74 0.03 0.24 0.03 0.02 0.30 0.02 0.06 0.30 0.02 0.34 0.36
75 0.02 0.17 0.02 0.01 0.33 0.04 0.05 0.33 0.04 0.39 0.40
76 0.02 0.17 0.02 0.01 0.33 0.04 0.04 0.33 0.04 0.39 0.40
77 0.02 0.18 0.02 0.01 0.35 0.01 0.04 0.35 0.01 0.39 0.40
78 0.02 0.44 0.01 0.01 0.22 0.03 0.03 0.22 0.03 0.26 0.25
79 0.02 0.18 0.02 0.01 0.36 0.01 0.01 0.36 0.01 0.40 0.40
80 0.02 0.17 0.02 0.04 0.34 0.01 0.04 0.34 0.01 0.37 0.40 Min 0.14 0.14 Max 0.42 0.40
3.3.3.2 Conditional Epistemic Irrelevance/Independence
To deal with the problem under conditional epistemic irrelevance/independence,
Campos and Cozman (2007) proposed an algorithm to compute lower and upper
78
expectations within a ‘Credal network’, which consists of ‘parent’ nodes and their
‘descendant’ nodes, connected by some directed arrows, as shown in Figure 3-3. Campos
and Cozman’s algorithm can deal with the unconditional epistemic
irrelevant/independent problem, but is more cumbersome compared to the algorithm in
Section 3.3.3.1. As for dealing with the conditional epistemic irrelevant/independent
problems, Campos and Cozman’s algorithm is designed for Credal network. In this
section, we are going to expand Campos and Cozman’s algorithm to a general form,
which can be applied to any problem with conditional epistemic
irrelevance/independence.
Figure 3-3 An example of Credal network
Let S1, S2, S3 be three random variables. Consider the case that S1 is epistemically
irrelevant to S2 given S3, and thus Eq. (3.25) can read as follows: ( ) ( )
( ) ( ) ( )
2 1 3 2 3
2 32 3
2 3
| |
|
i i
ii
T i
S s s S s
S sS s
S s
× =
×=
×
P q
1 q
(3.67)
where ( ) ( ) ( )( ) ( )312 3 2 3 2 2 33, , , niS s S s S s S S× × × = ×q q q q… … , and ( )2 3S S×q is a
valid measure over S2, S3, defined by
( ) ( )2 3 2 3S S S S× ∈Ψ ×q (3.68)
X1
X2
X3 X4
79
The reason why we do not write the constraints on ( )2 3| iS sq in terms of
( )2 1 3| iS s s×P is that set ( )2 3| iS sΨ is bounded by both constraints on ( )2 3| iS sq
and on ( )2 3S S×q , and that we cannot write the constrains on ( )2 3S S×q in terms of
( )2 1 3| iS s s×P . Therefore, here we create the new optimization variables ( )2 3S S×q ,
and all constraints on sets ( )2 3S SΨ × and ( )2 3|S SΨ could be expressed in terms of
( )2 3S S×q .
If S1 is epistemically independent of S2 given S3, constraints are obtained as
follows: ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )
( )
2 1 3 2 3 1 2 3 1 3
2 3 1 32 3 1 3
2 3 1 3
| | , | |
| , |
i i i i
i ii i
T i T i
S s s S s S s s S s
S s S sS s S s
S s S s
× = × =
× ×= =
× ×
P q P q
q qq q
1 q 1 q
(3.69)
where ( ) ( )( ) ( )312 3 2 2 33, , nS s S s S S× × = ×q q q… , and ( ) ( )( )31
1 3 1 3, , nS s S s× ×q q… =
( )1 3S S×q . ( )1 3S S×q and ( )2 3S S×q are defined by
( ) ( )1 3 1 3S S S S× ∈Ψ ×q
( ) ( )2 3 2 3S S S S× ∈Ψ ×q
(3.70)
The algorithm is illustrated by the following example, adapted from Campos and
Cozman (2007).
Example 3-4. Let X1, X2 and X3 be three binary random variables. Variable Xi takes values i and i . The
following information is available on the probability measure: 0.1 ≤ P(1) ≤ 0.3; 0.2 ≤ P(2|1) ≤ 0.5,
0.6 ≤ P(2| 1 ) ≤ 0.7; 0.7 ≤ P(3|2) ≤ 0.8, 0.3 ≤ P(3| 2 ) ≤ 0.4. Given X2, X1 is epistemic independent of X3. We are
interested to the upper and lower values of P(1,2,3)+P(1, 2 , 3 )+P( 1 ,2, 3 ).
Let P be the joint probability measure over X1, X2, X3, and thus its eight elements are defined by
p1=P(1,2,3), p2=P(1,2, 3 ), p3=P(1, 2 ,3), p4=P(1, 2 , 3 ), p5=P( 1 ,2,3), p6=P( 1 ,2, 3 ), p7=P( 1 , 2 ,3),
80
p8=P( 1 , 2 , 3 ). Let 1 2,3,3x xq and 1 2,
ˆ3,3x xq be two probability measures over X1, X2 for X3 = 3 and X3 = 3 ,
respectively, and 2 3,1,1x xq and 2 3,
ˆ1,1x xq be two probability measures over X2, X3 for X1 = 1 and X1 = 1 ,
respectively. Then the complete problem can be written as
Minimize 1 4 6p p p+ + Subject to
1 2 3 40.1 0.3p p p p≤ + + + ≤ 1 2
1 2 3 4
0.2 0.5p pp p p p
+≤ ≤
+ + +; 5 6
5 6 7 8
0.6 0.7p pp p p p
+≤ ≤
+ + +
1 5
1 2 5 6
0.7 0.8p pp p p p
+≤ ≤
+ + +; 3 7
3 4 7 8
0.3 0.4p pp p p p
+≤ ≤
+ + +
( ) ( )2,31,11
ˆ2,3 2,31 2 1,1 1,1
qpp p q q
=+ +
; ( ) ( )
2,31,13
ˆ ˆ ˆ2,3 2,33 4 1,1 1,1
qpp p q q
=+ +
;
( ) ( )2,3
ˆ1,15ˆ2,3 2,3
5 6 ˆ ˆ1,1 1,1
qpp p q q
=+ +
; ( ) ( )
2,3ˆ1,17
ˆ ˆ ˆ2,3 2,37 8 ˆ ˆ1,1 1,1
qpp p q q
=+ +
;
( ) ( )1,23,31
ˆ1,2 1,21 5 3,3 3,3
qpp p q q
=+ +
; ( ) ( )
ˆ1,23,33
ˆ ˆ ˆ1,2 1,23 7 3,3 3,3
qpp p q q
=+ +
;
( ) ( )1,2
ˆ3,32ˆ1,2 1,2
2 6 ˆ ˆ3,3 3,3
qpp p q q
=+ +
; ( ) ( )
ˆ1,2ˆ3,34
ˆ ˆ ˆ1,2 1,24 8 ˆ ˆ3,3 3,3
qpp p q q
=+ +
;
ˆ1,2 1,23,3 3,30.1 0.3q q≤ + ≤ ; ˆ1,2 1,2
ˆ ˆ3,3 3,30.1 0.3q q≤ + ≤
1,23,3
ˆ1,2 1,23,3 3,3
0.2 0.5q
q q≤ ≤
+;
1,2ˆ3,3
ˆ1,2 1,2ˆ ˆ3,3 3,3
0.2 0.5q
q q≤ ≤
+
1,23,3
ˆ ˆ ˆ1,2 1,23,3 3,3
0.6 0.7q
q q≤ ≤
+;
1,2ˆ3,3
ˆ ˆ ˆ1,2 1,2ˆ ˆ3,3 3,3
0.6 0.7q
q q≤ ≤
+
2,31,1
ˆ2,3 2,31,1 1,1
0.7 0.8q
q q≤ ≤
+;
2,3ˆ1,1
ˆ2,3 2,3ˆ ˆ1,1 1,1
0.7 0.8q
q q≤ ≤
+
2,31,1
ˆ ˆ ˆ2,3 2,31,1 1,1
0.3 0.4q
q q≤ ≤
+;
2,3ˆ1,1
ˆ ˆ ˆ2,3 2,3ˆ ˆ1,1 1,1
0.3 0.4q
q q≤ ≤
+
, ,3,3 3,30; 1i j i j
i jq q≥ =∑∑ ; , ,
ˆ ˆ3,3 3,30; 1i j i j
i jq q≥ =∑∑
, ,1,1 1,10; 1i j i j
i jq q≥ =∑∑ ; , ,
ˆ ˆ1,1 1,10; 1i j i j
i jq q≥ =∑∑
0; 1i ii
p p≥ =∑
(3.71)
81
Since all probabilities are positive, the 10 fraction constraints in Eq. (3.71) could be re-written in terms of
linear ones by using the techniques introduced in Section 3.3.3.1. The upper and lower probabilities of
P(1,2,3)+P(1, 2 , 3 )+P( 1 ,2, 3 ) are equal to 0.346 and 0.503. Solutions are summarized in Table 3-12.
Table 3-12 Example: Solutions of the optimization problems (3.71) for upper and lower conditional probabilities P(1,2,3)+P(1, 2 , 3 )+P( 1 ,2, 3 ).
Solution for P1 P2 P3 P4 P5 P6 P7 P8
Min 0.014 0.006 0.032 0.048 0.432 0.108 0.113 0.247
Max 0.108 0.027 0.050 0.116 0.441 0.189 0.028 0.042
Solution for
2 3,1,1x xq 2 3,
ˆ1,1x xq 1 2,
3,3x xq 1 2,
ˆ3,3x xq
Min 0.15 0.150.62 0.08⎛ ⎞⎜ ⎟⎝ ⎠
0.15 0.150.62 0.08⎛ ⎞⎜ ⎟⎝ ⎠
0.09 0.210.63 0.07⎛ ⎞⎜ ⎟⎝ ⎠
0.09 0.210.63 0.07⎛ ⎞⎜ ⎟⎝ ⎠
Max 0.02 0.080.62 0.28⎛ ⎞⎜ ⎟⎝ ⎠
0.02 0.080.62 0.28⎛ ⎞⎜ ⎟⎝ ⎠
0.03 0.070.54 0.36⎛ ⎞⎜ ⎟⎝ ⎠
0.03 0.070.54 0.36⎛ ⎞⎜ ⎟⎝ ⎠
If available information are given in terms of the upper and lower probabilities of P(2), P(1|2),
P(1| 2 ), P(3|2), and P(3| 2 ). Alternatively, we could use the algorithms described in Section 3.3.3.1 to find
all extreme distributions for sets ( )1 3, | 2X XΨ and ( )1 3ˆ, | 2X XΨ , then obtain extreme joint
distributions on X1, X2 and X3 by using Theorem 3-5, and finally find the min/max values of objective
function on some extreme joint distributions on X1, X2 and X3.
3.3.3.3 Strong Independence
The optimization problems under strong independence in Eqs. (3.31) and (3.32)
are NP-hard. To reduce the computational effort, Theorem 3-5 below addresses the
82
calculation of the extreme joint distributions and measures of the set of joint distributions,
sΨ . When strong independence is assumed, sΨ is not a convex set, as illustrated in
Figure 3-4 and explained in the Theorem 3-6.
Figure 3-4 Non-convex set of joint distribution, ΨS, under strong independence
Theorem 3-5 Under strong independence, the set of extreme joint distributions
(measures), EXT, is the set of product distributions (measures), each taken from the
extreme distributions (measures) of the marginals, ETXi: 1 2 1 1 2 2= : ,EXT P P P P EXT P EXT= ⊗ ∈ ∈ , i.e. (3.72)
( )1 2 1 2
,1 2= : ,
TEXT EXT EXT EXT EXTEXT EXT EXTξ η ξ η ξ η⎧ ⎫= ∈ ∈⎨ ⎬
⎩ ⎭P p p p p
(3.73)
Proof : Let the extreme points of the convex set of probability distributions on S1 and S2 be
1 1, 1,...EXTξ ξ ξ=p , and
2 2, 1,...EXTξ ξ ξ=p , respectively. Any p1 and p2 can be written as a
linear combination of extreme points:
( )( )1 11 1 1
1 11 1 1 1... ... ... ...
TEXT EXT EXT
ξ ξξ ξλ λ λ=p p p p ;
( )( )2 22 2 2
1 12 2 2 2... ... ... ...
TEXT EXT EXT
ξ ξξ ξλ λ λ=p p p p 1
2
11 11
22 21
0 1, 1,..., ; 1
0 1, 1,..., ; 1
ξξ ξ
ξ
ξξ ξ
ξ
λ ξ ξ λ
λ ξ ξ λ
=
=
≤ ≤ = =
≤ ≤ = =
∑
∑
(3.74)
sΨ2EXT
1EXT
3EXT
EXTξ
1EXTξ −
iEXT
83
Since 1 2= ⊗P p p , any joint probability distribution may be written as follows:
( ) ( )
( )
( )
( )
2
1 2 21 1 1
12
2
111
1 112 2 2
1
... ... ... ...
TEXT
T
EXTEXT EXT EXT
T
EXT
ξ ξξ ξξ ξ
ξξ
λ
λλ λ λ
λ
⎛ ⎞⎜ ⎟⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
⎝ ⎠ ⎜ ⎟⎜ ⎟⎝ ⎠
p
pP p p p
p
(3.75)
Extreme points of P are achieved if and only if 1
11
1, 0,
mm
ξ ξλ
ξ=⎧
= ⎨ ≠⎩and 2
22
1, 0,
mm
ξ ξλ
ξ=⎧
= ⎨ ≠⎩, 1 11,...,m ξ= , 2 21,...,m ξ= .
(3.76)
Therefore, EXT = ,PEXTξ η = ( )1 2 1 1:
T
EXT EXT EXT EXTξ η ξ ∈p p p , 2 2EXT EXTη ∈p . ◊
Theorem 3-6 Under strong independence, the set of joint distributions (measures), sΨ , is not convex.
Proof : Consider a counterexample with 1 3ξ = , 2 2ξ = in Eq.(3.73) and let us proceed
by contradiction by assuming that sΨ is a convex set . Let 1P and 2P be two joint
distributions in sΨ such that 1P is generated by taking ( )111 1 1... ... ξξλ λ λ = ( )1,0,0 ,
( )212 2 2... ... ξξλ λ λ = ( )0,1 , and thus ( ) ( )1 21 1
1 21 1 2 2... ... ... ...Tξ ξξ ξλ λ λ λ λ λ =
0 10 00 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
, and 2P is
generated by taking ( )111 1 1... ... ξξλ λ λ = ( )0,0,1 , ( )21
2 2 2... ... ξξλ λ λ = ( )1,0 , and thus
( ) ( )1 21 11 21 1 2 2... ... ... ...
Tξ ξξ ξλ λ λ λ λ λ =0 00 01 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
. The mid-point, mP , between 1P and 2P is
( )1 21 / 2 +P P . Consequently, ( ) ( )1 21 11 21 1 2 2... ... ... ...
Tξ ξξ ξλ λ λ λ λ λ for mP is equal to 0 1/20 0
1/2 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
,
which could be written in the form ( ) ( )1 2 3 1 21 1 1 2 2, , ,
Tλ λ λ λ λ based on the assumption that
84
mP is in the convex set sΨ . Thus, ( ) ( )1 2 3 1 21 1 1 2 2, , ,
Tλ λ λ λ λ =
1 1 1 21 2 1 22 1 2 2
1 2 1 23 1 3 21 2 1 2
,
,
,
λ λ λ λ
λ λ λ λ
λ λ λ λ
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
= 0 1/20 0
1/2 0
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
,
subject to 0jiλ ≥ , 1j
ijλ =∑ . Since 1 2
1 2λ λ =1/2 and 3 11 2λ λ =1/2, 1 2 3 1
1 2 1 2λ λ λ λ =1/4. However,
we also have 1 11 2λ λ =0 and 3 2
1 2λ λ =0, so 1 1 3 21 2 1 2λ λ λ λ = 1 2 3 1
1 2 1 2λ λ λ λ =0, which contradicts the
previous result 1 2 3 11 2 1 2λ λ λ λ =1/4. Therefore, there are no j
iλ that satisfy all requirements,
i.e., 1 2m ≠ ⊗P P P . Therefore, sΨ is not convex. ◊
When Option (2) is used, Eq.(3.73) provides an efficient algorithm to find all
extreme distributions of sΨ , as compared to finding the extreme distributions by setting
up the quadratic constraints based on the definition of strong independence in Eq.(3.29).
Although sΨ is not convex, Theorem 3-1 insures that maxima and minima of
linear functions are always achieved at extreme points of sΨ . There are several options
for carrying out calculations on the joint space (Bernardini and Tonon 2010):
1. If extreme distributions are available for the marginals (Eq.(3.3)), first
calculate all ξ1×ξ2 extreme joint distributions ,EXTξ ηP as indicated in
Eq.(3.55) and then calculate the objective function (e.g., prevision or
probability) on all extreme joint distributions, which is much easier to
solve than the non-convex and NP-hard optimization problems that follow.
2. If expectation (prevision) bounds are given on the marginals (Eq. (3.3)),
use these constraints to directly obtain quadratic constraints in the n1×n2
components pi,j and n1 + n2 components Pi(sik). This option is obviously
less efficient than the first one:
Subject to:
85
( )
( )1 2 0
; 1,..., ; 1,2
T
Tk k kLOW i i i UPP i iE f E f k k i
− =
⎡ ⎤ ⎡ ⎤≤ ≤ = =⎣ ⎦ ⎣ ⎦
P p p
f p
(3.77)
3. If expectation (prevision) bounds are given on the marginals (Eq. (3.3)),
first use constraints to obtain extreme distributions for the marginals, and
then proceed as stated in the first algorithm.
Example 3-5. Consider again the situation and knowledge available in Example 3-1, but now suppose
that a bolt and a nut are picked from each box in a stochastically independent way. We want to know the
upper and lower probabilities for the event T=(A, A), (B, B), (C,C), and the conditional probability that
the bolt is Type B given the type of nut.
Let us solve this problem by using Theorem 3-5 and the first option above. There are 3 extreme
points on Ψ1 and 4 extreme points on Ψ2, thus the number of extreme points on joint Ψs would not exceed
3× 4=12. Extreme points are listed in Table 3-13, which are the products of two extreme marginal
distributions. The lower and upper probabilities for T are 0.14 and 0.4, respectively; and the conditional
probability that the bolt is Type B given the type of nut is from 0.14 and 0.4, which is the same as the
bounds of the marginal of Ψ1.
Table 3-13 Example 3-5: Lower and upper probabilities for T and lower and upper conditional probability ( )1|2 1 2|P S B S B= = on all extreme distributions of Ψ (optimal solutions are highlighted.)
Extreme point of Ψ p1,1 p1,2 p1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,2
1,1 2,2 3,3p p p+ + p2,2/( p1,2+ p2,2+ p3,2)
1 0.03 0.23 0.03 0.01 0.11 0.01 0.06 0.46 0.06 0.20 0.14 2 0.05 0.40 0.05 0.03 0.20 0.03 0.03 0.20 0.03 0.28 0.25 3 0.02 0.16 0.02 0.04 0.32 0.04 0.04 0.32 0.04 0.38 0.40 4 0.06 0.23 0.00 0.03 0.11 0.00 0.11 0.46 0.00 0.17 0.14 5 0.10 0.40 0.00 0.05 0.20 0.00 0.05 0.20 0.00 0.30 0.25 6 0.04 0.16 0.00 0.08 0.32 0.00 0.08 0.32 0.00 0.36 0.40 7 0.00 0.26 0.03 0.00 0.13 0.01 0.00 0.51 0.06 0.19 0.14 8 0.00 0.45 0.05 0.00 0.23 0.03 0.00 0.23 0.03 0.25 0.25
86
9 0.00 0.18 0.02 0.00 0.36 0.04 0.00 0.36 0.04 0.40 0.40 10 0.00 0.29 0.00 0.00 0.14 0.00 0.00 0.57 0.00 0.14 0.14 11 0.00 0.50 0.00 0.00 0.25 0.00 0.00 0.25 0.00 0.25 0.25 12 0.00 0.20 0.00 0.00 0.40 0.00 0.00 0.40 0.00 0.40 0.40
Min 0.14 0.14
Max 0.40 0.40
3.3.3.4 Conditional strong Independence
Let S1, S2, S3 be three random variables. Consider the case that S1 is conditional
strong independent of S2 given S3, and thus Eq. (3.35) can read as follows:
( )( ) ( )
( )( )
2 1
1 2
1 2
1 3 1 32 21 1
1 3 1 32 21 1
1 321 1
, , , ,
, , , for , , 0, ,
n ni j l i j l
n nj ii j l i j l
n ni ji j l
i j
P s s s P s s s
P s s s P s s sP s s s
= =
= =
= =
⎛ ⎞⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠= >∑ ∑
∑∑∑∑
(3.78)
The algorithm is illustrated by the following example:
Example 3-6 Redo Example 3-4, but with the information that given X2, X1 is conditional strong
independent to X3. Then the complete problem can be written as
87
Minimize 1 4 6p p p+ + Subject to
1 2 3 40.1 0.3p p p p≤ + + + ≤ 1 2
1 2 3 4
0.2 0.5p pp p p p
+≤ ≤
+ + +; 5 6
5 6 7 8
0.6 0.7p pp p p p
+≤ ≤
+ + +
1 5
1 2 5 6
0.7 0.8p pp p p p
+≤ ≤
+ + +; 3 7
3 4 7 8
0.3 0.4p pp p p p
+≤ ≤
+ + +
( )( )1 2 1 51
1 2 5 6
p p p pp
p p p p+ +
=+ + +
; ( )( )1 2 2 62
1 2 5 6
p p p pp
p p p p+ +
=+ + +
;
( )( )3 4 3 73
3 4 7 8
p p p pp
p p p p+ +
=+ + +
; ( )( )3 4 4 84
3 4 7 8
p p p pp
p p p p+ +
=+ + +
;
( )( )5 6 1 55
1 2 5 6
p p p pp
p p p p+ +
=+ + +
; ( )( )5 6 2 66
1 2 5 6
p p p pp
p p p p+ +
=+ + +
;
( )( )7 8 3 77
3 4 7 8
p p p pp
p p p p+ +
=+ + +
; ( )( )7 8 4 88
3 4 7 8
p p p pp
p p p p+ +
=+ + +
;
0; 1i ii
p p≥ =∑
(3.79)
The upper and lower probabilities of P(1,2,3)+P(1, 2 , 3 )+P( 1 ,2, 3 ) are equal to 0.172 and 0.399. Solutions
are summarized in Table 3-14.
Table 3-14 Example: Solutions of the optimization problems (3.79) for upper and lower conditional probabilities P(1,2,3)+P(1, 2 , 3 )+P( 1 ,2, 3 ).
P1 P2 P3 P4 P5 P6 P7 P8 P1+P4+P6
min 0.016 0.004 0.032 0.048 0.432 0.108 0.144 0.216 0.172
max 0.042 0.018 0.072 0.168 0.441 0.189 0.021 0.049 0.399
If available information are given in terms of the upper and lower probabilities of P(2), P(1|2),
P(1| 2 ), P(3|2), and P(3| 2 ). Alternatively, we could use Theorem 3-5 to find all extreme distributions for
sets ( )1 3, | 2X XΨ and ( )1 3ˆ, | 2X XΨ , then obtain extreme joint distributions on X1, X2 and X3 by using
Theorem 3-5 again, and finally find the min/max values of objective function on some extreme joint
distributions on X1, X2 and X3.
88
3.3.4 Uncertain Correlation
According to the problem formulation in Eq.(3.41), the problem under uncertain
correlation reads as follows:
Minimize(Maximize) 1 2; , ,1; 1
i n j n i j i ji j
a P= =
= =∑
Subject to ( ) ( ) ( ) ( ) ( )
1 2 1 21 2 1 2 1 2S S S SE S E S D D E S S E S E S D Dρ ρ+ ≤ ≤ +
( )( )
1 2
1 1 1 1 1
2 2 2 2 2
( ) 1 ( ) 2
1 2
; 1,..., ;
; 1,..., ;
1 1
0 ; 0
Tk k kLOW UPP
Tk k kLOW UPP
T Tn n
E f E f k k
E f E f k k
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⋅ = ⋅ =
≥ ≥
f p
f p
1 p 1 p
p p
(3.80)
where ( ) ( )22 , 1,2iS i iD E S E S i= − = , ( )E S is the expected value of variable S.
Obviously, the constraint for uncertain correlation in (3.80) is nonlinear.
Algorithms of solving non-linear programming problems could be found in Luenberger
(1984).
Example 3-7. Consider again the situation and knowledge available in Example 3-1. Suppose that now
the type of the picked bolt from box 1 and the type of the picked nut in Box 2 has some uncertain
correlation: 0.1 < ρ < 0.5. We want to write down the optimization problems for finding upper and lower
expectations on the joint space and then calculate the upper and lower probabilities for the case in which
the same type of bolt and nut is selected, i.e. event T=(A, A), (B, B), (C, C). Finally, calculate upper and
lower conditional probabilities that the bolt is Type B given the type of nut, and contrast to upper and lower
conditional probabilities that the nut is Type B given the type of the bolt.
The nonlinear programming problem in Eq. (3.80) becomes:
89
Minimize (Maximize) 1,1 2,2 3,3p p p+ + (2,2
1,2 2,2 3,2p
p p p+ +,
2,2
2,1 2,2 2,3p
p p p+ +)
Subject to
1, 1, ,1 ,1 1,1 1, ,1
1, 1, ,1 ,1
0.5 1 1
0.8 1 1
i i j j i j
i i j j i j
i i j j
i i j j
p p p p p p p
p p p p
⎛ ⎞⎛ ⎞− ⎜ − ⎟ ≤ −⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
⎛ ⎞⎛ ⎞≤ − ⎜ − ⎟⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
∑ ∑ ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑
2, 2, ,2 ,2 2,2 2, ,2
2, 2, ,2 ,2
0.5 1 1
0.8 1 1
i i j j i j
i i j j i j
i i j j
i i j j
p p p p p p p
p p p p
⎛ ⎞⎛ ⎞− ⎜ − ⎟ ≤ −⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
⎛ ⎞⎛ ⎞≤ − ⎜ − ⎟⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
∑ ∑ ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑
3, 3, ,3 ,3 3,3 3, ,3
3, 3, ,3 ,3
0.5 1 1
0.8 1 1
i i j j i j
i i j j i j
i i j j
i i j j
p p p p p p p
p p p p
⎛ ⎞⎛ ⎞− ⎜ − ⎟ ≤ −⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
⎛ ⎞⎛ ⎞≤ − ⎜ − ⎟⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
∑ ∑ ∑ ∑ ∑ ∑
∑ ∑ ∑ ∑
( ) ( )( ) ( )( ) ( )
1,1 1,2 1,3 2,1 2,2 2,3
1,1 1,2 1,3 3,1 3,2 3,3
2,1 2,2 2,3 3,1 3,2 3,3
2 0
2 0
0
p p p p p p
p p p p p p
p p p p p p
− + + + ⋅ + + ≥
⋅ + + − + + ≥
− + + + + + ≥
1,1 2,1 3,1
1,2 2,2 3,2
1,3 2,3 3,3
1,1 1,2 1,3 2,1 2,2 2,3 3,1 3,2 3,3
,
0.0 0.2
0.8 1.0
0.0 0.1
1
0i j
p p p
p p p
p p p
p p p p p p p p p
p
≤ + + ≤
≤ + + ≤
≤ + + ≤
+ + + + + + + + =
≥
(3.81)
Lower and upper probabilities for the event T = (A, A), (B, B), (C, C) are equal to 0.178 and
0.565, respectively. Lower and upper conditional probabilities P(S1=B|S2=B) are 0.151 and 0.438 , and the
lower and upper conditional probabilities P(S2=B|S1=B) are 0.849 and 1. Solutions are detailed in Table
3-15 through Table 3-17.
Table 3-15 Example: Solutions of the optimization problems (3.81) for upper and lower probabilities for event T = (A, A), (B, B), (C, C).
Solution for P1,1 P1,2 P1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,3
Min 0.022 0.264 0.000 0.000 0.143 0.000 0.022 0.536 0.013Max 0.068 0.131 0.001 0.002 0.398 0.000 0.030 0.271 0.099
90
Table 3-16 Example: Solutions of the optimization problems (3.81) for upper and lower probabilities for P(S1=B|S2=B).
Solution for P1,1 P1,2 P1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,3
Min 0.042 0.244 0.000 0.000 0.143 0.000 0.001 0.557 0.013
Max 0.000 0.200 0.000 0.000 0.400 0.000 0.000 0.313 0.087
Table 3-17 Example: Solutions of the optimization problems (3.81) for upper and lower probabilities for P(S2=B|S1=B).
Solution for P1,1 P1,2 P1,3 P2,1 P2,2 P2,3 P3,1 P3,2 P3,3
Min 0.032 0.167 0.001 0.042 0.340 0.019 0.027 0.293 0.080
Max 0.032 0.168 0.000 0.000 0.400 0.000 0.068 0.254 0.078
3.4 SUMMARY
The constraints that define the five sets of probability measures are summarized
as follows:
Unknown interaction: ΨU: (3.7)
Epistemic irrelevance: |E
isΨ : (3.7) + (one of (3.21))
Conditional epistemic irrelevance: 3
|E|
issΨ : (3.7)+(3.26) (or (3.28))
Epistemic independence: ΨE: (3.7) + (3.21)
Conditional epistemic independence: 3E|sΨ : (3.7)+ (3.26)+ (3.28)
Strong independence: ΨS: (3.7) + (3.21) + (3.30)
Conditional strong independence: 3S|sΨ : (3.7)+(3.36)
Uncertain correlation: ΨC: (3.7) +(3.40)
91
Since constraints are consecutively added, the sets of probability measures are
nested, i.e. ΨS ⊆ ΨE ⊆ |E
isΨ ⊆ ΨU; ΨC ⊆ ΨU; and 3S|sΨ ⊆
3E|sΨ ⊆ 3
|E|
issΨ ⊆ ΨU. As a
consequence, the upper and lower probability bounds are also nested: | |
, , , , , , , ,i is s
U low E low E low S low S upp E upp E upp U uppP P P P P P P P≤ ≤ ≤ ≤ ≤ ≤ ≤
, , , ,U low C low C upp U uppP P P P≤ ≤ ≤
3 3 3 3 3 3
| |, | , | , | , | , | , | , ,
i is sU low E s low E s low S s low S s upp E s upp E s upp U uppP P P P P P P P≤ ≤ ≤ ≤ ≤ ≤ ≤
(3.82)
As for conditional probability on joint space, the upper and lower conditional
probability bounds are: | |
, , , , , , , ,i is s
U low E low E low S low S upp E upp E upp U uppP P P P P P P P≤ ≤ = ≤ = ≤ ≤
, , , ,U low C low C upp U uppP P P P≤ ≤ ≤
3 3 3 3 3 3
| |, | , | , | , | , | , | , ,
i is sU low E s low E s low S s low S s upp E s upp E s upp U uppP P P P P P P P≤ ≤ = ≤ = ≤ ≤
(3.83)
This is exemplified by the probability of set T=(A, A), (B, B), (C,C) in
Example 1 through Example 4:
Unknown interaction: , ,0.0; 0.60U low U uppP P= =
Epistemic irrelevance: 1 1| |, ,0.11; 0.48s s
E low E uppP P= =
Epistemic independence: , ,0.14; 0.42E low E uppP P= =
Strong independence: , ,0.14; 0.40S low S uppP P= =
Uncertain correlation: , ,0.178; 0.565C low C uppP P= =
Also, the conditional probabilities that the bolt is type B given the type of nut are
Unknown interaction: , ,0; 0.50U low U uppP P= =
Epistemic irrelevance: 1 1| |, ,0.12; 0.45s s
E low E lowP P= =
Epistemic independence: , ,0.14; 0.40E low E uppP P= =
Strong independence: , ,0.14; 0.40S low S uppP P= =
Uncertain correlation: , ,0.151; 0.438C low C uppP P= =
92
Table 3-18 summarizes all algorithms discussed in this chapter and their
applications to the cases under different types of independence. The algorithms are
divided into two groups by the type of constraints: prevision bounds or extreme points of
marginals. Both Option (1) and Option (2) are applied to solve problems under each type
of constraints. Every option includes two algorithms, written in terms of the joint
probability or marginals. The applicability of each algorithm to a specified type of
independence is listed in the table. For example, if prevision bounds are given on the
marginals, consider the case in which one is interested in the extreme values of the
conditional probability and decides to use Option (1) written in terms of the joint
probability. Then one has to solve an optimization problem with linear objective
functions and linear constraints, i.e. N/L as stated in Table 3-18.
Note: E: Enumerate all joint extreme points UI: Unknown Interaction; EIR: Epistemic Irrelevance
EIN: Epistemic Independence SI: Strong Independence NA: Not available L: Nonlinear N: Nonlinear Q: Quadratic
93
Table 3-18 Summary of all algorithms for different types of independence.
Independence Constraints Given as Algorithm
UI EIR EIN SI Objective
L/L L/L L/L L/Q Prevision on joint distributionConstraints written
in terms of P N/L N/L N/L N/Q Conditional
Probability
Q/Q Q/Q Q/Q Prevision on joint distribution
Option (1): Optimization
problem Constraints written in terms of p1 and
p2 NA
N/Q N/Q N/Q Conditional Probability
L/L L/L L/L Prevision on joint distributionConstraints written
in terms of P N/L N/L N/L NA Conditional
Probability Prevision on
joint distribution
Bounds on Previsions of
marginals
Option (2): Find all extreme
joint distributions. Constraints written in terms of p1 and
p2 NA Conditional
Probability
L/L L/Q L/Q L/Q Prevision on joint distributionConstraints written
in terms of P N/L N/Q N/Q N/Q Conditional Probability
L/L Q/Q Q/Q Q/Q Prevision on joint distribution
Option (1): Optimization
problem Constraints written in terms of p1 and
p2 N/L N/Q N/Q N/Q Conditional Probability
L/L Prevision on joint distributionConstraints written
in terms of P N/LNA Conditional
Probability Prevision on
joint distributionConstraints written in terms of p1 and
p2 NA Conditional
Probability
L/L Prevision on joint distributionCombination of
EXTs on p1 and p2NA
N/L NA NA Conditional
Probability Prevision on
joint distribution
Extreme points of marginals
Option (2): Find all extreme
joint distributions.
Enumeration of all joint extreme
points E Conditional
Probability
94
Chapter 4 Failure and Decision Analysis
This chapter focuses on the application of Imprecise Probability to failure analysis
(i.e. Event Tree Analysis and Fault Tree Analysis) and decision analysis. As explained in
Chapter 3, the available information is used to construct a convex set of probability
distributions, which are then considered during failure analysis and decision making. In
the failure analysis, our aim is to determine the upper and lower bounds of a prevision
(expectation of a real function) or of the probability of failure; in the decision analysis,
our objective is to determine the optimal action(s) to take. Corresponding algorithms are
developed and illustrated by examples. All theorems presented in this chapter are
developed and proved by the author.
4.1 INTRODUCTION
In risk analysis, we often face the problem of evaluating the prevision or the
probability of some undesirable consequence (i.e. failure), and of making the best
decision based on the available information. To answer the first question, two common
methodologies in failure analysis are Event Tree Analysis (ETA) and Fault Tree Analysis
(FTA). As for the second question, Decision Tree Analysis is a usual technique. For
example, the International Tunnelling and Underground Space Association published
guidelines (Eskesen et al., 2004) on risk analysis and recommended using Event Tree
Analysis, Fault Tree Analysis and Decision Tree Analysis as risk analysis tools in
tunneling.
An event tree presents an inductive logical relationship, starting from a hazardous
event, which is called as “initiating event” and followed by all the possible outcomes. By
analyzing all outcomes at the bottom of the tree, the failure paths are determined. The
95
failure probability is obtained by summing up the probabilities calculated on each failure
path. For example, Figure 4-1 shows an event tree for the case of “Pedestrian walks
against red light”. By assigning probabilities to all outcomes, the probability of
occurrence of an accident could be evaluated quantitatively. Detailed introduction of
Event Tree Analysis with precise probability in Civil Engineering could be found in
Benjamin and Cornell (1970).
Figure 4-1: Example of Event Tree (Eskesen, 2004).
A fault tree analysis is a deductive analytical technique. It starts from a specified
state of the system as the “Top event”, and includes all faults which could contribute the
top event. At the bottom of the fault tree, the basic initiating faults, which could not be
further developed, are called as “Basic events” linked by fault tree gates, including AND-
, OR-, and Exclusive OR- gates etc. Figure 4-2 illustrates a fault tree to analyze the
failure of a toll sub-see tunnel project. For further introduction and application of Fault
Tree Analysis with precise probability in Civil Engineering, see Ang and Tang (1984).
96
Figure 4-2: Example of Fault Tree (Eskesen, 2004).
A decision Tree presents all choices in a tree-like structure with the information
of consequences and probabilities. Figure 4-3 shows an example of decision tree. With
specific probabilities and consequences, one could first make a decision between Choice
1a and 1b. The selected action between these two becomes the representative of Choice 1.
Then Choice 1 and 2 are compared based on their expected value. Finally, the one with
maximum expected value is selected as the best option. Detailed introduction of Decision
Tree Analysis with precise probability in Civil Engineering may be found in Benjamin
and Cornell (1970) and Ang and Tang (1984).
97
Figure 4-3: Example of Decision Tree (Eskesen, 2004).
In the conventional fault analysis and decision analysis, probabilities are usually
treated as precise values. By assigning a single probability distribution to all events in the
tree (Whitman, 1984), the prevision or probability of some outcomes is evaluated
quantitatively.
In tunneling projects, ground and groundwater conditions are affected by large
uncertainty, and thus past experience or data may not be reliably used for evaluating the
probabilities precisely. As a result, the available evidence in the event-tree analysis relies
on the judgment of experienced engineers and experts. Those judgments might be some
probability intervals, or more generally, upper and lower previsions (expectations) of
gambles (bounded real functions), as the constraints expressed in Section 2.1. In general,
these constraints generate convex sets of probability distributions. Thus, the following
sub-sections will solve event trees, fault trees, and decision trees with convex sets of
probabilities.
Several references allowed for imprecision in probabilities when dealing with
event-tree analysis, such as Tonon et al. (2000), Huang et al. (2001), and Kenarangui
98
(1991); random sets or interval probabilities were used to evaluate the input probabilities,
but they still did not consider the most general case, i.e. convex sets of probability
distributions.
To illustrate, a random set on a finite universe Ω is composed of n non-empty
subsets i ⊆ΩA with associated probability assignment ( )im A : ( )im A > 0,
( )1
n
ii
m=∑ A = 1. Bernardini and Tonon (2010) explains that a random set
( )( ) , , 1,...,i im i n=A A generates a convex set of probability distributions bounded by
belief (or plausibility) functions, which are special cases of upper and lower probabilities.
Upper and lower probabilities are special cases of previsions when the gamble is the
characteristic function of an event (Tonon et al. 2001).
4.2 FAILURE ANALYSIS WITH IMPRECISE PROBABILITY
4.2.1 Event Tree Analysis (ETA)
In the event tree analysis, the available information on probabilities could be
given in three different types: (1) probabilities conditional to the event at the upper level;
(2) total probabilities of occurrences, i.e., probabilities not conditional to other events; (3)
the combination of the previous two types. In the following sub-sections we will discuss
the solutions in these three cases.
4.2.1.1 ETA with conditional probabilities
This sub-section deals with event-tree analysis when the probabilities conditional
to the upper level events are given. Let N be the number of levels of the event tree (see
Figure 4-4). The subscript ( )21, ,..., Ni i is used to index an event in the tree: if the event
is located at level k, then ij = 0 for j = k+1,…,N. Let 21, ,..., ,0,..,0ki iS be an event at level k
99
with 21, ,..., ,0,..,0ki in possible states (or outcomes). It is located on the ik-th branch of event
2 11, ,..., ,0,...,0ki iS−
at level k-1, which is on the ik-1-th branch of event 2 21, ,..., ,0,...,0ki iS
− at Level k-
2, …, and so on till event 21, ,0,0,...,0iS on the i2-th branch of the event 1,0,...,0S at level 1.
For example, if N = 4 in Figure 4-4, then 1,2,1,0S is an event at level 3, which is located
on the 1st branch of event 1,2,0,0S , which is on the 2nd branch of event 1,0,0,0S .
Figure 4-4: Event-tree with N levels.
In the middle of the tree, 21, ,..., ,0,...,0ki ip denotes the probability vector for event
21, ,..., ,0,..,0ki iS , with its j-th component 21, ,..., , ,0,...0ki i jp being the probability of occurrence of
the j-th state 21, ,..., , ,..,0ki i jS of event
21, ,..., ,0,..,0ki iS . Let 21, ,..., ,0,...,0ki iΩ be the set of outcomes
of event 21, ,..., , ,..,0ki i jS . Let
21, ,..., ,0,...,0k
mi if :
21, ,..., ,0,...,0ki iΩ → , 21, ,..., ,0,...,01,...,
ki im m= be a set
of limited functions (gambles) on 21, ,..., ,0,...,0ki iΩ .
21, ,..., ,0,...,0k
mi if denotes a
21, ,..., ,0,..,0ki in -
1
1,0,0,...,0N
S−
1,1,0,...,0S
……
1,2,0,...,0S
1,4,0,...,0S
1,1,1,...,0S
1,1,2,...,0S
1,1,3,...,0S
……
1,4,1,...,0S
1,4,2,...,0S
1,4,3,...,0S
1,2,1,...,0S
1,2,2,...,0S
1
1,1,...,1,1N
s+
1,1,...,1,2s1
1,1,...,1,1N
a+
1,1,...,1,2a 1,1,...,1,0N
a
……
……
……
……
Level 1 Level 2 Level 3 …… Level N Outcome Consequence i2 i3 …… iN
1,1,...,2,2s 1,1,...,2,2a 1,1,...,2,0a1,1,...,2,1s 1,1,...,2,1a
Tree Bottom
Tree Top
1,1,...,1N
S
1,1,...,2S
……
100
column vector, whose j-th component is the function value for the j-th state 21, ,..., , ,..,0ki i js
for event 21, ,..., ,0,..,0ki iS .
21, ,..., ,0,...,0ki ip denotes one element in set 21, ,..., ,0,...,0ki iΨ of
probability distributions whose expectations (previsions) for 21, ,..., ,0,...,0k
mi if fall between
assigned upper and lower bounds 21, ,..., ,0,...,0k
mUPP i iE f⎡ ⎤
⎣ ⎦ and 21, ,..., ,0,...,0k
mLOW i iE f⎡ ⎤
⎣ ⎦ ,
respectively. 21, ,..., ,0,...,0ki iΨ is thus defined as follows:
( )( )
2
2 2 2
2
2 2 2
1, ,..., ,0,...,0
1, ,..., ,0,...,0 1, ,..., ,0,...,0 1, ,..., ,0,...,01, ,..., ,0,...,0
1, ,..., ,0,...,0 1, ,..., ,0,...,0 1, ,..., ,0,...,0
:
;
;
k
k k k
k
k k k
i i
Tm mi i i i LOW i i
i i Tm mi i i i UPP i i
E f
E f
⎡ ⎤≥ ⎣ ⎦Ψ =⎡ ⎤≤ ⎣ ⎦
p
f p
f p
21, ,..., ,0,...,0 1,...,ki im m
⎧ ⎫⎪ ⎪⎪ ⎪⎪ ⎪⎨ ⎬⎪ ⎪⎪ ⎪⎪ ⎪=⎩ ⎭
(4.1)
By applying the algorithm for finding extreme points of set 21, ,..., ,0,...,0ki iΨ described in
Section 2.4, the extreme distributions for 21, ,..., ,0,...,0ki ip (vertices of
21, ,..., ,0,...,0ki iΨ ) may be
obtained:
1, ,..., ,0,...,022 1, ,..., ,0,...,0 1, ,..., ,0,...,0 1, ,..., ,0,...,02 2 2
1 21, ,..., ,0,...,0 , ,..., i ik
k i i i i i ik k ki i EXT EXT EXTEXT ξ
= p p p (4.2)
Let 21, ,..., ,0,...,0ki ia be the consequence vector for event
21, ,..., ,0,..,0ki iS , where its j-th
component 21, ,..., , ,0,...0ki i ja is the consequence of the occurrence of the j-th state
21, ,..., , ,..,0ki i js
of event 21, ,..., ,0,..,0ki iS . Let
2 11, ,..., , ,...,0k ki i iE+
be the prevision (expectation) of event
2 11, ,..., , ,0,..,0k ki i iS+
conditional to its upper level events. If all 2 11, ,..., , ,...,0k ki i iE
+
(21 1, ,..., ,0,...,01,...,
kk i ii n+ = ) are calculated, consequence vector 21, ,..., ,0,...,0ki ia at level k is
2
2
2 1, ,..., ,0,...,02
1, ,..., ,1,0,...,0
1, ,..., ,0,...,0
1, ,..., , ,0,...,0
k
k
k i ik
i i
i i
i i n
E
E
⎛ ⎞⎜ ⎟
= ⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
a (4.3)
At the bottom of the event tree, let 21, ,..., ,Ni i js be the j-th possible state of event
21, ,..., Ni iS , and let 21, ,..., ,Ni i ja be the numeric value of the consequence of its outcome. Thus,
vector 21, ,..., Ni ia is the consequence vector and
21, ,..., ,Ni i ja is its j-th component. Let
101
21, ,..., Ni ip be the associated probability vector where its j-th component 21, ,..., ,Ni i jp is the
probability of outcome 21, ,..., ,Ni i js conditional to the occurrence of upper level events, i.e.,
2
2
2 1, ,...,2
1, ,..., ,1
1, ,...,
1, ,..., ,
N
N
N i iN
i i
i i
i i n
a
a
⎛ ⎞⎜ ⎟
= ⎜ ⎟⎜ ⎟⎝ ⎠
a ; 2
2
2 1, ,...,2
1, ,..., ,1
1, ,...,
1, ,..., ,
N
N
N i iN
i i
i i
i i n
p
p
⎛ ⎞⎜ ⎟
= ⎜ ⎟⎜ ⎟⎝ ⎠
p (4.4)
Each outcome at the bottom of the event tree has one probability path. Assume
the ζ -th outcome 21, ,..., ,Ni is ζ of event
21, ,..., Ni iS is the j-th outcome at the bottom of the
event tree. The probability path is composed of N events: 1,0,...,0S , 21, ,0...,0iS , …,
21, ,..., Ni iS ,
with associated N conditional probabilities equal to 21, ,0,...,0ip ,
2 31, , ,0,...,0i ip , …, and
21, ,..., ,Ni ip ζ , respectively. The consequence for outcome 21, ,..., ,Ni is ζ is
21, ,..., ,Ni ia ζ , and thus
the prevision contributed by this probability path is given by the product ( )( ) ( )
( )2 2 2 3 21, ,..., , 1, ,0,...,0 1, , ,0,...,0 1, ,..., ,N Ni i i i i i i ja p p pζ ζ⎡ ⎤⎣ ⎦ . Assume that there are a total
of M events at the bottom of the tree: the total prevision of the whole event tree is
( )( ) ( )( )
1, ,...,2
2 2 2 3 21, ,..., , 1, ,0,...,0 1, , ,0,...,0 1, ,..., ,1 1
i iN
N N
nM
i i i i i i i jjE a p p pζ ζ
ζ= =
⎡ ⎤= ⎣ ⎦∑ ∑ .
To calculate the upper and lower previsions from the event tree, the following
constrained optimization problems must be solved: Minimize (Maximize)
( )( ) ( )( )
1, ,...,2
2 2 2 3 21, ,..., , 1, ,0,...,0 1, , ,0,...,0 1, ,..., ,1 1
i iN
N N
nM
i i i i i i i jja p p pζ ζ
ζ= =
⎡ ⎤⎣ ⎦∑ ∑
Subject to
( )( )
2 2 2
2 2 2
2
1, ,..., ,0,...,0 1, ,..., ,0,...,0 1, ,..., ,0,...,0
1, ,..., ,0,...,0 1, ,..., ,0,...,0 1, ,..., ,0,...,0
1, ,..., ,0,...,0
;
;
1,..., ; 1,...,
k k k
k k k
k
Tm mi i i i LOW i i
Tm mi i i i UPP i i
i i
E f
E f
m m k N
⎫⎡ ⎤≥ ⎪⎣ ⎦⎪⎪⎡ ⎤≤ ⎬⎣ ⎦⎪
= = ⎪⎪⎭
f p
f p (Eq.(3.1))
( ) 21, ,..., ,0,...,0 22
2 2
1, ,..., ,0,...,0 1, ,..., ,0,...,0
1, ,..., , ,...,0 1, ,..., ,0,...,0
1 must be
a probability distribution0; 1,..., ; 1,...,
ki i kk
k k
Ti in i i
i i j i ip k N j n
⎫⋅ = ⎪⎬⎪≥ = = ⎭
1 p p
(4.5)
102
The objective function in these optimization problems is the sum of products of
M N× variables subject to linear constraints. The optimization problems are thus non-
linear and non-convex. Besides being computationally intensive (Luenberger, 1984), non-
linear and non-convex problems may yield local minima and maxima.
We now show how problem (4.5) can be efficiently solved by a sequence of linear
programming problems (when prevision bounds UPPE and LOWE are known
(Eq.(3.1)), Case 1) or by enumeration (when the extreme points for iΨ are known
(Eq.(4.2), Case 2).
At level N, given the probability distributions 21, ,..., Ni ip conditional to the
occurrence of event 2 11, ,..., ,0Ni iS
−, the upper and lower previsions 21, ,..., Ni iE and
21, ,..., Ni iE of
event 21, ,..., Ni iS can be obtained by solving the following optimization problems:
2 2 2
1, ,..., 1, ,..., 1, ,...,maxN N N
Ti i i i i iE = pa
2 2 21, ,..., 1, ,..., 1, ,...,minN N N
Ti i i i i iE = pa
Subject to
( )2 2 2 2 21, ,..., 1, ,..., 1, ,..., 1, ,..., 1, ,..., ; 1,...,N N N N N
Tm m mLOW i i i i i i UPP i i i iE f E f m m⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦f p
( ) 21, ,...,2
2 2
1, ,...,
1, ,..., , 1, ,...,
1
0; 1,...,
Ni iN
N N
Ti in
i i j i ip j n
⋅ =
≥ =
1 p
(4.6)
In Case 1, problems (4.6) can be solved as linear programming problems. In Case 2, the upper and lower previsions 21, ,..., Ni iE and
21, ,..., Ni iE of event
21, ,..., Ni iS are determined by enumerating the extreme probability distributions of 21, ,..., Ni iΨ ,
which is the set of probability distributions 21, ,..., Ni ip . All extreme probability distributions
in 21, ,..., Ni iΨ should satisfy constraints (4.6), and are calculated by the algorithm in
Chapter 2 Section 2.4, and collected in set
103
1, ,...,22 1, ,..., 1, ,..., 1, ,...,2 2 2
1 21, ,..., , ,..., i iN
N i i i i i iN N Ni i EXT EXT EXTEXT ξ
= p p p . Then 21, ,..., Ni iE and 21, ,..., Ni iE may be
obtained by searching in 21, ,..., Ni iEXT :
2 2 1, ,...,21, ,..., 1, ,...,maxN N i iN
Ti i i i EXTE ξ= pa
2 2 1, ,...,21, ,..., 1, ,...,min
N N i iN
Ti i i i EXTE ξ= pa
Subject to
1, ,..., 22 1, ,...,i i NNEXT i iEXTξ ∈p
(4.7)
Once the upper and lower bounds 21, ,..., Ni iE and
21, ,..., Ni iE (2 11, ,..., ,01,...,
NN i ii n−
= ) are
all obtained, then the upper and lower consequence vectors 2 11, ,..., ,0Ni i −a and 2 11, ,..., ,0Ni i −
a are
written as
2 1
2 1
2 1 1, ,..., ,02 1
1, ,..., ,1
1, ,..., ,0
1, ,..., ,
N
N
N i iN
i i
i i
i i n
E
E
−
−
− −
⎛ ⎞⎜ ⎟
= ⎜ ⎟⎜ ⎟⎝ ⎠
a ; 2 1
2 1
2 1 1, ,..., ,02 1
1, ,..., ,1
1, ,..., ,0
1, ,..., ,
N
N
N i iN
i i
i i
i i n
E
E
−
−
− −
⎛ ⎞⎜ ⎟
= ⎜ ⎟⎜ ⎟⎝ ⎠
a (4.8)
Now, we continue the same procedure at level N-1 with conditional probability vector
2 11, ,..., ,0Ni i −p , and consequence vectors 2 11, ,..., ,0Ni i −a and
2 11, ,..., ,0Ni i −a in Eq.(4.8). The
upper and lower previsions 2 11, ,..., ,0Ni iE − and 2 11, ,..., ,0Ni iE
− of event
2 11, ,..., ,0Ni iS−
can be
obtained as
2 1 2 1 2 11, ,..., ,0 1, ,..., ,0 1, ,..., ,0maxN N N
Ti i i i i iE − − −
= pa
2 1 2 1 2 11, ,..., ,0 1, ,..., ,0 1, ,..., ,0minN N N
Ti i i i i iE
− − −= pa
Subject to
( )2 1 2 1 2 1 2 1
2 1
1, ,..., ,0 1, ,..., ,0 1, ,..., ,0 1, ,..., ,0
1, ,..., ,0
;
1,...,N N N N
N
Tm m mLOW i i i i i i UPP i i
i i
E f E f
m m− − − −
−
⎡ ⎤ ⎡ ⎤≤ ≤⎣ ⎦ ⎣ ⎦=
f p
( ) 2 11, ,..., ,02 1
2 1 2 1
1, ,..., ,0
1, ,..., , 1, ,..., ,0
1
0; 1,...,
Ni iN
N N
Ti in
i i j i ip j n
−−
− −
⋅ =
≥ =
1 p
(4.9)
After this second step, the upper and lower consequence vectors 2 21, ,..., ,0,0Ni i −a and
2 21, ,..., ,0,0Ni i −a are equal to:
104
2 2
2 2
2 2 1, ,..., ,02 2
1, ,..., ,1,0
1, ,..., ,0,0
1, ,..., , ,0
N
N
N i iN
i i
i i
i i n
E
E
−
−
− −
⎛ ⎞⎜ ⎟
= ⎜ ⎟⎜ ⎟⎝ ⎠
a ; 2 2
2 2
2 1 1, ,..., ,0,02 2
1, ,..., ,1,0
1, ,..., ,0,0
1, ,..., , ,0
N
N
N i iN
i i
i i
i i n
E
E
−
−
− −
⎛ ⎞⎜ ⎟
= ⎜ ⎟⎜ ⎟⎝ ⎠
a (4.10)
Then, one would repeat the analysis at level N-2 with conditional probability vector
2 21, ,..., ,0,0Ni i −p , and consequence vectors 2 21, ,..., ,0,0Ni i −a and
2 21, ,..., ,0,0Ni i −a . After N-1
iterations, one obtains the upper and lower previsions 1,0,...,0E and 1,0,...,0E at level 1,
which are the upper and lower previsions E and E of the event tree:
1,0,...,0 1,0,...,0 1,0,...,0maxT
E = pa
1,0,...,0 1,0,...,0 1,0,...,0min TE = pa Subject to
1,0,...,0 1,0,...,0∈Ψp
(4.11)
Let us now illustrate the aforementioned algorithm for Event Tree Analysis with
imprecise probabilities with an example.
Example 4-1 Consider a leaking water-conveyance tunnel. The analysis is meant to evaluate the failure
probability with the initiating event S1,0: construction void behind the unreinforced concrete lining. Here
‘failure’ is defined as either structural collapse failure or service failure of the lining (i.e., wide cracks).
Three states are considered for the construction void: (1) large void (s1,1,0: Diameter ∅ > 3 ft); (2) small
void (s1,2,0: Diameter ∅≤ 1 ft); and (3) intermediate size void (s1,3,0: Diameter 1 ft<∅≤ 3 ft). Information
on the initiating event S1,0 is as follows: (1) 10% of voids are either large (s1,1,0) or intermediate (s1,3,0); (2)
80% of voids are small (s1,2,0); (3) the remaining 10% of voids is indeterminate. This information can be
condensed in a random set: ( s1,1,0, s1,3,0, 0.1), (s1,1,0, s1,2,0, s1,3,0, 0.1), (s1,2,0, 0.8), whose set Ψ1,1 is
depicted in Figure 4-6(a) and has four vertices 1,0
1EXTp = (0, 0.9, 0.1)T,
1,0
2EXTp = (0, 1, 0)T,
1,0
3EXTp = (0.1,
0.8, 0.1)T, and 1,0
4EXTp = (0.2, 0.8, 0)T (Bernardini and Tonon, 2010).
105
If the void is large, it is possible that structural collapse (Event S1,1) occurs because of non-
uniform load distribution on the lining. Thus, there are only two outcomes, Yes ( 1,1,1s ) or No ( 1,1,2s ) with
conditional probability evaluated as 0.8≤P( 1,1,1s |s1,1,0) 0.9≤ . Set Ψ1,1 has two extreme points: 1,1
1EXTp =
(0.8, 0.2)T and 1,1
2EXTp = (0.9, 0.1)T.
If the void is small, although the lining will not collapse, the void may cause some cracks (Event
S1,2), which may reduce the serviceability. Given the existence of small voids, there are three outcomes for
crack development (Event S1,2): (1) 1,2,1s : wide cracks; (2) 1,2,2s : small cracks; (3) 1,2,3s : no cracks. Here
we assume that only wide cracks will cause leakage through the lining. The following information is
available: in the case of small void, the probability of wide cracks ( 1,2,1s ) is less than twice the probability
of small cracks ( 1,2,2s ); the probability of no cracks ( 1,2,3s ) is less than twice the probability of wide cracks
( 1,2,1s ); having a small crack ( 1,2,2s ) is less probable than having no cracks ( 1,2,3s ). This information may
be expressed in terms of the following inequalities: P( 1,2,1s |s1,2,0) ≤ 2P( 1,2,2s |s1,2,0); P( 1,2,3s |s1,2,0)
≤2P( 1,2,1s |s1,2,0); P( 1,2,2s |s1,2,0) ≤ P( 1,2,3s |s1,2,0), whose set Ψ1,2 is depicted in Figure 4-6(b), and has three
extreme points: 1,2
1EXTp = (0.29, 0.14, 0.57)T,
1,2
2EXTp = (0.5, 0.25, 0.25)T, and
1,2
3EXTp = (0.2, 0.4, 0.4)T.
If the size of the void is intermediate, structural collapse (Event S1,3) is still possible, i.e., two
outcomes - Yes ( 1,3,1s ) or No ( 1,3,2s ) are possible with conditional probabilities assigned as:
0.5≤P( 1,3,1s |s1,3,0) 0.6≤ , and Ψ1,3 has two vertices: 1,3
1EXTp = (0.5, 0.5)T and
1,3
2EXTp = (0.6, 0.4)T.
The event tree is depicted in Figure 4-5. Since we are interested in the probability of failure, we
set a1,1,1 = a1,2,1 = a1,3,1 = 1, and all other 1 2 3, ,i i ia are equal to 0.
106
Figure 4-5: Event-tree in Example 4-1.
(a) (b)
Figure 4-6: Example 4-1: sets Ψ1,0 and Ψ1,2 in the 3-dimensional spaces of the probability of the singletons.
To determine the extreme values of the expectation, we first consider the sub-trees for S1,1, S1,2,
and S1,3, respectively. In set Ψ1,1, function 1,1,1 1,1,1 1,1,2 1,1,2a p a p+ achieves its maximum value 0.9 at 1,1
2EXTp
and minimum value 0.8 at 1,1
1EXTp . For the sub-tree of S1,2, the extreme values of function
1,2,1 1,2,1 1,2,2 1,2,2 1,2,3 1,2,3a p a p a p+ + are 0.5 and 0.2, achieved at 1,2
2EXTp and
1,2
3EXTp , respectively.
107
Likewise for S1,3, the maximum value for 1,3,1 1,3,1 1,3,2 1,3,2a p a p+ is 0.6 at 1,3
2EXTp and minimum value 0.5
at 1,3
1EXTp .
Next, we conduct the analysis for S1,0. The maximized objective function
1,1,0 1,2,0 1,3,00.9 0.5 0.6p p p+ + achieves its maximal value (i.e., 0.58) at 1,0
4EXTp , and the minimized
objective function is 1,1,0 1,2,0 1,3,00.8 0.2 0.5p p p+ + , whose minimum value 0.2 is achieved at 1,0
2EXTp .
Therefore, the upper and lower probabilities of leakage are equal to 0.58 and 0.2, respectively,
which conveys the initial imprecision on the input values.
4.2.1.2 ETA with total probabilities
In the last sub-section, we discussed the case when the probabilities are given
conditional to the occurrences of upper level events. That is to say, we are confident
about how the occurrence of upper level events affects the probabilities of the lower level
events. However, sometimes this information is not available. If the information is only
given in terms of total probabilities, the analysts have to assume some interaction
between the upper and lower level events, including unknown interaction, epistemic
irrelevance, epistemic independence, strong independence, and uncertain correlation.
To set up the event-tree, one needs to consider that the interaction between events
are partially defined; therefore, the outcomes at the bottom of the event tree should
include all combinations of events. Let 1S , 2S , …, mS be m events, where event iS
has in possible states. 1S is the initiating event, and mS is the event at the bottom of
the tree. Then there are a total of 1 2 mn n n× × outcomes at the bottom of the event tree,
collected in matrix a , whose ( 1 2, , , mi i i )-th entry is 1 2, , , mi i ia . The probability
distribution of all outcomes (i.e., the joint distribution of 1E through mE ), P , is also a
108
1 2 mn n n× × matrix whose ( 1 2, , , mi i i )-th entry is the probability of the
( )1 2, , , -thmi i i outcome, i.e. 1 2, , , mi i iP .
Under the assumed interaction between events 1S and 2S , the constraints on the
probabilities of events are as follows (Section 3.3):
(1) Unknown interaction: ΨU:
( )2 1P S⋅× ∈Ψ ; ( )1 2P S ×⋅ ∈Ψ (4.12)
(2) Epistemic irrelevance: | isEΨ :
( )2 1P S⋅× ∈Ψ ; ( )1 2P S ×⋅ ∈Ψ
( ) ( ) ( ) ( ) ( ) ( )
2 2
1 1
| |1 1 1 2 1 1 2 2
| |2 2 1 2 2 2 1 1
: , 2
: , 1
s s
s s
P P U s P U P s i OR
P P s U P U P s i
∃ ∈Ψ × = =
∃ ∈Ψ × = = (4.13)
(3) Epistemic independence: ΨE:
( )2 1P S⋅× ∈Ψ ; ( )1 2P S ×⋅ ∈Ψ
( ) ( ) ( ) ( ) ( ) ( )
2 2
1 1
| |1 1 1 2 1 1 2 2
| |2 2 1 2 2 2 1 1
:
:
s s
s s
P P U s P U P s AND
P P s U P U P s
∃ ∈Ψ × =
∃ ∈Ψ × = (4.14)
(4) Strong independence: ΨS:
( )2 1P S⋅× ∈Ψ ; ( )1 2P S ×⋅ ∈Ψ
2|2 2 1 1: ss S P P∀ ∈ = and 1|
1 1 2 2: ss S P P∀ ∈ = (4.15)
(5) Uncertain correlation: ΨC:
( )2 1P S⋅× ∈Ψ ; ( )1 2P S ×⋅ ∈Ψ
( ) ( ) ( ) ( ) ( )1 2 1 21 2 1 2 1 2S S S SE S E S D D E S S E S E S D Dρ ρ+ ≤ ≤ + (4.16)
Various algorithms for finding the extreme points of the set of joint distributions
under different interactions were proposed and exemplified in Chapter 4.
It should be noted that there are two ways to interpret the interactions between
events. The first one is to interpret them as pair-wise interactions between Events iS
109
and jS ( i j≠ ). The second way is to interpret them as interactions between Event iS
and the new joint event composed of Events 1iS − through 1S with 1 2 1in n n −× ×
possible states. For example, in the event tree shown in Figure 4-5, one may define
unknown interaction between Events 2S and 1S , and strong independence between
Events 3S and 2S . One may also define the interactions using the second way:
unknown interaction between Events 2S and 1S , and strong independence between
Event 3S and the joint event composed of 2S and 1S , which has 2 1n n× possible
states. These two different ways of defining interactions results in different problem
formulations and final results.
If the interactions between events iS and jS are defined pair-wise, the
optimization problems are rewritten in terms of P as follows (see the Appendix A for
explicit formulations):
Minimize (Maximize) 1
1 2 1 21
, , , , , ,1 1
m
m mm
nn
i i i i i ii i
a P= =∑ ∑
Subject to (1) Unknown interaction between iS and jS ,: ^i j U∈ΨP ; or
(2) iS is epistemically irrelevant to jS : |^
isi j E∈ΨP ; or
(3) Epistemic independence between iS and jS : ^i j E∈ΨP ; or
(4) Strong independence between iS and jS : ^i j S∈ΨP ; or
(5) Uncertain correlation between iS and jS : ^i j C∈ΨP .
(4.17)
where UΨ , | isEΨ , EΨ , SΨ , and CΨ are defined in Eqs.(4.12)-(4.16) and detailed in
Chapter 3 (Sections 3.3.2, 3.3.3, and 3.3.4, respectively), and ^i jP is the joint
probability matrix for Events iS and jS , which may be expressed in terms of P as
follows:
110
1 11 11
1 21 1 1 1 1
^ , , ,1 1 1 1 1 1
j ji i m
mi i j j m
n nn n nn
i j Pξ ξ ξξ ξ ξ ξ ξ ξ
− +− +
− + − += = = = = == ∑ ∑ ∑ ∑ ∑ ∑P (4.18)
In this case, we solve the problem directly in terms of the 1 2 mn n n× × matrix P , and
the computational difficulty will dramatically increase with the dimensions of the matrix.
Alternatively, the interactions could be defined between Event iS and the joint
event composed of Events 1iS − through 1S . The solution for this problem proceeds in a
top-to-bottom pattern. Here is the general algorithm:
1. Start the calculation from the initiating event 1S . Only consider the first two
Events, 1S and 2S . Set 2i = .
2. According to the assumed interaction between Events 1iS − and iS , set up the
corresponding constraints for the probability distribution combP , which is a
( )1 2 1i in n n n−× × × matrix, i.e. the joint distribution of 1iS − and iS , as
detailed in Chapter 4.
3. Find all extreme distributions of combP that satisfy the constraints set up in the
last step.
4. Consider Events 1iS − and iS as a joint event combS with ( )1 2 1i in n n n−× × ×
possible states subject to the probability distribution combP .
5. Replace 1iS − with combS and replace iS with 1iS + . Then set 1i i= + .
6. Repeat Step 2 through Step 5 till 1i m= + . All extreme distributions of combP
are obtained, and P = combP .
7. Find the upper and lower bounds of the prevision of the event tree analysis by
enumerating the extreme distributions of P .
Example 4-2 illustrates the algorithm for pair-wise interactions in Event Tree
Analysis. In Example 4-3, the interactions are defined the second way, i.e. interaction
111
with the joint event at the upper level. Comparison of the results in Example 4-2 and
Example 4-3 highlights the differences due to the different ways of defining the
interaction.
Example 4-2 Consider three fire alarm systems in a tunnel: Systems I, II, and III. If none of them work,
the tunnel is in a dangerous situation. Let S1, S2, and S3 denote failure of Systems I, II, and III, respectively.
The event tree for the fire alarm system is depicted in Figure 4-7. Considering the imprecision of the alarm
sensors, the evacuation system will be activated only when at least two fire systems are activated. The
objective of the analysis is to evaluate the probability of failure, which is defined as “not activating the
evacuation system when a fire is really happening in the tunnel”.
Figure 4-7: Event Tree when total probabilities are assigned to all events.
Assume that we have no information about the correlation or interaction between the failures of
Systems I and II. Thus it is safe to model the interaction between failures of the two systems as unknown
Event 1 Event 2 Event 3
1P
2P
1,1P
1,2P
2,1P
2,2P
1,1,1 1,1,1 P a
1,1,2 1,1,2 P a
1,2,1 1,2,1 P a
1,2,2 1,2,2 P a
2,1,1 2,1,1 P a
2,1,2 2,1,2 P a
2,2,1 2,2,1 P a
2,2,2 2,2,2 P a
112
interaction. However, we are confident that System III works stochastically independently to System II,
thus failures of Systems III and II are strongly independent. As for the failure probability of each system,
the available information is 0<P(S1)<0.1, 0.05<P(S2)<0.1, and 0<P(S3)<0.15. Thus, the extreme vertices for
Ψ1, Ψ2, and Ψ3 are 1
1EXTp = (0, 1)T,
1
2EXTp = (0.1, 0.9)T;
2
1EXTp = (0.05, 0.95)T,
2
2EXTp = (0.1, 0.9)T;
3
1EXTp =
(0, 1)T, 3
2EXTp = (0.15, 0.85)T.
According to the objective of the analysis, let 1,1,1a = 1,1,2a = 1,2,1a = 2,1,1a =1, and all the other a’s
are equal to 0. The complete optimization problem in Eq.(4.17) reads
Minimize (Maximize) 1 2 3 1 2 3
1 2 3
2 2 2
, , , ,1 1 1
i i i i i ii i i
a P= = =∑∑∑
Subject to
2 32 3
2 2
1, ,1 1
0 0.1i ii i
P= =
< <∑∑ ;
1 31 3
2 2
,1,1 1
0.05 0.1i ii i
P= =
< <∑∑ ;
1 21 2
2 2
, ,11 1
0 0.15i ii i
P= =
< <∑∑
( )( )
2 3 2 3 2 2 2 2
3 3 3 3
1, , 2, , 1, ,1 2, ,1 1, ,2 1, ,2
1,1, 2,1, 1,2, 2,2,2 3
2 3
Strong independencebetween and
1, 2; 1,2;
i i i i i i i i
i i i i
P P P P P P
P P P PS S
i i
⎫+ = + + +⎪⎪⋅ + + + ⎬⎪
= = ⎪⎭
1
1 21
1 2
, , ,1 1
, , ,
1
0
m
mm
m
nn
i i ii i
i i i
P
P= =
=
≥
∑ ∑
(4.19)
Under the assumptions of unknown interaction between E1 and E2 and strong independence between E2 and
E3, the upper and lower failure probabilities are equal to 0.115 and 0, respectively, and solutions are
detailed in Table 4-1. One can check that the joint distribution of E2 and E3 satisfies the assumption of
strong independence, i.e., 2^3P = 2 3Tp p .
113
Table 4-1: Solutions for the optimization problems (19) for the upper and lower probabilities of failure.
1,1,1 1,1,2 1,2,1 1,2,2
2,1,1 2,1,2 2,2,1 2,2,2
; ; ; ;
; ; ;
P P P P
P P P P⎛ ⎞⎜ ⎟⎜ ⎟⎝ ⎠
P = 1p 2p 3p 2^3P 1 2 3 1 2 3
1 2 3
2 2 2
, , , ,1 1 1
i i i i i ii i i
a P= = =∑∑∑
max 0; 0.015; 0.085; 0;
0.015; 0.12; 0; 0.765⎛ ⎞⎜ ⎟⎝ ⎠
0.10.9⎛ ⎞⎜ ⎟⎝ ⎠
0.10.9⎛ ⎞⎜ ⎟⎝ ⎠
0.150.85⎛ ⎞⎜ ⎟⎝ ⎠
0.015 0.0850.135 0.765⎛ ⎞⎜ ⎟⎝ ⎠
0.115
min 0; 0; 0; 0.1;
0; 0; 0.05; 0.85⎛ ⎞⎜ ⎟⎝ ⎠
0.10.9⎛ ⎞⎜ ⎟⎝ ⎠
0.050.95⎛ ⎞⎜ ⎟⎝ ⎠
01⎛ ⎞⎜ ⎟⎝ ⎠
0 0.050 0.95⎛ ⎞⎜ ⎟⎝ ⎠
0
Example 4-3 Consider again the situation and knowledge available in Example 2. Suppose that now the
interactions are defined in the second manner, i.e. interaction between an event and the joint event at its
upper level. We still assume unknown interaction between S1 and S2, but we assume strong independence
between S3 and the joint event Scomb composed of S1 and S2.
We start by finding all the extreme points in the set of joint distributions for S1 and S2, subject to
the linear constraints on 1p and 2p , as shown in Eq. (4.20) below. All 7 extreme distributions of combP
are listed in Table 4-2.
1,1 1,20 0.1P P< + < ; 1,1 2,1
2 2
,1 1
,
0.05 0.1
1
0
i ji j
i j
P P
P
P= =
< + <
=
≥
∑∑ ;
(4.20)
Theorem 3-5 regarding the extreme joint distributions in Chapter 4 states that: “Under strong independence,
the set of extreme joint distributions (measures) is the set of product distributions (measures), each taken
from the extreme distributions (measures) of the marginals”. Accordingly, the extreme points for the joint
distributions of Scomb and S3 are found by taking the product of one of the 7 extreme joint distributions in
Table 4-2 and one of the two extreme distributions of S3 at a time. A total of 14 different extreme joint
distributions are obtained for Scomb and S3 as listed in Table 4-3. By enumerating all the extreme points in
Table 4-3, the maximum and minimum of the objective functions are found to be 0.1 and 0, respectively,
114
which are different from the results obtained in Example 4-2 (i.e., 0.115 and 0) because of the two different
interpretations on the interaction between events.
Table 4-2: Extreme points for the set of joint distributions of E1 and E2.
Extreme points of Ψcomb
P1,1 P1,2 P2,1 P2,2
1 0.10 0.00 0.00 0.90
2 0.00 0.00 0.05 0.95
3 0.05 0.00 0.00 0.95
4 0.00 0.00 0.10 0.90
5 0.00 0.1 0.05 0.85
6 0.05 0.05 0.00 0.90
7 0.00 0.10 0.10 0.80
Table 4-3: Extreme points for the set Ψ of joint distributions for Scomb and S3. combP
Extreme points of Ψ
P1,1,1 P1,1,2 P1,2,1 P1,2,2 P2,1,1 P2,1,2 P2,2,1 P2,2,2 1 2 3 1 2 3
1 2 3
2 2 2
, , , ,1 1 1
i i i i i ii i i
a P= = =∑∑∑
1 0.000 0.000 0.000 0.000 0.100 0.000 0.000 0.900 0.100 2 0.000 0.000 0.000 0.000 0.000 0.000 0.050 0.950 0.000 3 0.000 0.000 0.000 0.000 0.050 0.000 0.000 0.950 0.050 4 0.000 0.000 0.000 0.000 0.000 0.000 0.100 0.900 0.000 5 0.000 0.000 0.000 0.000 0.000 0.100 0.050 0.850 0.000 6 0.000 0.000 0.000 0.000 0.050 0.050 0.000 0.900 0.050 7 0.000 0.000 0.000 0.000 0.000 0.100 0.100 0.800 0.000 8 0.015 0.000 0.000 0.135 0.085 0.000 0.000 0.765 0.100 9 0.000 0.000 0.008 0.143 0.000 0.000 0.043 0.808 0.008 10 0.008 0.000 0.000 0.143 0.043 0.000 0.000 0.808 0.050 11 0.000 0.000 0.015 0.135 0.000 0.000 0.085 0.765 0.015 12 0.000 0.015 0.008 0.128 0.000 0.085 0.043 0.723 0.023 13 0.008 0.008 0.000 0.135 0.043 0.043 0.000 0.765 0.058 14 0.000 0.015 0.015 0.120 0.000 0.085 0.085 0.680 0.030
115
4.2.1.3 ETA with Combination of conditional probabilities and total probabilities
Suppose that the information on some events in the tree is available in terms of
conditional probabilities, and that the information on the remaining events are given in
terms of total probability with partially known interaction to the upper events, i.e.
unknown interaction, epistemic irrelevance, epistemic independence, strong
independence, or uncertain correlation. The algorithm for this case is as follows:
1. Start the calculation from Events on which total probabilities are assigned, for
example iS and jS .
2. Set up the constraints based on the defined interaction, find all extreme
distributions for ,comb ijP , which is the joint distribution of iS and jS .
3. Replace Events iS and jS in the event tree with the joint event ,comb ijE ,
whose probability distribution is defined by ,comb ijP .
4. Proceed the analysis with the equivalent tree where information is now only given
in terms of probabilities conditional to the occurrences of upper level events.
As an example, consider the event tree shown in Figure 4-8. The information
available assigns probabilities of 5S conditional to 11s and the probabilities of 2E
through 4E conditional to the occurrence of their upper level events, respectively. As
for the interaction between 2E and 5E , it is assumed that they are strongly
independent.
Since 2S and 5S are strongly independent, we first combine 2E and 5E as a
new event combE , which has 4 outcomes. Then we find all extreme distributions of
combP conditional to 1E , as shown in Figure 4-9.
116
Figure 4-8: Event Tree with mixed information consisting of conditional probabilities and total probabilities.
1S
2S
3S
4S
1 11 1 s P
2 21 1 s P
3 31 1 s P
13P
14P
24P
1a
2a
3a
4a
5a
6a
7a
8a
9a
5S
5S
23P
33P
117
Figure 4-9: Event Tree equivalent to the tree in Figure 4-8 that contains only probabilities conditional to the upper level events.
4.2.2 Fault Tree Analysis
In Fault Tree Analysis, the failure event (major fault) is logically connected with
the sub-events (alternative faults) by gates. Here we only consider two basic gates: OR-
and AND-gates (See Figure 4-10 and Figure 4-11). Any event in the fault tree has only
two possible states: occurrence or not occurrence. The occurrence probabilities for the
failure events are assumed to be given imprecisely, i.e. as interval probabilities [PLOW,
PUPP].
1E
combE
3E
4E
1 11 1 s P
2 21 1 s P
3 31 1 s P
13P
14P
24P
1a
2a
3a
4a
5a
6a
7a
8a
9a
23P
33P
1combP
2combP
3combP
4combP
118
In the probability evaluation for Figure 4-10 and Figure 4-11, let P be the joint
probability distribution for sub-events 1E through nE , and thus P is a matrix of
dimensions 2 2 2n
× × × , where the i-th index indicates the states of Event iE ; let the
subscript “1” denote the probability of occurrence, and “2” denote the probability of non-
occurrence. For the OR-gate in Figure 4-10, the occurrence probability for the failure
event E is ( ) 2,2, ,21
n
P E P= − (4.21)
where 2,2, ,2P is the probability of that none of the n events occurs, and thus 2,2, ,21 P−
is the probability that the complementary event (any of the n events ) occurs.
For the AND-gate in Figure 4-11, the occurrence probability for the event E is ( ) 1,1, ,1
n
P E P= (4.22)
where 1,1, ,1P is the probability that all the n events occur.
Figure 4-10: Sub-tree with OR-gate.
119
Figure 4-11: Sub-tree with AND-gate.
Let ^i jP be the joint probability distribution for sub-events iE and jE , which
can be written in terms of the joint distribution P for 1E through nE :
1 21 1 1 1 1
2 2 2 2 2 2
^ , , ,1 1 1 1 1 1
n
i i j j n
i j Pξ ξ ξξ ξ ξ ξ ξ ξ− + − += = = = = =
= ∑ ∑ ∑ ∑ ∑ ∑P (4.23)
According to the assumed interaction between iE and jE , the joint distribution ^i jP
should fall in one of the following cases:
1. Unknown interaction: ^i j U∈ΨP ; or
2. Epistemic irrelevance: |^
isi j E∈ΨP ; or
3. Epistemic independence: ^i j E∈ΨP ; or
4. Strong independence: ^i j S∈ΨP ; or
5. Uncertain correlation: ^i j C∈ΨP .
Take three sub-events E1, E2, and E3 as an example. Assume epistemic
independence between E1 and E2 and strong independence between E2 and E3. Lower and
upper 1,1,1P and 2,2,2P are obtained by solving the optimization problems below.
120
Minimize (Maximize) 1,1,1P ( 2,2,2P ) Subject to
1^2 E∈ΨP ; 2^3 S∈ΨP 2 2 2
, ,1 1 1
; 1i j li j l
P= = =
≥ =∑∑∑P 0
(4.24)
Then lower and upper probabilities for event E are obtained by inserting lower and upper
1,1,1P and 2,2,2P into Eqs.(4.21) and (4.22), respectively.
The objective of Fault Tree Analysis is to determine the upper and lower bounds
of the failure probability of the top event. The Fault Tree Analysis with imprecise
probabilities proceeds in a bottom-to-top pattern:
1. Start the analysis from the bottom of the fault tree, where the basic failure events
cannot be decomposed further.
2. Set up the constraints according to the defined interaction between sub-events
(Section 3.4). Express the occurrence probability of the upper level event ( )P E
by Eq. (4.21) or (4.22).
3. Calculate 1,...,1P or 2,...,2P as described in Eq. (4.24).
4. Determine the upper and lower bounds of the upper level event ( )P E .
5. Repeat Step 2 through Step 4 with all events in the same sub-tree till the top event
is included.
It is important to notice that epistemic irrelevance, epistemic independence, and
strong independence all lead to the same results in the fault tree analysis. Here we will
explain it for the case in Figure 4-10. In Eq. (4.21):
( )max P E = ( )2,2, ,2max 1 P− = ( )2,2, ,21 min P− (4.25)
and thus
( )min P E = ( )2,2, ,21 max P− (4.26)
121
By observing Eqs. (4.25) and (4.26), we find that the optimization problems of
maximizing ( and minimizing) ( )P E are equivalent to the problems of minimizing (and
maximizing) 2,2, ,2P . It should be noted that 2,2, ,2P is an element in matrix P . Under
either epistemic irrelevance or strong independence, any element of P could be written
in terms of the product of marginals. For example, if n=2 and Event E1 is epistemically
irrelevant to Event E2, the (2,2)-th component in the joint distribution P , 2,2P , is equal
to 2 21 2p p⋅ . Therefore, under epistemic irrelevance or strong independence, the extreme
values of 2,2, ,2P coincide and are obtained as
( )22,2, ,2
1max max
n
n
ii
P p=
⎛ ⎞=⎜ ⎟⎜ ⎟
⎝ ⎠∏
( )22,2, ,2
1min min
n
n
ii
P p=
⎛ ⎞=⎜ ⎟⎜ ⎟
⎝ ⎠∏
(4.27)
where 2ip is the 2nd component of the probability vector iP for event Ei.
Since | iSE E SΨ ⊇Ψ ⊇Ψ (Section 3.4),
| |, , , , , ,
i is sE low E low S low S upp E upp E uppP P P P P P≤ ≤ ≤ ≤ ≤ (Section 3.4) (4.28)
Since Eq. (4.27) is true under epistemic irrelevance and strong independence, | |
, , , , and i is sE low S low E upp S uppP P P P= = (4.29)
From Eqs. (4.28) and (4.29), one obtains: |
, , ,is
E low E low S lowP P P= = (4.30) |
, , ,is
E upp E upp S uppP P P= = (4.31)
Likewise for the case of Figure 4-11 and Eq.(4.22).
Example 4-4 Let us consider the example of the fault tree adapted from Eskesen (2004) (Fig. 2), where
the failure of a toll sub-sea tunnel project (Event E) is caused by two sub-events: technical failure (Event
122
E1) or economical failure (Event E2), which are here assumed to be strongly independent. Technical failure
may happen due to the occurrence of at least one of two epistemically independent events:
(1) total collapse: seawater fills the tunnel (Event E1,1), as a result of the occurrence of both too small rock
cover (Event E1,1,1) and insufficient investigations (Event E1,1,2). E1,1,1 and E1,1,2 are here assumed to be
linked by uncertain correlation with coefficient [ ]0.6,0.8ρ ∈ ;
(2) the tunnel cannot be built (Event E1,2) because of difficult rock conditions (Event E1,2,1) and poor
investigation (Event E1,2,2) occurring at the same time. Events E1,2,1 and E1,2,2 are assumed to be linked by
unknown interaction.
Figure 4-12: Fault tree analysis for the failure probability of sub-sea tunnel project with imprecise probabilities.
The economical failure is triggered by two strongly independent events: either too small toll
revenue (Event E2,1) or too high construction and maintenance costs (Event E2,2). The structure of the fault
123
tree is the same as Figure 4-2, but the probabilities for the events at the bottom of the fault tree are assigned
as imprecise probabilities, as shown in Figure 4-12. Our objective is to determine the upper and lower
probabilities for the failure of the sub-sea tunnel project, given the assumed interactions and imprecise
probabilities in the fault tree.
We first consider the sub-tree for Event E1,1 together with sub-events E1,1,1 and E1,1,2, and
determine the bounds on the occurrence probability of Event E1,1 by solving the optimization problems
(4.32) written in terms of the joint distribution P of E1,1,1 and E1,1,2 (Section 3.2.3): Minimize (Maximize) 1,1P Subject to
1,1 1,2
1,1 2,1
0.01 0.05
0 0.01
P P
P P
< + <
< + <
( ) ( ) ( )( ) ( ) ( )
2 2
,1 1
,
1,1,1 1,1,2 1,1,1 1,1,2 1,1,1 1,1,2
1,1,1 1,1,2 1,1,1 1,1,2 1,1,1 1,1,2
1
0
0.8
0.6
i ji j
i j
P
P
E E E E E E E D D
E E E E D D E E E
= ==
≥
≤ +
+ ≤
∑∑
(4.32)
where (Section 4.2.3):
( ) ( ) 1,1 1,21,1,1 1,1 1,2
2,1 2,21 0
P PE E P P
P P+⎛ ⎞
= = +⎜ ⎟+⎝ ⎠;
( ) ( ) 1,1 2,11,1,2 1,1 2,1
1,2 2,21 0
P PE E P P
P P+⎛ ⎞
= = +⎜ ⎟+⎝ ⎠
( )1,1,1 1,1,2 1,11 00 0
E E E P⎛ ⎞
= =⎜ ⎟⎝ ⎠
P , where , ,i j i ji j
P a= ∑∑P a ;
( )( )1,1,1 1,1 1,2 2,1 2,2D P P P P= + + ;
( )( )1,1,2 1,1 2,1 1,2 2,2D P P P P= + + .
The extreme values of the occurrence probability of E1,1 are found to be 0.01 and 0.0036. Detailed
solutions are listed in Table 4-4.
124
Table 4-4: Solutions for the optimization problems (4.32) for the upper and lower probabilities of Event E1,1.
P: joint dist. of E1,1,1 and E1,1,2 P(E1,1,1) P(E1,1,2) P(E1,1) = P1,1
max 0.01 0.0055
0 0.9845⎛ ⎞⎜ ⎟⎝ ⎠
0.0155 0.01 0.01
min 0.0036 0.0064
0 0.99⎛ ⎞⎜ ⎟⎝ ⎠
0.01 0.0036 0.0036
Next, we calculate the bounds on the probability of Event E1,2, and the optimization problems read
as follows (Section 3.2.1) Minimize (Maximize) 1,1P Subject to
1,1 1,2
1,1 2,12 2
,1 1
,
0 0.02;
0 0.1;
1;
0
i ji j
i j
P P
P P
P
P= =
< + <
< + <
=
≥
∑∑
(4.33)
where matrix P is now the joint distribution of E1,2,1 and E1,2,2. Solutions detailed in Table 4-5 show that the
upper and lower occurrence probabilities for E1,2 are equal to 0.02 and 0, respectively.
Table 4-5: Solutions for the optimization problems (4.33) for the upper and lower probabilities of Event E1,2.
P: joint dist. of E1,2,1 and E1,2,2 P(E1,2,1) P(E1,2,2) P(E1,1) = P1,1
max 0.02 0
0 0.98⎛ ⎞⎜ ⎟⎝ ⎠
0.02 0.02 0.02
min 0 00 1⎛ ⎞⎜ ⎟⎝ ⎠
0 0 0
The next step is to determine the extreme values of the occurrence probability of Event E1.
Replace the probabilities of E1,1 and E1,2 with intervals [0.0036, 0.01] and [0, 0.02], respectively. Since E1
is connected with E1,1 and E1,2 by an OR-gate in Figure 4-12, the optimization problems are
125
Minimize (Maximize) 2,21 P− Subject to
1,1 1,2
1,1 2,1
0.0036 0.01;
0 0.02
P P
P P
< + <
< + <
2 2
,1 1
,
;
1;
0
E
i ji j
i j
P
P= =
∈Ψ
=
≥
∑∑
P
(4.34)
where matrix P is the joint distribution of E1,1 and E1,2.
The upper and lower probabilities for E1 in (4.34) are 0.0298 and 0.0036, respectively. By
observing the solutions shown in Table 4-6, it is easy to check that both maximum and minimum solutions
belong to set ΨS, the set of joint distribution under strong independence. For example, the maximal solution 0.0002 0.00980.0198 0.9702⎛ ⎞
= ⎜ ⎟⎝ ⎠
P can be written as ( )0.0002 0.0098 0.010.02 0.98
0.0198 0.9702 0.99⎛ ⎞ ⎛ ⎞
= =⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
P , i.e. product of two extreme
distributions of E1,1 and E1,2, respectively. Thus, in the fault tree analysis there is no difference between
epistemic independence and strong independence. This observation is consistent with the previous
theoretical derivation.
Table 4-6: Solutions for the optimization problems (4.34) for the upper and lower probabilities of Event E1.
P: joint dist. of E1,1 and E1,2 P(E1,1) P(E1,2) P(E1)
max 0.0002 0.00980.0198 0.9702⎛ ⎞⎜ ⎟⎝ ⎠
0.01 0.02 0.0298
min 0 0.00360 0.9964⎛ ⎞⎜ ⎟⎝ ⎠
0.0036 0 0.0036
As for the sub-tree of Event E2, we determine the bounds on the occurrence probability by first
multiplying all the extreme distributions of E2,1 by all extreme probability distributions of E2,2 (Section
126
3.3.3) and then by calculating P(E2) = 1- P2,2 on all the extreme joint distributions as shown in Table 4-7.
The upper and lower occurrence probabilities for E2 are 0.012 and 0, respectively.
Table 4-7: Extreme joint distributions of E2,1 and E2,2.
Extreme points of Ψcomb
P1,1 P1,2 P2,1 P2,2 P(E2)
1 0.000 0.000 0.000 1.00 0.000
2 0.000 0.000 0.010 0.990 0.010
3 0.000 0.002 0.000 0.998 0.002
4 0.000 0.002 0.010 0.988 0.012
Finally we come to the top event of the fault tree E, i.e. the failure of the sub-sea tunnel project.
The extreme values of the occurrence probabilities are achieved by the same procedure as E2. Table 4-8
lists all extreme points with their values of P(E), and the upper and lower occurrence probabilities for E are
0.041 and 0.004, respectively.
Table 4-8: Extreme joint distributions of E1 and E2.
Extreme points of Ψcomb
P1,1 P1,2 P2,1 P2,2 P(E)
1 0.000 0.004 0.000 0.996 0.004
2 0.000 0.030 0.000 0.970 0.030
3 0.000 0.004 0.012 0.984 0.016
4 0.000 0.029 0.012 0.959 0.041
4.2.3 Combination of Event Tree Analysis and Fault Tree Analysis
Usually, it is difficult to directly evaluate the occurrence probability of some
events at the bottom of the fault tree, or such events cannot be decomposed further due to
the lack of knowledge. Then, one may consider an event tree analysis to evaluate the
127
probabilities of those events and input the calculated upper ad lower probabilities into the
fault tree analysis.
For example, consider Event E1,2,2 “Poor investigation” in the fault tree of
Example 4-4. It might not be easy to directly evaluate the probability for the occurrence
of this event, but one may first consider several common initiating events for E1,2,2 “Poor
investigation” first, and then obtain the probability bounds of E1,2,2 by an event tree
analysis as explained in Section 4.2.1.
It is common to incorporate Fault Tree Analysis with Event Tree Analysis in risk
analysis with precise probabilities, and the procedure for analysis with imprecise
probabilities is not much different. One can just carry out the analysis step by step for
each Fault Tree and Event Tree with imprecise probabilities as explained in the previous
two sub-sections 4.2.1 and 4.2.2.
4.3 DECISION ANALYSIS WITH IMPRECISE PROBABILITIES
In the conventional decision analysis, precise values are assigned to evaluate the
occurrence probabilities of events. By calculating the expected utility on each decision
branch, the branch with maximal value is chosen and kept in further analysis. However,
as explained previously (Section 2.1), when the information is limited, assigning precise
probabilities may not be sensible or practical. Usually we have to use imprecise
probabilities to describe uncertainty. In this section the author is going to develop a
methodology of decision analysis with imprecise probabilities.
128
4.3.1 Standard form of decision tree
A decision tree consists of decision nodes, chance nodes, and end nodes.
Typically, a decision node is represented by a square node, with branches representing
feasible alternatives, choice, or options; a chance node is represented by a circle, from
which possible outcomes radiate; an end node indicates a solution where the decision is
made and uncertainty has be resolved. Huntley and Troffaes (2008) introduce the
notation used here and suggest a standard form of decision trees, where
- A decision tree must start from a decision node;
- Successive decision nodes must be separated by a chance node;
- Successive chance nodes must be combined as one chance node;
- The number of nodes must be the same for all paths; otherwise, dummy decision
and chance nodes are added to the shorter paths.
This standard form is used because it enables us to discuss algorithms and methodologies
for decision analysis based on a general format of the decision tree. In the following
subsections, all decision trees are in the standard form.
Below is an example of a decision tree and its standard form. In the standard
form, a circle represents a chance node labeled as S and a square represents a decision
node labeled as ς . When at a decision node, we may refer to a choice also as a gamble;
for example, in Figure 4-13b at decision node 11ς , if the decision maker chooses option
111d , then in terms of gambles, the decision maker chooses gamble 1
11X , which has value
is 1111a if 11
11E occurs, and 1211a if 12
11E occurs.
129
(a) (b) (c)
Figure 4-13: (a) Decision tree, (b) its standard form (Huntley et al. 2008), and (c) reduced decision tree with only the optimal choices.
In Figure 4-13(b), d1 and d2 are decisions at the root node, and Ei1, Ei
2, … are the
events at the chance node Si. In the conventional decision analysis with precise
probabilities, by calculating the expected utilities on each decision branch, a unique
optimal choice is selected because of the maximal expected utility value. In Figure
4-13(c), the reduced decision tree shows only the optimal choices when the following
conditions are satisfied in sequence: (1) 111d is preferred to 1
12d given E11; (2) 2
12d is
preferred to 211d given E1
2; and (3) 1d is preferred to 2d . However, when the input
consists of imprecise probabilities, the optimal choice might not be unique, or it may
even be indeterminate, as explained in Section 2.6 and illustrated by Example 4-5 in the
context of decision trees.
1111a1211a
2112a2212a
21ς
11ς
ς
111X
1211a
1111a
1212a
1112a
2211a
2111a
2212a
2112a
1221a
1121a
1222a
1122a
11ς
21ς
12ς
ς
2S
130
4.3.2 Algorithm of decision analysis with imprecise probabilities
Let **ς and *
*iS be a decision node and a choice node, respectively. Here we use
super-script ‘*’ and sub-script ‘*’ to refer to any indices, and thus **iS means the i-th
choice node at decision node **ς . Let *
*A be the union of all events leading to decision
node **ς and chance node *
*iS ; let **N and *
*iN be the normal forms of sub-trees at
decision node **ς and chance node *
*iS , respectively; and let **
jiE be the j-th event at
the chance node **iS . Here we also use *
*j
iE to refer to its characteristic function, i.e., **
jiE = 1 if it occurs; otherwise, *
*j
iE = 0. Let opt(·) be the set of optimal choices. Huntley
and Troffaes (2008) show that at a decision node **ς , the set of optimal choices (or
gambles) is a subset of the union of the sets of optimal choices at its children **iS :
( ) ( )* * * ** * * *| |i
iopt N A opt N A⊆∪ ,
(4.35)
and that at an intermediate chance node **iS , any optimal choice could be written in the
form * ** *
j ji i
jE X∑ for ( )* * *
* * *|j j ji i iX opt N A∈ , where *
*j
iX is the gamble at the j-th
branch of chance node **iS , and *
*j
iE is the characteristic function for event **
jiE :
( ) ( )* * * * * * ** * * * * * *| : |j j j j ji i i i i i
jopt N A E X X opt N A
⎧ ⎫⎪ ⎪⊆ ∈⎨ ⎬⎪ ⎪⎩ ⎭∑ (4.36)
If **iN is the final chance node, then there is only one gamble *
*iX at this node.
Thus the optimal choice is
( ) * * * * ** * * * *| j ji i i i
jopt N A X E a
⎧ ⎫⎪ ⎪= = ⎨ ⎬⎪ ⎪⎩ ⎭∑ (4.37)
where **
jia is the reward (or utility) at the j-th branch of chance node *
*iS .
Given an event A and a set of gambles N, references Walley (1991, page 161) and
Huntley and Troffaes (2008) introduce the following formula to determine the optimal
gambles in set N: ( ) ( ) | : and , | 0LOWopt N A X N Y N Y X E Y X A= ∈ ∀ ∈ ≠ − ≤ (4.38)
131
If there is only one gamble X satisfying the constraints in Eq.(4.38), then X is preferred to
any other gamble in set N and thus only gamble X is kept and considered in further
analysis. If there are more than one optimal gamble in set opt(N|A), then the preference is
indeterminate and all optimal gambles in opt(N|A) should be kept. Section 2.5 deals with
preference between two gambles within the context of imprecise probability. It also gives
algorithms to calculate ( )|LOWE X Y A− and presents the relaxed constraints and the
strict constraints.
The general algorithm for decision analysis with imprecise probabilities is as
follows:
(1) Determine the set of optimal choices at the final chance nodes by Eq.(4.37);
(2) At the parent decision nodes of the chance nodes in last step, determine the set of
optimal choices by Eqs.(4.35) and (4.38).
(3) At the parent chance nodes of the decision nodes in Step 2, determine the set of
optimal choices by Eqs.(4.36) and (4.38).
(4) Recursively apply Step 2 and Step 3 untill the root decision node is reached.
Let us illustrate the algorithm by an example of tunnel construction strategy,
which is adapted from Karam et al. (2007), where precise probabilities were used.
Example 4-5 Consider a tunnel being constructed in a ground where two uncertain geologic states may
occur: either geologic state 1G or 2G. The current available information for the probability of geologic state
1G is: 0.35 ≤ P(1G) ≤ 0.45. Two construction strategies are considered. Table 4-9 lists cost of
construction strategies C1 and C2 in different geologic states: construction strategy C1 costs $1,890,000
under geologic state 1G and $1,935,000 under geologic state 2G; construction strategy C2 costs $1,125,000
under geologic state 1G and $2,520,000 under geologic state 2G. As a consequence, the cost for C1 is less
132
sensitive to the geological state than C2, but C2 is much less expensive than C1 if 1G occurs. The selection
of the construction strategy must be made between C1 and C2.
Table 4-9: Construction Cost Matrix.
Geologic state
Construction strategy 1G (×$1,000) 2G (×$1,000)
C1 1,890 1,935
C2 1,125 2,520
Figure 4-14: Decision tree for the tunnel, adapted from Karam et al. (2007).
First let us construct a decision tree for the problem, as shown in Figure 4-14, where jia is the
reward if construction strategy Ci is chosen and the geology state is jG, and jiE is a characteristic
function given that construction strategy Ci is chosen, i.e., jiE = 1 if geologic state is jG, otherwise, j
iE =
0. At the final chance nodes S1 and S2, Eq.(4.37) gives the sets of optimal solutions as
( )
( )
1 1 2 21 1 1 1 1 1
1 1 2 22 2 2 2 2 2
opt N E a E a X
opt N E a E a X
= + =
= + = (4.39)
At the decision node ς , the set of optimal solutions is a subset of the union of ( )1opt N and ( )2opt N ,
and is obtained by Eq. (4.35):
( ) 1 2,opt N X X⊆ (4.40)
By using Eq.(4.38), the preference between X1 and X2 is determined by ( )1 2LOWE X X− and
( )2 1LOWE X X− .
Construction Strategy
Geologic State
Cost(x $1,000)
d1=C1
d2=C2
a11= - 1890
a12= - 1935
a21= - 1125
a22= - 2520
1S
2S
ς
11 1GE =
21 2 GE =12 1GE =
22 2 GE =
1X
2X
133
We know the fact that uncertainty in the geologic states does not change with the selection of
construction strategies, thus two options to reflect this fact may be chosen: either the relaxed constraint or
the strict constraint in Section 2.5.
Let ijP be the probability of geologic state iG if construction strategy Cj is selected, i.e. event
ijE occurs. When the relaxed constraint is applied, the following optimization problems are obtained
(notice that 1 2X XΨ = Ψ )
Minimize ( )1 2X X−E = ( )( )
( )( )1 2
1 2 1 21 1 2 21890 1935 1125 2520
X X
P P P P− − − − −
E E
Minimize ( )2 1X X−E = ( )( )
( )( )2 1
1 2 1 22 2 1 11125 2520 1890 1935
X X
P P P P− − − − −
E E
Subject to (relaxed constraint) 1
11 2
1 1
1
0.35 0.45;
1;
0, 1, 2i
P
P P
P i
⎫≤ ≤⎪⎪+ = ⎬⎪≥ = ⎪⎭
1XΨ
12
1 22 2
2
0.35 0.45;
1;
0, 1,2i
P
P P
P i
⎫≤ ≤⎪⎪+ = ⎬⎪≥ = ⎪⎭
2XΨ
(4.41)
where ( ) ( ) ( ) ( ) ( )1 1 2 2 1 1 2 21 2 1 2 1 1 1 1 2 2 2 2X X X X E a E a E a E a− = − = + − +E E E E E = ( ) ( )( )1 1 2 2
1 1 1 1a E a E+E E -
( ) ( )( ) ( ) ( )( ) ( ) ( )1 1 2 2 1 1 2 2 1 1 2 2 1 1 2 21 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2a E a E a E a E a P a P a P a P+ − + = + − +E E E E . Likewise for ( )2 1X X−E .
Since ( )1 2 $27,000 0LOWE X X− = − < and ( )2 1 $117,000 0LOWE X X− = − < , then the preference
between X1 and X2 is indeterminate when the relaxed constraint is used.
When the strict constraint is applied, ( )1 2LOWE X X− and ( )2 1LOWE X X− are equal to -
$22,500 and -$112,500, respectively, and are obtained by solving the optimization problems in Eq.(4.42)
below. Note that these values are different from the results obtained when the relaxed constraint is used;
actually the lower expectation values become higher due to the strict constraint. However, the preference
between X1 and X2 is still indeterminate because both have negative signs.
134
Minimize ( )1 2X X−E = ( ) ( )1 2 1 21 1 2 21890 1935 1125 2520P P P P− − − − −
Minimize ( )2 1X X−E = ( ) ( )1 2 1 22 2 1 11125 2520 1890 1935P P P P− − − − −
Subject to
1 11 2P P= (strict constraint)
11
1 21 1
1
0.35 0.45;
1;
0, 1, 2i
P
P P
P i
⎫≤ ≤⎪⎪+ = ⎬⎪≥ = ⎪⎭
1XΨ
12
1 22 2
2
0.35 0.45;
1;
0, 1,2i
P
P P
P i
⎫≤ ≤⎪⎪+ = ⎬⎪≥ = ⎪⎭
2XΨ
(4.42)
Therefore, the allowable information: 0.35 ≤ P(1G) ≤ 0.45 is not enough to decide between
construction strategies C1 and C2. As Walley (1991, page 2) has stated,
“An inevitable consequence of admitting imprecise probabilities is that probabilistic reasoning may produce indeterminate conclusions (we may be unable to determine which of two events or hypotheses is more probable), and decision analysis may produce indecision (we may be unable to determine which of two actions is better). When there is little information on which to base our conclusions, we cannot expect reasoning (no matter how clever or thorough) to reveal a most probable hypothesis or a uniquely reasonable course of action. There are limits to the power of reason.”
If further information is available and narrows down the interval, then a preference may be
chosen; for example, in the original example in Karam et al. (2007), P(1G) = 0.4, and this precise
information allowed a unique optimal choice to be made (C1 was preferred to C2).
Let us assume that new or additional information has been acquired that leads to a tighter bound:
0.38 ≤ P(1G) ≤ 0.42. Section 4.3.3 that follows discusses how to combine prior and new information.
Here we are going to redo Example 4-5 and see if the tighter bound leads to a preferred option.
When the relaxed constraints are applied, problems (4.41) become:
135
Minimize ( )1 2X X−E = ( ) ( )1 2 1 21 1 2 21890 1935 1125 2520P P P P− − − − −
Minimize ( )2 1X X−E = ( ) ( )1 2 1 22 2 1 11125 2520 1890 1935P P P P− − − − −
Subject to (relaxed constraint)
11
1 21 1
1
0.38 0.42;
1;
0, 1, 2i
P
P P
P i
⎫≤ ≤⎪⎪+ = ⎬⎪≥ = ⎪⎭
1XΨ
12
1 22 2
2
0.38 0.42;
1;
0, 1, 2i
P
P P
P i
⎫≤ ≤⎪⎪+ = ⎬⎪≥ = ⎪⎭
2XΨ (4.43)
By solving problems (4.43), ( )1 2LOWE X X− and ( )2 1LOWE X X− are found to be equal to $16,200 and
-$73,800, respectively. Because ( )1 2LOWE X X− is positive and ( )2 1LOWE X X− is negative, then X1 is
preferred to X2 with the relaxed constraints, i.e., C1 is the optimal choice.
When the strict constraints are applied, problems (4.42) become
Minimize ( )1 2X X−E = ( ) ( )1 2 1 21 1 2 21890 1935 1125 2520P P P P− − − − −
Minimize ( )2 1X X−E = ( ) ( )1 2 1 22 2 1 11125 2520 1890 1935P P P P− − − − −
Subject to
1 11 2P P= (strict constraint)
11
1 21 1
1
0.38 0.42;
1;
0, 1, 2i
P
P P
P i
⎫≤ ≤⎪⎪+ = ⎬⎪≥ = ⎪⎭
1XΨ
12
1 22 2
2
0.38 0.42;
1;
0, 1, 2i
P
P P
P i
⎫≤ ≤⎪⎪+ = ⎬⎪≥ = ⎪⎭
2XΨ
(4.44)
Then ( )1 2LOWE X X− and ( )2 1LOWE X X− are found to be equal to $18,000 and $72,000− ,
respectively. Again, when the strict constraint is used, ( )1 2LOWE X X− is positive and ( )2 1LOWE X X−
is negative, thus X1 is preferred to X2, i.e., C1 is still the optimal choice. This shows that due to the
additional information, the tighter bound (0.38 ≤ P(1G) ≤ 0.42) allows a unique optimal choice to be
made.
136
4.3.3 Decision analysis with uncertain new information
Usually, new information is obtained by sampling or experiments. However, tests
and experiments are not perfect, and test results may not be accurate. With the new
information, uncertainty on possible states will change and thus their relevant
probabilities should be updated by taking into account the probability that the test result
is not accurate. This subsection deals with the decision analysis with probabilities
updated by new information. Both prior probabilities and reliabilities of new information
are assigned as imprecise probabilities as described next in Section 4.3.3.1. Then, Section
4.3.3.2 describes how to calculate the lower prevision of the difference between two
gambles conditional to an observation result. Such a prevision is necessary to construct
the set of optimal gambles in Eq. (4.38) that is used in Steps 2 and 3 of the algorithm on
page 131.
4.3.3.1 Input Data
Let S be a state variable with n possible states is , i = 1,…,n. The prior probability
measure of possible states is a vector p of size n with the i-th entry pi denoting the
probability of is . Let pΨ be the convex set of prior probabilities p. Assume that new
information is provided by a test, which predicts the realization of state variable S with
reliability matrix XR of size n n× , where ijXr , the (i, j)-th entry of XR , is the
conditional probability of test result being †is conditional to the choice X and real case
js . Here we use superscript ‘ † ’ to denote the test result. A perfect test is only a special
case with a unit reliability matrix.
Because each column of reliability matrix XR is a probability measure
conditional to the real case, and such conditional probability measure is given
imprecisely, the j-th column of XR may be written as an element of a convex set of
137
conditional probability measures, jΨ . Let j
Ψ be a convex set of reliability matrices,
which is defined as 11 1 1 1
1
:
j n jX XX X
j X j
n nj nn njX XX X
r r r r
r r r r
⎧ ⎫⎛ ⎞ ⎛ ⎞⎪ ⎪⎜ ⎟ ⎜ ⎟
Ψ = = ∈Ψ⎨ ⎬⎜ ⎟ ⎜ ⎟⎪ ⎪⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎩ ⎭
R (4.45)
Let rXΨ be the intersection of all convex sets
jΨ :
1
n jrX
j=Ψ = Ψ∩ (4.46)
rXΨ is convex because it is the intersection of convex sets, and each element in r
XΨ is
a reliability matrix XR : 11 1 1 1
1
: , 1,...,
j n jX XX X
r X jX
n nj nn njX XX X
r r r rj n
r r r r
⎧ ⎫⎛ ⎞ ⎛ ⎞⎪ ⎪⎜ ⎟ ⎜ ⎟
Ψ = = ∈Ψ =⎨ ⎬⎜ ⎟ ⎜ ⎟⎪ ⎪⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎩ ⎭
R (4.47)
Matrix XR should satisfy the constraints below: X r
X∈ΨR
( ) ( )T X
n n=1 R 1 ; 0X ≥R (4.48)
Let jEXT be the set of extreme points of jΨ . According to the definition of set rXΨ in Eq.(4.47), the set of extreme points of r
XΨ , denoted as rXEXT , is obtained as
11 1 1 1
1
: , 1,...,
j n jX XX X
r X jX
n nj nn njX XX X
r r r rEXT EXT j n
r r r r
⎧ ⎫⎛ ⎞ ⎛ ⎞⎪ ⎪⎜ ⎟ ⎜ ⎟
= = ∈ =⎨ ⎬⎜ ⎟ ⎜ ⎟⎪ ⎪⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎩ ⎭
R (4.49)
i.e., rXEXT is composed of matrices whose j-th column is in jEXT .
Let us prove this statement in Eq. (4.49) by a contradiction. Assume that XR is
a matrix in rXEXT and the k-th row of XR is a point in set kΨ but not an extreme
point of kΨ , i.e.,
138
11 1 1 1
1
1 1
where , 1,..., 1, 1,... ;
j k nX X XX
X
n nj nk nnX X XX
j kXX
j k
nj nkXX
r r r r
r r r r
r rEXT j k k n
r r
⎛ ⎞⎜ ⎟
= ⎜ ⎟⎜ ⎟⎝ ⎠⎛ ⎞ ⎛ ⎞⎜ ⎟ ⎜ ⎟
∈ = − + ∈Ψ⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
R
(4.50)
Assume m extreme points in set kΨ : 1
1
11
,...,
m
m
kkX X
k
nk nkX X
r rETX
r r
⎧ ⎫⎛ ⎞⎛ ⎞⎪ ⎪⎜ ⎟⎜ ⎟⎪ ⎪= ⎨ ⎬⎜ ⎟⎜ ⎟⎪ ⎪⎜ ⎟⎜ ⎟⎪ ⎪⎝ ⎠ ⎝ ⎠⎩ ⎭
(4.51)
Any point of set Ψj that is not an extreme point can be written in terms of a convex
combination of extreme points in kΨ : 1
1
111
1 ...
m
m
kkkX X X
mnk nk nkX X X
r r r
r r r
λ λ
⎛ ⎞⎛ ⎞⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟
= + + ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠
, 1 ... 1; 0, 1,...,m i i mλ λ λ+ + = ≥ = (4.52)
Thus, the extreme point XR can be written as a convex combination of m extreme
points in set rXΨ :
1
1
1111 1 11 1
11 1
...
m
m
kk n nX X X XX X
X mn nk nn n nk nnX X X XX X
r r r r r r
r r r rr r
λ λ
⎛ ⎞⎛ ⎞⎜ ⎟⎜ ⎟
= + + ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
⎝ ⎠ ⎝ ⎠
R (4.53)
but this means that XR cannot be an extreme point because any extreme point cannot
be expressed as a convex combination of other extreme points. Therefore, the proposition
in Eq.(4.50) is not correct.
4.3.3.2 Algorithm to calculate ( )†|LOW iE X Y s−
In order to calculate ( )†|LOW iE X Y s− , one needs to define the feasible domain
Ψ . In particular, in this case the feasible domain is identified by the extreme
139
probabilities of the state variable conditional to the test result †is . The train of thoughts is
as follows: the joint probability distribution over the space of test results and real cases is
first defined based on Bayes theorem applied to the reliability matrix and the prior
probabilities. The joint distribution if finally conditioned on †is . All of the above steps
are carried out on the extreme distributions.
Recall that jp is the prior probability of real case js , p is a vector whose j-th
entry is jp , and pΨ is a convex set composed of vectors p . Let QX be the joint
probability measure over the space of test results and real cases. The (i, j)-th entry ijXq =P(test = †
is , real case = js | Choice X) is obtained by using Bayes’ Theorem: ij ij
jX Xq r p= (4.54)
In matrix form, QX is obtained as
( ) X X Diag=Q R p (4.55)
Let ΨX be the set of probability measures over the joint space of test results and
real cases. Theorem 4-7 below shows that ΨX is a convex set and that its extreme points
may be obtained in an efficient way.
Theorem 4-7 Let pΨ be the set of prior probabilities p ; and let pEXT be its set of
extreme points. Let rXΨ be the set of reliability matrices and r
XEXT be the set of
extreme points of rXΨ (Eq. (4.49)). Then the set of joint distribution ΨX defined by
Eq.(4.56) is convex: ( ) : ,r
X X X X X pDiagΨ = ∈Ψ ∈Ψ= Q R p R p (4.56) The set of extreme points
XEXTQ of ΨX is:
( ) = : ,X
rX EXT X X X pEXT Diag EXT EXT= ∈ ∈Q R p R p (4.57)
Proof : Based on Eq.(4.56), , jX⋅R ,the j-th column of reliability matrix XR , may be
written as
140
,,
,
jj X
X T jX
⋅⋅
⋅=QR
1 Q, 1,...,j n= (4.58)
where , jX⋅Q is the j-th column of , j
X⋅Q and 1 is a unit vector of size n.
Since , jX⋅R is an element in the convex set jΨ , there are matrix 0j ≠A and
vector 0j ≠b such that set jΨ may be constructed by linear constraints as follows:
( ) , ,: 0j j jj jX X
⋅ ⋅Ψ = + ≤R A R b , 1,...,j n= (4.59)
By inserting Eq. (4.58) into the constraints in Eq. (4.59) and multiplying the constraints
in Eq. (4.59) by ,T jX⋅1 Q , one obtains:
( ) ( ), , 0j T jj jX X
⋅ ⋅+ ≤A Q b 1 Q , 1,...,j n= (4.60)
Thus, set ΨX may be obtained as ( ) ( ) , ,: 0, 1,...,j T j
X j jX X j n⋅ ⋅Ψ = + ≤ =Q A Q b 1 Q (4.61)
Since the constraints used to construct set ΨX in Eq. (4.61) are all linear, set ΨX is
convex.
Regarding the extreme points of set ΨX, any p∈ pΨ and RX∈ rXΨ can be
written as a linear combination of extreme points in pΨ and rXΨ , respectively:
( )
( )
( )
( )
1
1 ... ...p
p
TEXT
Tp p pEXT
T
EXT
ξξ ξ
ξ
λ λ λ
⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
= ⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠
p
p p
p
10 1, 1,..., ; 1
pp p
p
ξ
ξ ξξ
λ ξ ξ λ=
≤ ≤ = =∑
(4.62)
141
( ) ( ) ( ) ( )( ) 11 11
11
1,1 ,1 , ,
1
0
0
0
0
n
n
n nnX EXT EXT EXT EXT
n
ξ ξ
ξ
ξ
λ
λλ
λ
⋅ ⋅ ⋅ ⋅
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥=⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
R R R R R
1
0 1, 1,...,
1, 1,...,j
jj
j j n
ξ
ξ
ξξ
λ ξ ξ
λ=
≤ ≤ =
= =∑
where ,i
jEXT⋅R is the i-th extreme point in set jΨ .
(4.63)
By inserting Eqs.(4.62) and (4.63) into Eq. (4.56), one obtains:
( ) ( ) ( ) ( )( )
( )
( )
( )
( )
11 11
11
1,1 ,1 , ,
1
1
1
0
0
0
0
... ...
n
n
p
p
n nnX EXT EXT EXT EXT
n
TEXT
Tp p pEXT
T
EXT
Diag
ξ ξ
ξ
ξ
ξξ ξ
ξ
λ
λλ
λ
λ λ λ
⋅ ⋅ ⋅ ⋅
⎡ ⎤⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥= ⋅⎢ ⎥⎢ ⎥⎢ ⎥⎢ ⎥⎣ ⎦
⎛ ⎞⎛ ⎞⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠
Q R R R R
p
p
p
(4.64)
Extreme points of XΨ are achieved if and only if 1, 0,
p mmξ
ξλ
ξ=⎧
= ⎨ ≠⎩ and
1,
0, jj
jξ
ξ ηλ
ξ η
=⎧⎪= ⎨ ≠⎪⎩,
1,..., pm ξ= , 1,...,j jη ξ= .
Therefore, ( )X pX
mEXT EXTEXT Diagη=Q R p and the upper limit for the number of
extreme joint distribution is 1
n
p jj
ξ ξ=∏ . ◊
142
After obtaining the extreme points of the joint distributions XQ , at gamble X, the extreme probabilities of the state variable conditional to the test result †
is , †| i
XEXT sP , are
obtained by applying Theorem 3-2, i.e.:
†
,
,|X
iX
iEXTX
iEXT sEXT
⋅
⋅=Q
PQ 1
(4.65)
where ,X
iEXT⋅Q is the i-th row of
XEXTQ and 1 is a unit vector of length n.
Likewise for †| i
YEXT sP .
Let †| i
XsEXT and †| i
YsEXT be the sets of extreme probabilities †| i
XEXT sP and
†| i
YEXT sP , respectively. The lower prevision of the gamble difference X-Y conditional to †
is
can be obtained by enumerating all of the extreme conditional probabilities:
( ) ( ) ( )( )† † †
,| min | |
X YLOW i i iE X Y s X s Y s− = −
P PE E
Subject to †| i
XX sEXT∈P , †| i
YY sEXT∈P
(4.66)
If the choice between gambles X and Y does not affect the uncertainty of the state variable
S, then the relaxed constraint could be applied to reflect this fact, i.e., in addition to the
constraints in Eq. (4.66), one imposes that the sets of extreme conditional distributions
are the same: † †| |i i
X Ys sEXT EXT= (4.67)
Alternatively, the strict constrain could be applied as well, i.e., in addition to the
constraints in Eq. (4.66), the extreme conditional distributions are forced to be the same: † †| |i i
X YEXT s EXT s=P P (4.68)
Example 4-6 Consider again the situation and information in Example 4-5. Suppose that now we could
obtain new information by performing additional exploration and we want to combine prior and new
information. It is assumed that construction strategies do not affect uncertainty in exploration reliability,
143
which is quantified by the reliability matrix assigned as in Table 4-10. In this example, the gamble X is the
construction strategy, i.e., X can be either C1 or C2. However, set rXΨ is unique because the reliability
matrix is independent of the construction strategy. Set S is composed of the two geological states: 1s = 1G,
2s = 2G. The extreme points of set rXΨ (i.e., r
CjΨ , 1, 2j = ) are obtained by taking extreme distributions
in each column of the reliability matrix as described in Eq.(4.49): 0.85 0.35 0.85 0.45 0.95 0.35 0.95 0.45
, , ,0.15 0.65 0.15 0.55 0.05 0.65 0.05 0.55
rXEXT
⎧ ⎫⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎪ ⎪= ⎨ ⎬⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎪ ⎪⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎩ ⎭
(4.69)
The prior probabilities are assigned as 0.35 ≤ P(1G) ≤ 0.45, which gives the following extreme points
for ΨP: 1EXTp = (0.35, 0.65), and 2
EXTp = (0.45, 0.55). Construction strategies need to be selected between
C1 and C2 conditional on the exploration results.
Table 4-10: Exploration Reliability Matrix, where the following values are assigned 0.85≤ r11≤ 0.95; 0.35≤ r12≤ 0.45. Hence, 0.05≤ r21≤ 0.15; 0.55≤ r22≤ 0.65.
Real state Exploration indicates
geologic state given reality 1G 2G
1G r11 r12
2G r21 r22
Figure 4-15: Decision tree for the tunnel with exploration, adapted from Karam et al. (2007).
Construction Strategy
Geologic State
Cost(x $1,000)
Exploration Results
1Sς
1 111 1GE =
11ς
21ς
1 211 2GE =1 112 1GE =
2 11 1 1GE =
2 112 1GE =
1 21 2 2GE =
2 212 2 GE =
11 1S
11 2S
21 1S
21 2S
111 1d C=
11 2 2d C=
212 2d C=
211 1d C=
1111 1890a = −
1211 1935a = −
1112 1125a = −
1212 2520a = −2111 1890a = −
2211 1935a = −
2112 1125a = −
2212 2520a = −
2 211 2 GE =
111X112X211X212X
11 1E G=
21 2E G=
144
First, let us construct a decision tree for the tunnel (shown in Figure 4-15), where jiE is the event
that the exploration result is jG, jeikE is the event that the real geologic state is eG given the test result jG
and construction strategy Ck is chosen, and jeika is the reward obtained if event je
ikE occurs. At the final
chance nodes 111S , 1
12S , 211S and 2
12S , Eq.(4.37) gives the sets of optimal solutions as
( ) 1 11 11 12 12 111 11 11 11 11 11opt N E a E a X= + =
( ) 1 11 11 12 12 112 12 12 12 12 12opt N E a E a X= + =
( ) 2 21 21 22 22 211 11 11 11 11 11opt N E a E a X= + =
( ) 2 21 21 22 22 212 12 12 12 12 12opt N E a E a X= + =
(4.70)
At the decision nodes 11ς and 2
1ς , based on Eq.(4.35), the sets of optimal solutions are the subsets of the
union of ( )111opt N and ( )1
12opt N and the union of ( )211opt N and ( )2
12opt N , respectively,
( ) 1 1 11 11 12,opt N X X⊆
( ) 2 2 21 11 12,opt N X X⊆ (4.71)
By using Eq.(4.38), the optimal decisions at nodes 11ς and 2
1ς are determined by ( )1 111 12LOWE X X− ,
( )1 112 11LOWE X X− and ( )2 2
11 12LOWE X X− , ( )2 212 11LOWE X X− .
With the assigned prior probabilities of geological states 0.35 ≤ P(1G) ≤ 0.45 and the
exploration reliability matrix shown in Eq. (4.69), the set of extreme joint probabilities ( XEXT ) can be
obtained by applying Theorem 4-7 and is given in Table 4-11. For example, the first extreme joint
probability distribution is obtained by 0.85 0.35 0.350.15 0.65 0.65
Diag⎛ ⎞⎛ ⎞ ⎛ ⎞⎜ ⎟⎜ ⎟ ⎜ ⎟⎜ ⎟⎝ ⎠ ⎝ ⎠⎝ ⎠
, i.e., the product of the first extreme
point of rXΨ and the first extreme point of pΨ . XEXT has 8 elements because pΨ has 2 extreme
points and rXΨ has 4 extreme points. Note that XEXT is the same for both construction strategies
because the reliability matrix is independent of the construction strategy.
145
Table 4-11: Set of extreme joint probabilities XEXT .
,XEXT mQ Extreme point
number, m 1,1
,XEXT mq 1,2,XEXT mq 2,1
,XEXT mq 2,2,XEXT mq
1 0.298 0.228 0.053 0.423
2 0.298 0.293 0.053 0.358
3 0.333 0.228 0.018 0.423
4 0.333 0.293 0.018 0.358
5 0.383 0.193 0.068 0.358
6 0.383 0.248 0.068 0.303
7 0.428 0.193 0.023 0.358
8 0.428 0.248 0.023 0.303
Finally, Eq.(4.65) gives the extreme probabilities conditional to the exploration results, listed in
Table 4-12. For example, the first extreme conditional probability distribution is obtained by taking
following operations on the first extreme joint probability distribution: 0.298 0.228 0.053 0.423, , ,
0.298 0.228 0.298 0.228 0.053 0.423 0.053 0.423⎛ ⎞⎜ ⎟+ + + +⎝ ⎠
. Again, these extreme points are the same for
both construction strategies C1 and C2.
Table 4-12: Extreme probabilities conditional to the exploration results.
†|1m
CjEXT G
P †|2m
CjEXT G
P Extreme point number, m
( )†|11
m
CjEXT G
P G ( )†|12
m
CjEXT G
P G ( )†|21
m
CjEXT G
P G ( )†|22
m
CjEXT G
P G
1 0.567 0.433 0.111 0.889 2 0.504 0.496 0.128 0.872 3 0.594 0.406 0.040 0.960 4 0.532 0.468 0.047 0.953 5 0.665 0.335 0.159 0.841 6 0.607 0.393 0.182 0.818 7 0.690 0.310 0.059 0.941 8 0.633 0.367 0.069 0.931
146
Let 1lijP be the probability of geologic state iG if the exploration result is geologic state †
ls lG=
and construction strategy Cj is selected. Either the relaxed constraint in Eq.(4.67) or the strict constraint in
Eq.(4.68) should be chosen to reflect the fact that selection of construction strategies will not change
uncertainty of geologic states.
At decision node 11ς (i.e., given exploration result †1G ), when the relaxed constraint is applied,
( )1 111 12LOWE X X− and ( )1 1
12 11LOWE X X− are equal to -$354,200 and $87,400, respectively, obtained by
solving the following optimization problems:
Minimize ( )1 1
11 12X X−E = ( ) ( )11 12 11 1211 11 12 121890 1935 1125 2520P P P P− − − − −
Minimize ( )1 112 11X X−E = ( ) ( )11 12 11 12
12 12 11 111125 2520 1890 1935P P P P− − − − −
Subject to (relaxed constraint)
†
11 811 1
1, |112111
8
1, 1,1
;
0; 1
m
Cm EXT G
m
m mm
P
Pλ
λ λ
=
=
⎫⎛ ⎞⎪⎜ ⎟ =
⎜ ⎟ ⎪⎝ ⎠ ⎬⎪
≥ = ⎪⎭
∑
∑
P
†1
|1C
GΨ †
11 812 2
2, |112112
8
2, 2,1
;
0; 1
m
Cm EXT G
m
m mm
P
Pλ
λ λ
=
=
⎫⎛ ⎞⎪⎜ ⎟ =
⎜ ⎟ ⎪⎝ ⎠ ⎬⎪
≥ = ⎪⎭
∑
∑
P
†2
|1C
GΨ
(4.72)
Since ( )1 111 12LOWE X X− is negative and ( )1 1
12 11LOWE X X− is positive, then gamble 112X is preferred
to 111X under the relaxed constraint.
Alternatively, if the strict constraint is applied, ( )1 111 12LOWE X X− and ( )1 1
12 11LOWE X X− are
obtained by solving the optimization problems in Eq. (4.73) and they are equal to -$345,800 and $95,700,
respectively, which are higher than the results obtained with the relaxed constraints. Thus, gamble 112X is
preferred to 111X under the strict constraint as well.
147
Minimize ( )1 111 12X X−E = ( ) ( )11 12 11 12
11 11 12 121890 1935 1125 2520P P P P− − − − −
Minimize ( )1 112 11X X−E = ( ) ( )11 12 11 12
12 12 11 111125 2520 1890 1935P P P P− − − − −
Subject to
11 1111 1212 12
11 12
P P
P P
⎛ ⎞ ⎛ ⎞⎜ ⎟ ⎜ ⎟=⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
(strict constraint)
†
11 811 1
1, |112111
8
1, 1,1
;
0; 1
m
Cm EXT G
m
m mm
P
Pλ
λ λ
=
=
⎫⎛ ⎞⎪⎜ ⎟ =
⎜ ⎟ ⎪⎝ ⎠ ⎬⎪
≥ = ⎪⎭
∑
∑
P
†1
|1C
GΨ †
11 812 2
2, |112112
8
2, 2,1
;
0; 1
m
Cm EXT G
m
m mm
P
Pλ
λ λ
=
=
⎫⎛ ⎞⎪⎜ ⎟ =
⎜ ⎟ ⎪⎝ ⎠ ⎬⎪
≥ = ⎪⎭
∑
∑
P
†2
|1C
GΨ
(4.73)
Likewise for the case at node 21ς (i.e., given exploration result †2G ). ( )2 2
11 12LOWE X X− and
( )2 212 11LOWE X X− are equal to $332,300 and -$537,700 with the relaxed constraint, and they are equal to
$338,700 and -$531,300 with the strict constraint, respectively. As a result, gamble 211X is preferred to
212X in both cases.
Therefore, with the prior probability 0.35 ≤ P(1G) ≤ 0.45 and the imprecise reliability of the
exploration, (1) if the exploration predicts geologic state †1G , construction strategy C2 (i.e. 112X ) is
always preferred to construction strategy C1 (i.e. 111X ) no matter whether the relaxed constraint or the
strict constraint is applied; (2) if the exploration predicts geologic state †2G , construction strategy C1 (i.e.
211X ) is always preferred to construction strategy C2 (i.e. 2
12X ).
Compared to the indeterminate result in Example 4-5, the new information provided by further
exploration allowed us to decide between C1 and C2, despite the fact that all information was provided in
terms of imprecise probabilities. ♦
Now let us see an example with perfect information.
148
Example 4-7 Consider again the situation and information in Example 4-5. Suppose now that we could
obtain perfect information by performing additional exploration. The exploration reliability is shown in
Table 4-13 in this case, it is an identity matrix because perfect information is assumed. Thus, set rXΨ
(i.e., rCjΨ , 1,2j = ) collapses to a single point, i.e.,
1 00 1
rX
⎧ ⎫⎛ ⎞⎪ ⎪Ψ = ⎨ ⎬⎜ ⎟⎪ ⎪⎝ ⎠⎩ ⎭
. As in Example 4-6, the prior
probabilities are assigned as 0.35 ≤ P(1G) ≤ 0.45, which give extreme distributions for ΨP: 1EXTp =
(0.35, 0.65), and 2EXTp = (0.45, 0.55). Constructions strategy needs to be selected between C1 and C2 given
the exploration results.
Table 4-13: Exploration Reliability Matrix.
Real state Exploration indicates
geologic state given reality 1G 2G
1G 1 0
2G 0 1
The decision tree with perfect information is the same as in Figure 4-15, and at the final chance
nodes 111S , 1
12S , 211S and 2
12S , the sets of optimal solutions are again defined as in Eq.(4.70).
Since ΨP contains 2 extreme points and rXΨ contains 1 extreme point, there are only two
extreme joint probabilities in XEXT . Table 4-14 and both extreme joint probabilities yield the same
extreme conditional probability as listed in Table 4-15, which means no imprecision and consequently no
difference between the relaxed constraint and the strict constraint. Thus, at decision node 11ς (i.e., given
exploration result †1G ), ( )1 111 12LOWE X X− = ( )1 1
11 12X X−E = (-1890×1-1935×0)-(-1125×1-2520×0) =
-765 × $1,000 = -$765,000, and ( )1 112 11LOWE X X− = ( )1 1
12 11X X−E = ( )1 111 12X X− −E = $765,000,
indicating that gamble 112X is preferred to 1
11X ; at decision node 21ς (i.e., given exploration result
†2G ), ( ) ( )2 2 2 211 12 11 12LOWE X X X X− = −E = (-1890× 0-1935× 1)-(-1125× 0-2520× 1) = 585× $1,000 =
$585,000, and ( ) ( ) ( )2 2 2 2 2 212 11 12 11 12 11LOWE X X X X X X− = − = − −E E = -$585,000, indicating that gamble
211X is preferred to 2
12X .
149
Therefore, with the prior probability 0.35 ≤ P(1G) ≤ 0.45 and perfect information from the
additional exploration, (1) if the exploration predicts geologic state †1G , construction strategy C2 (i.e.
112X ) is preferred to construction strategy C1 (i.e. 1
11X ); (2) if the exploration predicts geologic state
†2G , construction strategy C1 (i.e. 211X ) is preferred to construction strategy C2 (i.e. 2
12X ). These
decisions are the same as the ones in Example 4-6. ♦
Table 4-14: Set of extreme joint probabilities, XEXT .
XEXTQ No. of extreme probabilities
1,1EXTq 1,2
EXTq 2,1EXTq 2,2
EXTq
1 0.35 0 0 0.65
2 0.45 0 0 0.55
Table 4-15: Extreme probabilities conditional to the exploration results.
†|1CjEXT G
P †|2CjEXT G
P No. of extreme probabilities ( )†|1
1CjEXT G
P G ( )†|12Cj
EXT GP G ( )†|1
1CjEXT G
P G ( )†|12Cj
EXT GP G
1 1 0 0 1
4.3.3.3 Discussion
The previous Section 4.3.3.2 studied updating probability measures with
imprecise new information. In this section the effect of the uncertainty of the new
information on the probability measures is investigated. Two indeterminate cases are
explained and discussed.
4.3.3.3.1 Uncertain new information
150
Recall the imprecise prior information in Example 4-5 and the imprecise
reliability matrix in Table 4-10. The sets of probability measures conditional to the
exploration results determined in Section 4.3.3.2 (extreme points of the set are listed in
Table 4-12) are depicted in Figure 4-16a, where pΨ is the set of prior probability
measures on the state variable S (geological states), ( )†|1S GΨ and ( )†| 2S GΨ are
the sets of probability measures conditional to exploration results †1G and †2G ,
respectively. Figure 4-16a illustrates the changes in the set of probability measures caused by new information: (1) sets ( )†|1S GΨ and ( )†| 2S GΨ are larger than pΨ ;
(2) sets ( )†|1S GΨ and ( )†| 2S GΨ are closer to the two ends (i.e., zero-uncertainty,
where P(1G) = 1 and P(2G) = 1, respectively) than pΨ . The first observation is a result
of the uncertainty of the new information, which compounds with the original
imprecision. As for the second observation, the reason is that the new evidence increases
the probability of observing one geological state or the other, and the increase depends on
the reliability of the evidence.
Consequently, the new information assigned in terms of imprecise probabilities
acts independently on the imprecision and on the probability of geological states.
Although the updated probability measures are more imprecise than the prior probability
measures, it may be enough to make a decision, which could not have been obtained with
the prior information. For instance, decision between construction strategies C1 and C2 is indeterminate within pΨ in Example 4-5 but become determinate within ( )†|1S GΨ
and ( )†| 2S GΨ in Example 4-6, though both ( )†|1S GΨ and ( )†| 2S GΨ are more
imprecise than pΨ . On the other hand, it is also possible that the updated information is
not enough to make a decision, because of the increase in imprecision possibly leading to
a negative lower value of the new information. Thus, more information may not always
151
be better than less. An example of such a situation will be shown and illustrated in the
case history of Section 5.3.
It is worth comparing the effect of additional information using precise
probabilities, where imprecision cannot be taken into account. With precise probabilities,
only the second phenomenon can be observed. Thus, the value of information within
precise probabilities can never be negative, and more information is always better than
less.
Below is the discussion about the effects of the imprecision and the reliability of
the uncertain new information on the set of probability measures on the state variable S:
(1) Imprecision of new information We construct two additional convex sets ( )†|1S GΨ and ( )†| 2S GΨ (as
shown in Figure 4-16b), which are obtained from a more precise exploration, to study the
effect of imprecision of the new information and compare it with the original case in Table
4-10 (as shown in Figure 4-16a).
In Figure 4-16a, where the reliability interval width is 0.1, sets of probability measures conditional to exploration results (i.e. ( )†|1S GΨ and ( )†| 2S GΨ ) are larger
than the set of prior probability measures (i.e. pΨ ). In Figure 4-16b, where the
reliability interval width is 0.05, the size of ( )†|1S GΨ and ( )†| 2S GΨ is similar to
pΨ and is obviously smaller than the corresponding sets in Figure 4-16a. Thus, the
imprecision of probability measures on the state variable is highly correlated with the
imprecision of the new information.
152
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0
P (1G)
P(2
G)
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0
P (1G)P
(2G
)
(a) (b)
Figure 4-16: Effect of imprecision: a) Table 4-10: 0.85≤ r11≤0.95; 0.35≤ r12≤0.45; b) 0.85≤ r11≤0.90; 0.3≤ r12≤0.35
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.2 0.4 0.6 0.8 1.0
P (1G)
P(2
G)
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
P (1G)
P(2
G)
(a) (b)
Figure 4-17: Effect of reliability: (a) Table 4-10: 0.85≤ r11≤0.95; 0.35≤ r12≤0.45; (b) 0.88≤ r11≤0.98; 0.15≤ r12≤0.25
( )†| 2S GΨ
pΨ
( )†|1S GΨ
∑P(iG)=1
A
B
( )†| 2S GΨ
pΨ
( )†| 1S GΨ
∑P(iG)=1
A
B
( )†| 2S GΨ
pΨ
( )†| 1S GΨ
∑P(iG)=1
A
B
( )†| 2S GΨ
pΨ
( )†| 1S GΨ
∑P(iG)=1
A
B
153
(2) Reliability of new information
Figure 4-17a reproduces Figure 4-16a, and Figure 4-17b presents a case with
higher reliabilities but same interval width, i.e. 0.1. Compared to the original case in Figure 4-17a, sets ( )†|1S GΨ and ( )†| 2S GΨ in Figure 4-17b are closer to the two
ends (where P(1G) = 1 and P(2G) = 1, respectively). Higher reliabilities ultimately
generate a higher probability of observing one geological state or the other.
4.3.3.3.2 Indeterminate cases
Because of imprecision in probabilities, we often have indeterminate problems,
such as Example 4-5, where a decision between construction strategies C1 and C2 could
not be made. Generally, there are two kinds of cases in which indeterminate problems
occur:
Case (1): Intersection point
In this case, the problem is indeterminate when the expected values of the utility
of the two gambles attain the same value (and thus intersect) somewhere in the set of
probability measures on the state variable. As a result, the preference cannot be
determined. Take Example 4-5 as an instance. The horizontal axis in Figure 4-18 and
Figure 4-19 is the line with equality P(1G)+P(2G)=1 in Figure 4-16 and Figure 4-17;
their coordinates are set in Figure 4-16a. The vertical axis in Figure 4-18 and Figure 4-19
are the expected utilities for gambles. The set of prior probabilities (set pΨ in Figure
4-18a ) contains the intersection point of two expected utility curves for gambles X1 and
X2 (X1 = C1 and X2 = C2). Therefore, the preference between C1 and C2 is indeterminate
based on prior information. When new evidence is provided by additional exploration, the probability measures are updated. As shown in Figure 4-18a, sets ( )†|1S GΨ and
154
( )†| 2S GΨ do not contain the intersection point. When the exploration result is †1G ,
C2 is preferred to C1; when it is †2G , C1 is preferred to C2 (see case (2) for details).
However, if the new information cannot move the set of probability measures far from the intersection point, the problem is still indeterminate, for example, set ( )†|1S GΨ in
Figure 4-18b.
(a) (b)
Figure 4-18: Case (1): Intersection point in set Ψ
Case (2): No intersection point
Figure 4-19 shows the case in which there is no intersection point within the set of
probability measures on the state variable. Recall that strict constraints use the same
probability measure for different choices (Figure 2-10). Thus, we can always select the
unique option as long as its utility curve is higher than all the others. As shown in Figure
4-19, since curve for X2 (i.e., construction strategy C1) is above that for X1 (i.e.,
construction strategy C2) in Ψ , C2 is preferred to C1 under the strict constraints.
However, the relaxed constrains take the whole set of probability measures into
consideration (Figure 2-9). Even if there is no intersection point, problems may still be
indeterminate. Figure 4-19 illustrates this situation, since Emin(X2) < Emax(X1) and then
X2
X1
( )†| 2S GΨ pΨ ( )†|1S GΨ
∑P(iG)=1
E(Xi)
A B
X2
X1
( )†| 2S GΨ ( )†|1S GΨpΨ
∑P(iG)=1
E(Xi)
A B
155
Emin(X2- X1) = Emin(X2) - Emax(X1) < 0, we cannot draw the conclusion that C2 is preferred
to C1.
Figure 4-19: Case (2): No intersection point in set Ψ
To summarize, in Case (1) indeterminate problems occur regardless of the type of
constraints imposed, while indeterminate problems such as Case (2) can occur only when
relaxed constraints are applied.
4.3.4 Lower and upper values of information
Oftentimes, the decision maker wants to know the maximum cost to be allowed
for collecting additional information. In conventional decision analysis with precise
probabilities, the maximum cost is determined by the ‘value of information’ (VI); for
perfect information, it is the ‘value of perfect information’ (VPI). Let X be the choice
made without additional information. Let Y be the choice made with imperfect additional
information, and Z be the choice made with perfect additional information. In precise
probabilities, if consequence is a linear function of the cost of information, the value of
X2
X1
Ψ
Emax(X1)
Emin(X2)
∑P(iG)=1
E(Xi)
A B
156
imperfect information (VI) and the value of perfect information (VPI) are obtained as
follows (e.g. Ang and Tang, 1984, Page s 30 - 31):
( ) ( )VI Y X= −E E (4.74)
( ) ( )VPI Z X= −E E (4.75)
Accordingly, the updated value (VUP) from imperfect information to perfect information
is
( ) ( )VUP Z Y= −E E (4.76)
Because in precise probability the probability measure is unique, the value of (perfect)
information is a single value, which is the maximum cost to be allowed for collecting
additional information and also the minimum selling price of (perfect) information for
which we could give away the new information under the scenario that we acquired the
new information and someone is willing to buy it from us. In other words, a reasonable
price for imperfect (or perfect) information should be less than VI (or VPI); on the other
hand, if an amount higher than VI (or VPI) is offered to the decision maker, the decision
maker is willing to sell the new information and not use it anymore.
When dealing with imprecise probabilities, the probability measures are not
unique, and therefore the value of (perfect) information can no longer be a single value:
indeed, it is an interval, which is bounded by lower and upper values of (perfect)
information. The lower value of (perfect) information determines the maximum allowed
cost for collecting the information, and the upper value of (perfect) information is the
minimum selling price for the new information. There is no reason why the maximum
buying price should be the same as the minimum selling price. For example, consider a
contractor that has to bid on a project: the cost that he is willing to pay for additional
information on the project is different (much smaller) than the amount for which he
157
would sell this information to his competitors without using the information. Here we are
going the explain how to obtain these bounds on the value of information.
In imprecise probabilities, the bounds on VI and VPI are calculated as follows: ( ) ( )( ) ( )( ) ( )VI min minLOW LOWY X Y X E Y X= − = − = −E E E
( ) ( )( ) ( )( ) ( )VI max minUPP LOWY X X Y E X Y= − = − − = − −E E E (4.77)
( ) ( )( ) ( )( ) ( )VPI min minLOW LOWZ X Z X E Z X= − = − = −E E E
( ) ( )( ) ( )( ) ( )VPI max minUPP LOWZ X X Z E X Z= − = − − = − −E E E (4.78)
Here VILOW ( VPILOW ) is the least value of the imperfect (perfect) information.
Therefore, VILOW is the maximum cost allowed for collecting the information. VIUPP
(VPIUPP) is the maximum value of the imperfect (perfect) information. Once the price is
higher than VIUPP (VPIUPP), we should be willing to sell the information without using it.
If any lower value of (perfect) information is found to be negative, then the preference
among the involved gambles is indeterminate, indicating that it may not be worth buying
the new (perfect) information even if the new information is free.
Similarly, lower and upper updated values from imperfect information to perfect
information (VUP) are as follows:
( )VUPLOW LOWE Z Y= −
( )VUPUPP LOWE Y Z= − − (4.79)
By observing Eq.(4.78), ( )VPILOW LOWE Z X= − = ( ) ( )LOWE Z Y Y X⎡ ⎤− + −⎣ ⎦ .
In imprecise probabilities, the following property hold (Walley 1991, page 76): ( ) ( )LOWE Z Y Y X⎡ ⎤− + −⎣ ⎦ ≥ ( ) ( )LOW LOWE Z Y E Y X− + − = VUPLOW + VILOW .
Therefore,
VPI VUP VILOW LOW LOW≥ + , or
VUP VPI VILOW LOW LOW≤ − (4.80)
158
indicating that the maximum allowed cost (i.e., lower value) for perfect information
( VPILOW ) is more than the sum of the lower updated value ( VUPLOW ) and the lower
value of imperfect information ( VILOW ); on the other hand, the lower updated value
( VUPLOW ) can not exceed the difference between the lower value of perfect information
( VPILOW ) and the lower value of imperfect information ( VILOW ). Similarly,
VPI VUP VIUPP UPP UPP≤ + , or
VUP VPI VIUPP UPP UPP≥ − (4.81)
which means that the upper value for perfect information ( VPIUPP ) may not exceed the
sum of the lower updated value ( VUPUPP ) and the lower value of imperfect information
( VIUPP ); also, the upper updated value ( VUPUPP ) is more than the difference between
the upper value of perfect information ( VPIUPP ) and the upper value of imperfect
information ( VIUPP ).
Let us now consider the case in which there are sets of gambles: iX and iY ,
and the problem with perfect information is always determinate and thus only contains
one gamble, i.e. Z . Then, the lower and upper VI, VPI, and VUP are as follows:
( ),
VI minLOW LOW j ii jE Y X= −
( )( ) ( ),,
VI max minUPP LOW i j LOW i ji ji jE X Y E X Y= − − = − − (4.82)
( )VPI minLOW LOW iiE Z X= −
( )( ) ( )( )VPI max minUPP LOW i LOW iiiE X Z E X Z= − − = − − (4.83)
( )VUP minLOW LOW jjE Z Y= −
( )VUP minUPP LOW jjE Y Z= − − (4.84)
Relations in Eqs.(4.80) and (4.81) still hold in this case.
159
Again, if the choice between no additional information (X), imperfect information
(Y) and perfect information (Z) does not affect the uncertainty on the state variables, then
both the relaxed constraint and the strict constraint could be applied.
Notice that a negative lower value of imperfect information may occur under both
the relaxed constraints and the strict constraints. Given the new evidence from imperfect
new information, preference among gambles (choices) is probably indeterminate. The
indeterminacy among multiple choices may lead to a negative lower value of imperfect
information (as will be exemplified in Figure 5-14 and Figure 5-15). The situation in
perfect information is different. With perfect information, problems are determinate. A
negative lower value of perfect information can never occur under the strict constraints;
however, it is possible under the relaxed constraints. Case(2) in Section 4.3.3.3.2 (Figure
4-19) may occur, where X2 represents the case with perfect information and X1 represents
no information. Thus, a negative lower value of perfect information is obtained (as will
be exemplified in Figure 5-17). Within precise probabilities, because perfect information
is always preferred to no information, we can never obtain a negative value of
information.
Here a simple example illustrates the methodology to determine lower and upper
values of information.
Example 4-8 Consider again the situation and information in Example 4-5, the imperfect exploration in
Example 4-6, and the perfect exploration in Example 4-7. We now want to determine the lower and upper
values of VI, VPI, and VUP.
The reduced decision tree for exploration is depicted in Figure 4-20, which only includes the
optimal choices obtained from Example 4-5 through Example 4-7:
160
(1) At chance node S1 (i.e. the choice of no exploration), the set of optimal gambles is
( ) 1 11 11 12,opt N X X= ;
(2) At chance node S2 (i.e. the choice of imperfect exploration), the set of optimal gambles is
( ) 1 1 2 22 2 22 2 21opt N E X E X= + ;
(3) At chance node S3 (i.e. the choice of perfect exploration), the set of optimal gambles is
( ) 1 1 2 23 3 32 3 31opt N E X E X= +
Figure 4-20: Decision tree for the tunnel exploration, adapted from Karam et al. (2007).
To determine the lower and upper values of imperfect information (VI), Eq.(4.82) reads as
follows:
( ) ( )( )1 1 2 2 1 1 1 2 2 12 22 2 21 11 2 22 2 21 12VI min ,LOW LOW LOWE E X E X X E E X E X X= + − + −
( ) ( )( )1 1 1 2 2 1 1 1 2 211 2 22 2 21 12 2 22 2 21VI min ,UPP LOW LOWE X E X E X E X E X E X= − − − − − (4.85)
Construction Strategy
Geologic State
Cost(x $1,000)
Exploration Results
E31=1G†
E32=2G†
E21=1G†
E22=2G†
No Exp
lorati
on
Imperfect Exploration
Perfect
Exploration
Exploration Method
1S
ς
11 1S
11 2S
12 2S
22 1S
2S
3S
11ς
12ς
22ς
13ς
23ς
13 2S
23 1S
1111 1890a = −
1211 1935a = −
1112 1125a = −
1212 2520a = −
1122 1125a = −
1222 2520a = −
1132 1125a = −
2121 1890a = −
2221 1935a = −
2231 1935a = −
111 1d C=
112 2d C=
122 2d C=
22 1 1d C=
1 11 1 1GE =
1 21 1 2 GE =1 11 2 1GE =
1 21 2 2 GE =
1 122 1GE =
1 22 2 2 GE =
1 132 1GE =
2 12 1 1GE =
2 22 1 2 GE =
2 231 2 GE =
132 2d C=
231 1d C=
111X112X122X221X132X
231X
161
As for ( )1 1 2 2 12 22 2 21 11LOWE E X E X X+ − , it could be rewritten as
( ) ( ) ( )( )1 1 2 2 1 1 1 2 2 12 22 2 21 11 2 22 2 21 11minLOWE E X E X X E X E X X+ − = + −E E (4.86)
Note that both the relaxed constraint and the strict constraint may be applied because the decision of
performing exploration will not change the uncertainty on the geologic states. Since gambles 122X and
221X are affected by uncertainty on the geologic states and the exploration results may not be completely
reliable, calculating ( )1 1 2 22 22 2 21E X E X+E requires the joint probability measure Q over the exploration
results and the real geologic states (Section 4.3.3), which is constructed by the extreme distributions
XEXTQ listed in Table 4-11. Let ijq be the joint probability of that the exploration result is Gi and that the
geologic state is Gj. Accordingly, ( )1 1 2 22 22 2 21E X E X+E = 11 12 21 22
22 11 22 12 21 21 21 22a q a q a q a q+ + + . To calculate
( )111XE , only the uncertainty on the geologic states is required because it is under the option of no
additional exploration. Let ip be the prior probability of geologic state Gi, and thus
( )111XE = 11 12
11 1 11 2a p a p+ . If the relaxed constraint is applied, ( )1 1 2 2 12 22 2 21 11LOWE E X E X X+ − is equal to
$52,000, which is obtained by solving the following optimization problem:
Minimize
( ) ( )1 1 2 2 12 22 2 21 11E X E X X+ −E E =
( ) ( )11 12 21 22 1 21125 2520 1890 1935 1890 1935q q q q p p− − − − − − − Subject to (relaxed constraint)
8
1,1
8
1, 1,1
;
0; 1
mm EXTm
m mm
λ
λ λ
=
=
=
≥ =
∑
∑
Q Q
12,1 2,2
22
2, 2,1
0.35 0.45;
0.65 0.55
0; 1i ii
pp
λ λ
λ λ=
⎛ ⎞ ⎛ ⎞ ⎛ ⎞= +⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎝ ⎠
≥ =∑
(4.87)
where mEXTQ are given in Table 4-11. Notice that X1 in Eq.(2.37) (i.e.
1 2X XΨ = Ψ ) is here
1 1 2 22 22 2 21E X E X+ , and X2 in Eq. (2.37) is here 1
11X . The equality constraint in Eq. (2.37): 1 2X XΨ = Ψ is
satisfied because if Q is marginalized to the state variable, the marginal is an element of pΨ , the convex
set of prior probabilities, i.e. 1X pΨ = Ψ . On the other hand, 1
2
pp
⎛ ⎞⎜ ⎟⎝ ⎠
is also constrained to be an element of
set pΨ , i.e. 2X pΨ = Ψ . Thus, the relaxed constraint is applied here by enforcing
1 2X X pΨ = Ψ = Ψ .
162
Similarly, ( )1 1 2 2 12 22 2 21 12LOWE E X E X X+ − is $29,500; ( )1 1 1 2 2
11 2 22 2 21LOWE X E X E X− − and
( )1 1 1 2 212 2 22 2 21LOWE X E X E X− − are equal to -$218,900 and -$331,400, respectively. Thus, when the relaxed
constraint is applied, VILOW = min($52,000,$29,500) = $29,500, and VIUPP =
( )min $331,400, $218,900− − − = $331,400.
If the imperfect exploration in Example 4-6 costs less than $29,500, we should perform the
exploration; if we are offered more than $331,400, we are willing to give the imperfect exploration
information away without using it. As for any price between VILOW and VIUPP, the preference between no
exploration and the imperfect exploration is indeterminate.
Alternatively, if the strict constraint is applied, ( )1 1 2 2 12 22 2 21 11LOWE E X E X X+ − is calculated as
$56,500 by solving the following optimization problem:
Minimize
( ) ( )1 1 2 2 12 22 2 21 11E X E X X+ −E E =
( ) ( )11 12 21 22 1 21125 2520 1890 1935 1890 1935q q q q p p− − − − − − − Subject to
1
2
T pp
⎛ ⎞= ⎜ ⎟⎝ ⎠
1 Q (strict constraint)
8
1,1
8
1, 1,1
;
0; 1
mm EXTm
m mm
λ
λ λ
=
=
=
≥ =
∑
∑
Q Q
12,1 2,2
22
2, 2,1
0.35 0.45;
0.65 0.55
0; 1i ii
pp
λ λ
λ λ=
⎛ ⎞ ⎛ ⎞ ⎛ ⎞= +⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎝ ⎠
≥ =∑
(4.88)
where the strict constraint is used by imposing that the marginal of Q is the same as the prior probabilities.
Similarly, ( )1 1 2 2 12 22 2 21 12LOWE E X E X X+ − = $125,300; ELOW ( )1 1 1 2 2
11 2 22 2 21X E X E X− − = -$214,400,
and ( )1 1 1 2 212 2 22 2 21LOWE X E X E X− − = -$233,800. Thus, when the strict constraint is applied, VILOW =
min($56,500, $125,300) = $56,500, and VIUPP = -min(-$214,400, -$233,800) = $233,800.
Compared to the results with the relaxed constraints, we have a higher VILOW and lower VIUPP
because the strict constraint reduces the imprecision.
163
When the relaxed constraint is applied, VPILOW =$240,800, VPIUPP=$461,300, VUPLOW =$48,800,
and VUPUPP =$292,300; when the strict constraint is used, VPILOW =$267,800, VPIUPP =$380,300, VUPLOW
=$129,800, and VUPUPP =$211,300. One may easily check that the inequalities in (4.80) and (4.81) are true
under both the relaxed constraint and the strict constraint, i.e., VPILOW > VUPLOW + VILOW, VPIUPP <
VUPUPP + VIUPP.
Now let us consider a example which may yield a negative lower value of information.
Example 4-9 Consider again the situation in Example 4-5 and Example 4-6. Instead of the original
imprecise reliability matrix shown in Table 4-10, here we consider another exploration with reliability
matrix such that 0.45≤ r11≤ 0.55; 0.45≤ r12≤ 0.55. Hence, 0.45≤ r21≤ 0.55; 0.45≤ r22≤ 0.55. We want to
determine the lower and upper values of information for this exploration.
Now, the extreme points of set rXΨ (i.e., r
CjΨ , 1, 2j = ) are obtained by taking extreme
distributions in each column of the reliability matrix as described in Eq.(4.49): 0.45 0.45 0.55 0.45 0.45 0.55 0.55 0.55
, , ,0.55 0.55 0.45 0.55 0.55 0.45 0.45 0.45
rXEXT
⎧ ⎫⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎪ ⎪= ⎨ ⎬⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎪ ⎪⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎩ ⎭
(4.89)
The prior probabilities are still assigned as 0.35 ≤ P(1G) ≤ 0.45, and thus the extreme points
for ΨP is 1EXTp = (0.35, 0.65), and 2
EXTp = (0.45, 0.55). Construction strategies need to be selected between
C1 and C2 conditional on the exploration results.
The first step is to determine the optimal construction strategies given exploration results.
Decision tree is the same as depicted in Figure 4-15. By applying Theorem 4-7, the set of extreme joint
probabilities ( XEXT ) is obtained, as shown in Table 4-16. Finally, Eq.(4.65) gives the extreme
probabilities conditional to the exploration results, listed in Table 4-17.
At decision node 11ς (i.e., given exploration result †1G ), when the relaxed constraint is applied,
( )1 111 12LOWE X X− and ( )1 1
12 11LOWE X X− are equal to -$98,700 and -$180,900, respectively. Since both
are negative, then gambles 112X and 1
11X are optimal under the relaxed constraint. Alternatively, if the
164
strict constraint is applied, ( )1 111 12LOWE X X− and ( )1 1
12 11LOWE X X− are equal to -$90,000 and -
$172,100, respectively, which are still both negative. Thus, gambles 112X and 1
11X are optimal under the
strict constraint as well.
Likewise for the case at node 21ς (i.e., given exploration result †2G ). ( )2 2
11 12LOWE X X− and
( )2 212 11LOWE X X− are equal to-$98,700 and -$180,900 with the relaxed constraint, and they are equal to -
$90,000 and -$172,100 with the strict constraint, respectively. As a result, gambles 211X and 2
12X are
optimal in both cases.
Therefore, with the prior probability 0.35 ≤ P(1G) ≤ 0.45 and the imprecise reliability of the
exploration, given the exploration result, the preference between construction strategy C1 and C2 are still
indeterminate.
Table 4-16: Set of extreme joint probabilities XEXT .
,XEXT mQ Extreme point
number, m 1,1
,XEXT mq 1,2,XEXT mq 2,1
,XEXT mq 2,2,XEXT mq
1 0.158 0.358 0.193 0.293 2 0.158 0.293 0.193 0.358 3 0.193 0.358 0.158 0.293 4 0.193 0.293 0.158 0.358 5 0.203 0.303 0.248 0.248 6 0.203 0.248 0.248 0.303 7 0.248 0.303 0.203 0.248 8 0.248 0.248 0.203 0.303
165
Table 4-17: Extreme probabilities conditional to the exploration results.
†|1m
CjEXT G
P †|2m
CjEXT G
P Extreme point number, m
( )†|11
m
CjEXT G
P G ( )†|12
m
CjEXT G
P G ( )†|21
m
CjEXT G
P G ( )†|22
m
CjEXT G
P G
1 0.306 0.694 0.397 0.603 2 0.350 0.650 0.350 0.650 3 0.350 0.650 0.350 0.650 4 0.397 0.603 0.306 0.694 5 0.401 0.599 0.500 0.500 6 0.450 0.550 0.450 0.550 7 0.450 0.550 0.450 0.550 8 0.500 0.500 0.401 0.599
Figure 4-21: Decision tree for the tunnel exploration.
111X112X121X122X221X222X132X
231X
Construction Strategy
Geologic State
Cost(x $1,000)
Exploration Results
E31=1G†
E32=2G†
E21=1G†
E22=2G†
No
Expl
orati
on
Imperfect Exploration
Perfect
Exploration
Exploration Method
2231 1935a = −
2231 1125a = −
1211 1935a = −
1111 1890a = −
1212 2520a = −
1112 1125a = −
ς
S1
S2
S3
11ς
12ς
22ς
13ς
23ς
111S
112S
121S
122S
221S
222S
132S
231S
1221 1935a = −
1121 1890a = −
1222 2520a = −
1122 1125a = −
2221 1935a = −
2121 1890a = −
2222 2520a = −
2122 1125a = −
111 1d C=
112 2d C=
121 1d C=
122 2d C=
221 1d C=
222 2d C=132 2d C=
231 1d C=
1G
2G1G
2G
1G
2G1G
2G1G
2G1G
2G1G
2G
166
The next step is to determine the lower and upper values of imperfect information (VI). Decision
tree is shown in Figure 4-21. Eq.(4.82) reads as follows:
( )( )1 1 2 2 12 2 2 2 1, ,
VI minLOW LOW i j ki j kE E X E X X= + −
( )( )1 1 1 2 21 2 2 2 2, ,
VI minUPP LOW k i ji j kE X E X E X= − − − (4.90)
By alternating on the extreme distributions XEXTQ listed in Table 4-16, if the relaxed constraint
is applied, VILOW = –$139,500, and VIUPP = $161,600; if the strict constraint is applied,
VI $112,500LOW = − , and VIUPP = $112,500.
As explained on Page 159, given the new evidence from imprecise new information, preference
among gambles is probably indeterminate. The indeterminacy among multiple choices may lead to a
negative lower value of information. Here the negative lower values of information are obtained.
Recall that the value of information can never be negative if the inputs are precise probabilities.
Someone may argue that the value of information should never be negative even the inputs are imprecise,
and the idea is to solve the problem with a unique prior probability and a precise reliability matrix at a time,
which are picked from Ψp and Ψr, respectively. Here we illustrate this procedure by this example.
By marginalizing XEXTQ (Table 4-16) on exploration results and real geological states, the
corresponding marginals are obtained as listed in Table 4-18. Then the preference between C1 and C2 can
be determined within each precise inputs. For example, for the prior information, if P(1G) = 0.35, then
E( 111X ) = -$1,919,250, and E( 2
11X ) = -$2,031,750, and thus C1 is preferred to C2, i.e. the expected value
with no additional exploration is -$1,919,250.
When new information is provided, take the first extreme points in Table 4-17 as an example:
given that the exploration result is †1G , P(1G) = 0.306, P(2G) = 0.694, then the expected value for C1 is
( ) 18900.306 0.694
1935−⎛ ⎞⎜ ⎟−⎝ ⎠
= -$1,921,240, and the expected value for C2 is ( ) 11250.306 0.694
2520−⎛ ⎞⎜ ⎟−⎝ ⎠
= -
$2,093,370, thus, C1 is preferred to C2; given that the exploration result is †2G , P(1G) = 0.397, P(2G) =
0.603, then the expected value for C1 is ( ) 18900.397 0.603
1935−⎛ ⎞⎜ ⎟−⎝ ⎠
= -$1,917,140, and the expected value
167
for C2 is ( ) 11250.306 0.694
2520−⎛ ⎞⎜ ⎟−⎝ ⎠
= -$1,966,310, thus, C1 is preferred to C2. Since P(1G†) = 0.525 and
P(2G†) = 0.485 in Table 4-18, the expected value with exploration results is ( ) 19212400.525 0.485
1917140−⎛ ⎞⎜ ⎟−⎝ ⎠
=
-$1,919,250, which is the same as the expected value with no additional exploration (-$1,919,250). Thus,
the value of information is 0. Similarly for all other extreme points in Table 4-17 and Table 4-18. The
unique optimal construction strategies for each extreme reliability matrix and corresponding values of
information are listed in Table 4-19.
It should be noticed that this method actually deal with the problem with multiple precise
probabilities, thus problems are always determinate. Therefore, the value of information will never be
negative. In the previous results obtained by the algorithm developed in this study, the negative lower
values of information is the results of indeterminacy, i.e., the indeterminacy between construction strategy
is taken into account when calculating the value of information.
Table 4-18: Extreme marginals on exploration results and real geological states
( )1P G ( )2P G ( )†1P G ( )†2P G
1 0.35 0.65 0.515 0.485 2 0.35 0.65 0.450 0.550 3 0.35 0.65 0.550 0.450 4 0.35 0.65 0.485 0.515 5 0.45 0.55 0.505 0.495 6 0.45 0.55 0.450 0.550 7 0.45 0.55 0.550 0.450 8 0.45 0.55 0.495 0.505
168
Table 4-19: Unique optimal construction strategies and value of information
Exploration Result †1G †2G
VI
1 C1 C1 0
2 C1 C1 0
3 C1 C1 0
4 C1 C1 0
5 C1 C2 $22,050
6 C2 C2 0
7 C2 C2 0
8 C2 C1 $22,050
169
Chapter 5 Case Histories
This chapter revisits several case histories of risk analysis in tunneling by using
the methodologies developed in previous chapters. Section 5.1 applies event-tree analysis
with imprecise probabilities to obtain the bounds on the occurrence probability of
accidents during the construction of an underwater tunnel. Section 5.2 deals with the
probability of the environmental damage caused by the construction activities of the
Stockholm Ring Road Project. Section 5.3 revisits the Sucheon Tunnel by introducing the
imprecision of probabilities, and finally the optimal exploration plan is determined.
Section 5.4 introduces the application to the risk register of the East Side CSO Project in
Portland, Oregon. All results obtained based on imprecise probabilities are compared
with the results from precise probabilities.
5.1 ETA APPLIED TO THE DESIGN OF A UNDERWATER TUNNEL
The event-tree analysis of Section 4.2.1 was applied to analyze the potential risks
at the design stage of a 1.27 km long and 8.1-meter diameter TBM tunnel under crossing
the Han River in South Korea. Because the tunnel started and ended in the downtown
area (see Figure 5-1), the major concerns were the potential risks to neighborhoods and
local business, existing structures and facilities. A risk analysis was conducted to quantify
the occurrence probability of accidents (Hong et al. 2009). In this section, all precise
probability inputs and the project background information are from Hong et al. (2009).
170
Figure 5-1 Construction site plan (Hong et al. 2009).
Three important initiating events: poor ground conditions, high water pressure,
and heavy rainfall were identified after an extensive analysis of the available empirical
data. Without any mitigation measures, the three initiating events would lead to an
accident and cause an impact on schedule and cost, even tunnel failure. To avoid or
mitigate the impact, safety measures are proposed, and classified into five categories: A.
investigation/design; B. Process planning; C. Machine type; D. Construction
management; E. Reinforcement. Under each initiating event, the success probabilities of
safety measures A through E are obtained by averaging the probability evaluations from
four experts. The precise inputs used by Hong et al. (2009) are summarized in Table 5-1.
Table 5-1 Success probabilities of safety measures (Hong et al. 2009)
Safety measures Poor ground conditions High water pressure Heavy rainfall
A. investigation/design 0.02 0.15 0.40 B. Process planning 0.13 0.30 0.19 C. Machine type 0.65 0.68 0.73 D. Construction management 0.63 0.65 0.63
E. Reinforcement 0.38 0.56 0.28
171
Figure 5-2 Event tree for initiating event of poor ground conditions (Hong et al. 2009)
172
The event tree with the initiating event ‘Poor ground conditions’ is depicted in
Figure 5-2. The occurrence of accidents and their consequences are identified at the end
of each probability path. The consequences are evaluated at five levels: catastrophic,
critical, serious, marginal, and negligible. Hong et al. (2009) offered the analysis results
for the three initiating events. As shown in Table 5-2, the occurrence probabilities of
accidents are 0.59, 0.56, and 0.46, respectively. Based on the consequence, accidents are
classified into three risk levels: I, II, and II, where catastrophic and critical accidents are
grouped in level I, serious and marginal accidents belong to level II, and negligible
accidents are grouped in level III. For risk level I, Hong et al. (2009) suggest to apply
significant mitigation measures to reduce the risk level or to remove the causes of risk
reasons; for risk level II, proactive mitigation measures need to be considered; risks at
level III should be taken care of in the construction management. The objective of the
risk management is to ensure that the occurrence probability of risks at level I is smaller
than 5%. The probabilities of accidents at different levels are summarized in Table 5-3,
and the probabilities that the initiating events cause a level I Risk are 0.25, 0.11, and 0.09,
respectively. Among the three initiating events, event ‘Poor ground conditions’ is far
from the objective of 5% probability. Thus, to reduce the potential of risk I, mitigation
measures should be implemented carefully and their effects on reducing the occurrence
probability or the consequence of accidents should be monitored regularly.
173
Table 5-2 Probabilities of criticality and occurrence of accident (Hong et al. 2009)
Consequence Poor ground conditions High water pressure Heavy rainfall
Catastrophic 0.08 0.05 - Critical 0.17 0.06 0.09 Serious 0.34 0.45 0.36 Marginal 0.25 0.19 0.42 Negligible 0.16 0.25 0.13 Accident 0.59 0.56 0.46
Table 5-3 Probabilities of accident at different risk levels (Hong et al. 2009)
Risk level Poor ground conditions High water pressure Heavy rainfall
I 0.25 0.11 0.09 II 0.59 0.64 0.78 III 0.16 0.25 0.13
The author redid the analyses by applying the event tree in Figure 5-2 to all three
initiating events and by using the precise probabilities in Table 5-1; however, different
results were obtained in several occasions, and are highlighted in bold in Table 5-5 and
Table 5-6. It is probably because the event tree for ‘poor ground conditions’ is different
from the cases of ‘high water pressure’ and ‘heavy rainfall’, which are not provided in
Hong et al. (2009). The author contacted the corresponding author in the reference Hong
et al. (2009). Unfortunately, the corresponding author did not have time to confirm the
event-trees for the other two cases. Therefore, in the following part of this section, we
only compare our calculated results from precise probabilities to the results from
imprecise probabilities.
174
Since the precise probabilities are obtained by averaging the evaluations of four
experts, it would be better to admit the imprecision in the available data and use
imprecise probabilities. As shown in Table 5-4, the predefined success probabilities of
each safety measure are bounded by intervals, where the interval width are all equal to
0.1 and the precise probabilities in Table 5-1 are contained in these intervals. Results
obtained from the intervals in Table 5-4 are summarized in Table 5-5 and Table 5-6. It is
easy for one to check that the results from precise probabilities are also bounded by the
results from the imprecise probabilities.
Results from imprecise probabilities show that the case of ‘poor ground
conditions’ is more risky than other two cases. To ensure that the occurrence probability
of risk I is less than 5%, the upper probabilities of risk I must be less than 5%. It is a
stricter requirement than that in precise probabilities, where the probability of risks at
level I is represented by a single value. Therefore, by admitting the imprecision in
probability evaluation, more attention should be paid to risk management to ensure that
the upper probability be in the acceptable range.
Table 5-4 Success probabilities of safety measures in imprecise probabilities
Safety measure Poor ground conditions
High water pressure
Heavy rainfall
A. Investigation/design [0.00, 0.10] [0.10, 0.20] [0.35, 0.45] B. Process planning [0.10, 0.20] [0.25, 0.35] [0.10, 0.30] C. Machine type [0.60, 0.70] [0.60, 0.70] [0.70, 0.80] D. Construction management [0.60, 0.70] [0.60, 0.70] [0.60, 0.70]
E. Reinforcement [0.30, 0.40] [0.50, 0.60] [0.20, 0.30]
175
Table 5-5 Probabilities of criticality and occurrence of accident
Poor ground conditions High water pressure Heavy rainfall Consequence
Imprecise Precise Imprecise Precise Imprecise Precise
Catastrophic [0.0540, 0.1120] 0.08 [0.0360, 0.0800] 0.05 [0.0420, 0.0960] 0.07Critical [0.1112, 0.2080] 0.17 [0.0887, 0.1662] 0.12 [0.0507, 0.1187] 0.09Serious [0.2658, 0.3936] 0.34 [0.3293, 0.4528] 0.39 [0.2933, 0.4266] 0.38Marginal [0.2160, 0.3430] 0.25 [0.1440, 0.2450] 0.19 [0.2940, 0.4480] 0.33Negligible [0.1080, 0.1960] 0.16 [0.1800, 0.2940] 0.25 [0.0840, 0.1680] 0.13Accident [0.5100, 0.6400] 0.59 [0.5100, 0.6400] 0.56 [0.4400, 0.5800] 0.54
Table 5-6 Probabilities of accident at different risk levels
Poor ground conditions High water pressure Heavy rainfall Risk level
Imprecise Precise Imprecise Precise Imprecise Precise
I [0.1807, 0.2944] 0.25 [0.1337, 0.2302] 0.17 [0.1031, 0.1949] 0.16II [0.5208, 0.6819] 0.59 [0.5063, 0.6496] 0.58 [0.6488, 0.7955] 0.71III [0.1080, 0.1960] 0.16 [0.1800, 0.2940] 0.25 [0.0840, 0.1680] 0.13
5.2 FTA APPLIED TO THE STOCKHOLM RING ROAD TUNNELS
The Stockholm Ring Road project is a vast underground construction project, and
will provide a ring road around Stockholm to improve public transportation. The majority
of the alignment is in hard rock, but several sections are in soft ground (Sturk et al. 1996).
Figure 5-3 shows the project plan in 1992 based on “Dennis Agreement”, which
was a political agreement on transit and highway improvements in and around
Stockholm. However, the Stockholm Ring Road project has been opposed by green
parties and local residents mainly because of environmental concerns since the permitting
176
stage, which led to several project halts (Tollroads Newsletter, 1997). As of 2007, about
half of the ring road was built. After a new environmental evaluation was completed, the
project was resumed and is expected to be ready for opening in 2020 (Stockholm News,
2009).
Figure 5-3 Stockholm Ring Road project plan in 1992 (Stockholm ring road, from http://en.wikipedia.org/wiki/Stockholm_ring_road)
177
Because of the high environmental concerns, a ‘Review Team’ composed of
experts in geotechnical engineering was set up at the beginning of the project in 1996 to
provide and ensure high quality technical solutions, where a risk analysis was conducted
by using fault-trees to evaluate the environmental damage due to tunneling (Sturk et al.
1996). Figure 5-4 through Figure 5-7 show the fault-trees, where the top event is ‘the
lime trees are damaged due to the tunneling activities’. It should be noted that all events
in Sturk et al. (1996) are assumed to be independent to each other. Interaction noted in
Figure 5-4 through Figure 5-7, such as ‘unknown interaction’ etc, is applied only when
imprecise probabilities is considered later in this section. For the events at the bottom of
the fault-trees, Sturk et al. (1996) used precise probabilities, which are summarized in the
columns under ‘precise’ in Table 5-7. Finally, the occurrence probability for the top event
is equal to 0.105, which Sturk et al. (1996) thought acceptable. However, the current
status of the project tells us that it is not a good estimation. The probability might be
higher than 0.105 and thus it is not acceptable.
178
Table 5-7 Occurrence probabilities of events at the bottom of fault-trees
Event Precise Imprecise Event Precise Imprecise E02 0.50 [0.45, 0.55] E36 0.50 [0.45, 0.55] E08 0.05 [0.01, 0.10] E37 0.90 [0.85, 0.95] E10 0.05 [0.01, 0.10] E39 0.50 [0.45, 0.55] E12 0.50 [0.45, 0.55] E40 0.25 [0.20, 0.30] E13 0.25 [0.20, 0.30] E41 0.05 [0.01, 0.10] E14 0.25 [0.20, 0.30] E42 0.01 [0.01, 0.10] E15 0.05 [0.01, 0.10] E43 0.10 [0.05, 0.15] E16 0.10 [0.05, 0.15] E44 0.10 [0.05, 0.15] E18 0.25 [0.20, 0.30] E45 0.10 [0.05, 0.15] E23 0.25 [0.20, 0.30] E52 0.10 [0.05, 0.15] E24 0.01 [0.01, 0.10] E53 0.05 [0.01, 0.10] E25 0.10 [0.05, 0.15] E54 0.10 [0.05, 0.15] E26 0.417 [0.35, 0.45] E55 0.25 [0.20, 0.30] E27 0.25 [0.20, 0.30] E56 0.25 [0.20, 0.30] E28 0.207 [0.25, 0.35] E57 0.25 [0.20, 0.30] E29 0.25 [0.20, 0.30] E58 0.50 [0.45, 0.55] E33 0.01 [0.01, 0.10] E59 0.10 [0.05, 0.15]
179
Figure 5-4 Fault tree for damage to lime trees due to tunneling activities, adapted from Sturk et al. (1996)
The lime trees are damaged
Trees exposed to damage Measures
to reduce damage fail
Damage due to carelessness
Trees dried up
RootsDamaged due to
vibrations
AND
OR
Trees suffocating, no oxygen to roots
Contractor not interested in saving trees
Lack of control
Deficient information
concerning the problem
No financial incentives to save
trees
AND
OR
AND
Soil excessively compacted
No remediation
measure
AND
Other roots cannot
supply tree with enough water
Branch A Branch B
Branch C
Independence
0.5[0.45, 0.55]
0.05[0.01, 0.10]
0.50[0.45, 0.55]
0.25[0.20, 0.30]
0.25[0.20, 0.30]
Independence
Independence
[ ]0.5,0.8ρ∈ Unknown Interaction
Unknown Interaction
0.05[0.01, 0.10]
0.05[0.01, 0.10]
0.10[0.05, 0.15]
ET
E01 E02
E03 E04 E05 E06 E07
E08
E09 E10 E11
E12 E13 E14
E16E15
E28
180
Figure 5-5 Fault tree for Branch A, adapted from Sturk et al. (1996)
Branch A
Roots damaged mechanically
RiskReducing measures
fail
AND
OR
Minor roots damaged due to
settlements
Deficient information
concerning the problem
Tunnelling causes settlements that damage
minor roots
AND
AND
Hazardous settlement occur
Early monitoring and remediation of
settlements fail
Spilling/forepoling in
contact with roots
No survey of extension of
roots
Deep rootsexist
AND
Main roots damaged due to
settlements
AND
Tunnelling causes settlements that damage
main roots
Early monitoring and remediation of
settlements fail
Trees exposed to root damageE04
Independence
Independence
Independence
Independence
Unknown Interaction
Unknown Interaction
0.25[0.20, 0.30]
0.25[0.20, 0.30]
0.207[0.15, 0.25]
0.25[0.20, 0.30]
0.417[0.35, 0.45]
0.25[0.20, 0.30]
0.01[0.01, 0.10]
0.10[0.05, 0.15]
E17 E18
E19 E20 E21
E22 E23 E24 E25
E26 E27 E28 E29
181
Figure 5-6 Fault tree for Branch B, adapted from Sturk et al. (1996)
Branch B
Trees damaged by chemicals in the ground
Spillage on the ground
Petroleum products from
tunneling
AND
OR
Petroleum products reach
the roots
Hazardous products are
used
Petroleum products transported
to roots
Spillage in tunnel
AND
Trees damaged by Petroleum
products
Trees damaged by chemicals from grout
AND
Chemicals in grout soluble in
water
Chemicals reach roots
Chemicals affect trees
Trees damaged by direct contact
with grout
Grout affects the tree/rocks
AND
Grout reaches the roots
AND
Grouting procedure not
adapted to situation
Quality con-trol insufficient
(grouting)
OR
E05
E30 E31 E32
E33 E34 E35 E36
E37 E38
E41 E42
E39 E40
E43 E44 E45
Unknown Interaction
Independence
Independence
IndependenceIndependence
Independence
0.01[0.01, 0.10]
0.90[0.85, 0.95]
0.05[0.01, 0.10]
0.01[0.01, 0.10]
0.10[0.05, 0.15]
0.10[0.05, 0.15]
0.01[0.01, 0.10]
0.50[0.45, 0.55]
0.25[0.20, 0.30]
0.50[0.45, 0.55]
[ ]0.5,0.8ρ∈
182
Figure 5-7 Fault tree for Branch C, adapted from Sturk et al. (1996)
As stated by Sturk et al. (1996), “the probabilities were assessed subjectively,
based on expert knowledge and experience”, which is a major reason to use imprecise
probability instead of precise probabilities as explained in Section 4.2.2. The two
columns ‘imprecise’ heading in Table 5-7 list all imprecise probabilities which evaluate
the uncertainty of the events at the bottom of the fault-trees, where the interval widths are
equal to 0.1. As shown in Figure 5-4 through Figure 5-7, the interaction between events is
assumed to be ‘unknown interaction’, ‘independence’, or ‘uncertain correlation’ with the
correlation coefficient [ ]0.5,0.8ρ ∈ . Table 5-8 summarizes all calculated occurrence
[ ]0.5,0.8ρ∈ [ ]0.5,0.8ρ∈
183
probabilities obtained from both the precise and the imprecise inputs. The lower and the
upper probabilities for the top event are equal to 0.0189 and 0.3116, respectively.
Compare the two types of input: precise and imprecise. The only differences are
(1) that the former is precise and the latter is given in terms of intervals, and (2) relaxing
the constraint of independence and assuming different types of interaction. Finally, we
find that the probability of the top event (i.e., the lime trees are damaged by tunneling
activities) can be as high as 0.3116, which is much higher than the original estimation
(0.105) and might not be acceptable anymore. As a result, further proactive and effective
solutions should be considered to deal with the environmental concerns.
Table 5-8 Calculated occurrence probabilities of events
Event Precise Imprecise Event Precise Imprecise
E01 2.09E-01 [4.20E-02, 5.67E-01] E22 1.04E-01 [7.00E-02, 1.35E-01]
E03 7.25E-03 [1.56E-02, 1.00E-01] E30 1.00E-02 [1.42E-02, 1.70E-01]
E04 1.83E-02 [2.80E-03, 9.17E-02] E31 1.25E-02 [4.50E-03, 2.48E-02]
E05 2.29E-02 [1.42E-02, 2.03E-01] E32 5.00E-04 [2.25E-04, 8.25E-03]
E06 6.80E-02 [0, 2.51E-01] E34 4.50E-04 [4.25E-03, 7.79E-02]
E07 6.25E-02 [0, 3.00E-01] E35 1.00E-03 [5.00E-04, 1.50E-02]
E09 1.45E-01 [5.95E-03, 2.35E-01] E38 5.00E-04 [5.00E-03, 8.20E-02]
E11 1.36E-01 [7.19E-02, 2.61E-01] E46 5.00E-03 [5.00E-04, 1.50E-02]
E17 7.73E-02 [1.40E-02, 3.06E-01] E47 1.32E-01 [7.19E-02, 2.46E-01]
E19 2.61E-02 [1.40E-02, 4.05E-02] E48 2.50E-02 [5.88E-02, 1.50E-01]
E20 1.00E-03 [5.00E-04, 1.50E-02] E49 1.09E-01 [1.40E-02, 1.13E-01]
E21 5.18E-02 [0, 2.50E-01] E50 6.25E-02 [1.20E-01, 2.58E-01]
184
5.3 DECISION ANALYSIS: THE OPTIMAL EXPLORATION PLAN FOR THE SUCHEON TUNNEL
The Sucheon tunnel is a 2 km long rock tunnel, buried in micrographic granite
and diorite (see Figure 5-8). Based on the investigation results: RMR, resistivity, and Q
values, the geologic states were classified into five categories, as shown in Table 5-9.
Five construction strategies C1 through C5 were proposed (see Table 5-10). Costs for
each construction strategy under different geologic states are listed in Table 5-11, where
the bolded ones are the minimum cost per unit length under different geologic states.
According to the prior information of geologic states, the tunnel is divided into 18
sections (see Table 5-12). Table 5-13 shows the reliability matrix of an imperfect
additional exploration. The selection of the construction strategy with or without the
imperfect additional exploration must be made among C1 through C5 for each tunnel
section. Moreover, the value of information and the optimal exploration plan need to be
determined.
Figure 5-8 Geological profile and layout of the Sucheon Tunnel (Min et al. 2003)
185
Table 5-9 Description of Geologic States (Karam et al. 2007)
Description Geologic state RMR Resistivity (Ωm) Q value
1G > 81 > 3000 > 40
2G 60 - 80 1000 - 3000 4 - 40
3G 40 - 60 300 - 1000 1 – 4
4G 20 - 40 100 - 300 0.1 - 1
5G < 20 < 100 < 0.1
Table 5-10 Description of Construction Strategies (Karam et al. 2007)
Construction strategy Description
C1 Full face excavation with nominal support
C2 Full face excavation with extensive support
C3 Heading and bench excavation with nominal support
C4 Heading and bench excavation with extensive support
C5 Multi-heading and bench excavation
186
Table 5-11 Construction Cost (per meter) (Karam et al. 2007)
Geological state Construction strategy 1G 2G 3G 4G 5G
C1 $3,150 $4,500 $5,700 $7,200 $8,850 C2 $4,125 $3,525 $5,325 $7,350 $9,000 C3 $4,350 $4,875 $4,650 $6,375 $8,138 C4 $4,650 $4,500 $5,400 $6,285 $7,875 C5 $4,425 $4,575 $5,175 $6,450 $6,600
Table 5-12 Tunnel Section and Precise Prior Probabilities (Karam et al. 2007)
Geologic state Section Length (m)
1G 2G 3G 4G 5G 1 60 0.51 0.49 0.00 0.00 0.00 2 20 0.15 0.47 0.38 0.00 0.00 3 57 0.15 0.47 0.38 0.00 0.00 4 40 0.00 0.00 0.47 0.53 0.00 5 120 0.00 0.29 0.64 0.07 0.00 6 41 0.15 0.25 0.53 0.07 0.00 7 96 0.14 0.25 0.56 0.06 0.00 8 104 0.47 0.43 0.10 0.00 0.00 9 586 0.48 0.45 0.07 0.00 0.00
10 106 0.49 0.51 0.00 0.00 0.00 11 67 0.51 0.49 0.00 0.00 0.00 12 167 0.50 0.50 0.00 0.00 0.00 13 36 0.78 0.22 0.00 0.00 0.00 14 74 0.10 0.83 0.07 0.00 0.00 15 43 0.19 0.81 0.00 0.00 0.00 16 110 0.67 0.33 0.00 0.00 0.00 17 188 0.86 0.14 0.00 0.00 0.00 18 50 0.13 0.32 0.54 0.00 0.00
187
Table 5-13 Exploration Reliability Matrix (Karam et al. 2007)
Reality Exploration results 1G 2G 3G 4G 5G
1G 0.90 0.10 0.00 0.00 0.00 2G 0.10 0.90 0.10 0.00 0.00 3G 0.00 0.00 0.80 0.10 0.00 4G 0.00 0.00 0.10 0.80 0.10 5G 0.00 0.00 0.00 0.10 0.90
In Table 5-12 and Table 5-13, probabilities were assigned as precise values. With
the precise inputs, Karam et al. (2007) decided the optimal construction strategies and
exploration plan by using decision trees.
Now we are going to redo the decision analysis with imprecise prior probabilities
and imprecise reliability matrix, as shown in Table 5-14 and Table 5-15. The extreme
point form of Table 5-14 and Table 5-15 are given in the Appendix B (shown in Table
B-1 and Table B-2). To avoid division-by-zero errors in the analysis, all zero
probabilities in Table 5-12 through Table 5-15 are replaced by a very small number: 10-9.
The results obtained from both imprecise and precise probabilities will be compared later
in this section.
For Section i (i =1,…,18), we first use the decision trees in Figure 5-9 and Figure
5-10 to determine the optimal construction strategies in the cases of no additional
exploration and imperfect additional exploration, respectively. Both cases consider the
relaxed constraints and the strict constraints, respectively. Since the precise inputs (Table
5-12 and Table 5-13) are contained in the imprecise inputs (Table 5-14 and Table 5-15),
188
the results obtained with the precise inputs should also be contained in those with
imprecise inputs. When imprecise inputs are used, the sets of probability measures are the
same regardless the type of constraints (relaxed or strict). However, since the strict
constraints can reduce uncertainty, the results generated under the strict constraints are
included in those obtained by using the relaxed constraints. Accordingly, all results with
precise or imprecise inputs can be presented in one table (i.e. Table 5-16). The results
under the strict constraints are bolded in Table 5-16. The underlined construction
strategies are obtained with precise inputs. For instance, in Section 1, given the
exploration result is 4G, if the relaxed constraints are imposed, the optimal construction
strategies are C1 through C5; under the strict constraints, the optimal ones are C1 and C5;
if the precise inputs are used, the unique optimal construction strategy is C5.
For the case of perfect additional exploration, the optimal construction strategies
are the one with minimum cost under different geological states, as shown in Table 5-11.
Accordingly, for the case of perfect additional exploration, if the exploration result is iG,
the optimal construction strategy is Ci (i = 1, …, 5).
Table 5-14 Imprecise Prior Probabilities
Geologic state Section Length (m)
1G 2G 3G 4G 5G 1 60 [0.45, 0.55] [0.45, 0.55] 0.00 0.00 0.00 2 20 [0.10, 0.20] [0.40, 0.50] [0.30, 0.40] 0.00 0.00 3 57 [0.10, 0.20] [0.45, 0.50] [0.35, 0.40] 0.00 0.00 4 40 0.00 0.00 [0.45, 0.55] [0.45, 0.55] 0.00 5 120 0.00 [0.20, 0.30] [0.60, 0.70] [0.05, 0.15] 0.00 6 41 [0.10, 0.20] [0.20, 0.30] [0.50, 0.60] [0.00, 0.10] 0.00 7 96 [0.10, 0.20] [0.20, 0.30] [0.50, 0.60] [0.00, 0.10] 0.00 8 104 [0.40, 0.50] [0.40, 0.50] [0.05, 0.15] 0.00 0.00
189
Geologic state Section Length (m)
1G 2G 3G 4G 5G 9 586 [0.40, 0.50] [0.40, 0.50] [0.05, 0.15] 0.00 0.00
10 106 [0.48, 0.52] [0.48, 0.52] 0.00 0.00 0.00 11 67 [0.48, 0.52] [0.48, 0.52] 0.00 0.00 0.00 12 167 [0.48, 0.52] [0.48, 0.52] 0.00 0.00 0.00 13 36 [0.75, 0.80] [0.20, 0.25] 0.00 0.00 0.00 14 74 [0.05, 0.15] [0.80, 0.85] [0.05, 0.10] 0.00 0.00 15 43 [0.15, 0.20] [0.80, 0.85] 0.00 0.00 0.00 16 110 [0.65, 0.70] [0.30, 0.35] 0.00 0.00 0.00 17 188 [0.85, 0.90] [0.10, 0.15] 0.00 0.00 0.00 18 50 [0.10, 0.15] [0.30, 0.35] [0.50, 0.55] 0.00 0.00
Table 5-15 Imprecise Exploration Reliability Matrix
Reality Exploration results 1G 2G 3G 4G 5G
1G [0.85, 0.95] [0.08, 0.12] 0.00 0.00 0.00 2G [0.05, 0.15] [0.88, 0.92] [0.08, 0.12] 0.00 0.00 3G 0.00 0.00 [0.70, 0.85] [0.05, 0.15] 0.00 4G 0.00 0.00 [0.05, 0.15] [0.75, 0.90] [0.06, 0.12] 5G 0.00 0.00 0.00 [0.08, 0.15] [0.88, 0.94]
190
Figure 5-9 Decision tree for Section i of the Sucheon Tunnel without additional exploration
191
Figure 5-10 Decision tree for Section i of the Sucheon Tunnel with imperfect additional exploration
192
Table 5-16 Optimal Construction Strategies Obtained
With Exploration Result Section No exploration
†1G †2G †3G †4G †5G 1 C1, C2 C1 C2 C1, C2, C3 C1, C2, C3, C4, C5 C5 2 C2 C1 C2 C3 C3 C5 3 C2 C1 C2 C3 C3 C5 4 C3 C1,C3 C3 C3 C3, C4, C5 C4 5 C2, C3 C2 C2 C3 C2, C3, C4, C5 C4 6 C1, C2, C3, C5 C1 C2 C3 C1, C2, C3, C4, C5 C3, C4, C5
7 C1, C2, C3, C5 C1 C2 C3 C1, C2, C3, C4, C5 C3, C4, C5 8 C1, C2 C1 C2 C3 C3 C5 9 C1, C2 C1 C2 C3 C3 C5
10 C1, C2 C1 C2 C1, C2, C3 C1, C2, C3, C4, C5 C5 11 C1, C2 C1 C2 C1, C2, C3 C1, C2, C3, C4, C5 C5 12 C1, C2 C1 C2 C1, C2, C3 C1, C2, C3, C4, C5 C5 13 C1 C1 C2 C1, C2, C3 C1, C3, C4,C5 C5 14 C2 C1, C2 C2 C3 C3 C5 15 C2 C1 C2 C2 C2, C3, C4, C5 C5 16 C1 C1 C2 C1, C2, C3 C1, C3, C4, C5 C5 17 C1 C1 C1, C2 C1, C3 C1, C3, C5 C5 18 C2 C1 C2 C3 C3 C5
As shown in Table 5-16, the optimal construction strategies obtained with
imprecise probabilities are not necessarily unique, while the construction strategies under
precise probabilities are always unique, except the case of tunnel section 12 with no
additional exploration, where C1 and C2 are both optimal because C1 and C2 give
exactly the same expected cost.
As explained in Section 4.3.3.3.1, the uncertain new information may decrease or
increase the uncertainties. This statement is verified in Table 5-16. By comparing the
results under no additional exploration and imperfect additional exploration, we cannot
193
draw the conclusion that the new evidence provided by imperfect exploration always
helps reducing the range of optimal construction strategies. Take section 13 as an
example. There is a unique optimal strategy under no exploration and when the
exploration result is 1G, 2G, or 5G, while multiple optimal strategies occurs when the
exploration results is 3G or 4G. Only later calculations determine whether the new
evidence is useful in reducing the range of optimal choices
Uncertainty under the strict constraints is higher than that under the relaxed
constraints. As shown in Table 5-16, the results obtained by using the strict constraints
are always included in the results from the relaxed constraints. In some sections, the
relaxed constraints generate multiple selections while the strict constraints give a unique
selection, like the case of section 1 given that the exploration result is 3G, where the
optimal strategies are C1, C2 and C3 under the relaxed constraints and it becomes C2
when the strict constraints are imposed. It is because the strict constraints may reduce
uncertainty, as described in Case (2) in Section 4.3.3.3. However, the strict constraints
sometimes give multiple choices the same as the relaxed constraints. Such as the case in
section 6 given that the exploration result is 5G. Both the relaxed constraints and the
strict constraints give multiple optimal strategies:C3, C4 and C5. It is because the utility
curves for the three strategies intersect in the set of probability measures conditional to
the exploration result, which has been explained in Case (1) in Section 4.3.3.3.2 (Figure
4-18). The third situation is the combination of the previous two situations: both the
relaxed and the strict constraints give multiple choices, but the number of choices under
strict constraints is less than that under the relaxed constraints. This can be explained as a
combination of Case (1) and (2) in Section 4.3.3.3.
Sections 2, 3 and 18 always have a unique optimal construction strategy,
indicating that the imprecision of probability measures on state variable does not drive
194
any problem to be indeterminate. Similar as the case of precise probabilities, we may
expect a positive value of imperfect information for these sections in the calculations
later.
Because the inputs of probability measures listed in Table 5-12 through Table
5-15 are the same in the following sections: sections 2 and 3, sections 6 and 7, sections 8
and 9, and sections 10 through 12, respectively, the optimal construction strategies should
be the same as well, which is confirmed by the results in Table 5-16.
After obtaining the optimal construction strategies under no additional
exploration, the imperfect additional exploration, and the perfect additional exploration,
respectively, the next step is to determine the lower and the upper values of (perfect)
information for each tunnel section. For example, the decision trees in Figure 5-11
through Figure 5-13 are used to determine the lower and the upper values of (perfect)
information in tunnel section 1. Again, the relaxed constraints and the strict constraints
are considered, respectively.
No Exp
lorati
on
Perfect
Exploration
Figure 5-11 Decision tree for determining the value of information for Section 1 of the Sucheon Tunnel: No additional exploration branch
195
Figure 5-12 Decision tree for determining the value of information for Section 1 of the Sucheon Tunnel: Imperfect additional exploration branch
196
$3,150
$4,500
$5,700
$7,200
$8,850
$4,125
$3,525
$5,325
$7,350
$9,000
$4,350
$4,875
$4,650
$6,375
$8,138
$4,650
$4,500
$5,400
$6,385
$7,875
1G
2G
3G
4G
5G
1G
2G
3G
4G
5G
1G
2G
3G
4G
5G
1G
2G
3G
4G
5G
C1
C2
C3
C4
1G
2G
3G
4G
5G
$4,425
$4,575
$5,175
$6,450
$6,600
1G
2G
3G
4G
5G
C5
Construction Strategy
Geologic State
Cost(/m)
Exploration Results
Exploration Methods
No Exp
lorati
on
ImperfectExploration
Perfect
Exploration
Figure 5-13 Decision tree for determining the value of information for Section 1 of the Sucheon Tunnel: Perfect additional exploration branch
197
Values of information for each section along the tunnel are depicted in Figure
5-14 through Figure 5-19. The results in tabular forms are listed in Appendix B (Table
B-3 and Table B-4). By comparing the results with the relaxed constraints (Figure 5-14
through Figure 5-18) and the ones with the strict constraints (Figure 5-15 through Figure
5-19), one can easily find that the former has larger bounds than the latter. The value of
information obtained with precise inputs (Karam et al. 2007) are also included in the
figures, represented by a thin solid line, which is always located between the lower and
the upper value of information.
There are several negative lower values of imperfect information in Figure 5-14
and Figure 5-15 as a result of the imprecision in the probability measures. As illustrated
in Section 4.3.4, a negative lower value of information means the indeterminacy in
buying the information even if it is free. Thus, the imperfect exploration is indeterminate
in sections 4 – 7, 14, and 17 if the relaxed constraints are used and in sections 4 and 17 if
the strict constraints are imposed. As for perfect information, a negative lower value is
obtained in section 4 under the relaxed constraints (Figure 5-16); however, it turns to be
positive when the strict constraints are imposed (Figure 5-17), as explained in Section
4.3.4. Similarly for the case of updating value (Figure 5-18 and Figure 5-19). Because the
relaxed constraints do not reduce uncertainty, several sections obtain negative lower
updating values (Figure 5-18); under the strict constraints, perfect exploration is always
better than imperfect exploration, thus updating values can never be negative (Figure
5-19).
198
-$50,000
$0
$50,000
$100,000
$150,000
$200,000
$250,000
$300,000
$350,000
$400,000
$450,000
0 400 800 1200 1600 2000
Distance Along Tunnel (m)
Val
ue o
f Inf
orm
atio
n
Imperfect Exploration(lower)Imperfect Exploration(upper)Imperfect Exploration(precise)
Figure 5-14 Value of imperfect exploration (with relaxed constraints)
-$50,000
$0
$50,000
$100,000
$150,000
$200,000
$250,000
$300,000
$350,000
$400,000
0 400 800 1200 1600 2000
Distance Along Tunnel (m)
Val
ue o
f Inf
orm
atio
n
Imperfect Exploration(lower)Imperfect Exploration(upper)Imperfect Exploration(precise)
Figure 5-15 Value of imperfect exploration (with strict constraints)
Cost of Imperfect Exploration
Cost of Imperfect Exploration
199
-$50,000
$0
$50,000
$100,000
$150,000
$200,000
$250,000
$300,000
$350,000
$400,000
$450,000
0 400 800 1200 1600 2000
Distance Along Tunnel (m)
Val
ue o
f Inf
orm
atio
n
Perfect Exploration(lower)Perfect Exploration(upper)Perfect Exploration(precise)
Figure 5-16 Value of perfect exploration (with relaxed constraints)
-$50,000
$0
$50,000
$100,000
$150,000
$200,000
$250,000
$300,000
$350,000
$400,000
0 400 800 1200 1600 2000
Distance Along Tunnel (m)
Val
ue o
f Inf
orm
atio
n
Perfect Exploration(lower)Perfect Exploration(upper)Perfect Exploration(precise)
Figure 5-17 Value of perfect exploration (with strict constraints)
Cost of Perfect Exploration
Cost of Perfect Exploration
200
-$60,000
-$30,000
$0
$30,000
$60,000
$90,000
$120,000
$150,000
$180,000
0 400 800 1200 1600 2000
Distance Along Tunnel (m)
Upd
atin
g va
lue
Updating value(lower)Updating value(upper)Updating value(precise)
Figure 5-18 Value of updating to perfect exploration (with relaxed constraints)
$0
$10,000
$20,000
$30,000
$40,000
$50,000
$60,000
$70,000
$80,000
0 400 800 1200 1600 2000Distance Along Tunnel (m)
Upd
atin
g va
lue
Updating value(lower)Updating value(upper)Updating value(precise)
Figure 5-19 Value of updating to perfect exploration (with strict constraints)
Cost for Updating
Cost for Updating
201
Next, we are going to determine the optimal exploration plan for the Sucheon
tunnel. The costs of the imperfect exploration and the perfect exploration are assumed to
be $30,000 (Karam et al. 2007) and $50,000, respectively. Thus the cost of updating from
imperfect exploration to the perfect one is equal to $20,000. The three costs are
represented by three solid horizontal lines in Figure 5-14 through Figure 5-19. For a
considered tunnel section, only if the lower value of the exploration is above the cost line,
i.e. a sure positive saving, the exploration may be considered. If the upper value of the
exploration is below the cost line, which indicates a sure negative saving, the exploration
should not be considered. If the lower value and the upper value straddle the cost line,
which means both positive and negative saving are possible, then whether to conduct the
exploration or not is indeterminate. Finally, the optimal exploration plans with imprecise
probabilities and precise probabilities are summarized in Table 5-17. In the case of
imprecise probabilities, the relaxed constraints and the strict constraints are adopted,
respectively. For example, when the relaxed constraints are used, imperfect additional
exploration is warranted only in sections 9, 10, and 12; as for sections 2 – 4, 13 – 15, and
17 – 18, imperfect additional exploration is not considered; for the remaining section,
decisions cannot be made with current imprecise information.
By observing the optimal exploration plans in Table 5-17, one could see that the
set of sections determined to perform the additional exploration under the relaxed
constraints is a subset of the one under the strict constraints; and the set under the strict
constraints is a subset of the one with precise probabilities. Similarly for the sections
where the new exploration is denied. Take the case of the imperfect additional
exploration as an example. The set of the sections with the additional exploration
warranted under the relaxed constraints is section 9, section 10, section 12, as shown in
Table 5-17. The set becomes section 8, section 9, section 10, section 12 under the strict
202
constraints, and section 5, section 7, section 8, section 9, section 10, section 12 with
precise probabilities. This is because the problem with imprecise probabilities under the
relaxed constraints has the highest uncertainty, and the problem with precise probabilities
has the lowest uncertainty. The higher the uncertainty is, the more indeterminate the
problem will be. Therefore, only sections 9, 10, and 12 are determined to be worth
performing the imperfect additional exploration with imprecise probabilities under the
relaxed constraints; while sections 5, 7, 8, 9, 10, and 12 are worthy performing the
additional exploration with precise probabilities.
Next step is to calculate the minimum and the maximum savings resulting from
the optimal exploration plans. The minimum saving would be the sum of the lower values
of information deducted by the costs for the exploration, and thus the maximum saving
would be the sum of the upper values less the exploration costs. For example, in
imperfect exploration under the relaxed constraints, minimum saving =
($128,000 - $30,000) + ($34,100 - $30,000) + ($53,800 - $30,000) = $125,900. Similarly,
maximum saving = ($401,000 - $30,000) + ($48,700 - $30,000) + ($76,700 - $30,000) =
$436,400. The lower and the upper values of information for each tunnel section can be
found in Table B-3 through Table B-5.
203
Table 5-17 Optimal Exploration Plans and the Corresponding Savings.
Relaxed Constraints
Imperfect exploration Perfect exploration Update to perfect exploration
Y N I Y N I Y N I
Sections 9,10,12 2-4,13-15, 17,18
1,5-8, 11,16 9, 12 1-4,11,
13-18 5-8,10, None1-4,
10,11, 13-16,18
5-9, 12,17
Saving [$125,900, $436,400] [$175,700, $424,200] $0
Strict Constraints
Imperfect exploration Perfect exploration Update to perfect exploration
Y N I Y N I Y N I
Sections 8-10, 12 2-4,6,11,
13-15, 17,18
1,5,7,16 9, 12 1-6,11, 13-18 7,8,10 9
1-8,10, 11,13-16, 18
12,17
Saving [$201,300, $369,800] [$246,200, $333,700] [$15,700, $56,100]
Precise probabilities
Imperfect exploration Perfect exploration Update to perfect exploration
Y N I Y N I Y N I
Sections 5, 7-10, 12
1-4,6,11, 13-18 None 8-10, 12 1-7,11,
13-18 None 9 1-8, 10-18 None
Saving $297,605 $286,755 $35,904
Note: Y – Additional exploration is warranted;
N – Additional exploration is not warranted;
I – Indeterminate.
5.4 RISK REGISTER FOR THE EAST SIDE CSO PROJECT
The East Side CSO (Combined Sewer Overflow) Tunnel Project is located in
Portland, Oregon. The primary component of this project is constructed in soft ground
below the groundwater table. It is 20-foot in internal diameter and 6 miles long. During
the construction, this project needs to negotiate with existing infrastructures, including
204
historic structures, railroad, outfalls, and several major utilities. A comprehensive risk
analysis was conducted to reduce the potential impact to the existing infrastructure
(Pennington et al. 2006).
The risk register is one of the standard tools to control the risks. In the risk
register, risk items are identified for all construction activities. For each risk item, experts
estimate its occurrence probability and its consequence. A risk register (Gribbon, City of
Portland Bureau of Environmental Services) developed by the contractor during the
construction is showed in Appendix C, where the occurrence probability of the risk items
are classified into five categories, and rated by a number from 1 through 5, as shown in
Table 5-18. Each rating represents a corresponding probability interval. It is actually
imprecise probability! For example, for a rating of ‘4’, the occurrence probability is the
interval [51%, 70%]. In the original risk register, the contingency for mitigating the risks
was estimated as (consequence×probability rating/5), where the probability is replaced
by (probability rating/5) to calculate the expected value of the consequence. However,
since the occurrence probabilities are originally evaluated by imprecise probabilities, we
can use the imprecise probabilities to obtain the lower and the upper expected
consequence, i.e. the lower contingence and the upper contingence.
Table 5-18 Description of occurrence probability in the East Side CSO project.
Rating Description of occurrence probability 5 Almost Certain (>71% - Expected to occur) 4 Probable (51% - 70% - Will probably occur) 3 Likely (31% - 50% - Likely to occur) 2 Unlikely (11% - 30% - May occur) 1 Rare (<10% - Will rarely occur)
205
Let X1, …, Xn be n risk items. Because the contingency assigned to each risk item
is equal to the expected value of consequence, the total contingency is obtained as
E(X1+ … + Xn). According the property of expectation: E(X1+ … + Xn) = E(X1)+ …
+E(Xn), regardless of the type of interaction between the risk items. Then the lower
contingency is obtained by using the lower values of the occurrence probabilities use, and
the upper contingency is obtained when the upper values of the occurrence probabilities
are used. The values of contingency values are listed in Appendix C. The lower and the
upper total contingency are equal to $19,901,475 and $32,865,750, respectively, which
are the results of the original imprecise evaluations from experts and keep the
imprecision. Within the current available information, we can draw the conclusion that
the total contingency can be as low as $19,901,475 and as high as $32,865,750. However,
the simplification of probabilities as (probability rating/5) led to a probability even higher
than the upper value of the corresponding interval. Take probability rating ‘3’ as an
example, probability rating/5 indicates a value of 0.6, which is larger than the upper value
of interval [31%, 50%). As a result, this inappropriate simplification generated the
original contingency of $39,688,000, which is even higher than the upper contingency
estimated by imprecise probabilities. When a contractor is at the bidding stage, he may
lose the project because of the overestimation of contingency.
206
Chapter 6 Summary and Future Work
6.1 SUMMARY
6.1.1 Algorithms for different types of interaction in imprecise probabilities
Various algorithms for different types of interaction in imprecise probability are
proposed in this study. All algorithms were designed to accommodate two types of
constraints over marginal distributions: prevision bounds or extreme distributions. Each
algorithm was written in terms of both joint distributions and marginal distributions. All
algorithms developed in Chapter 3 have been summarized in Table 3-19.
Previsions on the joint space are linear functions of probability masses, thus
prevision bounds are always achieved at the extreme joint distributions. As for the non-
linear conditional probability on the joint space, we have shown that its upper and lower
bounds are obtained at the extreme joint distributions for all types of independence.
The constraints under unknown interaction are always linear. As a result, the set
of joint distributions on joint finite spaces ΨU is convex. In epistemic
irrelevance/independence, although the constraints stated by the definition are quadratic,
when constraints are given as bounds on marginal previsions, it is possible to rewrite
algorithms in terms of joint distributions, which turns quadratic problems into linear ones.
In strong independence, we have proved that the set of joint probability distribution ΨS is
not convex; an efficient algorithm to find all extreme distributions is presented. The
Pearson correlation coefficient is applied in the case of uncertain correlation, so the
problem becomes non-linear and non-convex. Non-linear programming techniques are
required for solving the problem.
Constraints are consecutively added with types of interaction, and thus the sets of
probability measures are nested, i.e. ΨS ⊆ ΨE ⊆ | isEΨ ⊆ ΨU and ΨC ⊆ ΨU.
207
6.1.2 Application to the standard tools in risk analysis
6.1.2.1 Event tree analysis
Novel methodologies for event tree analysis with imprecise probabilities are
developed in the dissertation. Three types of evidence on outcome probabilities were
considered: probabilities conditional to the occurrence of the event at the upper level,
total probabilities of occurrences, and the combination of the previous two types.
When evidence is given in terms of probabilities conditional to upper level events,
efficient recurrent algorithms were given either in terms of extreme points or as linear
programming problems.
If total probabilities are given, the interaction between any two events should be
assumed in the analysis, including unknown interaction, epistemic irrelevance, epistemic
independence, strong independence, and uncertain correlation. Two different ways to
interpret the interactions are discussed: pair-wise interaction and interaction between
event Si and a new combined event composed of events at its upper level.
In the case of a combination of conditional probabilities and total probabilities,
the event tree can be converted into an equivalent event tree, where all probabilities are
conditional to the upper level events, and then the equivalent tree is used to carry out the
analysis, as in the first case.
6.1.2.2 Fault tree analysis
In this study, the major fault is logically connected with the sub-events by OR-
and AND-gates. The occurrence probabilities for the failure events are not assumed to be
determined but are given imprecisely. Different types of interaction between sub-events
208
can be taken into account, including unknown interaction, independence, and uncertain
correlation.
We have observed and proved that epistemic irrelevance, epistemic independence,
and strong independence all lead to the same results in the fault tree analysis. An example
of a fault tree application has confirmed this claim.
6.1.2.3 Decision tree analysis
This dissertation proposes algorithms for decision analysis within the standard
form of a decision tree, where information on probability evaluation is provided
imprecisely. Considering that the uncertainty may not be changed by the decision, two
types of constraints: relaxed constraints and strict constraints are proposed to deal with
this issue.
Further, this study presents how to consider both prior and new implicit
information and show how to determine the value of new information. Relaxed and strict
constraints are implemented, respectively, while relaxed constraints always provide more
imprecise results than strict constraints.
In the decision-making process, we study lower and upper values of information
within the theory of imprecise probabilities. The lower value is the maximum buying
price for the new information and the upper value if the minimum selling price for the
new information. When the cost is less than the lower value of the new information,
indicating a positive gain, the decision maker should buy the information; when the cost
is even higher than the upper value of the new information, indicating a negative gain, the
decision maker should reject buying the new information.
The new imperfect information may increase or decrease the uncertainty on state
variables. Though the probability measures are updated by the new evidence, the problem
209
may become more indeterminate and then may probably lead to a negative lower value of
information. As for the perfect information, although problems can never be
indeterminate, this study has shown that a negative lower value of information may occur
under relaxed constraints, and will never occur under strict constraints. Generally, more
information are not always better than less information within imprecise probabilities.
6.1.2.4 Risk register
When the theory of imprecise probability is implemented to risk register and the
estimated total contingency is required, it is obtained as the prevision of the sum of all
risk items listed in the risk register. No matter which type of interaction between the risk
items is considered, the lower contingency is obtained when all occurrence probabilities
adopt their lower values, and the upper contingency is obtained when the upper value of
occurrence probabilities are used.
6.2 FUTURE WORK
6.2.1 Elicitation and Assessment with Imprecise Probabilities
In this dissertation, we proposed algorithms and methodologies for risk analysis
with imprecise probabilities. The proposed algorithms, however, require input
information after elicitation and assessment. Future research may develop elicitation and
assessment procedures in a tunnelling risk management process by using the theory of
imprecise probability. Methodologies to elicit and assess information can be developed
for typical applications: risk registers and event, fault, and decision trees. This
information is the input information needed by the algorithms developed in this
210
dissertation; probability intervals and risk intervals are then calculated. The
methodologies may be tested in brainstorming sessions for risk registers and trees.
6.2.2 Improvement on algorithms for different types of interaction
Different algorithms are developed to deal with unknown interaction,
independence, and uncertain correlation. By applying transformation or other
mathematical techniques, some non-linear, non-convex problems have been converted to
linear ones or more efficient algorithms are proposed. However, some problems, such as
uncertain correlation, still require non-linear programming techniques. Due to the
computational difficulty, much remains to be done to improve and simplify the
algorithms for non-linear, non-convex problems.
6.2.3 Cost/Contingency and Schedule Estimation
Cost and schedule estimation under uncertainty requires the use of continuous
variables, as opposed to discrete variables used in risk registers and trees. Similarly, when
the consequence in the risk register is provided imprecisely, the contingency estimation
should be carried out with continuous variables. Algorithms developed in this dissertation
should be extended to continuous variables and applied to cost and scheduling estimation.
This entails the extensive use of functional analysis. The methodologies should be tested
in several case histories and relative advantages/disadvantages should be evaluated.
211
Appendix A Explicit Formulation for Optimization Problems in Section 4.2.1.2
This appendix contains the explicit formulation for the optimization problems over ^i jP
(the joint probability distribution for sub-events iE and jE ) in Section 4.2.2.
1. Unknown interaction: ^i j U∈ΨP
Minimize (Maximize) 2; 2 , ,^1; 1 i ja Pξ η ξ η ξ η
ξ η
= =
= =∑
Subject to
( )( )
^
^
^
,^
; 1,...,
; 1,...,
1
0; 1, 2; 1, 2
Tk k kLOW i i i j UPP i i
Tk k T kLOW j j i j UPP j j
Ti j
i j
E f E f k k
E f E f k k
Pξ η ξ η
⎡ ⎤ ⎡ ⎤≤ ⋅ ≤ =⎣ ⎦ ⎣ ⎦
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⋅ ⋅ =
≥ = =
f P 1
f 1 P
1 P 1
(A.1)
2. Epistemic irrelevance: |^
isi j E∈ΨP
Minimize (Maximize) 2; 2 , ,^1; 1 i ja Pξ η ξ η ξ η
ξ η
= =
= =∑
Subject to
( ) ^
2 2, , ,
^ ^ ^1 1
^
,^
; 1,..., ;
; 1,..., ;
1
0; 1,2; 1, 2
Tk k kLOW i i i j UPP i i
k k kLOW j j UPP j ji j i j i j
Ti j
i j
E f E f k k
E f P E f P k k
P
ξ η ξ ξ η
η η
ξ η ξ η
⋅
= =
⎡ ⎤ ⎡ ⎤≤ ⋅ ≤ =⎣ ⎦ ⎣ ⎦⎛ ⎞ ⎛ ⎞
⎡ ⎤ ⎡ ⎤⋅ ≤ ⋅ ≤ ⋅ =⎜ ⎟ ⎜ ⎟⎣ ⎦ ⎣ ⎦⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
⋅ ⋅ =
≥ = =
∑ ∑
f P 1
P f
1 P 1
(A.2)
where ,^i jξ ⋅P is the ξ -th row of matrix ^i jP .
3. Epistemic independence: ^i j E∈ΨP
Minimize (Maximize) 2; 2 , ,^1; 1 i ja Pξ η ξ η ξ η
ξ η
= =
= =∑
Subject to (A.3)
212
( )2 2
, , ,^ ^ ^
1 1
2 2, , ,
^ ^ ^1 1
^
,^
; 1,..., ;
; 1,..., ;
1
0;
Tk k kLOW i i UPP i ii j i j i j
k k kLOW j j UPP j ji j i j i j
Ti j
i j
E f P E f P k k
E f P E f P k k
P
ξ η η ξ η
ξ ξ
ξ η ξ ξ η
η η
ξ η
⋅
= =
⋅
= =
⎛ ⎞ ⎛ ⎞⎡ ⎤ ⎡ ⎤⋅ ≤ ≤ ⋅ =⎜ ⎟ ⎜ ⎟⎣ ⎦ ⎣ ⎦⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎛ ⎞ ⎛ ⎞
⎡ ⎤ ⎡ ⎤⋅ ≤ ⋅ ≤ ⋅ =⎜ ⎟ ⎜ ⎟⎣ ⎦ ⎣ ⎦⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠
⋅ ⋅ =
≥
∑ ∑
∑ ∑
f P
P f
1 P 1
1, 2; 1, 2ξ η= =
4. Strong independence: ^i j S∈ΨP
Minimize (Maximize) 2; 2 , ,^1; 1 i ja Pξ η ξ η ξ η
ξ η
= =
= =∑
Subject to
( )( )( )
^ 0
; 1,...,
; 1,...,
1; 1
0; 0
Ti j i j
Tk k kLOW i i i UPP i i
Tk k kLOW j j j UPP j j
T Ti j
i j
E f E f k k
E f E f k k
− =
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⋅ = ⋅ =
≥ ≥
P p p
f p
f p
1 p 1 p
p p
(A.4)
5. Uncertain correlation: ^i j C∈ΨP .
Minimize (Maximize) 2; 2 , ,^1; 1 i ja Pξ η ξ η ξ η
ξ η
= =
= =∑
Subject to ( ) ( ) ( ) ( ) ( )i j i ji j S S i j i j S SE S E S D D E S S E S E S D Dρ ρ+ ≤ ≤ +
( )( )( )
^ 0
; 1,...,
; 1,...,
1; 1
0; 0
Ti j i j
Tk k kLOW i i i UPP i i
Tk k kLOW j j j UPP j j
T Ti j
i j
E f E f k k
E f E f k k
− =
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⎡ ⎤ ⎡ ⎤≤ ≤ =⎣ ⎦ ⎣ ⎦
⋅ = ⋅ =
≥ ≥
P p p
f p
f p
1 p 1 p
p p
(A.5)
where ( ) ( )22 , ,mS m mD E S E S m i j= − = , ( )E S is the expected value of variable S.
213
Appendix B Input Data and Results of Optimal Exploration Plan
Table B-1 Extreme Prior Probabilities for each tunnel section
Section Length (m)
No. of Ext. points 1G 2G 3G 4G 5G
1 60 2 0.45 0.55 0.00 0.00 0.00 0.55 0.45 0.00 0.00 0.00
2 20 3 0.20 0.40 0.40 0.00 0.00 0.10 0.50 0.40 0.00 0.00 0.20 0.50 0.30 0.00 0.00
3 57 4 0.20 0.45 0.35 0.00 0.00 0.10 0.50 0.40 0.00 0.00 0.15 0.50 0.35 0.00 0.00 0.15 0.45 0.40 0.00 0.00
4 40 2 0.00 0.00 0.45 0.55 0.00 0.00 0.00 0.50 0.50 0.00
5 120 6 0.00 0.30 0.60 0.10 0.00 0.00 0.20 0.70 0.10 0.00 0.00 0.30 0.65 0.05 0.00 0.00 0.25 0.70 0.05 0.00 0.00 0.20 0.65 0.15 0.00 0.00 0.25 0.60 0.15 0.00
6 41 6 0.20 0.20 0.60 0.00 0.00 0.10 0.30 0.60 0.00 0.00 0.20 0.30 0.50 0.00 0.00 0.20 0.20 0.50 0.10 0.00 0.10 0.30 0.50 0.10 0.00 0.10 0.20 0.60 0.10 0.00
7 96 6 0.20 0.20 0.60 0.00 0.00 0.10 0.30 0.60 0.00 0.00 0.20 0.30 0.50 0.00 0.00 0.20 0.20 0.50 0.10 0.00 0.10 0.30 0.50 0.10 0.00 0.10 0.20 0.60 0.10 0.00
8 104 6 0.50 0.40 0.10 0.00 0.00 0.40 0.50 0.10 0.00 0.00 0.50 0.45 0.05 0.00 0.00 0.45 0.50 0.05 0.00 0.00 0.40 0.45 0.15 0.00 0.00
214
Section Length (m)
No. of Ext. points 1G 2G 3G 4G 5G
0.45 0.40 0.15 0.00 0.00 9 586 6 0.50 0.40 0.10 0.00 0.00 0.40 0.50 0.10 0.00 0.00 0.50 0.45 0.05 0.00 0.00 0.45 0.50 0.05 0.00 0.00 0.40 0.45 0.15 0.00 0.00 0.45 0.40 0.15 0.00 0.00
10 106 2 0.48 0.52 0.00 0.00 0.00 0.52 0.48 0.00 0.00 0.00
11 67 2 0.48 0.52 0.00 0.00 0.00 0.52 0.48 0.00 0.00 0.00
12 167 2 0.48 0.52 0.00 0.00 0.00 0.52 0.48 0.00 0.00 0.00
13 36 2 0.75 0.25 0.00 0.00 0.00 0.80 0.20 0.00 0.00 0.00
14 74 4 0.15 0.80 0.05 0.00 0.00 0.05 0.85 0.10 0.00 0.00 0.10 0.85 0.05 0.00 0.00 0.10 0.80 0.10 0.00 0.00
15 43 2 0.15 0.85 0.00 0.00 0.00 0.20 0.80 0.00 0.00 0.00
16 110 2 0.65 0.35 0.00 0.00 0.00 0.70 0.30 0.00 0.00 0.00
17 188 2 0.85 0.15 0.00 0.00 0.00 0.90 0.10 0.00 0.00 0.00
18 50 3 0.15 0.30 0.55 0.00 0.00 0.10 0.35 0.55 0.00 0.00 0.15 0.35 0.50 0.00 0.00
215
Table B-2 Extreme conditional probabilities of exploration result given real geological states
Reality Exploration Result 1G 2G 3G 4G 5G
No. of Ext. Pts 2 2 5 5 2 1G 0.85 0.95 0.08 0.12 0.000.00 0.00 0.000.00 0.00 0.000.00 0.00 0.00 0.00 0.002G 0.15 0.05 0.92 0.88 0.080.12 0.10 0.080.12 0.00 0.000.00 0.00 0.00 0.00 0.003G 0.00 0.00 0.00 0.00 0.850.83 0.85 0.770.73 0.15 0.050.15 0.05 0.10 0.00 0.004G 0.00 0.00 0.00 0.00 0.070.05 0.05 0.150.15 0.75 0.870.77 0.80 0.75 0.06 0.125G 0.00 0.00 0.00 0.00 0.000.00 0.00 0.000.00 0.10 0.080.08 0.15 0.15 0.94 0.88
Table B-3 Value of information (with relaxed constraints)
Value of Information (Relaxed Constraints) VI VPI VUP Section
LOW UPP LOW UPP LOW UPP 1 $17,600 $33,300 $26,100 $37,300 $1,590 $10,900 2 $2,840 $10,300 $4,950 $11,600 -$1,700 $5,110 3 $11,500 $24,400 $17,500 $28,000 -$1,710 $11,400 4 -$6,140 $4,410 -$1,470 $5,250 -$2,430 $7,940 5 -$16,300 $100,000 $14,700 $105,000 -$28,400 $64,100 6 -$6,400 $52,600 $4,370 $55,000 -$15,000 $28,200 7 -$15,000 $123,000 $10,200 $129,000 -$35,100 $66,100 8 $22,700 $71,100 $35,500 $77,600 -$9,130 $28,400 9 $128,000 $401,000 $200,000 $437,000 -$51,500 $160,000
10 $34,100 $48,700 $48,000 $55,300 $5,070 $15,500 11 $21,600 $30,800 $30,500 $35,000 $3,200 $9,780 12 $53,800 $76,700 $75,700 $87,200 $7,980 $24,400 13 $1,340 $7,480 $6,340 $9,450 $1,290 $5,680 14 -$4,340 $12,400 $4,160 $17,800 -$1,580 $1,540 15 $1,070 $5,280 $6,290 $8,380 $2,300 $6,030 16 $15,200 $33,300 $30,100 $39,600 $4,270 $17,000 17 -$12,700 $21,300 $14,800 $31,000 $6,190 $31,000 18 $15,400 $25,600 $20,400 $28,700 -$669 $8,760
216
Table B-4 Value of information (with strict constraints)
Value of Information (Strict Constraints) VI VPI VUP Section
LOW UPP LOW UPP LOW UPP 1 $19,900 $30,800 $28,500 $34,900 $4,020 $8,650 2 $5,240 $8,050 $7,350 $9,300 $1,250 $2,240 3 $14,900 $20,900 $20,900 $24,600 $3,630 $6,280 4 -$720 $1,210 $1,800 $1,980 $774 $2,520 5 $25,700 $44,900 $33,500 $49,700 $4,670 $7,810 6 $12,400 $24,000 $16,400 $26,400 $2,370 $4,300 7 $29,000 $56,100 $38,300 $61,900 $5,540 $10,100 8 $34,600 $55,500 $47,600 $62,000 $6,340 $13,500 9 $195,000 $313,000 $268,000 $349,000 $35,700 $76,100
10 $35,600 $47,100 $49,600 $53,700 $6,660 $14,000 11 $22,500 $29,800 $31,400 $34,000 $4,210 $8,860 12 $56,100 $74,200 $78,200 $84,700 $10,500 $22,100 13 $1,970 $6,760 $7,020 $8,780 $1,970 $5,050 14 $103 $7,960 $8,600 $13,300 $4,010 $11,100 15 $1,070 $5,280 $6,290 $8,380 $3,100 $5,280 16 $17,100 $31,000 $32,200 $37,500 $6,330 $15,100 17 -$8,620 $17,500 $18,300 $27,500 $9,710 $27,500 18 $18,400 $22,900 $23,400 $25,900 $3,020 $5,170
217
Table B-5 Value of information (with precise probabilities)
Value of Information Section VI VPI VUP
1 $24,716 $31,054 $6,338 2 $6,333 $8,055 $1,722 3 $18,049 $22,957 $4,908 4 $307 $1,908 $1,601 5 $41,872 $47,736 $5,864 6 $18,178 $21,476 $3,298 7 $41,329 $49,003 $7,674 8 $44,694 $54,522 $9,828 9 $244,274 $300,179 $55,904 10 $40,306 $50,642 $10,335 11 $25,477 $32,009 $6,533 12 $65,130 $81,413 $16,283 13 $4,212 $7,722 $3,510 14 $3,652 $10,712 $7,060 15 $3,773 $7,966 $4,193 16 $24,667 $35,393 $10,725 17 $7,332 $25,662 $18,330 18 $20,777 $24,837 $4,060
218
Appendix C Risk register for East Side CSO Project, Portland, Oregon
Note: adapted from the risk register from Gribbon, City of Portland Bureau of Environmental Services Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
No. RISK ITEM MITIGATION
ACCESS / PERMIT RISKS
101 Delay in Obtaining Property
101.1 Opera Parking Lot
(1) Start discussion early with Portland Opera; (2) Condemnation of property; (3) Accelerate shaft / tunnel; (4) Move shaft within OMSI site; (5) Alternate mining site
3 3 9 $1,000,000 $600,000 $310,000 $500,000
101.3 Alder Shaft - Corno Building (1) Condemnation of property; (2) Multi-shift / increase work-days per week 2 4 8 $1,000,000 $400,000 $110,000 $300,000
108 Emergency Surface Access Issues
108.2 Micro-Tunnel Machine Recovery Shaft(1) Maintain reasonable bore lengths; (2) reduce number of long bores; (3) MTBM preparation for longer drives
3 3 9 $500,000 $300,000 $155,000 $250,000
110 Crew Parking
110.2 Alder Shaft (1) Phase work to minimize workers; (2) Mandate street parking; (3) Bus crew 3 3 9 $950,000 $570,000 $294,500 $475,000
219
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
112 Barging Permit (1) Start permit during pre-construction; (2) Qualify under Slopes; (3) Perform work during first summer work window (2006)
3 3 9 $800,000 $480,000 $248,000 $400,000
114 Noise Variances
114.2 Alder Shaft (1) Apply for noise variance; (2) Equipment modifications; (3) Limit multi-shift work 4 2 8 $500,000 $400,000 $255,000 $350,000
114.3 Pipeline Shaft Sites (1) Apply for noise variance; (2) Equipment modifications; (3) Limit multi-shift work 4 2 8 $250,000 $200,000 $127,500 $175,000
114.3 Microtunnel Pipelines (1) Apply for noise variance; (2) Equipment modifications; (3) Limit multi-shift work 4 2 8 $250,000 $200,000 $127,500 $175,000
115 Other Permits
115.1Corno's Building Demolition Permits -requirement to removal concrete basement structure
(1) Work with City permitting agency 3 3 9 $750,000 $450,000 $232,500 $375,000
TUNNEL CONSTRUCTION RISKS
202 Labor Availability (1) Monitor availability of experienced tunnel personnel (Brightwater concern); 3 3 9 $1,000,000 $600,000 $310,000 $500,000
220
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
203 TBM Initial Launch
203.1 Initial Launch (1) Redundant systems - cement wall, steel can, double seal; (2) Detailed planning; 3 4 12 $3,300,000 $ 1,980,000 $ 1,023,000 $ 1,650,000
203.2 Shaft Break-In / Out (1) Redundant systems - flooded shaft; (2) Detailed planning; 1 4 4 $3,200,000 $640,000 $- $320,000
204 TBM Productivity (Discrete Items)
204.1 Longer Learning Curve (1) Detailed step-step work plan; (2) TBM supplier tech. representation; (3) Over-staff initial start-up
3 4 12 $1,600,000 $960,000 $496,000 $800,000
204.11 Electrical Plant Upgrades - PGE (1) Early discussions with PGE; (2) Detailed ramp-up analysis; (3) Peer review of PGE provided equipment
3 3 9 $500,000 $300,000 $155,000 $250,000
204.13 Main Bearing Replacement
(1) Design life at 15,000-hr; (2) Maintenance program; (3) Inspection after north drive; (4) Main bearing stored in US / under warranty
2 4 8 $5,600,000 $ 2,240,000 $616,000 $ 1,680,000
204.14 Power Drop at Opera Site (1) Continue discussions with PGE; (2) Look at alternative power access routes 5 3 15 $350,000 $350,000 $248,500 $350,000
205 TBM Productivity (excluding interventions and time through shafts)
205.2 Production at 35 feet per day (1) Increase to 6 wd per week; (2) Increase to 7 wd per week 3 5 15 $12,250,000 $ 7,350,000 $ 3,797,500 $ 6,125,000
221
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
206 Slurry Separation System
206.1 Difficult Separation in Sand/Silt Alluvium
(1) Truck material or pay barging premium, do not impact tunneling operation. 4 2 8 $250,000 $200,000 $127,500 $175,000
206.3 Availability of Slurry Separation Plant (1) Procure new plant; (2) Maintenance program; (3) Experienced personnel 2 4 8 $1,600,000 $640,000 $176,000 $480,000
207 Wear on TBM and Parts
207.1 River Street - Increased Maintenance in Preparation of Long North Drive
(1) Increase shaft rehab time; (2) Procure 2nd cutterhead 3 3 9 $840,000 $504,000 $260,400 $420,000
208 Interventions
208.2 Increased Duration of Interventions (1) Machine maintenance at shafts; (2) Wear detection systems; (3) TBM design 3 4 12 $1,600,000 $960,000 $496,000 $800,000
208.3 Inability to Perform Compressed Air Intervention
(1) Ensure intervention in Troutdale; (2) Jet grout zone; 3 3 9 $1,330,000 $798,000 $412,300 $665,000
209 Segmental Lining
222
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
209.4 Excessive Water Leaks (1) QC Program for segment fabrication; (2) Lining installation quality control inspection
3 3 9 $600,000 $360,000 $186,000 $300,000
210 Banfield Interchange Piles (Option B)
210.1 Underpinning Program for Bridge Bent(1) Perform pile investigation program; (2) Perform engineering analysis of down-drag on pile group
3 3 9 $750,000 $450,000 $232,500 $375,000
210.2 Jet Grouting for Pile Removal (1) Perform pile investigation program 2 3 6
210.3 Alternate Alignment - Option C (1) Perform pile investigation program; (2) Consider 210.1 and 210.2 risks 2 4 8 $2,300,000 $920,000 $253,000 $690,000
211 Port Center Tunneled Connection
211.3 Redundant Bulkhead System in Case of Early Tie-In 4 3 12 $500,000 $400,000 $255,000 $350,000
GROUND IMPROVEMENT RISKS
223
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
303 Compensation / Compaction Grouting
303.1 Structure protection grouting - Reach 2
(1) Perform detailed engineering settlement analysis; (2) Instrumentation and monitoring program; (3) Pre-drill grout holes
3 3 9 $600,000 $360,000 $186,000 $300,000
305 Scope, Complexity of Instrumentation Installation and Monitoring
4 3 12 $500,000 $400,000 $255,000 $350,000
MAIN SHAFT CONSTRUCTION RISKS
401 Slurry Wall Construction
401.1 Difficulty of Excavation in Cobbles and Boulders
(1) Reasonable GBR values; (2) Evaluate appropriate equipment; (3) Multiple shift and added work days per week
4 3 12 $525,000 $420,000 $267,750 $367,500
$401 Difficulty of Excavation in Obstructions (Logs, Buried Objects, etc.)
(1) Reasonable GBR values; (2) Evaluate appropriate equipment; (3) Multiple shift and added work days per week
4 3 12 $525,000 $420,000 $267,750 $367,500
403 Shaft Excavation
224
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
403.1 Main Mining Shaft Not Ready Ahead of TBM Assembly
(1) Early procurement of slurry wall; (2) Multiple shift and added work days per week for slurry wall and shaft excavation
4 2 8 $100,000 $80,000 $51,000 $70,000
405 Permanent Site Restoration at Shaft Locations
(1) Complete design/drawings with defined scope of work; (2) Receive approval from O&M staff
3 3 9 $700,000 $420,000 $217,000 $350,000
406 Shaft Concrete
406.1 Scope of Structural Concrete -Quantities and Complexity
(1) Complete design/drawings with defined scope of work; (2) Design to ERC budget 4 3 12 $1,800,000 $ 1,440,000 $918,000 $ 1,260,000
MICROTUNNELING CONSTRUCTION RISKS
502 MTBM Operation Set-Up / Move
502.1 Difficulty With Set-Up and Relocation of Microtunnel TBM and Support Systems
(1) Detailed planning; (2) Utilize same crews - learning curve benefit. 4 2 8 $337,500 $270,000 $172,125 $236,250
502 MTBM Break-In / Out(s)
225
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
502.1 Shaft Break-In / Out of MTBM (1) Detailed planning; (2) Utilize same crews - learning curve benefit. 4 2 8 $315,000 $252,000 $160,650 $220,500
503 MTBM Productivity (Discrete Items)
503.1 Obstructions - logs, buried objects, etc. (1) Reasonable GBR values; (2) Evaluate appropriate equipment 4 2 8 $320,000 $256,000 $163,200 $224,000
503.2 Conflicts With Existing Utilities (1) Adequate microtunnel depth; (2) Investigation of existing utilities; (3) Locates
4 2 8 $320,000 $256,000 $163,200 $224,000
503.5Excessive Friction Due to Drive Lengths Too Long For Ground Conditions
(1) Jacking force and use of stations; (2) Pipe lubricants; (3) Non-stop operations 4 2 8 $160,000 $128,000 $81,600 $112,000
505 Interventions
505.4Increased Number / Duration of Interventions (MTBM Stuck Beneath Railroad Tracks)
(1) Geological investigation; (2) Proper machine design; (3) Non-stop operations 4 2 8 $320,000 $256,000 $163,200 $224,000
505.5 Rescue Shaft for MTBM (1) Maintain reasonable bore lengths; (2) reduce number of long bores; (3) MTBM preparation for longer drives
3 3 9 $1,200,000 $720,000 $372,000 $600,000
OPEN-CUT PIPELINE CONSTRUCTION RISKS
226
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
601 Scope of Relocation/Protection of Existing Utilities
(1) Completeness of drawings; (2) Perform utility locates; (3) Test pits 5 2 10 $400,000 $400,000 $284,000 $400,000
605 Scope of Surface Restoration (1) Discussion / agreement with PDOT and businesses; (2) Reasonable quantities in ERC.
5 2 10 $250,000 $250,000 $177,500 $250,000
606 Upgrade of Surface Storm System (1) Discussion / agreement with PDOT and businesses; (2) Reasonable quantities in ERC.
2 4 8 $2,000,000 $800,000 $220,000 $600,000
607 Difficulties With Final Tie-Ins to Existing Outfall Structures
(1) Detailed planning; (2) Perform work during low flow period; (3) Ground stabilization program; (4) Complete maximum amount of work prior to final tie-in.
4 2 8 $200,000 $160,000 $102,000 $140,000
PIPELINE SHAFT CONSTRUCTION RISKS
701 Support of Excavation System -Soldier Pile Shafts Revised to Secant Pile
701.1 Structures - OF-30, OF-43, OF-46 (1) Detailed geotechnical assessment; (2) Adequate dewatering provisions; (3) Structure instrumentation
3 4 12 $2,300,000 $ 1,380,000 $713,000 $ 1,150,000
701.2 Increased Scope of Dewatering Work Anticipated for Soldier Pile Shafts
(1) Detailed geotechnical assessment; (2) Adequate dewatering scope; (3) Dewatering duration
4 2 8 $300,000 $240,000 $153,000 $210,000
704 Pipeline Shaft Excavation
227
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
704.1 Secant Pile Drilling - Encountering Cobbles and Boulders
(1) Reasonable GBR values; (2) Evaluate appropriate equipment; (3) Multiple shift and added work days per week
3 3 9 $555,000 $333,000 $172,050 $277,500
705 Pipeline Shaft Concrete
705.1 Scope of Structural Concrete -Quantities and Complexity
(1) Complete design/drawings with defined scope of work; (2) Design to ERC budget 3 3 9 $1,225,000 $735,000 $379,750 $612,500
FINANCIAL / OTHER RISKS
901 Equipment
901.1 Purchase More Equipment Than Planned
(1) Perform detailed equipment analysis; (2) Control of purchasing; (3) Sequence operations to avoid extra concurrent activities
3 3 9 $1,500,000 $900,000 $465,000 $750,000
901.2 Purchase Cost of Equipment Exceeds Budget
(1) Perform detailed equipment analysis; (2) Control of purchasing; (3) Obtain competitive pricing
3 4 12 $2,500,000 $ 1,500,000 $775,000 $ 1,250,000
901.3 Less Than Expected Equipment Salvage
(1) Control equipment hours; (2) Proper equipment maintenance program; (3) Corporate equipment marketing and sales
3 3 9 $500,000 $300,000 $155,000 $250,000
228
Probability of Risk Degree of Risk Risk Ratings
5 = ( >70% - Expected to occur) 5 = Very Serious (>$10m - >6 months) Prob. Degree Total Potential Cost
Precise Contingency
Lower Contingency
Upper Contingency
4 = (51% - 70% - Will probably occur) 4 = Major ($2m - $10m - 3-6 months)
3 = (31% - 50% - Likely to occur) 3 = Moderate ($0.5m - $2m - 1-3 months)
2 = (10% - 30% - May occur) 2 = Minor ($0.1m - $0.5m - 1-4 weeks)
1 = ( <10% - Unlikely to occur) 1 = Insignificant (<$0.1m - <1 week)
Cost$ X Probability/5
901.4 Foreign Exchange Rates (1) All purchases in US$; (2) Receive quotes in US$ or secure Euro conversion 4 3 12 $200,000 $160,000 $102,000 $140,000
902 Subcontractors
902.1 Actual Subcontract Contract Amounts Differ From Plug Prices in Final ERC
(1) Obtain quotations for ERC pricing (absent 100% design); (2) Use reasonable and actual unit (past) costs; (3) Package work to receive best value bids
3 4 12 $5,000,000 $ 3,000,000 $ 1,550,000 $ 2,500,000
902.2Subcontract Change Orders Not Easily Resolved (Beyond those Identified in This Risk Assessment
(1) Defined Subcontract Agreements and Bid Documents; (2) Unit Price Items (where applicable); (3) Contract administration
3 3 9 $1,000,000 $600,000 $310,000 $500,000
TOTAL $39,688,000 $19,901,475 $32,865,750
229
Reference
AACE International Recommended Practice No. 41R-08, Risk Analysis And Contingency Determination Using Range Estimating, 2008. Available at http://www.aacei.org/technical/rps/41R-08.pdf.
Ang, A.H.-S. and Tang, W.H., 1984. Probability concepts in engineering planning and design. Decision, Risk and Reliability vol. II, Wiley, New York.
Barton, N.R., Lien, R. and Lunde, J. Engineering classification of rock masses for the design of tunnel support. Rock Mech. 6(4), 189-239. 1974.
Benjamin, R.J., Cornell, A.C., 1970. Probability, Statistics and Decision for Civil Engineers. McGraw-Hill, New York.
Berleant, D. and J. Zhang, Using Pearson Correlation to improve envelopes around the distribution s of functions, Reliable Computing 10(2), 139-161, 2004.
Bernardini, A. and Tonon, F. Bounding Uncertainty in Civil Engineering – Theoretical Background. Springer. 2010
Caltrans, Project Risk Management Handbook, 2007. Available at http://www.dot.ca.gov/hq/projmgmt/documents/prmhb/caltrans_project_risk_management_handbook_20070502.pdf
Campos, C.P., and F.G. Cozman, Computing lower and upper expectations under epistemic independence, International Journal of Approximate Reasoning, 44(3):244-260, 2007.
Campos, L.M. and S. Moral. Independence concepts for convex sets of probabilities. In Ph. Besnard and S. Hanks, eds, Proceedings of the 11th Conference on Uncertainty in Artificial Intelligence, 108–115. Morgan & Kaufmann, San Mateo, 1995.
Cano A., and S. Moral, Algorithms for Imprecise Probabilities, Handbook of Defeasible Reasoning and Uncertainty Management Systems, V 5, 369 - 420, Springer, 2000.
Couso, I., S. Moral, and P. Walley. (1999). Examples of independence for imprecise probabilities. ISIPTA '99 - Proceedings of the First International Symposium on Imprecise Probabilities and Their Applications, Zwijnaarde, Belgium.
Couso, I., S. Moral, and P. Walley. A survey of concepts of independence for imprecise probabilities. Risk Decision and Policy, Vol. 5, pp. 165-181, 2000.
230
Durrett R., Probability: theory and examples, second or third ed., Duxbury Press, Belmont, CA, 1996.
Eskesen, S.D., P. Tengborg, J. Kampmann, T.H. Veicherts, Guidelines for tunnelling risk management: International Tunnelling Association, Working Group No. 2, Tunnelling and Underground Space Technology, Volume 19, Issue 3, May 2004, Pages 217-237.
Ferson, S., R.B. Nelsen, J. Hajagos, D.J. Berleant, J. Zhang, W. T. Tucker, L. R. Ginzburg and W. L. Oberkampf, Dependence in probabilistic modeling, Dempster-Shafer theory, and probability bounds analysis, SAND2004-3072, 2004. Available at http://www.ramas.com/depend.pdf.
Ferson S., V. Kreinovich, L. Ginzburg, D. S. Myers, and K. Sentz, Constructing Probability Boxes and Dempster-Shafer Structures, Sandia National Laboratories, Technical Report SAND2002-4015, Albuquerque, New Mexico. Available at http://www.sandia.gov/epistemic/Reports/SAND2002-4015.pdf.
Fetz, T. and M. Oberguggenberger (2004). Propagation of uncertainty through multivariate functions in the framework of sets of probability measures. Reliability Engineering & System Safety 85(1-3): 73-87.
Gribbon, P. “Risk Register for East Side CSO Project, Portland”, City of Portland Bureau of Environmental Services, personal communication.
FHWA, Risk Assessment and Allocation for Highway Construction Management, 2006. Available at http://international.fhwa.dot.gov/riskassess/.
FTA, Project & Construction - Management Guidelines, 2003. Available at http://www.fta.dot.gov/publications/reports/other_reports/publications_3875.html.
Harrison, J.P., J.A. Hudson, Engineering rock mechanics: part 2: illustrative worked examples, Elsevier Science, 2000.
Hong, E. S., I. M. Lee, H. S. Shin, S. W. Nam, and J. S. Kong (2009), “Quantitative risk evaluation based on event tree analysis technique: Application to the design of shield TBM.” Tunnelling and Underground Space Technology, 24(3). 269-277.
Horst, R., P.M. Pardalos and N.V. Thoai (2000) Introduction to Global Optimization, 2nd Edn. Dortrecht: Kluwer
Huang D., Toly Chen, Mao-Jiun J. Wang, A fuzzy set approach for event tree analysis, Fuzzy Sets and Systems ,Volume 118, Issue 1, 16 February 2001, Pages 153-165
231
Huntley, N., Troffaes, M., An Efficient Normal Form Solution to Decision Trees with Lower Previsions, Soft Methods for Handling Variability and Imprecision, ASC 48, Page 419-426, 2008.
International Tunnel Insurance Group, The Code of Practice for Risk Management of Tunnel Works, 2006
Karam K. S., J. S. Karam, and H. H. Einstein (2007), “Decision Analysis Applied to Tunnel Exploration Planning I: Principles and Case Study.” Journal of Construction Engineering and Management, 133(5), 344-353.
Kenarangui R., Event-Tree Analysis by Fuzzy Probability, IEEE TRANSACTIONS ON RELIABILITY, VOL. 40, NO. 1, 1991 APFUL, Page 120-124
Luenberger D.G. Linear and nonlinear programming, 1984
Matheron G. Random sets and integral geometry. New York: Wiley, 1975.
Min, S. Y., H. H. Einstein, J. S. Lee, and T. K. Kim (2003). Application of decision aids for tunneling (DAT) to a drill & blast tunnel. KSCE Journal of Civil Engineering, 7(5). 619-628
Pennington, T.W., R.F. Cook, D.P. Richards, J. O'Carroll, and T. Cleys, A tunnel case study in Risk Management, Proceedings of North American Tunneling 2006, Chicago, IL, USA.
Rockafellar, R.T., Convex Analysis, Princeton University Press, Princeton, NJ, 1997.
Simiu E., N. A. Heckert, J. J. Filliben and S. K. Johnson, Extreme wind load estimates based on the Gumbel distribution of dynamic pressures: an assessment, Structural Safety, 23(3), 2001, Pages 221-229
Stockholm ring road. (2010). Retrieved June 6, 2010, from http://en.wikipedia.org/wiki/Stockholm_ring_road
Stockholm News, "Controversial road project to be approved", Sept. 3 2009, http://www.stockholmnews.com/more.aspx?NID=3898
Sturk, R., L. Olsson, and J. Johansson (1996), “Risk and Decision Analysis for Large Underground Projects, as Applied to the Stockholm Ring Road Tunnels.” Tunnelling and Underground Technology, 11(2), 157-164.
Tonon F., A. Bernardini, A. Mammino, Determination of parameters range in rock engineering by means of Random Set Theory, Reliability Engineering & System Safety Volume 70, Issue 3, December 2000, Pages 241-261
232
Tollroads Newsletter, “SETBACK FOR TOLLS Stockholm tunnel ring stopped”, issue 13, Mar 1997, http://www.tollroadsnews.com/node/1978
TunnelTalk, “Fatal collapse on Cologne’s new metro line”, Available at http://www.tunneltalk.com/Cologne-collapse-Mar09-Deadly-collapse-in-Cologne.php, retrieved on June 1, 2010.
Vicig, P. (1999). Epistemic independence for imprecise probabilities, ISIPTA '99 - Proceedings of the First International Symposium on Imprecise Probabilities and Their Applications, Zwijnaarde, Belgium.
Walley, P. Statistical reasoning with Imprecise Probabilities. London, New York, Tokyo, Melbourne, Madras, Chapman and Hall, 1991.
Wannick, H. The code of Practice for Risk Management of tunnel works - Future Tunnelling Insurance from the insurer's point of view, ITA-AITES 2006 World Tunnel Congress, Seoul, Korea, 2006.
Whitman, R.V. Evaluating calculated risk in geotechnical engineering. Journal of Geotechnical Division, ASCE 1984;110(2):143-188.
WSDOT, Project Risk Management Guide, 2010. Available at http://www.wsdot.wa.gov/publications/fulltext/cevp/ProjectRiskManagement.pdf.
Zadeh, L.A. (1965) "Fuzzy sets". Information and Control 8 (3) 338–353. Available at http://www-bisc.cs.berkeley.edu/Zadeh-1965.pdf.
233
Vita
Xiaomin You was born in Nankang, Jiangxi Province, China, on December 02,
1982, the daughter of Weihua You and Xiangying Liu. After completing her work at
Ganzhou No.3 Middle School, she entered Tongji University, where she received her
Bachelor and Master degrees in Civil Engineering in 2004 and 2007, respectively.
Afterward, Xiaomin entered The University of Texas at Austin for her doctoral degree
under the supervision of Dr. Fulvio Tonon. Her research interests include tunnel
engineering, underground excavation, and risk analysis.
Permanent Address: 11 Dahuae Street
Ganzhou, Jiangxi 341000
China
This disseration was typed by the author.