Ang A. H-S, Probability Concepts in Engineering Planning and Design, 1984

Probability Concepts
in Engineering
Planning and Design
VOLU1l1E II DECISI01V, RISK,
AND RELIABILITY
ALFREDO H-S. ANG

Professor of CiLil Engineering
Unhersity of Illinois at Urbana-Champaign
WILSON H. TANG
Professpr of Civil Engineering
University oflllinois at Urbana-Champaign.
s
r
r
JOHN WILEY & SONS f

New York Chichester Brisbane Toronto Singapore :,
Dedicated to Myrtle and Bernadette
Copyright l!:l 1984 by John Wiley & Sons. Inc.

All righis reserved. Published simultaneousi) in Canada.
Reproduction or translation of any pan "of
this worl; beyond that permined by Sectic>ns
107 and 108 of the 1976 United States Copyrij!ht
Act without the permission of the copyright
owner is unlawful. Requests for permission
or further information should be addressed to
the Permissions Department. John Wiley & Sons.
c
Librtzry of onxrtss cllttzloxinx in Publictztion DIIUI:
!Revised for volume 2)
Ans. Alfredo Hua-Sin~;. 193(}-
Probability concepts in engineering planninl! and
design.
Includes bibliographical references and in~exes.
Contentl: , .. I. Basic principles-v. ::!. Decision.
risl.:. and reliability.
I. Engineering-Statistical methods. !!. Probabilities.
I. Tang. Wilson H.
TA340.AS 620 .. 00422"01519::! 75-589.:::
Printed in the Republic of Sinppore
10 9 8 7 6 s4 3 2
preface. _
Like Volume I, this volume emphasizes the applications of probability and statistics
in engineering. It is built on the basic principles contained in Volume I but
additional more advanced tools are developed. including statistical decision
analysis. Markov and queueing models. the statistics of extreme values, Monte
Carlo simulation. and system reliability. Again, the necessary concepts are
introduced and illustrated within the context of engineering problems. Indeed, the
developrr.ent of the logical concepts and the illustration of established principles
in engineering applications are the main objectives of the volume. The tools
provided should form the quantitative bases for risk evaluation, control, and
management, and intelligent decision making under uncertainty.
In some instances. problems in several diverse areas are used to illustrate the
same or similar principles. This was purposely done to demonstrate the universal
application of the pertinent concepts and also to provide readers with wider
choices of illustrative appii~ations.
Much of the mathematica( material is, of course, available in the literature and
some of it is in text form: for example, Gumbel (1958) on extreme value statistics;
Rubinstein ( 1981) on Monte Carlo simulation; Raiffa (1970) and Schlaifer ( 1969)
on statistical decision analysis: and Parzen (1962) and Saaty (1961) on Markov
and queueing processes. However, the concepts that are especially useful for
engineering planning and design are emphasized: moreover, these are presented
purposely in the context of engineering significance and in terms that are more
easily comprehensible to engineers (supplemented with numerous illustrative
examples). Much of the material in Chapters 6 and 7 on basic and systems relia-
bility is. for the most part, only currently available in scientific and technical
journals-in this regard. only typical references to the available literature are cited. .
as there is no intention to be exhaustive.
The material is suitable for two advanced undergraduate or graduate lrvel
courses. One would cover engineering decision and risk analysis, based on Chapters
2. 3, and 5 of the present volume supplemented by Chapter 8 of Volume I; the
other would focus on system reliability and design and may be developed using
the material contained in Chapters 4, 6. and 7. Both courses would require a pre-
requisite .background in introductory applied probability and statistics, such as the
material in Volume I.
During the preparation a:nd development of the material for this volume, the
authors are indebted in many ways to colleagues and students-to former students
who had to endure imperfect and incomplete versions of the material; to our
colleague Professor Y. K. Wen for numerous discussions and suggestions over
the years; and to Professors A. der Kiureghian, M. Shinozuka, E. Vanmarcke,
and J. T. P. Yao for constructive suggestions and incisive reviews of the manuscript.
A number of former graduate students also contributed to the development or
some of the material including numerical calculations of the examples; in par-
ticular. the contributions and assis.tance of R. M. Bennett, C. T. Chu, H . Pearce,
v
vi PREFACE
J. Pires, I. Sidi, M. Yamamoto, and A. Zerva are greatly appreciated. Finally, our
thanks to Claudia Cook for her patience and expert typing of several versions of
the manuscript material, and to R. Winburn for his professional artwork of the
figures.
A. H-S. Ang AND W. H. Tang
\
contents
l. Introduction
1.1 Aims and Scope of Volume II
1.2 Essence and Emphasis I

1.2.1 Decision Analysis I
1.2.2 Markov, Queueing, and Availability Models 2
1.2.3 Statistical Theory of Extremes 3
1.2.4 Monte Carlo Simulation 3
1.2.5 Reliability and Reliability-Based Design 3
1.2.6 Systems Reliability 4
2. Decision AJIO:lysis 5
.......
2.1 Introduction 5
2.1.1 Simple Risk-Decision Problems 5
2.1.2 Characteristics of a General Decision Problem 9
2.2 The Decision Model 10

2.2.1 Decision Tree 10
2.2.2 Decision Criteria 13
2.2.3 Decision Based on Exis~ing Information-Prior Analysis 16
2.2.4 Decision with Additio'nal Information-Terminal Analysis 25
2,.2.5 Preposterior Analysis 28
2~.6 Value of Information 30
2.2.7 Sensitivity of Decision to Error in Probability Estimation 32
2.2.8 More Examples 33
2.3 Applications in Sampling and Estimation 47

2.3.1 Bayes Point Estimator 47
2.3.2 Optimal Sample Siz~ 50
2.4 Elementary Concepts of Utility Theory 56

2.4.1 Axioms of Utility Theory 56
2.4.2 Utility Function 57
2.4.3 Determination of Utility Values 60
2.4.4 Utility Function of Monetary Value 63
2.4.5 Utility Function of Other Variables 66.
2.4.6 Maximum Expected Utility 67
2.4.7 Common Types of Utility Functions 70
2.4.8 _ ~nsitivity of Expected Utility to Form of Utility Function 73
viii CONTENTS
2.5 Assessment of Probability Values 74

2.5.1Bases of Probability Estimation 74
2.5.2Subjective Probability 75
'2.5.3~ubjective Distribution of a Random Variable 76
2.6 Decisions with Multiple Objectives 77

2.6.1 Weighted Objective Decision Analysis 78
2.6.2 Multiattribute Utility Approach 84
Problems for Chapter 2 96
3. . Markov, Queueing, ~nd Availability Models 112
3.1 The Markm Chain 112

3.1.1 Introduction 112
\ 3.1.2 The Basic Model 112
3.1.3 State Probabilities 113
3.1.4 Steady State Probabilities 118
3.1.5 First Passage Probabilities li2
3.1.6 Recurrent, Tran~ient, and Absorbing States 125
3.1.7 Random Time Between Stages 136
... 3.1.8 Continuous Parameter Markov Chain 138
....
3.2 Queueing Models 140
3.2.2 Poisson Process Arrivals and Depanures 140
3.2.3 Poisson Arrivals and Arbitrarily Distributed Service Time 149
3.2.4 Distribution of Waiting Time 154
3.3 Availability Problems 156

3.3.2 Continuous Inspection-Repair Process 156
3.3.3 Inspection and Repairs at Regular Intervals 166
3.3.4. Deteriorating Systems 176
Problems for Chapter 3 . 178
4. Statistics of Extremes 186
4.1 Introduction 186

4.1.1 Engineering Significance of Extreme Values 186
4.1.2 Objective and Coverage 187
4.2 Probability Distribution of Extremes 187

4.2.1 Exact Distributions 187'
COiHENTS _;~
4.2..2 Asymptotic Distributions 194

4.2.3 The Symmetry Principle .:!04
4.3 The Three Asymptotic Forms 206

4.3.1 Gumbel's Classification 206
4.3.2 The Type I Asymptotic Form 207
4.3.3 The Type II Asymptotic Form 216
4.3.4 The Type III Asymptotic Form 222
4.3.5 Convergence Criteria 228
4.4 Extremal Probability Papers 235

4.4.1 The Gumbel Extremal Probability Paper 235
4.4.2 The Logarithmic Extremal Probability Paper 238
4.5 Exceedance Probability 246

4.5.1 Distribution-Free Probability 246
4.5.2 Significance of Probability Distribution 249
4.6 Estimation of Extremal Parameters 255

4.6.1 Remarks on Extreme Value Estimation 255
..
4.6.2 The Method of Moments 256
4.6.3 The Method of Order Statistics 256
4.6. Parameters of Other Asymptotic Distributions 261
4.6.5 Graphical Method 266
5. Mont~ Carlo Simulation 274
\ 5.1 Introduction 274

5.1.1 An Example of Monte Carlo Simulation 275
5.2 Generation of Random Numbers 278

5.2.1 Random Numbers with Standard Uniform Distribution 279
5.2.2 Continuous Random Variables 281
5.2.3 Discrete Random Variables 288
5.2.4 Generation of Jointly Distributed Random Numbers 290
5.2.5 Error Associated with Sample Size 291
5.3 Variance Reduction Techniques 292

5.3.1 Antithetic Variates 292
5.3.2 Correlated Sampling 296
5.3.3 Control Variates 298
5.4 Application Examples ' 301

X CONTENTS
6. Reliability and Reliability-Based Design 333
6.1 Reliability of Engineered Systems 333
6.2 Analysis and Assessment of Reliability 334

6.2.1 Basic Problem 334
6.2.2 Second-Moment Formulation 340
6.2.3 Linear Perforniance Functions 347
6.2.4 Nonlinear Performance "Functions 359
6.3 Modeling and Analysis:ofUncertainty 383

6.3.1 Sources and Types of Uncertainty 384
6.3.2 Quantification ofUncertainty Measures 387
6.3.3 Analysis of Uncertainty 390
6.3.4 An Alternative Representation of Uncenaimy 406
6.4 hobability-Based Design Criteria 412

6.4.1 On Design and Design Criteria. 412
6.4.2 Second-Mc;>ment Criteria 414
Prob_lems for Chapter 6 < . 434
7. Systems Reliability 448
7.1 Introduction 448
7.2 Multiple Failure Modes 448

7.2.1 Probability Bounds 450
7.3 Redundant and Nonredundant Systems 471

7.3.1 Series Systems 473
7.3.2 Parallel Systems 474
7.3.3 Combined Series-Parallel Systems 478
7.4 Fault Tree, Event Tree Analyses 485

7.4.1 The Fault Tree Diagram 486
7.4.2 Probability Evaluation 495
7.4.3 Event Tree Analysis 498
7.5 Approximate Methods 504

7.5.1 The PNET Method 505

CONTElVTS xi"
Appendix A-Tables 531
Table A."I Values of the Gamma Function 532
Table A.2 CDF of the Standard Extremal VariateS 533

'
Table A.3 The Three Types of Asymptotic Extremal Distributions 537
Table A.4. Weights a1 and b1 for Leiblein's Order Statistics Estimator 538
Table A.S Relations Among Parameters of the T)i:,e III Asymptotic

Distribution 539
Appendix B-TransforTTUJtion of Nonir.orrnal Variates

to Independent Normal Variates 540
B.l The Rosenb_latt Transformation 540

B.2 Determinationof Safety Index 544
References 549
Inde;c 559
-
.
1. Introduction
1.1 AIMS AND SCOPE OF VOLUME II
The principal aims of the present volume parallel those of Volume I; namely, the
modeling of engineering problems containing uncertainties and the analysis of
their effects on system performance, and the development of bases for design and
decision making under conditions of uncertainty.
The basic principles and elementary tools introduced in Volume I are supple-
mented here with additional advanced tools and concepts, including statistical
decision theory, Markov and queueing processes, and the statistical theory of
extreme values. The discussion of numerical tools for probability calculations
would not be complete without mention ofthe Monte Carlo simulation methods;
one chapter summarizes the elements of this calculathnal procedure. One of the
most important applications of probability concepts in engineering is in the
evaluation of the safety and reliability of engineering systems, and the formulation
of associated design criteria. Recent developments for these purposes are the topics
of the final two chapters, with the last chapter devoted specifically to the analysis
of the reliability of technological systems.
The role of judgment is emphasized when dealing with practical problems;
engineering judgments are often necessary irrespective of theoretical sophistica-
tion. However, the proper place and role for judgments are delineated within the
overall analysis of uncertainty and of its effects on decision, risk, and reliability.
Moreover, consistent with a probabilistic approach,judgments have to be expressed
in probability or statistical terms; wherever expert judgments are expressed in
conventional or det'erministic terms (which is often the case), they have to be
translated into appropriate probabilistic terms. Methods for these purposes are
suggested.
1.2 ESSENCE AND EMPHASIS
The main thrust and emphasis in each of the ensuing chapters may be summarized
as follows.
1.2.1. Decision Analysis (Chapter 2)

The goal of most engineering analysis is to provide information or the basis for
decision making. Engineering decision making could range from simply selecting
the size of a column in a structure, to selecting the site for a major dam, to deciding
whether nuclear power is a viable energy s~urce. Unfortunately, uncertainties are
invariably present in practically all facets of engineering decision making. In this
I
2 INTRODUCTION
light, some measure of risk is unavoidable in any decision made (i.e., alternative
selected) during the planning and design of an engineering system. A systematic
framework for decision analysis under uncertainty, in which the feasible alternatives
. are identified and the respective consequences evaluated, could be very useful. The
tools for such decision analysis are the subject of Chapter 2
The decision tree model is introduced to identify the necessary components of a
decision problem, consisting of the feasible alternatives, the possible outcomes
associated with each alternative and respective probabilities, and the potential
consequences associated with each alternative. In short, the decision tree provides
an organized outline of all the information relevant to a systematic decision analysis.
.Various engineering examples are used to illustrate the concepts involved in
formal decision analysis. The concept of value of informarion is discussed with
respect to whether or not additional i~onnation should be gathered before making
a final decision. The probiem assOciated with sampling, riamely, determining the
parameter estimates for design, and the development of optimal sampling plans
may also be formulated as a decision problem. ,
Elementary concepts of utility theory are introduCed a~ a generalizi:d measure of
value, on the basis of which the relative significance of various potential conse-
quences of a decision may be evaluated. Specific c:ase studies of complex engineering
decision problems are discussed, including those itivolving multiple objectives.
l.l.2_i- Markov, Queueing, and Availability Models (Chapter 3)

1be ,performance characteristics of an engineering system may frequently be
classified into different discrete states. For example, the potential effects of an
eartbquake on a structure may be classified under several states corresponding to
di:>tinct damage levels; the financial status (e.g., _in terms of cash flow) of an
engineering contractor may represent the various states; including bankruptcy ;
. and the number of vehicles waiting (queu,e length) at a toll booth could represent
a state of a toll system. The conditions of these systems may change or move from
one state to another in accordance with some probability law. The probability
that the system will be in a particular state after a given number of moves may then
be of interest.
The Markov chain is a probability model specifically designed to analyze a
system with multiple states. The basic concepts of the discrete parameter homo-
geneous Markov chain model are first introduced. Queueing models are then
shown to be examples of a continuous parameter Markov chain. Steady state
queue length probabilities are derived for specific queueing systems with Poisson
arrivals.
The performance of an engineering system may often be divided into two states;
for example, safe and unsafe states, or operating and nonoperating states. The
study of the transient behavior of this two-state system is known as the availability
problem. Problems of system availability (system in a safe state) may include a
maintenance program, that is, inspection and repair at regular intervals, in which
case the renewal theory is a useful model. Availability problems of interest in
engineering would include the availability at a given time for a system with or
without maintenance.
ESSENCE AND ntPHASIS 3
A number of engineering applications of the Markov process, including the

queueing and availability models. are illustrated in Chapter 3.
1.2.3 Statistical Theory of Extremes (Chapter 4)

Extreme values or extremal conditions of physical phenomena are of special
interest in many engineering problems. especially in those concerned with the
safety and reliability of engineering systems, and/or involving natural hazards.
The statistical theory of extreme values. therefore; is of special significance in
engineering problems concerned with risk and reliability.
Much of the theory of statistical extremes is available in the literature; however,
the available material is primarily of a mathematical nature and is hardly accessible
to many engineers. One .Jf the purposes of Chapter 4is to summarize and h,ighlight
the essential concepts of statistical extremes, and to emphasize the practical
significance of these concepts in engineering problems. In particular, the asymptotic
property of the statistics of extremes enhances the significance of the extremal
theory; understanding this concept is essential for many applications. The asymp-
totic property and its associated concepts are, therefore, stressed throughout the
chapter, and illustrated with numerous examples of engineering problems.
1.2.4 Monte Carlo Simularloo~Cbapter 5)

As the complexity of an engineering system increases, the required analytical model
may become extremely difficult to formulate mathematically unless gross idealiza-
tion and simplifications are invoked; moreover, in some cases even if a formulation
is possible, the required solution may be analytically intractable. In these'instances,
a probabilistic solution may be obtained through Monte Carlo simulations. Monte
Carlo simulation is simply a repeated process of generating deterministic solutions
to a given problem; each solution corresponds to a set of deterministic values of the
underlying random variables. The main element of a Monte Carlo simulation
procedure is the generation .of random numbers from a specified distribution;
systematic and efficient methods for generating such 'random numbers from several
common probability distributions are summarized and illustrated.
Because a Monte Carlo solution generally requires a large number of repetitions,
particularly for problems i.tivolving very rare events, its application to complex
problems could be costly. There is, therefore, good reason to use the Monte Carlo
approach with some. caution;_generally, it should be used only as a last re5ort, that
is, when analytical or approximate methods are unavailable or inadequate. Often,
Monte Carlo solutions may be_the only means for checking or validating an.
approximate method of probability calculations.
A number of case study problems are illustrated, to demonstrate the simulation
procedure as well as the type of information that may be derived from a Monte
Carlo calculation. ,
1.2.5 Reliability_aiid Reliability-Based De5ign (Chapter 6)

. .
The safety and/or performance of an engineering system is invariably the principal
technical objective of an engineering design. In order to achieve some desired level
of reliability, proper methods fo_r its evaluation are, of course, required. As this
4 INTRODUCTION
must invariably be done in the presence of uncertainty, the proper measure of

reliability or safety may only be stated in the context of probability. Indeed,
consistent levels of safety and reliability may be achieved only if the criteria for
desigri are based on such probabilistic measures of reliability.
Engineering reliability and its .significance in engineering design is a -rapidly
growing field; the most recent and practically useful developments are presented
in this chapter. Applications in the various fields of engineering are illustrated
with emphasis on civil engineering. It goes without saying that problems of safety
and reliability arise because of uncertainty in design. The quantification and
analysis of uncertainty are, therefore, central .issues in the evaluation of reliability
and the development of associated reliability-based design. The statistical bases and
methods for these purposes are widely illustrated in this chapter.
1.2.6 Systems Reliability ((:bapter 7) ..

. .
Although the reliability of the total system is of principal concern, system reliability
is nonetheless a function of.the reliabilities of it!' constituent components ithat are
covered in Chapter 6); that is, the determination of system reliability must in-
variably be based on reliability information for the components. Moreov~.
engineering designs are invariably performed at the component level, from which
the reliability of the system . is evaluated through analysis; techniques for this
purpose are presented in Chapter 7.
From a reliability standpoint, a system is characterized by multiple modes of
failure, in which each of the potential failure modes may be composed of component
failure events that are in series or inparallel, or combinations thereof(especially for
a redilndant system). In the case of a general redundant system, the redundancy may
be of;the standby or active type. Depending on whether the redundancies are active
. ~.nriby, the reliability of the system as well.~!:~ its analysis will be different.
- In the case of complex systems, the identification of the potential modes of
failure may be quite involved, requiring a systematic procedure for identification,
such as the fault tree model. Moreover, the potential consequences of a failure
(or initiating event) of a system may vary, depending on the subsequent events that
follow the particular initiating event. The event rree model may be used to facilitate
the systematic identification of all potential consequencr-~. The applications of the
fault tree and event tree models are illustrated with a variety o( engineering examples
in this final chapter.
' ....
2. Decision Analysis
2.1 INTRODUCTION
Making technical decisions is a .necessary part of ~~gineering planning and design :

in fact, the primary responsibility of an engineer is to make decisions. Often, such
decisions have to be based on predictions and information that invariably contain
uncertainty. Under such conditions, risk is virtually unavoidable. Through.
probabilistic modeling and analysis, uncertainties may be modeled and assessed
properly, and their effects on a given decision accounted for systematically. In
this manner, the risk associated with each decision alternative may be delineated
and, if desired or necessary, measures taken to control or minimize the corres-
ponding possible consequences.
Decision problems in engineering planning and design often also require the
consideration of nontechriidi.l:factors, such as social preference or acceptance.
environmental impact, and sometimes even political implications. In these latter
cases, the selection of the .. best" decision alternative ~annot be governed solely by
technical considerations. A systematic framework that will permit the considera-
tion of all facets of a decision problem is the decision 11UJdel.
The elements of the decision model and the analyses involved in a decision
problem are developed in this chapter. Both single- and multiple-objective
engineering decision problems are discussed and illustrated.
2.1.1 Simple Risk-Decision Problems \

In any decision analysis, the set of decision (or design) variables should first be
identified and defined. For example, in the design of storm sewers, the engineer
may select a pipe diameter to provide a desired Bow capacity. Because of variability
in the future storm runoff; there is a probability that the ftow capacity of this pipe
may not be sufficient. Since this probability is related to the pipe size, the proba-
bility level may also be the decision variable instead of the pipe diameter.
In order to rank various designs, an objective function is usually defined in
terms of the decision variables. This function is frequently-expressed in monetary
units representing total benefit or total cost. It is obvious that the optimal design
will be determined by the values of the decision _variables that will maximize the
benefit or minimize the loss function. In some cases, in which the objective function
is a continuous function of the decision variables, the calculus of maximization and
minimization provides a convenient tool for optimization. Formally, if
X 1, X 2 , , XN denote the set of decision variables, the optimal design will be
5
6 DECISION ANALYSIS
given by values of the decision variables satisfying the following set of equations:
oF(Xl, ... ,X.)=O
i = 1, 2, ... , n (2.1)
ax,
where F(X ~o , X.) is the objective function. The second partial derivatives may
be examined to determine whether Eq. 2.1 yields the maximum or minimum
objective function. The procedure is illustrated in the following examples.
EXAMPLE 2J (excerpted from Shuler,1967)
A contractor is preparing a bid for a construction project. Based on experience, it is his judgment
that the probability of winning a bid depends on his bid ratio. R, as follows:
- . p = f.6 - R; 0.6 ~ R ~ t.6
. .. ' . .
in which R is the ratio of lris !;lid. price to the total estimated cost (Fig. E2.1 ).
O_bviously. th~ decision ,arlable here is the bid price B that the contract~r will submit The
corresponding objective function may be the expected profh from the project;
\ X = (B - C)p + 0 (I - p)
B-C
=-c-.pC
= (R -._l)pC
= (-Rz + 26R- 1.6)C; for0.6 ~ R S 1.6
'where C is the estimated cost o( construction. The optimal decision requires the maximization
:or.x with respect to R; thus, setting
dX
..
~
dR = C(-2R + 2.6) 0 =
we obtain R = 1.3. Since dl X fdR 2 < 0 at R = i.3,' ihe bid 'f'atio of 1.3 will maximize X.
Therefore, the optimal bid for the contractor should be 1.3 times the cstiinated cost.
The above solution assumes that there is no monetary gain or loss if the contractor does not
win the job. Suppose that the construction crew would be idle and the contractor will be
operating at a loss L if he fails to win ~is job. ln such a Case, the objective function should
include an expected loss; suppose L = O.lC. Then, the expected overall gain would be
X = (R - l)pC- (l - p)L
= [(R - 1)(1.6- R)- (R - 0.6) x O.l)C
= [-R' + 2SR- 1.54)C
1.0
p = l6- R
0.2
07-~~~--~~--~----~
0.6 QB lO 1.2 1.6 ,: R, .
Figur~ 2.1 Probability of winning versus bid 'ratio.
INTRODUCTION - .7
Then,
dX
-
dR
= -2R + 2.5 =0
yields R = 1.25. Hence, the optimal bid becomes 1.25 times the estimated cost. A lower bid is
required here so that the contractor will increase his chance of winning the bid, thus reducing
the chance of idling his construction crew.
EXAMPLE 2.2
The construction of a bridge pier requires the installation of a cofferdam in a river. Suppose the
occurrence of floods follows a Poisson distribution with a mean occurrence rate of 1.5 times per
year, and the elevation of each flood is exponentially distributed with a mean of 5 feet above
normal water leveL Each time the cofferdam is ovcrtoppcd,"the expected loss resulting from a
possible delay in construction and pumping cost is estimated to be $25,000. Since the flood
elevation can be predicted sufficiently ahead of time to effect the evacuation of personnel
working inside the cofferdam, the possibility of loss due !O workmen being trapPed may be
neglected. Suppose that the construction cost of the cofferdam is given by
c. = c. + 3000h
where c. includes the cost of a colfcrdam foundation and ncccssary construction to dam the
river to nonnal water level, arid "h is the height of the cofferdam (in feet) above normal water
level. Determine the optimal height of the cofferdam if it is expected to be used over a two-year
period.
A convenient choice of the objective function here is the expected monetary loss Cr. which
would consist of the construction cost and the expected total Joss from the flooding of the
cofferdam during the two years of service. The expected loss in each flood is
C = E(loss Iovertopping)P(overtopping)
Hence, the expected total loss from all floods during the two-year period {the time value of
money neglected) is
..
C1 = l:E(lossli floods)P(i floods in 2years)
10
"' 31e- 3
= L:i(25,000e- 1115) - -
10 I!
., Jle-J
= L:
zs;oroe-111' 10
i - .,-
I.
= 75,000e- 1115
The total expected monetary loss becomes
Cr= C.+ C1
= c.+ 3000h + 75,000e~lll'
Clearly, his the only deCiSion variable; whereas c. is a constant independent of h. Differentiating
Cr with respect to h, we obtain the optimal cofferdam heigh~ as foUows:-
d~r = 3000 + 75,000e- 1115( -!) =0

8 DECISION ANALYSIS
-
E
Construct ion
Cost, Cc
Pumpino + Oetoy
Cost, Ct
0~~~~--~~~--~~~--~~--~
0 2 3 'I 5 6 7 B 9 10 II
h (in feet)
Figur~ 2.1 Costs as functions of cofferdam elevation above normal water level.
or
Thus,
h.,.. = 5 In (5) = 8.05 ft
Therefore. the optimal height of the cofferdam should be approximately 8 feet above the normal
wat~ level of. the river. Figure E22 shows the various cost components as functions of the
decision variable h. It may be observed that as h increases, the construction cost increases;
whctcas the expected flood Joss decreases. The total expected loss takes on a minimum when
the slopes of these two functions arc equal but of opposite signs. At this point, the marginal
increase in Cc is balanced by the marginal decrease inC1 ,)be value of c., which was assumed
to be S40,000 in Fig. E2.2; simply adds a constant value to the ordinate at each value of h, and
does not affect the determination of the optimal value of h. The probability that a cofferdam
with h = 8.05 It will be overtopped in a flood is given by
- . .. 1 . .
P(flood level > 8.05') = J.
Se-;tJ' dx = 0.2
1 .0$
Alternatively, the objective function may be formulated using the probability of overtopping,
p, as the decision variable. Since the construction cost is a function of h, we need to express h in
terms of the overtopping probability p. Based on the exponential distnbution assumed for the:
flood level it can be shown that
p=
1..
-e-~" dx
J. 5
or
h = -5 Inp
Hence, the total expected cost becomes
CT = c. + 3000(- 5 In p) + 3(25.000)p
=c.- 15,000 In p +15,000p
Differentiating CT with respect to p, we obtain the optimal probability as follows :
. .
dC 1 .
-dT = -15,000 '- + 75,000 = 0 I I .
p p
INTRODC/CT/Ol'i 9
from which
p.;_. = 0.2
which is the same as obtained earlier. Based on this value of p... . the corresponding optimal
elevation of the cofferdam (above normal water level) is determined as
h.,.= -5 In p.,..
= - 5 In (0.2) ,
= 8.05 ft
The first approach described above directly gives the optimal value of the design variable,
whereas the second formulation directly yields the optimal probability level In cases involving
several cofferdams, each subject to a different distribution of floods, expressing the . overall
objective function iD. terms of a common rislt level could simplify the optimization process. In
any case, the two methods should lead to identical optimal designs, provided there is a cor-
respondence between the design variables and the probability of overtopping.
2.1.2 Characteristics of a General Decision Problem

The applicability of the optimization procedure presented above, based on calculus,
is limited; first' of all; the p~jective function must be expressed as a continuous
function of the decision variables. Unfortunately, this may not be the case for many
engineering decision problems. Consider a dam that has been proposed as a:
possible solution for flood control in a certain drainage basin. Its elevation may be
a decision variable; the darn site may also need to be determined. Moreover, other
forms of flood control strategies such as diversion channels, levees, multiple
reservoirs may be possible alternatives to be considered in the decision analysis.
Thus, it will be difficult, if not impossible, to obtain a continuous objective function
in terms of all the decision variables.
Sometimes, a decision may not be based solely on the available information. If
time permits, additional information may be collected prior to the final selection
among the feasible design alternatives. In engineering problems, additional data
could take the form oflaboratory or field tests, or further research. Since there could
be a variety of such schemes for collecting additional information, this would
further expand the spectrum of alternatives.
In short, in more general decision problems, a framework for systematic analysis
is required. Specifically, the decision analysis should at least include the-following
components:
(1) A list of all feasible alternatives, including the acquisition of additional in-
formation, if appropriate.
(2) A list of all possible outcomes associated with each alternative.
(3) An estimation of the probability associated with each possible outcome.
(4) An evaluation of the .consequences associated with each combination of
alternative and outcome.
(5) The criterion for decision.
(6) A systematic evaluation of all alternatives.
A decision model that considers all these basic components is presented in

Section42
10 DECISION ANAL YS/S
2.2 THE DECISION MODEL
l.l.l Decision Tree

The various components of a decision problem may be integrated into a formal
Ia yout in the form of a decision tree, consisting of the sequence of decisions-namely,
a list of feasible alternatives ;the possible outcomes associated with each alternative;
the corresponding probability assignments; monetary consequences and utility
evaluations (see Section 2.4).1n other words, the decision tree integrates the relevant
components of the decision analysis in a
systematic manner suitable for an
analytical evaluation of the optimal alternative. Probability models of engineering
analysis and design may be used to estimate the relative likelihoods of the possible
outcomes, and appropriate value or utility models evaluate the relative desirability
of each consequence. . .
Figure 2.1 shows a generic example of a decision tree with three ahernatives, in
which the third alternative involves performing an experiment to gather additional
information prior to any final decision. The word "experiment" should be" inter-
preted in a broad sense; covering any method of gathering additional data. The
. following notations are used in Fig. 21 :.
a1 = Alternative i. .
(Ji = Outcomej.
"""' e1 = Experiment k designed to gather additional information.
z, = Experimental outcome l.
u(-ri'1, 8) = Utility value corresponding tci alternative a1 and outcome 81; if the
:: utility depends pn experiment e1 and the corresponding experimental
~- outcome z, it will be denoted as u(e1 , z" a;. 81).
_,;..:..:...;...:=..::.:..!..::.:..!~:.... u(e 1 ,11,o2 .6, l

'-"--<~.:;c~:.r..:..:::.!.!;:.!.!;=._ u(e
1,1 1,o 2 ,62 1
r-"'--<:r---'-'-'-..:..:..:"""-';..:..:..;=-ule 1,z 2 ,o 1 ,6 1l
'-"'"~'-'-'..a..:.-'-'-'"'"""- ule 1,z 2 , o 1 ,62 l
Oecisoon Node
o Chonu Node
I, '
Figure 2.1 A generic decision tree.

THE DECISION .~IODEL J1
The decision tree in Fig. 2.1 begins with a square node, called a decision node. at
which point there are alternatives at> a 2 , and a 3 With each alternative, there may
be several possible outcomes. shown as branches originating from a circular node.
called a chance nocle. At each chance node, nature controls what will occur sub-
sequently. In this example, 8 1 , (} 2 are the two possible outcomes associated with
alternatives a 1 and a 2 These outcomes may or may not depend on the preceding
alternative; in general their probabilities are conditional on the preceding alter-
natives; for example, P(811a 1). Alternative a 3 is followed by experimental outcomes,
whose probabilities will generally depend on the experiment. Hence, the .prob-
abilities of (z1) are also conditional on et; that is, P(zd e.J. Another decision node
follows each experimental outcome, denoting that a decision between alternatives
a 1 and a 2 will be required after observing the specific additional information. The
probability of 81 in the subsequent branches would be updated based on the
particular experimental outcome. Hence, in general. this probability is expressed as
P(81 lz1, et, aJ.
The outcomes from a chance r.ode are mutually exclusive and collectively
exhaustive; thus, the sum of the conditional probabilities at each chance node
should add up to unity. Sometimes there may be a continuous spectrum of outcomes
originating from a chance node, such as those given by the values of a continuous
random variable. In such cases; the PDF of the random variable will be used to
denote the relative likelihoods of these branches. The desirability of the consequence
of each sequence or path in the tree is measured by its .. utility value" recorded at
the end of the sequence, such as u(a1, 81) or u(et, z., a~o 8).
The following examples illustrate the application of the decision tree model in
engineering. In Section 22.8, several in-depth examples of complete decision
analysis further illustrate the decisi~n tree model.
EXAMP~E 2.3
Two alternative designs arc considered for the structural scheme of a building. Design A is
based on a conventional procedure, whose probability of satisfactory performance is 99 ~.. and
costs $1.5 million. Design B is based on modem concepts and will reduce the cost of the building
to Sl million. The reliability ofdesign B is not known; however, the structural engineer estimates
that if his assumptions .arc correct, the reliability of satisfactory performance will be 0.99;
whereas, if his assumptioll5 arc not valid, the reliability is only 0.9. According to his judgment,
he is only 50% sure of the validity of his assumptions. Suppose the cost of unsatisfactory per-
formance is SlO million. The decision tree would be as shown in Fig. E23; the utilities, in terms
of losses _in million S. arc indicated for each sequence of the tree.
EXAMPLE 2.tl
In order to prevent or minimize thermal pollution from the operation of a nuclear power plant,
cooling lakes arc commonly used to allow the heat from the plant to dissipate through a
natural process. At one location, the dikes for the lake were constructed on a relatively pervious
ground where water seepage is possibli:. Such seepage can cause a loss ofwater as well as become
a nuisance to the surrounding a.rca.. As a remedy, pumps may be installed on the downstream
slope to pump back the water to the cooling lalcc, as shown schematically in Fig. E24a. Based on
a theoretical seepage model and soil data from the site. the ftow of water is estimated, for
simplicity, to be either Q1 = 100 gal/min or Ql = 120 gal/min, with relative likelihoods of9 to 1.
/2 DECISION ANALYSIS
Loss
(In Million $)
(0.99)
1.5
11.5
1.0
11.0
1.0
11.0
A Conventional Oesion
B s New Des ion
V = Assumptions Valid
V Anumptions Invalid
S Sotisfottory Performonte
S" Unsatisfactory Performance
Figur~ ~.3 Decision tree for structural design problem.
..
_......_
.. -4"
There are two pump system capacities that the engineer may order. namely pump system A with
100 gal/min or pump system B with 120 galjmin. which cost 17 and 20 units (I unit = S 100),
respectively. Ifpump system A is insuffieient, an additional small pump system C with 20 galfmin
will be..Drdered and the overall. additional expense. i11cluding inconvenience and temporary
disruption of performance, is estimated to be 8 units. Over the expected service period, the total
operatiOnal (OMR) cost. including maimcnance and repair costs. discounted to present values
are 22. 21. and 5 units for pump systems B. A, and c; resJ)ccthely. Two units of the OMR cost
of pump B could be discounted if the flow Q is actually 100 gal/niin.
Another feasible alternative is to install a barrier of bentonite seal beneath the upstream end
of the embankment. The total material and installation cost of the seal will be 41 units, and there
are negligible maintenance costs over the years.
The decision tree for this example is shown in Fig. E2.4b. The costs associated with each path
of the decision tree are clearly dependent on the particular alternative and the associated
outcome.
Coolino Loke
Pump System
Figure E2.4a Cross-section of embankm.e nt.

THE DECISION .\tO DEL JJ
Present
Worth of
Initial Operat111g Total
Cost Cost Cost
o,(0.9l
17 -21 -38
Pump System A Add
02(01) Pump System C
-25 -26 -51
0,(091
-20 -20 -40
Pump System 8
Q2(0.1l
-20 -22 -42
\:S::.:e:.::o~l_ _ _ __ _ __ _ __ __ -41
0 -41
Figur~ El.Jb Decision tree for seepage problem.
EXAMPLE 2.5 (excerpted from Haefner and Mor/ok, 1971)
Improving the Safetyof a short section of a rural road with a complex alignment through geo-
metric design changes is proposed: For simplicity, only one improvement scheme is considered:
namely, to improve the alignment and profile and install a median barrier. The decision alterna-
tives are simply whether to adopt this improvement scheme or to make no improvemeiu. The
decision tree is shown in Fig. E2.5 (following the branch indicating ~No Prediction Model").
The outcome is described by the various states of accic!ent and fatality rate encountered. To
simplify the problem, accident and fatality rates are assumed to be either high or low, thus
leading to four possible combinations of outcomes-namely, H . H, (high accident rate. high
fatality rate), HAL,, LA H ,, and L .. L,. The probabilities for each of these outcomes (as shown
in Fig. E2.5) have been assessed with existing information. The total cost associati:d with each
path consists of the road improvement cost, if any, and the accident and fatality losses.
It is common in highway engineering-to employ a prediction model, such as the K.ihlberg
and Tharp m6del or the Mann and Dart (Louisiana) model, to predict accident rates for a given
design alternative. The models pr~vide information on whether or not the proposed road
improvement will result in a significant reduction in total accidents and fatalities.
There are two stages in the decision process. The first decision is to select the particular
prediction model. Based on the model prediction result, namely H~H', or H~L',, or L~H', or
L~ L',, a decision is then made whether or not to improve the road condition. A representative
path of the decision tree for tbe Tharp model is shown in Fig. E25.
Observe that the probabilities associated with various fatality and accident combinations
do not change if no improvement is adopted. The Tharp model applies only to the prediction
associated with the improved.scheme. Depending on the model prediction, the updated prob-
ability for e:u:h fatality and accident combination will be different (see Problem 2.11). The cost
for performing the prediction model is also included in the total cost for appropriate paths.
2.2.2 Decision Criteria

The objective of decision analysis is to make the "best" decision. Only in rare
occasions would a decision maker be comparing alternatives whose monetary
values are known deterministically. In this case, he would obviously choose the
alternative with the highest value. However, most decisions are made under con-
ditions of uncertainty, in which the possible utility values of u 1, u2 , , u~ of a
given alternative may be stated only in terms of respective probabilities
P~o p2 , .._: ._ ~~ Also, depending on the temperament, experience, and degree of
14 DECISION ANALYSIS
Total Cost
(In~ Dollars)
HAHF (0.3)
19
(Q3)
12
(0.2)
9
(0.2)
6
(0.5)
15
(0.3)
B
(0.2)
5
(0)
2
(0.22)
21
(0.22)
14
(0.37)
II
(0.19)
B
(0.5)
17
(0.3)
. ::.L. 10
~ (0.2) "
7
...::: (0)
~ 4
.:::..
-~ .
.:....:
PrediCiive
Mocle I I
Model
Prediction
Results
I Alternative
Actions
Act uol Result
Figur~ E2.5 Decision tree for geometric ?esign in road i~provemcnt.
risk-aversiveness of the decision maker; the "best" decision may mean different
things to different peopleand at different times. Nevertheless, some general criteria
for a decision can
be identified. ..
Consider a simple decision problem, in which an owner has an option to install
an emergency electrical power system to guard against potential utility power
failure. The additional cost for the system will be $2000 per year; whereas if power
failure occurs and no emergency power supply is available, he could incur a loss
of $10.000. Based on his experience and judgment, he believes that the annual
probability of power failure is only 10%. Assume that the chance of more than one
power failure in a year is negligible. The decision tree for this example is shown in
Fig. 2.2 If the owner installs the emergency system, he would have to spend $2000;
THE DECISION .HODEL ]5
Power Failure (0.1)

-2000
No Power Failure (0.9)
-2000
n: No Power Failure (O.Il
-10000
Emergency System
No Power Failure (0.9)
0
Figure 2.2 Decision tree for installation of emergency power system.
if he does not, he has a 90% chance of not spending any money but also a 10%
chance of losing $10,000.
If the owner is extremely pessimistic (believing that nature always works against
him), he would try to minimize his loss by installing the emergency system, because
the maximum loss in this alternative is only $2000, relative to $10,000 for the other
alternative. On the other hand, if he is an extremely optimistic person (believing
that luck is always on his side), he would not install the emergency power system,
since this alternative will yield the maximum gain; namely, it may cost the owner
nothing compared to S2000 for the other alterruiiive. The first basis of selecting
the optimal alternative is calledthe mini-max criterion, in which the decision maker
selects the alternative that minimizes the maximum loss. In the. latter case, he tries
to maximize his maximum possible gain among the alternatives, referred to as the
maxi-max criterion.
Neither of these two criteria are practical in the long run. If the mini-max
decision maker consistently follows this criterion, he will never venture into a
decision that may result in substantial gain as long as there is a finite (though small)
chance that he may be incurring a loss, whereas the maxi-max decision maker will
never choose the realistic alternative provided that there is a finite chance for a
more attractive outcome. Any rational decision maker should consider the relative
likelihood associated with the gain or loss in each alternative, rather than strictly
following these criteria.
Mtuimum Ezpected Monetfl1'y Value Criterion When the consequences associated

with each alternative in a decision analysis can be expressed in terms of monetary
values, a widely used criterion for decision is the maximum expected monetary gain.
Suppose d 11 denotes the monetary value of the jth consequence associated with
alternative i, and p11 is the corresponding probability; then the expected monetary
value of the ith alternative is
E(a,) = LJ P1A1
.
. (2.2)
The optimal alternative according to this max EMV criterion is the one wh~se
expected monetary value "is
(2.3)
EXAMPLE 2.6
Suppose the owner referred to in Section 2.2.2 follows the max EMV decision criterion. He
would compute the EMV of the two alternatives as (see the decision tree in Fig. 2.2):
. E(I) = 0.1 X ( - 2000) + 0.9 X ( -2000)
= - 2000 dollars
and
E(ID = 0.1 X ( -10;000) + 0.9 X (0)
= -1ooo dollars
Comparing the two expected monetary values, the owner should not install the emergency
power supply system.
Both the maxi-max and mini-max criteria ignore the probabilities of the possible
outcomes of an alternative. The available information is, therefore, not fully
utilized in the decision analysis. With the maximum EMV criterion, a decision
maker is systematically weighing the value of each outcome by the corresponding
probability.
It is conceivable that an event with an occurrence probability of 99.99% may
still fail to occur, in spite of the extreme odd. Nevertheless, if a decision maker
consistently bases his decisions on the max EMV criterion, the total monetary
vaiJ~: obtained from all his decisions (in the long run) will be maximum
~other common decision criterion known as the maximum expecred uriliry
criterion is presented in Section 2.4.6.
2.2~ Decision Based on Existing Information-Prior Analysis

Once the decision tree has been established, with th: probabilities of the possible
outcomes at the chance nodes and the consequences ofall respective paths evaluated
based on existing information, the expected monetary value of each alternative
may be computed according to Eq. 22. A decision analysis ~ased entirely on
available prior information is called prior analysis; if this analysis is updated sub-
sequently with additional new information, the latter is called rerminal analysis
(described in Section 224).
Sometimes a given alternative may involve a series ohubsequent chance nodes
such as that shown in Fig. 2.3. Suppose. at chance node A, the outcomes are the
possible values of a discrete random variable X; whereas at chance .node B, the
outcomes involve a continuous random variable Y.
The expected value calculation for alternative a; starts at the last chance node;
namely, the chance node B in Fig. 23. At node B, the expected value of alternative
a;, given X = x,, is
E(adx1; = f:~ E(adx1; y)fr(J') dy

_ = J~., d~x,, y,arlfr(Y~ dy
. . . :
THE DEC/S/O:V .HODEL 17
d lj, y, a;)
Figuu lJ Decision tree with series of chance nodes.
where fr(Y) is the PDF of Y. Subsequently, at node A.

E(a,) = L E(atl.-cJ)PJ
j
where p1 is thePMF of X. Sometimes the complete probability distribution of the

random variable may not be necessary to calculate the expected value. For instance,
if the cost function is quadratic in y, namely
d(x1 , y, a1) = ay2 +by
.
+ c '
then, at node B,
E(ad_x!) = J~"' (ay 1

+by+ c)f~(y) dy
. ... ~ -
= aE(Y 2 ) + bE(Y) + c
= a[Var(Y) + E2(Y)] + bE(Y) + c
indicating that only the mean and variance of Y are required to compute the
expected cost at node B.
EXAMPLE 2.7
A new waste treatment process is designed for installation in '!- local community. Because the
proposed waste treatment process is based on a completely _r:1ew concept, its efficiency is yet
unknown. The project engi.necr has a choice qr designing a small unit A. or a larger unit B.
Assume that the efficiency could be classified irfo two levels: namely, EH and EL for high an?
low efficiencies, respectively. If the efficiency is actually high, unit A will be the appropriate
design, whereas unit B is basically designcdfor low efficiency. The engineer estimates a prob-
ability of 70% that the efficiency of the proposed treatment process will be high. The decision
tree is shown in Fig. E27, in which the relative monetary loss (in SIOOO) for 1=ach path.in the tree
has been evaluated.
Relative Monetary
Value ($1,0001
EH (0.7)
0
EL (0.3)
-100
EH (0.7)
-50
EL (0.3)
-20
Figarr E2.7 Decision tree for waste treatment system.

'!he expected monetary value for the two designs arc
E(A) = 0.7 X 0 + 0.3 X (-100) = -30
and
E(B) = 0.7 X (-SO)+ 0.3 X (-20) = -41
' With the given information, design A should be selected as the proposed design; the correspond-
ing expected monetary loss is $30,000.
EXAMPLE 2.8 (Sensitivity to Loss Functions)
After feasibility studies on the possible sites for a propOsed earth dam, two locations (sites A
and B) remain for final consideration. From a cost-benefit analysis of the project, it is shown
that locating the dam at site A will yield an overall return of S22 million, whereas the dam at site
B will-give an overall return of.S20 million. However, the eartbdams at these sites are also
susceptible to earthquake damages. From a seismic study of the area (that takes into account
the existing fault lines, geological .conditions; and available earthquake Tecords), the annual
mean occurrence rate of earthquakes with damaging intensity at. site A is estimated to be O.Ql
or 0.02 with equal likelihood, and the corresponding occurrence rate for site B is 0.005 or 0.01
per year with equal likelihood. The occurrence of an earthquake may be modeled by a Poisson
process. Assume that earthquake damages to the dam will be repaired after eacfi earthquake
and the dam restored to its original condition. Let X be the number of earthquakes occurring
in the life of the dam, assumed to be 50 years.
Determine the expected earthquake loss for each of the following cases. All losses are
measured in terms of million dollars.
Case)
The earthquake damage function is
L= 5X
The expected Joss given that site A is selected with a mean rate of earthquake occurrence v1 is
Cit" -
E[LIA, v1] = I 5x Pr[xlv.J

zO
= 5E(XIv.J
= 5(50vJ)
= 250v1
Hence,
2
E[I,.IA] = I_E[LIA, ''j]P[,1 i.A]
J l
'""' (250 X 0.01}(0.5) + (250 X 0.02){0.5)
= $3.75 million
Similarly.
2
[LIB)= .LE[LIB, ,.JP[v11B]
jJ
= (250 X 0.005}(0.5) + (250 X 0.01)(0.5)

= $1.875 million
THE DEClSlOt'i ."WODEL 19
Ca.se2
Suppose the earthquake damage function is
L = 5X 2
Then.
E[LjA, v1] =
"' 5x 2 Px[.,;iv ]
L 1
zO
= 5E[X 2 1vJ
= 5[Var(Xjv1) + E1 (X!v1)]
= 5[(50v1) + (5<iv1) 1 ]
= 250v1 + 12,500vj
Hence, in this case,
1
E[LIA] = L (250v1 + 12.500vJ)P[v1!A]
}I
= (250 . + 12,500 x-&.01 )(0.5) + (250 x

x o.ot 1
om + "12.500 x o.o21 )(0.5)
.
= $6.875 million -
E[LIB] = (250 X 0.005 + 12,500 X 0.005 1 )(0.5) + (250 X 0.01 + 12.500 X 0.01 1)(0.5)
= S1.fi56 million
Case3
Finally, if the earthquake damage function follows an exponentia1loss function,
. L=~
\ in which c = 3. Observe that L = 1 when X = 0; the small loss here could denote the mainten-
. ance cost through the years. Then,
E(LIA) = O.Se'o,o.olxl-lf + 0.5e'o(o.o1M3-Il

= 0.5e+ 0.5e1
= $5.06 million
Similarly,
E(LIB);,. O.Se'Oio.oo'K1l + 0.5e'OIO.OIM1l
= O.Se0 ' + 0.5e
= $2.186 million
The decision tree is shown in Fig. E2.8.
Average Total Earthquake Damage

CoM I Case II cosem
5X 5X 2 ~X
l =0
0 0 ..
x I 3
5 5
; 2
10 20 9
1 z 0
0 0 I
x=l - 5 5 3
. z 0
0 0 I
x= I
5 5 ~
z ('
0 0
,. QOI (0.5) x =I
5 5 3
' ' '--------
Figur~ 2.8 Decision tree for dam site selection.
Table E2.8 summarizes the expected benefit and damage for the three cases considered.
Depending on the specific loss function, A orB may be the optimal site for the dam.
Table E2.8 Expected Benefits and Damages for the Dam Sites
Damage Cost Net Benefit
Site Location Benefit Case I Case 2 -case 3 Case 1 Case2 Case3
A 22 3.75 6.875 5.06 18.25' 15.125 16.94

B 20 1.875 2.656 2.186 18.125 17.344' 17.814'
tndicates optimal alternative.
EXAMPLE 2.9
Suppose at a certain stage of construction, there are only 30 days left before a project's scheduled
completion date. Because of bad weatber in the first phase of the project, construction has not
been proceeding at the anticipated rate. Based on the contractor's judgment, the project could
still meet the deadline if exceptionally good weather (condition G) prevails in the next 30 days.
He estimates that this may occur with 20% probability. However, ifnormal weather (N) prevails.
completion will be delayed With. a mean delay of 5 days and a standard deviation of 3 days
beyond the scheduled completion date; if bad weather (B) continues. the corresponding mean
and standard deviation of delay will be 15 and 5 days,-respcctively. Assume that normal and bad
weather conditions are equally likely to occur in the remaining time required and the cost of
delay is (0.lx2 + x) in thousands of dollars, where x is the length of delay in days. This cost
function implies that the loss per day of delay increases as the delay becomes longer.
The contractor has an option to launch a crash progr.am. at an additional cost of $)0,000.
With this crash program. he estimates that he can improve the chances of completing the project
THE DECISION .~IODEI. 21
on schedule in the: events of normal (N) and bad (B) weather to 90 ~and 70 ~~. rc:spc:ctivc:ly. Also.
if a delay does occur. its mean and standard deviation will both be: reduced to I dav under
normal weather: and 5 and 2 days, respectively, under bad weather.
Should the contractor launch the crash program based on the maximum EMV criterion?
Solution
From the decision tree shown in Fig. E2.9, the expected loss if no crash program is launched is
. . .
E[LIC] = E[LiCG]P(G) + E[LICN]P(N) + E[LiCB]P(B)
= 0 X 0.2 + E[LICN] X 0:4 + E[LICB] X 0.4
where
E[LICN] = E[(O.IX 1 + X)/CN]
= 0.1{Var[XjCN] .+ E1 [XiCN]) + E[XiCN]
= 0.1{9 + 25} + 5
= $8.4 thousand
E[LICB] = 0.1{25 + 225} + IS= S40 thousand
Hence,
E[LI C] = 0 + 0.4 x 8.4 + 0.4 + 40 d$19.36 thousand
Similarly, for the: crash altemarive- -
E[LI CG] = $10 thousand
E[LiCN] = 0.9 x 10 + 0.1{10 + O.l[Var{X/CN) .+ E:(X/CN)] + E(XIC.V)}
= 0.9 X 10 + 0.1{10 + 0.1(1 +I)+ I}
= $10.12 thousand
E[LICB] = 0.7 X 10 + 0.3{10 + 0.1(4 + 25) + 5}
= $12.37 thousand
Delay ~
Good Weather G (0.2) =0

0 \
Normal Weather N (Q4l

x: =5,c:r=3 O.lx2 +x
Bad Weather B {Q4l

X: =15,c:r5 O.lx 2 + x .
x=O
10
(0.9) X =0
10
(0.1)
x: :z(,O":r.l IO+O.Ix 2 +x
(0.7) x=O
--c--- 10 .
(0.3)
x: =5,cr=2 iO+x+O.Ix 2
Figure E2.9 Decision tree for completing construction project.

and
E(LIC]- 0.2 X 10 + 0.4 X 10.12 + 0.4 X 12.37
= $11.00 thousand
The criterion of max~g the expected monetary value is equivalent to that of minimizing
the expected monetary loss. According to this analysis, the contractor should launch a crash
program to complete the construction project on-schedule.
EXAMPLE 2.10
A distance AB is to be measured. With instrument. (I), the surveyor must measure 10 separate
segments. The error in each segment is random with mean zero and standard deviation u1 ;
u 1 is estimated to have a mean oft inch and c.o.v. of 50%-Another instrument (II) is available,
whic;h only requires S'sCpar-ati:. segments to measure the distance AB; the error in each segment
has mean zero and standard deviation 112 with a mean of3 inches and c.o.v. of60%. The cost of
j:lerforming .the required measurements arc $150 and S100 with instruments (I) and (II),
respectively~ The penalty of inaccurate measurement is proportional to the square of the total
error, or
\ L = k(X 1 + X2 + + X.)2
where k = 2, and X 1 is the measurement error (inch) in segment i. Which of the instruments
should thf' surveyor usc based on the ~aximum EMV criterion?
Solution
The expected loss from inaccurate measurement based on the above quadratic penalty function
can be expressed in terms of the standard deviation of the error in each segment as
(Lju) = E[k(X, + X2 + + X.)2 lo-]
~ {~E(X?lu) + ~..t(X X lu)] 1 1
If each me~surement is assumed to be statisticall~ j;;d~dent and ide1)tically distributed,

(Ljo-) = kn(X 2 lu)
= knu2
and
(L).=. f}nu f(a) do-

2
. :
= kn(u2)
= kn[Varlo-) + 2 (o-)]
For instrument (I),.the expected toial cost, including operating and inaccuracy costs, is
(C,) ':" 150 + 2 X 10[0.52 + 11]
= $175
The corresponding total cost for instrument. (II) is

(C0 ) = 100 + 2 x 5[1.8 2 + 32 ]
= $2.22.5
Therefore, on the basis of the maximum EMV criterion, the surveyor should usc instrument (I).
THE DECISION MODEL 23
EXAMPLE 2.1 I (e:xcerptedfrom Revell, 1973)
(a) The existing Grapefruit Dam on the Grapefruit Ri_v er is a multipurpose project
including water supply. power supply; m:reation. and ftood control.
After a detailed economic analysis. the cost of the damage as a function of the discharge
le,el of the river is tabulated in Table E2lla. Determine the expected annual cost of damage.
Table El.lla Damage Cost at Given Discharge

Level
Discharge (I 000 cfs) Cost of Damage in S!OOO
60 0
80 1
100 101 .
124 5200
132 7635
170 17,000
185 24,500
222 40,180
240 59,500
~55 92,125
>255 92,125
Solution
The annual cost of damage depends on the level of maximum discharge in that year. Since the
level of maximum discharge is a random variable (denoted as X below), the expected annual
cost of damage is
E(C) = r.., c(x)fx{x) dx
Since the damage cost (Table E2lla) is given only for discrete levels of discharge, the expected
annual cost of damage may be computed approximately as follows:
E(C) = ~ c(x.) +lc(xl+ ,) P(x1< X~ xt+ 1)
300
:!
.... 2!)0

...c:
200
/
&" I!)() /
/
<.. I
100
!'!'
,......v
..
0
.<::
....
i5 !10 v
~ .....
~~
0
3 10 100 1000 '10.000
Averaqe Return Interval , years .
Figure 2.11 Flow-frequency, Grapefruit River at Podunk.
.24 DECISION ANALYSIS
where x 1 and x,.. 1 represent two successive discretized discharge levels. From tbe flow fre-
quency curve of Fig. E2.11, the probability that each tabulated discharge value will be exceeded
in a year can be obtained. Table E2.11b summarizes the computations for the expected cost;
these; results yield the expected annual damage cost of$32,810 for the existing Grapefruit Dam.
In the next part of this eximiple (see below). other alternatives for improving the spillway
capacity are considered. The expected 11Jlnual damage cost for each alternative is also computed
using the same technique.
~ . .. . ;,
Table El.llb Summary of Computations of Expected Annual Risk Cost
Average
Dis . Damage ~ Return Probability Damage Annual
charge Cost Period of Cost Risk Cost
(1000 cfs) (in SJOOO) (in'years) Exceedence P(x1 < x < x,+t) (in $1000) (in $1000)
60 o : .- 150 0.00667
0.00267 0.5 0
80 I 250 0.00400
0.00200 51.0 0.10
100 IOJ 500 0.00200 2650.5
0.00075 1.99
124 5200 800 0.00125
0.00025 6417.5 1.60
132 7635 1000 0.00100
0.00050 12,317.5 6.16
170 17,000 2000 0.00050
0.00017 20,750 3.53
1!!5 24,500 3000 . 0:00033
0.00013 32,340 4.20
222 40.180 .5000 0.00020
0.00006 49,840 2.99
240 59.500 7000 0.00014
0.00004 75,812 3.03
:!55 92.125 10,000 0.00010
0.00010 92,125 9.21
>:!55 92.125 a: 0.0
I:= 32.81.
""
EXAMPLE 2.11 (Continued)
(b) It is feared that the existing spillway capacii'y'oftite Grapefruit Dam may be inadequate
to handle the flood discharge. To improve the capacity. three alternative designs are considered:
a1 : Lengthening the spillway. ,
a2 : Lengthening the spillway, lowering the crest, and installing flashboard.
aJ: Lengthening the spillway, considerable crest lowering. and installing radial gates.
The capital costs required in the three alternatives arc S 1.04, 1.3, and 3.9 million, respectively.
Alternatives a 2 and a 3 would require annual operation and maintenance costs of$2000 anc:!
SlO.OOO, respectively. Each alternative desitm is expected to have a 50-year service life, and the
annual interest rate is assumed to be 6%.
The total annual expected damage cost, including losses in recreation, structural and
property damages, loss of power revenue, human injuries and deaths, etc., for alternatives a., a 2 ,
and aJ are S5.56, 1.95, 297 thousand, respectively: whereas, the damage cost for the existing
design (a 0 ) is $32.81 thousand:
Determine the best alternative based on the minimum expected cost criterion.
Solution
The equivalent annual cost for the initial investment of each alternative design can be obtained
by multiplying the capital recovery factor corresponding toa service life of 50 years at an annual
interest rate of 6%. which is 0.0634. Table E2.llc compares the annual costs for the various
alternatives. Ti)c existing design is as59(:iated with.the minimum total annual cost; hence, no
improvement should be implemented on the existing structure.
THE DECIS/0,\ . .HODEL 25
Table EZ.llc Summary of Annual Costs (in Dollars)
Annual
Average Annual Total Annual Total
Equivalent Operation and Expected Annual
Alternatives Capital Cost Maintenance Cost Risk Cost Cost
uo 0 0 32.810 32.810
at 66.oo0 0 5560 71.56(
al 82.000 2000 1950 85.950
UJ 2..7.000 10.000 2970 259.970
2.2.4 Decision with Additional Information-Terminal Analysis

In order to improve the existing state of information or to reduce the levels of
uncertainty. it is common in engineering to conduct research studies or perform
laboratory or field tests before a final decision is made. This occurs especially when
a decision involves unusual and innovative design concepts. or when time is
available for collecting additional information. The acquisition of additional
information. of. course. will require the expense of_time. energy, and financial
resources; in spite of these added.costs. the additionaiinformation may be valuable
and will generally improve the chance of making a better decision. The added cost
for this new information should be included or reflected in the decision analysis.
In general. the additional information would not eliminate all the uncertainties
in a decision problem. At best. based on the experimental outcome, the probabilities
in the decision tree would be updated by applying Bayes' Theorem (see Chapters
2 and 8, Vol. 1). A decision analysis with additional information (commonly
ref~:rred to as terminal analysis) is similar to the prior analysis, except that the
updated probabilities (probabilities conditional on the experimental outcome) are
used in the computation of the expected monetary value or utilily. The following
examples illustrate the terminal decision analysis.
EXAMPLE 2./2
Suppose the engineer in Example 2.7 feels that it is too rlsky .to adopt the smaller design, that is,
unit A; whereas choosing the conversative design. unit B, would mean a large initial inv.;stment
that may be wasted if the efficiency actually turns out to be high. After conferring withhis client,
the engineer decided to find out more about the proposed waste treatment process before
adopting a specific design. The engineer ordered the construction of a laboratory model to
simulate the proposed waste treatment process. The cost of constructing the experiment is
SIO,OOO, and the reliability of the experimental 'results is as follows: if tbe efficiency of tbe
proposed treatment process is actually high (EH), the probability that the experimental results
will indicate a high efficiency rating is 0.8, and the corresponding probabilities for medium and
low efficiency ratings are 0.15 and 0.05, respectively. On the other hand, if tbe efficiency of the
process is actually low, the probabilities that the experimental results will show high, medium,
. and low efficiency ratings are 0.1, 0.2 and 0.7, respectively. In short, the conditional probabilities
are as follows: . .
P(HRIEJi)';.._~.s; P(MRIE~ = 0.15, P(LRIEH) = 0.05 .
P(HRIEL) = 0.1, P(MRIEL) = 0.2. P(LRIEL) = 0.7
where H R, ~~ and LR denote events of high. medium. and low efficiency ratings, respectively.
Relative Monetary
Value ($1000)
EH (0.951
-10
A
El (Q05l
-110
Test HR
EH (0.95)
. B -60
El (0.05)
-30
Figrue 2.1211 Terminal analysis for HR test outcome.
Suppose ~be experimentwas performed and the results indicated a high rating. In such a case,
the updl!.ted probability of EH is given by Bayes' Theorem as (see Example 2 7 for prior prob-
abilities):
F(EH) = P(EHIHR)
P(HR IEH)P'(EH)
P(HR IEH)P'(EH) + P(HR!EL)P'(EL)
0.8 X 0.7
0.8 X 0.7 + 0.1 X 0.3
= 0.95
brother words, the probability that the process will actually have high efficiency is increased
from 0.7 to 0.95. The corresponding updated probability of EL is 0.05.
:x_The decision tree for this terminal analysis is shown in Fig. E212a, where the probabilities
for. aU the paths arc revised in accordance with the HR test outcome. .
; n.e expected monetary value of the two alternative desi!!Jls arc, respectively,
E<AIHR) = o.95 x <-IO) + o.Os"x "(-110) =. -t5

E(BIHR) = 0.95 X (-60) + 0.05 X (-30) = - 58.5
. .
Hence, if the test results indicate a high efficiency rating. design A should be chosen. It is
interesting to observe that the expected monetary loss is oilly $15,000, including the cost of the
. experiment (compare this with a $30,000 loss if no experiment is performed).
Similarly. if the test outcome indicates a medium rating (MR), the updated probabilities of
E!f and EL arc, respectively, 0.637 and 0.363 as indicated in Fig. E212b. The corresponding
Rei at ive Monetary

Value IS IOOOJ
EH (Q637l
-10
A
El (Q363l
-110
Test MR
" EHIQ637l .
B -60
EL 10.363)
-.30
Figurt 2J2b Terminal analysis for MR test ?utco_mc . .

THE DECIS/OiV .HODEL 2i
Relative Monetary
Value ISIOOOl
EH (0.143)
- 10
A
EL (0.857)
-110
Test LR
EH (0. 143)
-60
8
EL<0.857l
-30
Figure E2.12c Terminal analysis for LR test outcome.
revised expected monetary values for the two designs will be:
(.41 MR) = 0.637 X ( -10) + 0.363 x ( -110) = -46.3
E(BIMR) = 0.637 X (-60) + 0.363 X (-30) = -49.11
Although design A is still slightly preferable to design B, the expected loss of S46.300 exceeds the
expected loss with no experiment. In this case, the experimental evidence is not very useful; the
test results fall in a fuz:zy region. Of.course, the chance of this occurring may be reduced ifa more
discriminating experiment is performed.
Finally, in Fig. E2.12c. if the test indicates a low rating (LR), the updated probabilities of EH
and EL will be0.143 and 0.857, respectively; in such a case. design B should be chosen, with an
expected loss of $34,300.
EXAMI'LE 2.13
In Example 28, suppose that additional d!ita were obtained revealing that over a 20-year period.
one earthquake of damaging intensity had occurred in site A, whereas no earthquakes were
recorded at site B. Would the decision on the optimal dam sitebe different given this new piece
dr information? .
Assume that the average earthquake damage costs are 5X (million dollars}, where X is the
number of earthquake occurrences (i.e., Case I in Example 2.8).
First, the priorprobabilities on the mean rate of quake occurrences at sites A and B, namely
YA, Y should be updated with rio5pect to the additional earthquake data. For site A.
P"(vA = 0.01) = P(v,. = O.OliX = 1 in 20 yr)

P(X- 1 in 20 yrjv,..- 0.01)P'(v,.- 0.01)
P(X =- 1 in20yr!v..t = 0,01)P'(v,. = 0.01) + P(X = 1 in20yrlv..t =- 0.02)P'(v,. '""0.02)
(0.01 X 20\ll'-O.l
~---...:.'.c..
- - X 0.5
I!
- (O.Ql X .20)1!'- 0 1 .(0.02 x 20)1e- 0 4
1! x0.5+
1
! . x_O.S
0.08187
0.08187 + 0.134o6 ' .
= 0.378
P"(vA = 9~~) =1- 0.378 = 0.622

Similarly Cor site B,

P"( l'.a - 0.005}
PCX - 0 in 20 yrlv.a.;, 0.005)P'(v.a = 0.005)
P(X ""0 in 20 yrfv.a = O.OOS)P'(v.a ""0.005)+ P(X ""0 in 20 yr l".a-= O.Ol)P'(v1 = 0.01)
(0.005 X20)01!-o. t .
..:..__--:-,.....;---- X 0.5
0! .
= 0.524
P"(la.= 0.01) = I - 0.524 = 0.476

. The expected seismic damage ..loss for each alternative is now computed based on these
updated probabilities. For site A
. 2.
E"(LIA) = :[ E(LIA , 1'j)P"('JIA)
jt
= (25Q X 0.01)(0.378) + (250 X 0.02)(0.622)

= $4.055 million
Similarly, for site B,
E"(LIB) = (250 X 0.00~)(0.524) + (250 X 0.01)(0.476)
= $1.845 million
Combining the benefit data in Example 28. the net expected benefits for site A and site B are,
respectively. $17.945 and $18.155 mtllion. Therefore, with the additional information, site B
becomes the preferred location for the dam.
. ..........
2.2.5 Preposterior Analysis
In a terminal analysis. the analysis assumes that the additional information (or
.experimental outcome)ls available. However, of equal or greater interest may be the
question: ''Should additional information be obtained'r Of course, additional
information, such as performing an experiment, involves the added cost of time
and money: therefore, t~e answer to this question depends. on Jhe benefit of the
new infonna tion relative ~o the add!=ci cost. The add<;d cost rnay be justified if it
eliminates a significant part of the uncertainty, thus leading to a lower expected
loss. Decisions involving whether additional information should.be acquired calls
for preposrerior ono"/ysis.
Figure 2.4 shows a typical decision tree for preposterior analysis. A two-stage
decision problem is involved. At stage A, a decision on whether or not to proceed
with an experiment is required; that is, a choice between an experiment (a 3 ) and no
.experiment (a 1 o2 ). If alternative o 3 (experiment) i$. chosen, another decision at
stage B is needed, depending on the particular experimental outcome ZJ. A sub-
tree may be constructed for each experimental outcome zi at stage B, as shown in
Fig. 2.4.
In a preposterior analysis. terminal analyses are first .performed on each of the
subtrees as indicated in Fig. 2.4. The optimal alternative of each subtree is first
THE DECTSIO:V .HODEL 19
No Experiment
a3 (Experiment)
Staqe A
Figure 2.~ Decision tree for preposterior analysis.

-
~--- ~
identified as marked in the figure. The maximum expected monetary value associ-
ated with the optimal alternative of each subtree, namely dt, is then . used to
compute the EMV of the experiment alternative; thus,
E(a 3 , experiment) = I dt P(z 1) (2.4)

I
in which P(z1) may be obtained thro_ugh
P(z 1) =I P(z,l 91)P'(91) I= 1, 2, ... (25)

\ . j
where P'(9} is the prior probability of outcome (Jj (i.e. before the experimental
outcome is observed). Comparing the expected utilities of a h a 1 with that' of a 3
(experiment), the opt4rial alternative at stage A can be selected. If the alternative
a 3 is selected, the decision maker must wait until a specific experimental outcome
is obtained; at this point the corresponding optimal alternative (as indicated by an
asterisk) is selected.
EXAMPLE 1.14
For the decision problem of Example 2.12, should the experiment be condudtcd iri 'the first
place?
The decision tree for a prcposterior analysis is shown in Fig. E2.14. The probabilities that the
test will indicate HR. MR. and LR are computed by applying Eq. 2.5. For example,
.P(MR)=' P(MRIEH)P.(EH) + P(MRIEL)P'(EL)
= 0.15 X 0.7 + 0.2 X 0.J
= 0.165
Relative Monetary
Value ($ 1000)
EH{0.7)
0
"El (0.3)
-100
-50
-20
-10
-110
-60
-30 .
-10
-110
-60
-30
-10
-110
. EH(0.143l
-60
I El(O.B57)
I -30
I
I
Stage A Stage B _
. - . . ..
Figm2.14 Decision tree for preposterior analysis:
The subtrees following the respective test outcomes are also shown in Fig. E214. In each of
the subtrees, the optimal alternative is denoted by the asterisk and the corresponding optimal
monetary value is shown;for example, the EMV for the subtree with outcome HR is -15. The
expected monetary value of the test alternative is thus
(Test) = 0.59 X ( -15) + 0.165 X ( -46.3) + 0.245 X ( -34.3)
= -24.86
Comparing this with the expected monetary values of alternatives A and B assuming no test is
perfonned. namely - ~0 and ~41, ~:espectively (from Example 2 7), the engineer should opt for
the experiment. Then, depending on the test outcome. design A-orB may be selected accordingly,
as indicated in Fig. E214.
. ..
2.2.6 Value oflnformation
In purely expected monetary te~s, the value ~fan experiment may be measured by
(2.6)
where E(T) is the EMV of a test alternative excluding the cost of the experiment
THE DECISION .HODEL 31
and a is the optimal alternative if the- test was not performed. Hence, if VI exceeds
the experimental cost, the test' aitemative should be selected. The value of the
additional information, however, is bounded by a limit referred to as the .. value
of perfect information" (VPI). VPI is simply the difference between E(a*) and
the expected monetary value of a "perfect test"; that is, a test with 100% reliability.
Hence,
VPI = E(PT) - E(a*) (2.7)
Therefore, VPI represents the maximum ca'st that the decision maker should allow
for acquiring any additionaJ information.
EXAMPLE 2.15
Consider the waste treatment process design in Example 214. Suppose a perfect test is available
that will definitely tell whether the efficiency of the treatment process is high or low. The cor-
responding subtree is shown in Fig. E215; the cost of conducting the perfect test is not included.
The events EH 0 , EL 0 represent the possible perfect test results; accordingly, the subsequent
conditional probabilities of EH and EL are either one or zero. The expected monetary value for
the perfect test is thus
E(PT) = 0.7 X 0 1- 0.3 X ( -20)
= -6
and Eq. 2.7 yields
VPI = -6- (-30)
= 24
Therefore, the cost of acquiring any additional information should not exceed $24,000. For the
specific test program studied in Example 2.14, the value of information from that test can be
shown as
VI= E(D- E(a*~
= ( -24.86 + 10)- ( -30)
= 15.14
Relative Monetary
Value <11000)
EH(I)
0
EL(Ol
-100
EH(I)
-50
Perfect EL(Ol
Test
-20
EH!Ol
0
EL(I)
ELa (0.31 1-zol -100
EH(Ol
-50
EL (I)
-20
Figure 2.15 Subtree of a perfect test.

32 DECJSJON ANALYSIS
Sina: that test program costs only 510,000, which is less than its value of information, namely
$15,140. the test should be performed. Observe that the value of this test is still considerably
less thlJ.II the value of perfect information. Hence, more sophisticated test programs could be
considered for reducing the uncertainties of the-efficiency of the treatment process, as long
. as such test programs cost less than s24,000. Otherwise, the amount oT saving resulting from .
TCductions in expected loss through such additional information will not be sufficient to pay
for the cost of acquiring such_information in the !ifst place.
2.1.7 Sensitivity of Decision to ~or in Proba~bility Estimation

The prior probabilities in a decision analysis are often based on or supplemented
by the subjective judgments of the decision maker. Errors in the estimated proba-
bilities, therefore, are unavoidable. The effects of these errors on a decision are
naturally of interest. In particular, the sensitivity of the optimal alternative to the
values of the calculated probabilities is of interest. For instance, if the probability
estimates are off by 10 'Y.,. will this alter the op.timal alternative? The following
a
example illustrates the nature of this problem and shows how sensitivity analysis
may be performed. \
EXAMPLE 2.16
Consider the decision problem of Example 27 again. Instead of assigning 0.7 as the prior
probabihty of high efficiency, let this probability be p. Pte corresponding probability of low
efficiency is (I - p). Based on the decision tree ofFig. E2.7 (expressed in terms of p), the EMV
for the two alternative designs are
E(A) = p(O) +(I - p)( -100) = -100(1 - p)
E(B) = p(-50) +(I- p)(-20) = -(20 + 30p)
Figure. E2.16 shows a plot of these two EMV a:; functions of p. It can be observed that for
p < 0.61. E<B) > E(A). hence. design B is the preferrc:d .alternative. On the other hand. for
p > 0.62. design A would be preferred. Therefore. according to this analysis, design A should be
selected if p > 0.62; it is not necessary that p be determined very precisely.
Moreover. based on the decision tree in Fig. E2.15, the expected monetary value of a perfect
test is
E(PT) = p(O) +(I - p)( -20) = -20(1 - p)
which is also plotted in Fig. E:Z:l6. Observe that for p = 0 or 1, 2 perfect test
same conclusion: hence. a perfect test would be ofneg;ligible value.
rn indicate the
.. The .value ofperfect information (VPI) is given by the <;lifferenc:e in ordinates tween E(PT)
and the max EMV at aJ!ivcn value of p.lt attains the highest value at p ,., 0.62 in this elUmlple.
Around p = 0.62, the optimal alternative is quite sensitive top. and any additional information
can significantly increase the EMV of the decision. For the experiment proposed in Example
2.13, the EMV of the test alternative, excluding the experiment cost, is denoted by the curve
(Tl in Fig. E1.16. In generl!l. thc.EMV of any imperfect test wili not be linear with p and it
should lie between E(PT) and max EMV. that is. within the triangular shaded region shown in
Fig. ~- ~6. The difference in the ordimtes of (T) and the max .EMV is VI.
In short. if the number of alternatives is few. small errors in the estimated

probability may not affect the selection of the optimal alternative, especially if the
probabilitv is not close to the critical value (sucli as 0.62 in Example 2.16). The
sensitivity. analysis illustrated ab~ve can be extended to decision problems with
more alternatives; howc;ver. in such .cases. the 'arialysis w~uld naturally be more
complex.
THE DECIS/0.\" .HODEL 13
EMV -.
Figure 2.16 EMV of alternatives as functions of p.
2.2.8 More Examples
EXAMPLE 2.17 (excerpredfrom Howard era/, 1972)
The hurricane has long been a devastating force along the coastal regions .of the United States
around the Gulf of Mexico: the average annual property damage during the 1960s was ap-
proximately $440 million. A decision arialysis was performed to determine the possibility of
mitigating the destructive force of hurricanes by seeding them with silver iodide before a
hurricane hits the coast. The decision tree for the study is shown in Fig. E2.17a.
With the present seeding technology, the effect of seeding remains uncertain. Based on sub-
jective: judgment and limited data from 'the seeding of Hurricane Debbie. the probabilities for
given (discretiZed) changes in wind speed are estimate<~.. It can be observed from these prob-
abilities that wind speed tends to decrease with seeding; however, there is some probability
that the hurricane wind speed may actuaUy increase after seeding. For this reason, the con-
sequence of a seeding decision must include a ~government responsibility cost.~ If the govern-
ment decides not to seed. !he government will no~ be held responsible for damages, since people
will accept hurricanes as natural .disasters. On the other hand, if the government should order
seeding and the hurricane wind speed actually grows stronger because of it, the government
may be subjectto lawsuits for hurricane-induced damages. These costs have been estimated and
included in Fig. E217a. . .
Based on the information given.in the decision tree of Fig. E217a, the expected monetary
value for the seeding and no seeding alternatives are. respectively,
. .
(Seed)= 0.038 ~ ( -503.95) + 0.143
(-248.65)
X
+
+ 0.392 X ( -105.25) 0.255 X { -46.95) + 0.172 X" ( -16.55)
= -19.15-35.56-41.26- 11.97- 285
= -Sll0.8 million
c11onoe 1n PrapertJ Gcwer..,...llt Total

Ma1imum Damoqe R-sibility Cost
Sustained (Millions Cost l"-rtotit"' (Millions
Wind 5pHd o!Doltanl Pr-rfJ Oamaool of Dallora)
-+.32~ $335.8 +50~ $!503.95
......
-+16~ . 191.1 - .... -+30 :~~
248.65
0- 100.0 -+5 10525
-16~ 46.7 0 46.95
-32~ 16.3 0 16.55
+32'7. "335.8 335.8
+16~ 191.1 191.1
0 100.0 100.0
-16~ 46.7 46.7

\
-32~ 16.3 16.3
Figur~ E2.17a Decision tree. for seeding hurricane
E(No Seed)= 0.054 x ( -335.8) + 0.206 x ( -191.1)

+ 0.48 X ( - 100) + 0.2()6 X ( -46. 7) + 0.054 X ( -16.3)
= -18.1 - 39.3 - 48 - 9.62 - 0.88
= -SI15.9 million
According to 1his analysis, seeding should be implemcntea; in spitc.of the potential liability cost
to the t~ovemmcnt, because of the strong likelihood that property damages will be greatly
reduced.
A third alternative is to perform seeding experiments before deciding on whether or not to
seed. Figure E217b shows the corresponding decision tree that includes the seeding experiments
as an alternative. The possible changes in the wind speed resulting from the experiment are
discrctized into five outcomes-namely, +32% + 16%, 0, -16% -32%. Corresponding to
each experimental outcome, the probability of a particular change in wind speed in future
hurricanes will be updated from the previously estimated values. Representative paths or the
decision tree a~e shown in Fig. E2.17b; here the results of the prepostcrior analysis arc also
summarized. -
EXAMPLE 2.18
(a) For the structural design of Example 23, which scheme should be adopted?
(b) Suppose the performance of design B can be inveStigated with a mOdel test at a cost of
S50,000. If the assumptions of design Bare not valid, the model structure will fail in the test with
90% probability; if the assumptions are valid, the failure probability of the model structure is
only 30%. Should the test be performed?
(c) At most how much should be spent to check if the assumptions are valid?
THE DECJSIOiV MODEL 35-
Choice of Total
Whether To Operati<nll Cost
Perform Result of Seedinq (MilliCHII
Ex~Mrimet~t Experim.nt Decision Outcome of OoUonl
+32"' 5503.7
+16 248.4
0 10!1.0
-16 46.7
-32 16.3
+32" 503.7
+16 248.4
0 10!10
-16 46.7
-32 16_.3
+32.,.. 503.7
+16 248.4
Do Not
Perform 0 10!1.0
Experiment -J6 46.7
-32 16.3
+32.,.. 503.7
+16 248.4
,0 10!1.0
-16 46.7
-32 16.3
Figure E2.17b Decision analysis including value of seeding experiment.
Solutions
(a) The expected monetary loss associated with designs A and B are, respectively,
E(A) = 0.99 _x 1.5 + 0.01 x 11.5 = Sl.6 million .

E(i:J) = 0.5 X 0.99 X 1.0 + 0.5 X 0.01 X 11.0 0.5 + \
. X 0.9 X 1.0 + 0.5 X 0.1 X 11.0 = $1.55 million
Based on the max-EMV criterion, design B should be adopted. _

(b) For the test alternative. the corresponding decision tree is given in Fig. E218a. Observe
that the test result will only a1fcct the probabilities of whether the assumptions are valid or nilt,
whereas the probability of satisfactory performance of the actual structure depending on the
validity of the design assumptions are the same as those in part (a). With the given reliability
level of the model test, that is,
P(FI P') = 0.9, P(Fi J7) = 0.1, P(Fi Y) = 0.3, P(Fi V) = 0.7
and P(Y) := P(Ji} = 0.5, we._h ave
P(F} = P(FI V}P(V) + P(FI V)P(V)

= 0.3 X 0.5 + 0.9 X 0.5
= 0.6
P(FI V}P(Y) 0.3 X 0.5 025
P(Y IF) = P(F) =~=
Lass
(In Million Il
5 (0.99)
A U55
,...------<l s (0.01) 11.55
5(099)
1.05
,...:-=:=..:...~ S(O.OI)
11.05
1.05
11.05
5(0.99)
Test 1.55
r------....q s (0.01) 11.55
~- s (0.99)
1.05
s 10.01) 11.05
\
5(0.9)
1.05
11.05
F = Test Model Fails
F =Test Model Survives
Figuu E2.18a Decision tree for test alternative.
and P(Vif) = 0.75. Similarly, P(F} = 0.4.
P(VIF) = P(FIV)P(V) = 0.7 X 0.5 = 0.875

P(F) - 9.4 .
and P(ViFJ = 0.125.
If it is as~umed that the t~st model failed, the expected losses for the designs are, respectively,
E(A IF) = 0.99 x 1.55 + 0.01 x 11.55 = 51.65 million
E(B IF) = 0.25 X (0. 99 X 1.05 + O.QI X 11.05) + 0. 75 X (0.9 X 1.05 + 0.1 X 11.05)
= 0.25 X 1.15 + 0.75 X 205
= $1.825 million
Hence. A is the better design if the test model failed. On the other hand,
E(A iF) = S!.65 million
E(BiF) = 0.875 X 1.15 + 0.125 X 2.05
= $1.2625 million
indicating that B is the better design if the test model survives. The expected value of the test is
E(T) "" 0.6 X 1.65 + 0.4 X -1.2625 .
= 1.455
Since the expected loss of the test alternative is less than tbat of the optimal design in (a),
namely desip~ B. the test should be performed.
THE DECISION .HODEL Ji
Loss
{In Million $)
s (0.99) 1.5
s (0.01) 11.5
s (0.99) 1.0
Perfect s (0.01) 11.0
Test
s (0.99) 1.5
s (0.01) 11.5
s (0.9) 1.0
s<nn 11.0
Figure E1.18b Decision tree for perfect test.
(c) The maximum cost of the test that can be justifiably spent may be determined by
considering th~ perfect test. The decision tree for this alternative is shown in Fig. E218b.
Performing the ex~tecfloss calcllll!tion. we have

E(A 1V) = 0.99 + 0.01
x 1.5 x 11.5 =S1.6 million
E(BI V) = 0.99 + 0.01
x 1.0 x 11.0 = Sl.1.million
E(A 1V) = 0.~ x 1.5 + 0.01 x 11.5 = $1.6 million
E(B I V) = 0.9 x 1.0 + 0.1 x 11.0 = S20 million
Therefore.
E(PT) = E(BIV)P(V) + E(AjP)P(V)
= 1.1 X 0.5 + 1.6 X 0.5
= $1.35 million
The value of perfect information is \
. VPI = - [E(PT) - E(B)]
= -{1.35- 1.55)
= S0.20 million
Hence. $200.000 is the maxir..um amount that may be spent for checking the validity of the
assumptions.
EXAMPLE 2.19
The proposed site for a structure overlies a layer of saturated sand (see Fig. E2.19a) that is
vulnerable to liquefaction under earthquake excitation. To estimate the probability of lique-
faction, the following simplified model is suggested. LetS be the maximum earthquake load
over the proposed lifetime of-the structure, S = N(IOO, 40) at a critical point in.the stratum.
Let R be the resistance capacity of the sand against liquefaction, such th~t
R= (t + fo)x
Structure
!ipre E2.19a Profile of proposed site.
where X is N(150, 50) and d is the depth (in feet) of the water table below the ground surface.
Liquefaction occurs if S exceeds R. The present watenable is at the ground surface. However,
the water table may be lowered by as much as 10 feet to impro~ the soil resistance against
liqu~action. Suppose the cost of lowering the ~at~ table is
C, = O.Ola1 (inS million)

and the loss in the event of liquefaction is $5 million.
\
(a) If the water table iS to be changed, how much should it be lowered?
As the depth of the water table d increases, the resistance capacity of the sand against
liquefaction increases, thereby reducing its risk of liquefaction during earthquakes; however,
the associated cost oflowering the water table also increases rapidly. The optimal value of a
can be obtained by minimizing the sum of the cost oflowering the water table and the expected
loss from liquefaction. For a given value of d, the mean and standard deviation of R can be
determined as
JJa = (1 + :o)JJx = 150(1 + :o)
a,.= (1 + io)ax so(r ~ io)

=
Since RandS may be assumed to be statistically independent, the probability of liquefaction is
P(S > R) = P(R _ s < O) = 41 f0- (JJ,. - JJs)] = 41f_ JJs - 150(1 + d/10) J
. L Jai + ai ~~~ + 5020 + a11W
Substituting the values of JJs and as yields the expected total cost
100- 150(1 + df10) . J.
1600 + 2500(1 + d/10) 2
A trial and error solution. as shown below, reveals that CT is minimum with d "" 4 ft. Therefore,
if the water table is to be lowered, 4ft will be the optimal depth.
a c, Expected Liquefaction Loss Cr
0 0 1.09 1.09
2 0.04 0.67 0.71
3 0.09 0.54 0.63
4 0.16 0.44 0.60
5 0.25 0.36 0.61
6 0.36 0.29 0.65
THE DECISION :WODEL -39
Liquefaction (0.218)
No Liquefaction (0.7821
0
No Liquefaction ( 1.0)
-QB
Figure E2.19b Decision tree on excavation.
(b) Upon further investigation, any lowering of the water table will not be permitted by a
regulatory agency. However, the saturated sand can be excavated and replaced by material
with a negligible chance ofliquefaction. The excavation will cost SO.S million. Should the soil be
excavated or left untouched?
The decision tree for this case is shown in Fig. E2. 19b. If no excavation is performed, d = 0,
and the probability of liquefaction is
P(S > R) = ~
100 - ISO ] = 0218
1600 + 2500 .
The corresponding ex~i:ed mone~y value is
EMV(no excavation)= 02l8( -5) = -$1.~ million
which is a greater loss than the cost of excavation. Therefore, excavation is the preferred
alternative, at a cost of SO.S million.
(c) Suppose the.soil resistance parameter X consists of two components; namely,

X= Y+W
where Y = N(100, 40), W = N(SO, 30), and Y and Ware statistically independent Moreover,
the exact value of W may be obtained by conducting a test, with which the decision to excavate
or not is then made. Determine the value of this test alternative.
Since the test wiil yi~ld a given val~e of W, the random variable X becomes
X==Y+w
which is N(100 + w, _40). After thetc:St, if no excavation is performed, the probability of lique-
faction becomes
P(S > R) = >[100 - (100 + w)J

j1fi:X:J + 1600
The decision tree for the test alternative is shown in Fi~ E219c.
The distribution of W at the chance node A is N(50, 30). The decision analysis starts with
node B, where the expected monetary values of the two alternatives are
EMV(no ~~:ti~n) = -s~(~

and
EMV(cxcavation) = -0.8
No Excavation
0
Test A
Excovaticn No Liquefaction
-0.8
: ..:: . :-:-.:. ' -- -. . -. . (1.0)
Figuu 2.19~ Decision tree for test alternative.
Therefore, whether or not to excavate depends "on the value of,... obtained from the test. The
critical value of w occurs at
(~J
- S41\.to,jl--5J = -0.8 .
or
... = -40j2~-e~8) \
"'.56.29 .
Ifw is less than .56.29, EMV(excavation) > EMV(no excavation), and the optimal alternative is
to excavate, with a corresponding expected monetary value of -0.8. On the other hand, if w
exceeds .56.29, EMV(excavation) < EMV(no excavation), the preferred alternative is not to
excavate, with a corresponding expected monetary value of- 5C!l( -w;40j2). At node A, the
expected monetary value of the tes_t alternative is thus
EMV(test) = J' 6 19
-
- -0.8 f,.{w) dv. + f... -Sill(40.,.;'21
U.l9
"~\ fw{w) dw
where f,.{w) is the PDF of N(.50, 30). After simplification,

EMV(test) = -0.8~
.56.29- 50) - 5 f"" ~ ('--~:-)-
-- ~ 1 ---
exp[ - .50)1] dw
_1 ;("'-
--
(
30 sb.lt 4!:J.../2 .J1n 3(1 2 30
= -0.47 - 0.16
= -0.63
where !he integral has been evaluated numerically. When we compare this with the expected
monetary value of the optimal alternative without the test. from part (b). the value of the test in
monetary terms is
VI = -0.63 - ( -0.8)
= $0.17 million
EXAMPLE 2.10 (excerpted from Chamberlain, 1970)
The water quality in the Great Lakes is gradually being degraded because of industrial pollution.
Three alternatives have been proposed by the Great Lakes Management to improve the situa-
tion; namely:
a1 : Build a new efficient treatment plant for SJO million.
a1: Conduct research to develop a scc<;>ndary pr~cc;ssor: tht; <?Utcome of the research may be
one of the following:
THE DECISION .HODEL _41
(i) develoPed processor in 2nd year at a cost of $0.5 million with 0.3 probability:
(ii) developed processor in Jrd year at a cost of S0.75 million with 0.4 probability;
(iii) developed processor in 4th year at a .cost of SI.O million with 0.2 probability:
(iv) failure to de\elop processor in 4 years at $1.0 million with 0. 1 probability.
a1 : Stock the lakes with robust species of bacteria that arc more resistant to pollution at a
cost of $0.1 million.
A founh alternative. a 4 is to do nothing. Because of uncertainties in the environment, panic-
ularly with respect to the potential sources of pollutants from industry, the level offuture water
quality prior to treatment may be idealized as either q 1 or q2 , where q 1 is better 'than q1 .
The relative likelihoods of these two water qualities arc subjectively estimated to be 3 to 2 For
all possible combinations of alternatives and states of water quality prior to treatment. five
levels of improvement in water quality arc possible, namely I, II, III, IV, V, in decreasing degree
of improvement.
Suppose the monetary value associated with each of these five levels of water quality are
assessed by a group of planners as follows:
d(I) = .'C
d(II) = 0.9x if the processor is developed in 2nd year
= 0.8x if the processor _is developed in 3rd year
-= 0.6.'C if the processor .is dev.elopcd
~. . in 4th year.
d(III) = O.Si if the processor is developed in 2nd year
= 0.7.'C if the processor is developed in Jrtfyear
= O.Sx if the processor is developed in 4th year
d(IV) = 0.2x
d(V) = O.Oix
where the value of .'C in million dollars remains to be assigned.
Table E2.20 summarizes the conditional probabilities that the alternatives will achieve the
Table E2.20 Conditional Probabilities
Pretreatment Improved Water Quality Level

Water
Alternative Quality II III IV v
a, q, I 0 0 o 0
qz 0.8 0.2 0 0 0
a1 (with processor q, 0 0.7 0.3 0 0
developed) qz 0 0.5 0.5 0 0
a 2 (without q, 0 0 0 0 I
processor) qz 0 0 0 0 I
al , q. 0 0 0 1 0
qz o 0 0 .1 0
a,. q, 0 0 0 0 ...
ql. 0 0 0 0
corresponding water quality. For example, P(II Ia 1q1 ) denotes the probability of achieving level
II in water quality if alternative a 1 is chosen and the water quality prior to treatment is q 1 ;
P(llla 1qz.) = 0.2
The complete decision tree is given in Fig. E2.20a, wbere..T is the development time under
alternativea2 .Theexpectedmonetaryvaluesare -10 + 0.992x, -0.75 + 0.677x, -0.1 + 0.2.x,
and 0.01x for alternatives a 1 , a 3 , a 3 and a 4 , respectively. lt can be observed from Fig. E220b
that each of the four alternatives could be the optimal one, depending on the value of x. For x
less than 0.5, a 4 is the optimal alternative; a3 becomes optimalfor 0.5 < x < 1.4; beyond which
a 2 becomes optimal. For very large values of x(> 29.4). a~ will become the optimal alternative,
implying that if water quality is extremely valuable, it would pay to build a new and efficient
treatment plant even at a great cost.
EXAMPLE 2.21
A stream is subject to floods each spring. Because of frequent flood .damages to an adjacent
town, a plan to increase the present levee elevatfon of 10 feet to either 14 or 16 feet is being
investigated. The corresponding costs of construction arc $2 and $25 million, respectively, and
the reconstructed levee will last 20 years. The average annual damage caused by the inadequate
level of protection is $2 million. Based on the data collected from previous years, the annual flood
level in the stream can be estimated quite accurately as having a median of 10 feet and a c.o.v. of
20%; however, both the normal and Jognonnal distributions appear to fit the data equally well.
Suppose the true distribution could equally likely be normal or lognonnal. Assume a discount
rate of 7% per year.
(a) Determine the optimal decision based on EMV criterion:

(b) How much is it worth to verify the true distribution of the annual flood level?
THE DECISION MODEL - 43
Figure E2.10b Expected monetary value of alternatives as function of x.
Solution
a0 : Levee elevation remains at 10 feet.

..
The decision tree for the levee improvement is shown in Fi'&. E221a. The alternatives are
a1: Increase levee to 14 feet. \

a2 : Increase levee to 16 feet.
For each alternative the probability distribution of the annual maximum flood level, X, could
ctthcr be lognormal (LN) or normal (N); on these bases, the probability ofin'!-dequatc Hood
protection (event F) may be computed for each alternative as follows:
P(Fiao. LN) = 0.5
P(Fiao, N) = 0.5
. . .. . . {11114 - In 10)
P(Fiat- LN)_= 1 .- ell\-: 0.2 = 0.0462
. ,jl4
P(Flat- N) .= . 1 - .... \ --
2
-to) = 0.022~5
In 16 - In 10)
~Fia 2 ,LJY) = l .-ell ( O.l .. = O.!J0939
. . (16-
P(Fla 2 , N) = 1 -ell - -
10)
= 0.00135
2
Totol PreHn! Cast

!In Millil!n il
F (0.5)
21.2
f I0.5l
0
F 10.5) ..:.
21.~ ;,
f' I0.5l
0
F (0.0462)
2:3.2
F' I0.9538l
2.0
F (0.02275)
23.2
f (0.97725)
2.0
F (0.009:39)
2:3.7
f I0.99061l
2.5
F (0.00135)
23.7
F' 1099865l
Z.5
Figurt! E2.2Ja Decision tr~for levee improvement.
The total present cost of each alternative is the sum of the initial investment and the equivalent
present value uf the annual damatge cost. The present worth factor for 20 years at 7% interest
is 10.594. As an example, for ~th a 0 - LN- F of Fig. E2.2la, the total present cost is
C -LN-F = 0 + 2 X 10.5.94 :"' $21.188 million
whereas
C.,-i.N-F =2+2 X )0.594 = $23.188 million
C-.-LN - 1 = 25 +0 X 10.594 "".$2.5 million
The expected total present costs for each of the alternatives are:
E(C..,) = 0.25 X 21.2 + 0.25 X 21.2 = 10.594

. E(C.,) = 0.5 X (0.0462 X 23.2 + 0.9538 X 2.0) + 0.5(0.022?5 X 23.2 + 0.97725 X 20)
= 2731
E(C.,) = 0.5 X (0.00939 x. 23.7 + 0.99061 X 2.5) + _0.5(0.001~: X :p.7 + 0.99865 X 25)
= 2.614
Therefore. the levee should be increased to 16 feet if the expected total present cost is to be
minimized; the corresponding expected total present cost is $2614 million.
(b) For.perfect information on the probability distribution of the annual flood level. the
corresponding decision tree i$ shown in Fig. Ei2tb. Presumably the true .distribution will be
lognormal or normal with equal likelihood. Given that the distribution is lognormal, the
probabilities Of flooding are 0.5, 0.()462. and 0,00939 for alt.ematives. QO, Q 1, aJ_Jd Q 2 1 respectively.
The expected total costs for the respective ~ltematives are .. : . .. .
E(C.. ILN) = 0.5 x 21.2 = SI0.6.million .
E(C~,ILN) = 0.0462 x 23.2 + 0.9538 . x .20 =: S2.9.79 million
E(C.,ILN) = 0.00939 x 23.7 + 0.99061 x 1.5 = $2.699 million
THE DECISION ,'dODEL 4i
Total Present Cost

(In Millign ~l
F (0.51
21.2
. F co.5l
0
F (0.0462)
23.2
F (0.95381 2.0
F (0.00939) 23.7
'F' (0.99061)
2.5
Perfect
F (0.51
Information 21.2
F (0.51
0
F (0.022751
23.2
F (0.977251 :2.0
F (0.001351 2:3.7
F (0.998651 2.5
Figure 2.llb Decision tree for perfect information on distribution of annual flood level:
Hence. a 2 is optimal if the true di~tiiGLition is lognormal. Similarly, if we assume that the true
distribution is normal. the probabilities of floOding are0.5, 0.02275, and 0.001'35 for alternatives
a0 , a,, and al, respectively. The expected total cos~s are t~en
E(C..,IN) = 0.5 x 21.2 = $10.6 million
E(C.,IN) = 0.0.2275 x 23.2 + 0.97715 x 20 = $2482 million
(C.,INl = 0.00135 x 23.7 + 0.99865 x 25 = 52529 million
This implies that a 1 is optimal if the true distribution is normaL Hence, the expected present cost
of the perfect .information alternative is
(C~,) = 0.5 X 2.699 + 0.5 X 2482 = 2.590
Therefore. the maximum fund that may be spent to v.:riiy the true distribution or'the annual
flood level is .
VPI = 2614- 2590 = 50.024 million= $24,000
EXAMPLE Z.JZ (excerpted from Davis and Dvoranchik, 1971)
A 5CO..foot bridge is proposed over the flood dikes at Rillito Creek, near Tucson. Arizona. The
bridge will rest on four piers, each supported by 25 piles. Part of the bridge may be lost in a
flood as a result of scour undercUtting the piles. If this occurs, the cost is estimated to be S 150,000.
Suppose the cost of driving each pile is S4/fooL What is the optimal depth of the pier foundation?
Suppose p" is the probability that the bridge will need to be replaced in a year. The expected
annual replacement cost is thus 150,000p". Assuine an expected life span oft years for the bridge
and an interest rate of i% per annum. The present value of the expected total replacement cost
overt years is 150,000p" multiplied by the present worth factor (pwf) for interest rate i% per
annum. Therefore. the objective function is the sum of the cost for the pile foundation and the
expected replacement cost, given. by .
C = 4h(4 X 25) + (150,000p,) X pwf(i'}~ t)
where h is :he depth of the pier foundation.
The annual probability of failure, p,, is the probability that the annual maximum scour
depth D exceeds tpe pier foundation depth. To determine p,, the annual maximum stream flow
is first studied. Based on past records of stream flows, the annual maximum flow Y may be
c.
.assuiDed to be lognormally distributed with parameters A, Based on n years of record and a
uniform_(diffuse) prior distri?utio~, the post~riorjoint distn~ution of these parameters is
.1 (A- :X) 2J[(~rlP (r)-l/l ( ~-~ r)~

[. 1 n
L(A,O= ~c;Jnexp . -~. ~Jn r("_; ~) . r exp f'j --2
(see Table 8.1 of Vol 1). where x and

s2 are the sample mean and variance of the logarithm of
the n stream flow values. Hence, incorporating the effect of uncertainties in the distribution
parameters, ~e compute the probability of failure p, corresponding to a design pier depth has
~ .p,(~>- =
.. r rJ J~~~A,
111 .I '
OUA. o dA dCJ dy
where
fr<YIA. 0 = ;JC exp[- ~(In yC- A)]

y(h)is the stream flow that has a: scour depth of h. .
- The annual expected total costs based on 10 years of data (1920-1929, with x = 3.667 and
r =. Q.l314) and 40 years of data (1920-1959, with j = 3.684, ~ = 0.9i7) are computed ac-
. cordingly. The results are plotted in Fig. E222 as a function of the design depth h. The optimal
pier depth based on 10 years of data occurs at 16.5 feet with an annual expected cost of$7845,
whereas the optimal depth based on 40 years of data is 11 feet with an expected cost of $4680.
The difference in the two costs, $7845-$4680 = $3165, may be interpreted as the value of the
additional 30 years of data.
Results are also obtained assuming other -lengths of data period were available and are
summarized in Table E2.22. The optimal cost appears lo 'decrease with the increasing length of
available data period. In the extreme case, when there are in~ite y~s of data, no uncertainties
18
16
..
"'Q
14
12
.
10
w
8
4
10 12 14 16 18 20 22 24
Pier Depth, ft (h)
Figure 2.12 Annual total expected cost versus pier depth-based on 10 and 40 years of
annual peak flou data on Rillito Creek, Tucson, ~rizon~ (Davis and Dvoranchik, 1971).
APPLICATIONS IN SAMPLING AND ESTIMATION -n
. .
should exist in A. and C. The corresponding "value of the expected cost C under this condition
of perfect information is estimated to be $4349. Therefore, immediately after 1959, at most
$4680-$4349 = $331 will be gained if the statistical uncertainties in the parameters are
eliminated.
Table E2.22 Optimal Depth and Cost of Pier Foundation
Data Period 1920-1924 1920-1929 192Q.:.I939 1920-1959
Years 5 10 20 40
x 3.7014 3.6670 3.6943 3.6840
s'- 0.0977 0.1314 0.0890 0.0977
c.,. (S) 12,045 7845 5059 4680
hopt (ft) 21.2 -16.5 11.7 11.0
2.3 APPLICATIONS IN SAMPLING AND ESTIMATION
2.3,.1 Bayes Point Estimator

In Chapter 8 -of Vol.-~ we presented the Bayesian approach to estimation. This
approach assumes that the parameters of a probability distribution are random
variables. The uncertainty of a parameter is modeledby the corresponding prior
or posterior distribution. If a point estimate of the parameter is desired, it can be
formulated also as a problem of decision analysis. Consider the decision tree in
Fig. 2.5 that depicts a situation in which the point estiinator of the parameter e is
being selected. Depending on the particular choice of the estimator lJ and the actual
value of the parameter 8, a prediction error will result, generally followed by a loss.
By modeling this estimation process in the context of~ecision analysis, the Bayes
point estimator may be determined such that the expected loss associated with the
prediction etror is minimized. Mathematically, if lJ is the estimator of a parameter
whose actual value is governed by the distribution f(8), then the expected loss
resulting from the error in prediction is
L ..
= I~ g(~, 0)!(8) dO (28)
A .
9-9 g(9,9l
.Select ' . I Ac;uai.value

Estimator 9 of 9
IPredic:tianl
Error Loss
Figure 1.5 Selection of point estimator.

48 ' DECISION ANALYSIS
where g(9, 6) is the loss function. Generally, this loss function is expressed in terms
of the prediction error, namely (9- 6). Since the Bayes estimator 6 minimizes the
expecte~ loss, it s~ould satisfy the following equation:
-,.,c. . . . - .. -:- .. -= ~-..;~{f~: g(~. -6) f(O) ae] ~ 0 -.. ,-. (29)
The Bayes estimators for some of the common loss functions are derived in the
following ex~pl~-
EXAMPLE 2.23 (Linear Loss Function)
Suppo~ the loss function (see Fig. E223) is given by
.
; g(o, 6> {c(6-
c<e.:..
=
8);
D);
es6
8>6
where c is a conslanl.
r
The expected loss is
\
L = [., c(6 - -8) J(8) JB .+_ c(8 ; 6) /(8) J8
But
d[f~ .. c(B- 8) /(81 d8] = (" cf(B) do+ c{fl- 6) /(6) dO

. d9 J_., d6
=C l/(0)d8
Similarly.
Applying Eq. 29 leads to

d[ff' c{8 - :).f(B) J6] = .-.c.[.J(8) dB
dL
d
r f(8)d0- c i f(6)J0.= 0
. -
--
~ = c J.
. -a::~ .
or
~
j_ac
f(1) dO -
J,("' f(O) dll = 0
Fig11re 2.23 Linear loss function .

APPLICAT/Oi\"S IS SA,HPLING :4.\'D !STHIATION .J9_
But
.. . ...,
LJ<O> ~10 + J. f(O) JO =I
Hence.
J.,
-oc
f(O) dO = f.""' f(O) dO = t
I
Therefore. in the case of a linear loss function. the Bayes estimator 9 is the median of 1:1.
- . .
EXAMPLE 1.24 (Quadratic Loss Function) .
Suppose the loss function is

g(O. 6) = c(O - 0) 2 : -:X:< 0 <X
Applying Eqs. 28 and 29, we obtain
dl: = f"'d[c(O - 6)2] /(8) dO

dO ,~_"' dB ,
..= L"'..... --2c(IJ- 9) f(O) dO
= -2c
~
r. f(~)
.
0 c/0 + 2c0 r...J~o) dO
= -2cE(O) + 2c0 =0
Hence,
0 = E(O)
The Bayes estimator is the mean value of (J in the case o~ a quadratic loss function.
EXAMPLE 1.25 (Constant Loss Function)
Suppose the loss function is a opnstant over the entire range of 0. except that for the small
region where 19- 01 5. e(L theri: is no loss due to prediction error. Mathematically. the loss
function as shown in Fig. E2.25 is given as follows
E.
<'; .. -= < 9 ~ lJ-- .
. 2
g(O. 8) = 0:
&
c; 0 + < 0 < oo
2
. where & is small.
The expected loss is
[~ ~ cf(O~ dO +.[lrl11cf(~) dO
1 11
L=
., r.+(r/21
=
f -...
cf(O) dO -
-(r/21
cf(O) dO
""c- cif(6)
-,r-
11,..-------:-
.
8
8
Figurt! lp.25 Constant loss function.
Applying Eq. 29, we obtain
~
df1
= -a~J!La 0
tlf1 0 :
or
df({))- 0 \
dB- .
~en~ the Bay~ estimator for a co~tant loss function is the modal value of 8.
As shown in the above examples, the mean, median, mode, or perhaps other
values (see Problem 2.20) can be used as the point estimator of a parameter,
depending on the form of the Joss function appropriate for the problem. Unlike
the estimators in classical statistical ~estimation in which properties of unbiasedness,
sufficiency, and efficiency are used to define the quality of an estimator, Bayes point
estimators are determined on the basis of minimizing the Joss associated with
prediction errors. Physically, the Bayes poin(estimatOI: is more meaningful for
engineering purposes, and consistent with the decision maker's objectives.
2.3.2 Optimal Sample Size

One application of the pre posterior analysis is in the determination of the opt.iqlal
sample size in statistical sampling. In Chapters 5 and 8 l'lf Vol. 1, we showed that
the accuracy of the estimation of the statistical paramctc~:s of a random variable,
such as the mean value and variance, incre;ues with the sample size. A larger sample
size, of course, involves a higher cost; naturally, the optimal sample size will
involve a trade-off between accuracy and the cost of sampling.
Suppose X is a random variable with a paramett:r 8 to be estimated from
sampling observations. The decision model for determining .the optimal sample
size is shown in Fig. 2.6. Two phases of decision may be required. First; at B, a
decision on the sample size n is required. Observe that n = 0 represents the case of
no sampling needed. Then, a set of sample data {x} = '(x 1, x 2 , , x.) is observed;
the appropriate estimator for 8 is selected at C. The total i_ossfor each path following
the decision to sample generally depends on the ~pie size, the estimated and
actual values of 8.
APPLICATIONS IN SAMPLING AND ESTIMATION 51 -
8 L (n, {x}, 9,91
I
I
Select Observed Select Actual Loss
Sample
Size n I Sample
Values {x}
Estimator
For 9
Value
of 9 I Function
Figurl! 2.6 Decision tree for optimal sample size.
Decision analysis starts at the last node in which the expected loss for given n,
{x }, and estimate 9is calculated by weighing the loss over the posterior distribution
of 8; that is,
E(L In, {x }, 9) = J~.., L(n, {x}, 9, 8) J;(fJ) d8 (2.10)
At node C, the optimal estiinator 9 is the one that minimizes this expected loss
function; thus, -
dE(Ljn, {x}, 9) _
db - 0 (211)
yields {]p On this basis, the expected loss for a given sample size n (at D) is
. E(Lin)= f E[Lin,{x},9.pt].JiXJ({x})d{x} (212)

J,,.,
The optimal sample size "P' may then be evaluated from
dE(Lin) = O (2:3)
dn
As an illustration, consider the following loss function
L(n, {x}, 9, Jl) = c(Jl- 9)2 + kn . (214)
where c and k are both constants. This function represents a quadratic loss because
of an error in the estimation of Jl and a linear sampling cost.
Applying Eq. 210, we obtain
E(L!n, {x}, 9) = J~~ [c(Jl - 9) 2 + kn] f,." (Jl) dji

= cE;[(Jl - 9) 2] + kn
~ cE;[{(~ .:.. Jl") +.(jl" - + kn
9)} 2]
= cE;[(Jl - Jl") + 2(Jl - Jl"XJl" - 9) + (Jl"
2
-. 9Y2 ] + kn
= cVar"(Jl) + c(Jl" - 9) 2 + 1..71 (215)
where 11" and Var"(.u) arc the posterior mean and variance of J.1o respectively.
Applying Eq. 2.11 yields
.. dE(Lin. {x}, B) = 2c(J1" _ B) =0
- .. . d6 .
which gives
Bop! = 11" (2.16)
Hence, the optimal estimator is the posterior mc3.n value of JL However, the
optimal estimator of Jl is not always given by the posterior mean.p". For example,
n.
. if the less f!mction assumed .in Eq. 2.14 was not symrr.:trical about J.l, will not .
be equal to p." (see Problem 222).
Fonhe loss function of Eq. 2.14, the corresponding expected loss at node Cis
E[Lin, {x}, B~ = c{Var"(.u) + [Jl"- 6.] 2 } + kn
(2.17)
The poSterior variance, Var"(.u), depends on the distribution of the underlying
random variable (population). For the case in which the. population is Gaussian
with knov.:n a, the posterior variance (see Section 8.4.2, Vol. 1) assuming random
sampling is
(2.18)
where a' is the prior standard deviation of p.. Therefore, in this case, the expected
loss depends only on the sample size n, but not on any sample statistics of the
'.l: observed values; that is,
E(Lin) = J [c Var"(p.) + kn] Ji"xJ ({x}) d(x}

= c Var"(.u) + kn (2.19)
However, if the expected loss in Eq. 2.17 was also dependent on the posterior mean
of p., an integration over the distribution of the sample mean of the observed data
would be required (see Example 2.27).
For a Gaussian population X with known standard deviation u, if we substitute
Eq. 2.18 into Eq. 2.19, the expected loss for given n is
. (u')2 al/n
. E(Lin) = c (u') 2 + u 2 fn + kn (2.20)
Setting
dE(Lin) -cal(a') 2 (a') 2
-'--'--'- = + k =0
dn [a 2 + n(u') 2 ] 2
yields the optimal sample size
n~ = .k -(;.~2 (2.21)
APPLICATIONS fN SAMPLING AND ESTIMATION .53
This expression ~valid

only if~> rr/(tr')1 ; otherwise, nop< = 0 no ~d
sampling should be exercised. In the case for which no prior information is
available, u' may be assumed to be ::o: hence, the optimal sample size becomes
nopt = uji (222)
Extending the above concepts. Resendiz and Herrera ( 1969) studied a settlement
problem for a foundation in which they determined the optimal design pressure
(see Problem 223).
EXAMPLE 2.26
In a surveying project. suppose the loss due to error in a distance prediction is proportional to
the square of the error (in inches). The cost per unit square of error is $1200 and the cost for
. performing one measurement is $10. The standard error in each measurement with the proposed
device is I inch. Determine the optimal number of measurements for the following cases:
' .
(a) No orior information is known on the distance.
.
(b) A prlor information indicated that the distance is N(IOOO. 0.4) inches.
(c) The prior distribution <if~e-distancc is N(IOOO, 0.04) inches:
Solution
(a) The estimation of the actual distance is equivalent to the. estimation of p. (~ Section
5.23, Vol I). With Eq. 222, the optimal number of measurements with no prior information is
[1200
n.... = 1.0"10 = 10.95,. 11
(b) With the prior information on p. as N(IOOO, 0.4) the optimal number of additional
measurements (above and beyond the pnor measurements) is det~ned from Eq. ~I as
1200 . 12 \
n.,.. = W (l.W ~ (0.4)2
= 10.95 - 6.25
= 4.7
The expected loss for n = 4 and n = 5 can be computed with Eq. 220 giving respective losses of
. 157 and 156.5. Furthermore, observe that for the case of no sampling (n "" 0), the expected loss
is c Var'(u) = 1200(0.4)2 = 192 Hence, the optimal sample size is 5.
(c) If the prior information on.JJ. is N( 1000, 0.04),
fliOO 1 12
n... = v!0-(1.0) - (0.04)1
=II.:.. 625
= -614
which implies that realistically n.,. = 0. Hence, the best strategy in this case
is not to make any
additional measurements, but to estimate the distance using the available prior information
giving an estimated distance of 1000 inches.
EXAMPLE 2.27
An environmental engineer obtains water samples from a stream to estimate itS mean DO
(dissolved oxygen) conc:entration. If DO in the stream is normally distributed with mean ~ and
: standard- deviation 0.5, determine the optimal sample size n assuming the following Joss
~;:t~on: ..,_ . . . - ---:."~-~jOO(;'::_:B)l- ;~2 +~:~:-~ ~~-. -~,; .
where {J is the estimator of the mean DO concentration; lOO(u ..:_ fJ) 2 represents the quadratic
prediction loss; and - 5~ 2 denotes the benefit resulting from a higher mean DO concentration.
- .Solution
. Substituting the given loss function in Eq. 210 and using the result of Eq. 215, we get
. E(Lin, {x}. fJ) = [., [IOO(u- b)2 - 5~2 + nJi;<.u) d~

- . .:: tOO[Var-(,u) +(9 - p)2 ] - SE;-(u2 ) + n
'!" .9.5 Var~(.u)- 5~N + 100(fJ- ~) 2 + n
2
Differentiation with respect to fJ -relds

200({} - ~)-= 0
:Hence, fJ.pl is the posterior mean of"' The corresponding minimum expected loss is
E[Lin, {x}, fJ.] = 95 Var"(.u)- 5J.1d + n
Since no prior information is availilble for estimating J.l,
~ = x
and

0.25 cr2
Var (.u) =- =-
n n
Hence., the sample mean of the observed data is the only sample statistic needed for evaluating
the optimal sample size. Equation 2.12 becomes ... .
f., [ . 0.25
E(Lin) ~ -., 95 x -n-- 5x2 + n f 11 (x)dx
- ]
i3.75 X2
=---
. n
5( ) + n
.
23.75 . s[0.25,
=--- - +L _,.2 (X)] +n
n n
22.5
=11- - 5JJ2 + n
Differentiating E(L In) with respect ton, we obtain
. llopt = .jlli = 4.74
Since n must be an integer, observe the following :
Forn = 4, E(L~n) = 9.63- 5J.1 2
For n = 5, E(Lin) = 9.5 - 5J.1 2
The expected loss is less when n = 5; hence, ,
n.,.. = 5 ,
APPLICATIONS IN SAMPLING AND ESTIMA TIOJV 55
EXAMPLE 2.28 .
The occurrence of cracks along a highway pavement may be modeled as a Poisson process with
a parameter m, which is the mean occurrence rate of cracks. From a previous inspection of
similar pavements, the parameter m is estimated to have a mean of 2 per mile with a c.o.v. of
50%. Additional inspection may be conducted, at a cost ofS20 per mile, to improve the accuracy
of the estimated value of m. Suppose the loss due to error in the estimation of m is given by
L = IOOO(m - 0)1
where m is the actual mean rate of cracks and 6 is the estimated value of m. Assume that the
prior distribution of m is given by the conjugate distributio!l to the Poison random variable.
Determine the optimal distance t (in miles) to be inspected that will minimize the expected total
cost function.
Solution
From Table 8.1 ofVo.ll, we sec that the conju.gate concept requires a gamma distribution for
the parameter m. Hence, the prior distribution of M will .bc of the form
'
The parameters v' and It can be evaluated from
~- - . E(M) = 7k' = 2
and
It
Vai'(M) = v'1 = (0.5 X 2)1 =1
The solution of these equations gives y' = 2 and k' = 4. The decision model for this example is
shown in Fig. E218.
At node D, the expected loss corresponding to each estimate 6 is given by Eq. 210 as
E(Lit, x. 6) = [ [l~ (m - 6)1 + 20t] J;;(m) dm
= 1000 (Var"(M) + (m" - iJ)2 ] + 20r

Applying Eq. 211 again, we obtain 8..,. = m. The corresponding optimal loss at node C
becomes
E[Lit, x; 9.,.] = 1000 Var"(M) + 20t
I~~~~ I=~ I~:;tor I~~~of

Loss
Function
Figure 2.28 Decision tree for inspection disiance.

in which the posterior variance of the parameter m may be obtained from Table 8.1 ofVot 1 as
V k'+x 4+x
. ar"(M) .. (v' + rf = (2 + r)l
where xis the number of observed cracks in distance 1. Hence, from Eq. 212, the expected loss
weighted by the distribution of X yields ..,_ ' .
+'-~~-~-. ~ ' ... -.---. E~~lr) -~ rP~"~;: ~l ~ ~;] fx(~dx
= 1.NV'> 4 + E(X) 20
U\AI (2 + tf + r
. . 4 + E(M)t
= l<XX>_ (2 + t)2 + 20r
. 2000
=--+20r
2 +l .
Sc;ning the derivative to zero, ~e-obta.in . .
-dE(L!t~ = _ 2000 + 20 = O
til (2 + r)l
(2 + rf = 100
and
r.,. = 8 miles
Hence, the optimal distani:e to be inspected for cracks is 8 miles.
2.4 ELEMENTARY CONCEPTS OF UTIUTY THEORY
The desirability of an alternative may depend on several attributes such as monetary

value, time factors, prestige, social acceptance, pleasure, and others. Ranking the
feasible alternatives in a decision tree requires a scale for quantifying the degree
of preference among the attributes. A well estabii'shed scale of measure may exist
for some of the attributes such as monetary value; however, there is no obvious
way of quantifying the values for the majority of these attributes. The problem of
value measurement is further complicated when the consequences in the decision
tree require the evaluation of a combination of attributes. In order to establish a
uniform scale for measuring the overall value of an alternative, the concept of
utility may be introduced. Utility is defined as a true mea,sure of value to the
decision maker. Utility theory provides a framework whereby values may be
measured, combined, and compared consistently with respect to a decision maker.
Therefore, if the utility values of all the alternatives are available, the alternative
with the highest utility value will be preferred.
2.4.1 Axioms of Utility Theory

Introduce the following symbols defining the degrees of preference between two
events; namely,
A >- B denotes M A is preferred to B M
A - B denotes',' "indifferen:ce betV:.een ~-and B"

A ~B denotes . M A .is preferred at least as much as B"
.. . .
ELEJIENTARY CONCEPTS OF UTILITY THEORY 57
The following are the.i~portant axioms in utility theory;
(i) Orderabi/icy. The states of preference between A and B can be arranged in

order of increasing preference toward B as follows:
A>- B, A?= B, . A- B, A~B. A-<B
(ii) Transitivity. If A >- 8 and B >- C, then A >:- ~-
(iii) Continuity. If A >- B >- C. a value of p (between 0 and 1) exists such that a
person is indifferent between (a) obtaining B definitely and (b) getting A or C
with probabilities p and (1 :__ p). respectively. This indifference may be
presented schematically with the following lottery:
~~---B ~:
It is obvious from the lottery that if p is 0, getting C is certain, hence, B is
ol:learly preferred. Whereas if p is 1, the lottery implies getting A for sure, and - .
B is not pr~ferred. Therefore, for some vaiue of 0 < p < 1, the condition of
indifference between (a) aq<J.Jl?) will emerge. In mathematical tcrminologi,
B is the cercainty equivalent of the latter lottery.
(iv) Substitutability. A lottery and its certainty equivalent are interchangeal:lle
without affecting preference. This is obvious from (iii) that shows the two are
equally preferred.
(v) M onotonicity. If A >- B, then
p A p'
~8>- ~:
if and only if p > p'.
(vi) Decomposability. A series of branches may be replaced by a single branch as,
for example,
~A
~.
Most of these axioms are obvious, yet they form the backbone "of the field of
utility theory.
2.4.2 Utility Function

An individual whose preference characteristics satisfy the utility axioms of Section
24.1 may have his preferences encoded in a utility function. A utility flinction
quantifies the order of preferences. Mathematically, the function represents a
mapping of the degree of preference into the real line, thus permitting preference to
be expressed numerically.
Let the symbol u() denote the utility function of an event. The property of the
utility function requires that
If A>- B, then u(A) > u(B);
:if.. A- . -. B, . then u(A) = u(B);
~ .. -
if A ~ B, then u(A) ~ u(B).
Another important property of the utility function is that the utility of a lottery
iS equal to the expected utility of its prizes. For example, the utility of the following
lottery between A and C is
. .
u (~A) -pu(A)+(I-plu(C) (2.23)
c .
a
It can be demonstrated. that utility function is consistent with the axioms of
utility theory. As an example, consider the aiiom of monotonicity. For the case
A >- B, we have u(A) > u(B), and hence, u(A) - u(B) is a positive value. Further-
more, if p > p', then
p[u(A) - u(B)] > p'[u(A) - u(B)]
Adding u(B) to both sides ofthe inequality, we have
pu(A) -1:- (1- p)u(B) > p'u(A) + (1 - p')u(B)
or
> -\~a
'-~A)
or
~A
.
~
~
. 8 ---.:i__
. - 8
which states the axiom of monotonicity.

.. If we apply the property of the utility function in Eq. 223 to the following
preference statement,
. (i24)
the utility of the lottery on the left side is
u(L 1 ) = 1 x. u(B) =. u(!J) .

ELEMEiVTtlRY CONCEPTS OF UTILITY THEORY 59
whereas the utility of the lottery on the right side is

u(L 2 ) = pu(A) + (1 - p)u(C)
Since lotteries L 1 and L 2 are indifferent, we have
u(L 1) .= u(L2 )
or
u(B) = pu(A) + (1 - p)u(C) (225)
In other words, for given utility values of A and C, the utility value of B can be
determined from Eq. 225 by first establishing an indifference statement, Eq. 224,
with an appropriate choice of p.
Linear Transformation of Utility Function Another important property or the

utility function is that the relative utility values will remain unchanged if the
utility function undergoes a linear transformation. In other words, if a consistent
set of utility values u are transformed to u' through the following
..' u'(1) = a + bu(E;) i = 1, .. . , n (2i6)
where a is a constant-and b is a positive constant, the new set of utility values
u'(EJ, i = 1, . . . n will retain "the. decision maker's original preference. This
property can be easily proven as follows.
Suppose that 1 , 2 , 3 are in increasing. order of preference with respective
utility values u(E 1 ), u(E 2 ), and u(E 3 ). From the continuity axiom of utility theory,
a value of p exists such that
(227)
Multiply each term in Eq. 227 by b and add a to both sides of the equation; thus,
a+ bu(E:J = pbu(E 1) + (1 - p)bu(3) +a
= p[a + bu( 1)] + (1 - p)[a + bu( 3 )] (228)
From Eq. 226, we define
u'(1) = a + bu(E;) fori = 1, 2, 3
Hence, Eq. 228 becomes
.. (229)
By comparing Eqs. 227 and 229, it is obvious that the same value ofpis obtained
at the indifference point between.the lotteries involving 1, 1 , and 3 Therefore,
Eq. 229 implies that the utility values u'(E;) are also a consistent set of utility
assignments reflecting the decision maker's preference.
Since the values of a and b in the above transformation can be arbitrarily
assigned, it follows that in determining a consistent set of utility values, the first two
values can be arbitrarily assigned, as illustrated fu the subsequent section. As long
as all other utility values are consistently determined with respect to .these two
reference points, all the utility values will form a consistent set, In other words,
utility values are not unique. In fact, absolute utility values are not necessary for
decision making; relative utility values are sufficient
EXAMPLE 2.19
Given ~ following utility fmiction'
+ r; x ~ 0
u(x) = 10 + lOx
- :d~~c-;hc ;~rlo~cd- utiiii; f~ction s~~h ;h~t 'thc;;tiliiy values at;,;= 0 ~d x= 10
are, respectively, 0 and 1.
According to Eq. 226, a new utility function
u'(x) = a + bu(x)
"" a + b(lO + lOx.+ r )
can be defined such that u'(x) is still a valid utility function. Since a and bare constants to be
determined such that
u'(O) = 0 = a + lOb
~ :.
and
u'(lO) "" 1 = ll + 210b
the solution' to these-simultaneous equations yields a = -0.05 and b = 0.005. Hence, the
transformed utility function is
u'(x) = -0.05 + 0.005(10 + lOx + xl)
-=.0.05x +0.005x1
2.4.3 Detennination of Utility Valoes

... Suppose there are three events, say E 1, Ei, E~. to which utilities values are to be
;:.. assigned. First, arrange the events in decreasing.order of preference; for example,
E 1 >- E 2 >- 3 It is common to assign a utility of 1.0 to the most_ preferred event
and a zero to the least preferred event; that is, u(E 1 ) = 1.0 and u(E 3 ) = 0. This is
similar to the assignment of 1.0 and 0 in probability scale to the sample space and
the null set, respectively. To' determine the utility.of E 2 ', relative to those of E 1
and 3 , the decision maker is offered a choice between the following lotteries:
._
The value of p is adjusted until there is indifference between the two lotteries L 1
and L 2 An indirect way of obtaining the value of pis to usea probability wheel
(Spetzler and.Staelvon Holstein, 1972):The probability wheel is a disk with two
sectors, one blue and the other red, with a fixed pointer in the center of the disk.
The disk is spun, finally _stopping with the pointer either in the blue or red sector.
Simple adjustments may be made to change_the. relative sizes of the two sectors,
which in turn change the probabilities of the pointer landing on either sector as the
disk stops spinning. In determining the value of p. the decision maker is asked
whether he prefers L 1 or L 2 With L 1 he is certain to get E 2 In lottery L 2 , the disk
. is spun. If ~he pointer ends up in. the . E1. will. m;:Cur
. red. sector,. -. . ;if it' ends
. . up in..tbr:
ELEMENTARY CONCEPTS OF UTILITY THEORY 61
blue sector, 3 will occur instead. Starting with the absence of the red sector, the
decision maker would definitely choose L 1 because 3 will certainly occur in
lottery L 1 , and since 3 -< E~, u(L 1 ) > u(L 1 ). The red sector in the wheel is then
increased until indifference is observed. At that .point, the fraction of the red sector
in the wheel is simply the probability pat which there is indifference between L 1
and L 1 . Hence,
or
u( 2) = pu( 1) + (1 - p)u(3)
= p X + (1 - p) X 0
1.0
=p
The utility of 1 is simply pin this case. This value of pis used only to determine
u( 2); it is independent of the probabilities in the decision tree. .
Iq a complex decision tree, a large number of events will require the assignment .
of utility values. The following procedure may be used to systematically determine
the relative utility values.of n ~yents,
. .. ._ namely, 1 to E.:
~
(1) Ra~k E1 in order of preference such that 1 >:- 1 >-.>-E.

(2) Assign u(E 1} = 1.0 and u(E.) = 0.
(3) Using the method suggested above for three events, determine u(1 ) such
that indifference is obtained between the following lotteries:
Observe that the two extreme e:vents . 1 and E. are taken as the reference
points for the lottery.
(4) Repeat Step 3 (n - 3) times with 1 replaced each time by 3 , ... , E.- 1
respectively; of course. the value of p obtained each time will very likely be
different.
(5) At this stage, a set of utilities u(E 1). u(1 ). .. , u(E.) has been determined.
However, to provide cross checking on these values, repeat Step 3 using u(E 1)
and u(E.- 1) as the set of new reference points and determine a new utility"
value for 1 , namely u'( 2); such that indifference is achieved as follows:
~
P E,
o--- -''-- Ez - .. ., . 1-p
.,:.,
If the utility values are ~onsistent, u'(E 2 ) should be equal to u( 1 ).

(6} Repeat Step 5 (n- 4) times with E 2 replaced each time by 3 , , E.-z,
respectively.
. If any inconsistencies are found in the above procedure, repeat the process until
all utility values agree satisfactorily. If time permits, steps 5 and 6 may be repeated
with E.- 1 replaced by._ 2 or E~_ 3 , respectively, in each cycle. _
The ranking of preference may be difficult sometimes, especially if the E,'s arc
multiattributed; that is, they include Several measures of effectiveness such as the
combination of monetary vaJue and p<)llution level. As an alternative, multi-
attribute decision analysis (sec Section 2.6) may be used.
EXAMPLE 2.30
In a decision analysis that includes various alternatives to overcome air pollution from sulfur
dioxide (SOl) and carbon monoxide (CO) in a large city, the utilities of the following four
events are to be determined: (i) low CO. Iow S0 2 (ii) high CO.Iow S0 2 ; (iii) low CO, high S0 2 ,
(iv) high CO, high SOl concentrations. It is obvious that the desirability of the events would
vary inversely with the concentration of pollutants. Furthermore, the decision maker feels it is
more urgent to reduce sol than co at this time. Let
E 1 = {low-CO, low SOl}

El = {high CO, low SOl}
E 3 = {low CO, high SOl}
E., = {high CO, high SOl}
Then, the preference order is
-- In accordance with the procedure suggested above, assume
. and start with the questioning process. Suppose the decision maker is indifferent between each
of the following pairs of lottery, namely:
Then, from Eq. 2.25
u(.;l) = 0.8u(E 1 ) + 0.2u(E )-= 0.8

4
u(E~) = o.4u_(E 1) + 0.6~[E..t ~ 0~4-

_ln !Jrder.to provide cross checks on these utility assignments: the decision maker is asked for
the value ofp that he"would"be indifferent between . ' . . . . .
ELEMENTARY CONCEPTS OF UTILITY THEORY 6}
Suppose the answer is j,= 0.5. TheD. the new estimate of the utility of 2 is
u'( 2 ) = 0.5u(E 1 ) + 0.5u(E3)
= 0.5 X 1 + 0.5 X 0.4
= 0.7
This is not equal to the value of u(E 2 ) determined earlier. Therefore, inconsistency exists and the
decision maker is asked to reevaluate his assignment of probabilities. In the second round.
after careful consideration. the decision maker is found to be indifferent between each of the
following three pairs of lotteries:
~EI.
Et - t.
Es - ~~
~EI
Et - 5
Es
From the first two pain, we compp~c:_ ..

u(Ez) = 0.75 and u( 3 ) = 0.45
Substituting these into the tbird pair, we obtain
u'( 2) = 0.55 X 1 + 0.45 X 0.45
= 0.55 + 0.2025
= 0.7525
which js close to the vaiue of u(E2 ) = 0.75. Hence, the utility values of the four levels of air
pollution would be:
u(low CO, low S01 ) = 1.0
\ u(high CO, !ow S0 2) = 0.75
u(low CO, high SO,] = 0.45
u(high CO, higl:i. SO,) = 0.0
. !
2.4.4 Utility Function of Monetary Value

Consequences are often expressec,l in monetary terms. However, monetary value
is not always a consistent measure of utility; that is. the preference order of the
decision maker may depend on the magnitude or amount of money actually
involved. Consider the following. .
A decision maker is asked to choose between alternatives I and II from the
following pair oflotteries: . - . . ..
64 DECISION ANALYSIS . -..
In lottery 1, the outcome will be either E 1 or E 2 with equal likelihoods; whereas
in lottery II., the only outcome is E 3 for sure. If the monetary values of the respective
events are d('1 ) = 10, d(E2 ) = 0. and d(E 3 ) = 5, alternatives I and II are
indifferent to the decision maker. However, if d(E1 ) = $1000, d(E2 ) = $0, and
.d(E3 ) = $500, alternative II may be preferred as _ t here js a sure -gain _of $5_0 0;
whereas with alternative I there is aso% cnance of gaining nothing, and the chance
of gaining an additional $500 may not be as appealing. Observelhat the expected
monetary-values are the same for the two alternatives in all cases. .
The above illustrates the .fact that the order of_preference for a decision maker
may change with the amount of monetary value. Moreover, the preference order
may also depend on the financial status of the decision maker relative to the
monetary values of the alternatives; for example, for a millionaire, the above
lotteri6s may be indifferent, e~en with d(E 1 ) = $1000, d(E2 ) = $0, an d d( 3 ) = $500.
In general, therefore, ll utility function involving monetary . values for a specific
decision maker must be esublished. '
The construction of a utility function for money is similar to the method of
determining utilities for general outcomes E, as discussed in Section 2.4.3. The
procedure is illustrated by the following example. \ .
EXAMPLE 2.31
An engineer is performing a decision analysis involving consequences ranging from a loss of

$20 to a gain of $100. His utility function of money over this range is to be determined.
To establish the utility scale, the utilities ofSIOO and SO are assigned arbitrarily to be 1.0 and
0, respectively. Since the utility value of 0 is not assumed for the lowest dollar value involved,
the utility function will take on negative values. The first pair of lotteries is presented to the
decision maker as follo~s:
With lottery I, the engineer stands a 50-SO chance of gaining SlOO or nothmg; whereas he
will get Sx 1 for sure with lottery ll'At what value of xj will the f'vo lotteries be indifferent?
Suppose .x, = 30; then.
. .. .
u(30) = 0.511(100) + 0.5u(O) = 0.5
~e second pair of lotteries presented to t_he decision maker ~ ..
Now, suppose x 1 = ts; then,

u(l5) = O.supO) + O.Su(Oj_ "" 0.25
. . .. ......
ELE.'riENTARY CONCEPTS OF UTILITY THEORY 65
Similarly, with a third pair of lott~ries
and with x 1 = 55,

u(SS) = O.Su(lOO) + 0.5u(30) = 0.5 + 0.25 = 0.75
In order to establish the utility function for negative monetary values. the decision maker is
given the following pair of lotteries;
A value of x 4 is required to ~tablish indifference. Suppose x4 = -25; then.

u(O);;,u.Su(lOO) + O.Su( - 25)
and
"u(-25) = 2 x u(O)- u(lOO) = -1.0
At this stage, six pointson the utility function for the range of monetary values from - S25
to SlOO have been determined. Figure E231 graphically shows the utility function fitted
through the six utility values.
In Example 2.31, the probabilities of each lottery are set at 0.5 and the decision
maker is asked to determine the doll!lr values so that indifference is achieved. In
earlier sections, the events 1 were fixed and the decision maker was asked to
determine the probabili~ p s0 that indifference is established. Clearly, therefore,
there is more than one method of constructing the required utility function.
If the decision maker is a large firm or government organization, the utility
function over a range of monetary valu~ may be a straight_ line, for exampie; the
dotted line in Fig. E2.31. Mathema.tically, if d stands for the dollar value; the ~fility
function of money for a largo;~ may be given by the equation
u(d) = ex + fJd (2.30) .
where IX and fJ are constants. Inv.erting Eq. 2.30, we obtain
1
d = p[u(d) - !X]
= ri + fJ'u(d) (2.31?.
where IX' and fJ' are also constants. Since u(d) is a utility function, applying the
property of linear transformation, the dollar v~ue d is also. a consistent utility
function. In other words, monetary value may be a valid utility furiction for a large
finn.
. --..
u ldl
- 1.0
100
d.
Figur~ E2Jl A utility function of money.
2.4.5 Utility Function of Other Variables

Using.a si~ar approach, we can transform the value of other variables involved
in a decision &nalysis, such as time, settlement, or the pollutant level to a utility
scale. Mathematically, the construction of a -utility function involves the trans-
formation of the values of these continuous variables'mto respective consistent
scales in the utility space, representing the decision maker's actual .preference for
values of the original variables.
As an illustration, consider the decision analySis of the development of the
aii-port facilities in Mexico City (de Neufville and Keeney, 1972). One measure of
effectiveness is the average access time to the proposed airport. Possible values for
average access time ranged from 12 to 90 minutes. Sine~ desirability i~creases with
decreasing access time, utility values are arbitrarily aSsigned to these two extreme
times as follows :
u(12) = 1.0 and u(90) =0
After several pairs of indiffer.ent lott,er.ies ar~ established, the .utility function for
average access time is plotted in Fig. 2.7. ..
If access time is the only measure of cffectiyeness, then this utilityfunction alone
will be sufficient for determining the best a_ltemative. However, there are other
I?easures of effectiveness such as cost, capacity, noise pollution, social disturbance,
accident rates, etc. to be considered in the decisio~ analysis. The methods for
combining several utility functions into a single overall utility function .for decision
analysis is presented in, Section 2..~:~ ! . . .. . : .. :
ELWENTARY CONCEPTS OF UTILITY THEORY 67
-;;
-;
Q5
~
5
Average Access Time In Minules, x
Figurr 2.7'- Utility. function for average access time (from de Neufville and Keeney, 1972).
-.-- ~
2.4.6 Maximum Expected Utility
The Maximum Expected Utility Criterion It had been shown earlier that monetary
value may not necessarily represent a true utility scale. Accordingly, the max EMV
criterion may not always be the proper criterion for selecting alternatiyes that will
reflect the decision maker's actual preference. A more general criterion for decision
is~fu~~ .
Once the utility of each consequence is known, the expected utility .value of each
alternative is given by
i = l, ... ,n .(232)
\
. where u 11 is the utility ofthejth consequence ~sociated with alternative i and U 1 is
the utility of alternative i. MaximiZing the expected utility, therefore, yiel9s the
maximum expected utility oolue criterion. The optimal alternative according to
this max EUV criterion is the one whose expected utility value is
E~Uopt) = m~{ ~Pti"J} (233)
In the ~e in which the coriS~uences are expressed in monetary terms, d11 ,

Eq. 2.32 becomes
E(U1) = L piJu(diJ) i ,;,. 1! .... , n (234)
1
where u(d 11) is the utility function of money established for the pcirtiCular decision
maker. . . .
The procedures for decision analysis based on maXimum exPected-utility value
criterion will follow those described in Sections 22 and 2 3. The only modification
.68 DECISION ANALYSIS
is the replacement of diJ by u11 or u(d11) before taking the expectation at a chance
node.
EXAMPLE .2.32
. .
Consider the example discussed in Section 22..2 .(see Fig. 2.2) in which the owner is deciding on
whether or not to install an emergency power system. Sup~ his_utility function for money is
u(.x) ~ -.e-"''000; .. .x < 0
The expected utility value of the tint alternative (fcir installation) is thus
E(U1) = 0.1( -~OOOI'CI) + 0.9( -roooi'IIIKI)
:,_0.1( _,o") + 0.9( _,o")
= -1.49
. Silnilarly, for the no-installation alternative,

E(U11) ~ 0.1( -eo.ooo1sooo) + 0.9( -t'o,ooo)
\ = 0.1( -e2 ) + 0.9( -e~
= -1.64
According to the maximum EUV criterion, the 'owner should install the emergency power
supply system. Then he can avoid the potential risk oflosing $10,000, which is a considerable
disutility to him. Although he would have chosen not to install the system had he been a maxi-
mum expected monetary value decision maker (see Example 26), his risk-aversive behavior
as described by the exponential utility function causes him to select the more conservative
:arternativc.
:::: Itmay be worthwhile to observe that taking the expected value is only one
:method cf combining the utility and probability values of all possible outcomes.
Fer example, a conservative decision maker m'!y subconsciously choose to assign
more weights to those outcomes associated with the least- utility. In doing so, he is
overaccouining for his risk-aversive behavior, which should have already been
-properly accounted for in establishing his utility function. Since utility is a better
measure of value to the decision maker than monetary values, the max EUV
criterion should prove to be superior to the max EMV criierion. If the decision
maker is consistently using the max EUV criterion_iri all his decisions~ he will
always be maximizing the true utility, and in the long run ~ be assured of o~
taining the best value relative to decisions based on other
. . ... criteria.
Value of lnformlltion For decision analysis based on the maximum expected
utility value criterion, the value of information (e.g., from an experiment) may be
obtained through a trial and error procedure. Assuming a specific value for the
experiment cost. say x 1, we may calculate the expected utility of the experiment
alternative E(UT) based on the total monetary value incJuding the test cost. By
repeating this calculation with severl!-1 other cost values, a graph such as shown in
Fig. 2.8 can be drawn. E(U.,.) is the expected utility . ofthe optimal alternative with-
out the experiment. The point at which the E(UT) curve intersects E(U.,.) gives
the value of information, VI. accruable from the specific experiment. A similar
procedure may be followed to determine the value of perfect information, E(UPT),
as indicated in Fig. 28. ..
ELEMEiVTARY CO.'VCEPTS OF UTILITY THEORY 69
E(Ul
xz 3
Cost of Experiment
Figuu 1.8 Determination of VI and VPI in expected utility analysis.
EXAMPLE 2.33
The value a( a laboratory test and the value or perfe1:t information for the waste treatment
process design discussed earlier in Examples 214 and 215 can be reanalyzed using the maxi-
mum expected ut~ity criterion. Suppose the utility function of monetary value is
il(lt}_= ....;e-"'50 x< 0
where xis the monetary value inS 1000. Then, the expected utility value ordesigxi A. is
E(U,J = 0.7u{O) + 0.3u( -100)
= o.7( -e") + o.3( -r)
= -292
whereas that for design B is

E(U8) = 0.7u(-50) + 0.3u(-20)
=-ps
Therefore; without the test, t~ engineer should sele1:1 design- B with E(U 8 ) = -2jS. To
obtain the value of information from the proposed laboratory model tes.~ several coS{SOf the
.tes! are assumed. with the results for E(U T) summarized in Fig. E213. To illustrate the under-
lying calculations. consider a test cost of$10,000.
From the subtree of Fig. E214, we have
E(U.. I HR.) = 0.95u( -10) + O.OSu( -110)
= -1.61
E(U.IHR) = 0.9Su( -60) + 0.05u( -30)
.
.:. . = -3.25 . .
.:
. E(U .. iMR) =0.637u(-IO) + 0.363u(-ll0)
. . . . . = -4.05 .
E(U8 1MR) =0.637u( -60) + 0.363u( -30)

= -278
E(U.. I LR) = 0.14~u( -10) + 0.857u{ -110)
= -7.91
E(U8 iLR) = 0.143u{ -60) + 0.857u{ -30)
. = -204
70 .DECISION .4NALYSIS
E(UTI
-2
~
5
..
-o
u -3
.
a.
11.1
0 10 20 VI 30 VPI 40 50
Tn! Cos! !$10001
Figur~ 2.33 Value of information.
Hence, if the test indicates HR,, design A is optimal; otherwise, design B will be optimal. The
expected utility value ()f the test is
\
E(UT) = 0.59( -'-1.61} + 0.165{- 2 78} + 0.245{- 204)
:=:. -1.9_1
m
which is shown as point A Fig. E233. Compare this value with the expected utility of-235
of the optimal alternative without the test. Since this E(UT) is higher than -235, a larger test
cost is assumed in the next trial.
~ From Fig. E233~ the VI and VPI are determined to be $26,000 and S35,000. respectively.
These values far exceed those obtained earlier for EMV analysis (namely, SI 5,140 and $24,000,
respectively}. It appears that the risk-aversiveness of a decision maker increases the value of
additional information. ln other words,~the decision maker will value any reduction in the
le~l of uncertainty.
2.4.7 Common_ Types of Utility Functions ----

Most utility functi.o ns are convex such as that shown in Fig. E2.31; this means
that the marginal increase in utility decreases with increasing value of the attribute.
The preference behavior of a decision maker"exhibited by a convex utility function is
commonly referred to as "risk-a'\ersiveness.". Most people are risk-aversive to a
certain degree; some may be more risk-aversive than others. The mathematical
forms of utility functions commonly used to model such risk-aversive behavior
would include the following.
1: .. uponential Utility Function

u(x)= .a+be-'x y:<:::O (235)
where ;he parameter 'Y is the measure of the d~8ree ~frisk-av~rsion, anc;l a and bare
normalization constants. If the utility function is normalized such that u(O) = 0
and u(l) = 1, then the.normalized exponential. ...utility function
. becomes
1 . .: ..
u(x) = . _ e-r (1 - e- 7~ . }';;::: 0 (236)
1
as }' increases, me utility function becomes more convex (Fig. 29); indicating
higher risk-aversion.
ELEMENTARY CONCEPTS OF UTILITY THEORY 71
u (x)
1.0
05
1-e-Y
u(xl-- .
1-e-Y
Q2 0.4 06 QB 1.0 X
Figure.l..JJ. - Exponential utility function . .
1. Logarithmic Type
u(x) = a In (x + {J) + b x+fJ >0 (2.37)
where {J is a parameter, generally corresponding to the amount of capital reserve
of the decision maker; that is, as {J increases, the decision maker has more utility
to spare (such as money) and becomes less risk-aversive (see Fig. 2.10). A normalized
logarithmic function may be as follows: ...: ..
. .
u(x) ,.. In e;P)
1
~ex + {J) - 1n PJ x+fJ>O (2.38)
J. Quadratic Type
(2.39)
where a: .is the parameter related to the degree of risk-aversion. A normalized
quadratic utility function may be"sho~n to be;
1 .
u(x) = - - - (x
1 - rz/2
- ia:x2) (2.40)
The degree of risk-aversivencss of a decision maker is measured by the .. risk-

aversion coefficient"
r(x) = _ u"(x) (2.41)

u'(x)
.
u(x)
1.0
u (x) ~ In (lt+,Bl-ln.B
ln! 1 ;~)
Q2 0.4 0.6 08 1.0 X
. . Figare 2JO Logarithmic utility function.
where the prime denotes the derivative with respec~ to x. The coefficient measures
the negatiye curvature of the utility .function relative to the slope of the utility
function at a given value of .x. For the exponential utility function, the risk-
aversiveness can be shown, using Eq. 2.36 in Eq. 2.41. to be a constant; that is,
~ . ~~-7 ~~
Applying Eq. 2.41 with Eqs. 2.38 and 2.40, ~;~~ show the coefficients of risk-
aversion for the normalized logarithmi" and quadratic utility functions to be,
-- respectively,
1
r{x) = - - (243)
. .x +p
and
a .
r{x)=-- . (2.44)
1 -.ax
which are both functions of x . .
Observe that the coefficient of risk-aversion does not vary with the attribute in
the case of the exponential utility function ; whereas in the case of the logarithmic
. utility function the coefficient of risk aversi.on decreases with x, and in the case of
the quadratic utility function the risk-aversion increases with x.
The utility function, of course, may also be concave upward; that is, the marginal
increase in utility increases with increasing values of the attribute' x. In such cases,
the preference behavior of the decision maker is referred to as "risk-affinitive.~
The utility function of monetary value shown in Fig. 2.11 represents an example of
this behavior. It is believed that this preference -behavior is ordinarily not realistic.
ELEMENTARY CONCEPTS OF UTILITY THEOR'l 73
Monetary Value, d
Figrue 2.11 Utility function of money for a riskaffinitive person.
2.4.8 Sensitivity of Expected Utility to Form of Utility Function

Usually, it is difficult to ascertain what type of utility function is most appropriate;
for eumple, whether it should be of the exponential or quadratic form. The correct
choice of the form of the utility function, however, may not be very crucial, especially-
if the expected utility values are not sensitive to the form of the function.
In order to examine the sensitivity ofthe expected utility (associated with a given
action) to the above three forms of utility functions, consider the simple case in
which the possible outcomes from an action can be described by the value of a
random variable X. In this case, the expected utility of a given action may be
expressed as follows:
E(U) = J:..,u(x) fx(x) dx (2.45)
where fx{x) is the probability density function of X and u(x) is the utility function.
Using the second-order approximation to evaluate the expected utility (see Chapter
4, Vol. 1) yields the result
E(U):::: u(Jc) + t Var(X)uw(x) (2.46)
where x and Var(X) are the mean and variance of the random variable X, and u"(X)
is the second derivative of the utility function evaluated at x.
Applying Eq. 2.46 to the three types of utility functions described earlier, the
second-order approximation of the expected utility becomes, respectively. as
follows. For the exp<mential utility function,
.. . "i. ., . .
E(U) :::: - - - -7 [1 - e-i~..:. t/ Var(X). e-7z:J (2.47)
1- e
For the l<?garithmic utility function,
E(U)
..
~ ('I
ln - +
~ [1n .(~ + .p) -' ln p-. t .Var(X)
.
C
1
x.+ .
fl)Z] (2.48)
. -p- . .
1
E(U) ~ - - [x -1u{x2 + Var(X)}] (2.49)
1-~
.2
It has been shown (Ang and Tang, 1979) that the expected utility is relatively
insensitive to the form of the utility function at a given level of risk-aversion, and
that the expected utility docs not change significantly over a wide range of risk-
aversion coefficients. Hence, the exact form of. the utility function may not be a
crucial factor in the computation of an expected utility. Moreover, the risk-
aversiveness coefficient in the utility function need not be very precise; that is, any
error in the specification of the risk-aversiveness coefficient may not result in a
significant difference in the calculated expected utility. In short, the problem of
as<"Crtaining an accurate utility function would not be crucial in the application of
statistics] decision analysis.
As indicated in Eq. 2.46, an approximate expected utility value may be com-
puted on the basis of the mean and variance of the pertinent random variable. This
would suggest that the entireprobability density function may not be necessary
in most decision analysis problems. In practice, the first two statistical moments
could be all the information that may be available for a random variable; hence,
Eq. 246 provides a convenient approximate formula for computing the expected
utility of a given alternative.
2.5 ASSESSMENT OF PROBABILITY VALUES
2.5.1 Bases of Probability Estimation --- .. _

There are basically three approaches to the estimation of probabilities. The first
---approach is based entirely on observed information. Probability could be estimated
froni the relative frequency of observed events. Moreover, !l probability model may
be fitted to a set of observed data. With the model, the probabilities of the events or
interest can be computed The validity of this approach requires a large set of
observed data. . .
The second "approach is to eStimate probabilities ofevenis or distributions of
nuidom variables solely on the basis of subjective judgirierit of the decision maker.
This is necessary when direct observed_data are not available; or when the decision
. maker believes that he has integrated alfthe"relevant information in-his mind, thus
possessing a strong belief about the likelihoOd of th~ events of interest. The third
approach is commonly referred to as the Bayesian approach in which the judgmental
information of the decisio.n maker is combined with the observed information for
an updated estimation of the probabilities. Examples of these have been presented
in Chapter 8 of Vol. 1. In both the second and the third approaches, the accuracy of
the probability estimation will rely on how weil .the subjective probabilities can
be realistically extracted from the decision maker. In the foilowing sections, some
of the methods for obtaining these . subjective-probabilities
-. - . ~ ..are
. presented.
' .
ASSESSMENT OF PROBABILITY YALUES 75
~E, _
-~Ea
Figur~ 2.11. A simple branch in a decision tree.
2.5.2 Subjective Probability

Consider a set of branches as shown in Fig. 2.12. where the probability p of outcome
E 1 is to be estimated such that it will realistically represent the judgment of the
decision maker. .
Suppose the utilities of events E 1 and 1 are already known, for example, through
procedures presented in Section 2.4.3. If the decision maker is- indifferent between
the lottery in Fig. 2.12 and another alternative that guarantees an outcome Of
event E 3 , then fro.m Eq. 2.25;
or
. (2:50)
which could be easily determined provided that the utility of event 3 is also known.
EXAMPLE 1.34
A foundation engineer would like to estimate the probability of the presence of pockets of
soft material in a soil stratum that will affect the amount of differential settlement in a structure.
With the proposed design A, he; expects that the structure will have only 0.5 inch of differential
settlement if soft pockets are not. present; whereas the presence of soft pockets will lead to a
\ differential settlement of 2 inches. Suppose the following utility function is available:
u{x) = -Jxz
where x represents the differential settlement in inches.
The foundation engineer could be asked to choose between design A and a hypothetical
design Bat a comparable cost that will guarantee a differential settlement of x~ inch, as shown
in Fig. E2.34. . .
Let x 0 take on an initial value of 0.6 inch. If design B is preferred by the engineer, x 0 will be
increased slightly, say 0.7 inch, and the engineer is asked again to state his preference between-.
. . . .
Oilfereniial
Settlement (in.)
. Design A 2.0
QS
1.0
Figurr E2J4 Decision tree for evaluation of probability p.

designs A and B. This procedure continues until an indifference condition is achieved between
the two designs. Suppose .x0 ,. 1.0 inch at the indifference point From Eq. 2.50,
. u(l) - u(O.S}
P = ~2) - u(O.S)
(-3 )( 12 ) - (-3 )( 0:5 2 )
= (-3 x 2 2) - (-3 x 0.52)
= 0.2
Hence, the fo~dation engineer's subjective probability that a soft pocket of soil is present at
the site is 20%
1.5.3 Subjective Distribution of a Random Variable

Decisio~ analysis may sometimes require the knowledge of the entire. distribution
ofa-'random variable, i ncluding its inean.imd -v ariance.ln eases:where observed
data are not available, the distribution of the random variable may be established
on the basis of subjective judgments. A common method for this purpose is the
fractile method.
The decision maker would first estimate a value such that the random variable
will be equally likely to exceed . or fall below-that value. This value is the so per-
centile value .of the random variable. denoted ~s x 0 5 The decision. m~er would
then concentrate on the upper range of the values. He may ask himself the question:
"Supposethe value of the random variable is in this upper half range: what is the
value at which the random variable will be equally likely to be above or below it?"
This will give the value .x 0 75 ; that is, the 75 percentile value of the random variable.
Simii~rly, the procedure ~ould be repeated for the lower half range to derive ..x 0 .25 .
As a consistency check, the decision maker may ask himself, "Is it equally
likely that the random variable would be within the range x 0 . 25 to x 0 15 or outside
the range?" If the answer is no, he should go bac1< and re~se his previous assign-
ments on the percentile values .until consistency is obtained. Similarly, he can
-continue to subdivide each of-the four ranges, namely,< x 0 . 25 , ..x 0 .H to x 0 5 , x 0 .s
to .x0 . 75 > x 0 1 5 into two equally likely intervals, Again, he would cross check if
these eighth fractile values are consistent. Depending on. how far out in each tail
of the distribution function is .required, this .subdivision process can be repeated to
yield.low: and very high percentile values.. .
The fractile values. are then plotted to give th.e cum.u lative.distribution of the
random variable. As an illustration, Fig. 2.13 shows the cumulative distributions
of the mean friction.angle for a- sand that has been extracted from two engineers
using the fractile method. Curve B .is extracted from. an engineer who has more
experience with this particular type of sand. Based on this.c1,1mulative distribution,
the density function as well as the mean and variance of.the mean friction angle can
be computed. For a more detailed discussion and illustrations o(the fractile method,
readers may refer to Raiffa (1968) and Schlaifer (1969). . .. . .
The fractile method could be very time-consuming and mentally stren.uous for
most engineers. Quite often the available jnfotmation -is limited. to th.e upper and
lower bounds of the random variable. In this case. a distribution may be a,ssumed
over the suggested range. For instance, a uniform distd.bution woul~ be a convenient
' .. ... ......

.
~
' 1 ,...
DECISIONS WITH MULTIPLE OBJECTIVES 77.
1.0
,....
0
.;
.2
-;
~
o;
.
c
~
0
-;
E 02
u"'
0
25 .36
Mean friction Angle , ~
Figlll't: JJJ Cumulative distrib~tion of mean friction angle based on judgment.
"
choice. Based on this subjectively assumed distribution. the mean and variance of -..
the random variable can be estimated accordingly. The statistical properties of
some commonly assumed distribt,_~tion over the range {a. b} are summarized in .
Table 6.1 of Chapter 6 that contains a detailed treatment of the analysis of-un-
certainties.
2.6 DECISIONS WITH l\<IULTIPLE OBJECTIVES
With the increasing ~wareness for social relevance in engineering designs and
planning, many engineering decisions involve conflicting interests of various
groups: For example, suppose the mayor of a metropolit;m city is faced with a
decision on solid waste disposal for the city. The main objectives of the proposed
waste disposal scheme might include the minimization ofenvironmental pollution.
the minimizationof restrictions on future land use, the minimization of potential
adverse political consequence, the minimization of systems costs to the city, and
possibly others. Most of these objectives are nonmonetary, and methods to
quantify them accurately and consistently in monetary terms may hav.e serious
limitations: -
Decision analysis with multiple objectives requires the determination o("Inulti-
. attribute utility functioiis instead of a single utility function. These multiattribute
functions would evaluate the joirit utility value of several measures of effectiveness,
toward fulfilling the various objeCtives. However. manipulation of these multi-
attribute utility functions Can be.extremely complex. depending on the size of the
problem and the degree of dependence among the various objectives. An in-depth:
treatment of deciSion with multiple objectives is found in Keeney and Raiffa(l976).
Churchman, Ackoft and Arnoff (1957) suggested a simplified model.for the
treatment of multiobjective decision problems. A relatively simple and direct
method is provided for evaluating a diverse set of objectives in terms of Weighted
relative utility values, and selecting the alternative that best balances the competing
objectives. Stimson (1959) applied it to public health planning. The model bas also
78 DECIS/ONANALYSJS
been applied tq urban planning by Davidson (1971) and Morris (1972), and to
water resources development by Erskine and Shih (1972). The model may also
have implications in engineering decision problems; the basic concepts of this
niodei are presented in the next section.
2.6.1 Weighted Objective Decision Analysis

Consider a decision-requiring the resolution of several objectives versus a number
of feasible alternative engineering designs. The relative merits of importance of the
various objectives may be different. and the likelihoods of the different alternatives
achieving a stated objective may also be different. The information required for the
weighted objective decision analysis can be summarized as shown in Table 21.
..
~~..
. Table 2.1 . Tableau fo~ Weighted ObjectiVe Decision Analysis - :~ ..
Relative Weights
Overall
01 02 01 D. relative
uti!i,y
Pu Pu'PuPl U1 =2:Puw;
J
P21 P22 P2J P2o ul,;, L P2Jv.-J

J
Pn P12 PiJ Pu.
P.t P.. 2 -I'M1: : P-
_.. a1 = The ith altemativ~ i = 1 to m.

0 1 = Tbejth objcctivc,j = 1 ton. .
~j =The relative weight. of importance to objective 0 1,j = l ton.
Pli =The probability of attaining objective 0 1 through alternative c.;.
U, =The qvcrall relative utility ofaltemative a1
...........
. T~ im~lement the model, a set of relativ~ weights h~s tb be assigned to the various
objectives. This involves a two-step ranking procedure ~.suggested by Churchman
and Ackoff (1954, 1957). First, the n objectives a,re listed in. decreashig order of
.. importance; this is called ordinal ranking. In this step, a preference statement is also
solicited from the decision maker With respect to combinations of objectives. For
. .example, if 0 1, 0 1 , 0 37 0 4 represent four objectives in d'ecreasing importance, then
in addition to the preference statcmen( , - .
. 0, >- 02 >: 0~
.
~ -04., ... -.. ..
. .. .. . .. ~
the decision maker may also be asked if he prefers 0 1 to a comblnation of0 1 , 0 3 ,

and 0 4 ; that is, whether or not . . ,
o, >- 0~ + 03 + 04
- Other similar preference statements may also be. req~i;e"d_..
DECISIONS WITH MULTIPLE OBJECTIYES 79
Subsequently, a cardinal ranking of each of the objectives is established. The

relative importance of each objective with respect to . the other objectives is
evaluated by assigning numerical weights to each objective. Starting with the most
important objective, it is assigned an arbitrary weight, for example, 100. Next, the
numerical weights are assigned to each of the other objectives. The set of initial
weight assignment is then cross-checked for consistency with the preference
statements established in the ordinal ranking. For example, if w 1, w 2 , w 3 , w4
represent, respectively, the set of relative weights assigned to objectives 0 1, 0 2 , 0 3 ,
0 4 , then a consistency check would require
w 1 > w2 + w3 + w4
if 0 1 >- 0 2 + 0 3 + 0 4 ;his assumes implicitly that the overall relative weight of
a combination of objectives is equal to the sum of the weights of the individual
objectives. This assumption of linearity may not be valid in all d~sion problems.
For instance, in a water supply planning project, if 0 1 denotes the objective of
reducing the demand for water and 0 2 denotes the objective of increasing the
water supply, then the importance of achieving both objectives would. generally
be less th~ the sum of the importances of the individual objectives since the
problem of water shortage rna y be solved by either 0 1 or 0 2 and may n<?t necessarily
require the simultaneous.achievements of 0 1 and 0 2 Hence, either w 1 or w 2 will
be very close to the weight of the combination .w( 0 I + Oz). Because of diminishing
return, w(0 1 + Oz) # w 1 + w 2 However, this .. linearity" assumption does con-
siderably simplify the decision analysis, and facilitates the quantification of the
relative importance of each objective. As long as extreme care is exercised to
insure that the objectives are mutually exclusive with respect to the contributions
to the overall utility, linearity should yield useful results. For example, the problem
previously mentioned could be avoided if the two objectives are combined and
denoted as no water shortage. .
In the event that inconsistency in tb,e preference statement iii found, the decision
maker is asked to revise his preference statement or his assigm:ilent of relative
weights or both, until all inconsisten'(ies are eliminated.
The next step in the decision analysis includes a listing of feasible alternatives
and the assignment of probabilities Pii Some of these probabilities can be com-
puted based on probab~lity models, whereas others may have to be estimated sub-
jectively baaed on the decision maker's expCricncc and judgment (see Section 25 .. '
for methods ofassessing subjective probabilities).
Finally, the overall relatiye utility o(each alternative is computed as .
. . u.-= I pljw1 (2.51)
. . j .
The optimal alternative is the one.that has the maximum relative utility. Observe
that the absolute numerical value of ui is not important; the relative value of ul
is sufficient for the selection of the optimal alternative.
~-
EXAMPLE 2.35 (excerpted from Davit/so,; 1971)
A planning group was asked to devise a general land usc plan for a large section of a rural
county in a southeastern state of the U.S.
lJU UG'- l.:liV/1 IUlAL I ..11..1
There: are four objectiveS that are of paramount importance in the planning. namely:
0 1: Provide for Regional Recre~tion. This includes.theprovision offacilities for both weekend
-and day recreati"n, such llli campsites, cabins, picnic areas, and hiking trails. The concept of a
regional.recreation area is within a 50 to 100-miJeradius of the land under consideration.
0 1': Pror;ide for Day Recrearion. This will primarily serve the local residents who are within
sliort driving distances. and involves no overnight facilities: ., .
0~: Encourage "Employme,;t Opponunities. The region is primarily a poor agricultural area
with a minimum number of industrial concerns. Improving job opportunities for loc:al resi-
dents is desired.
0 4 : Con5eTI!ation. Because of the natural beauty and character of this mountain region, the
planners felt that the conservation of the natural state is imperative.
Weighing of Objecrives.
In-ordinal ranking, the foilowmg preference ~tatements ~e obtained from the initial round of
questioning.
o.. >- o, >- Oz >- o3

o.. >- o, + Oz + ol \
. o,~ o2 + <?3
ln the cardinal ranking step..numeric:al ~eights are assigned to each of the four objectives. The
initial assignment of relative wei$1Jts is as follows:
w, =so
w 2 = 40
w 3 :=:_ 20
w.. = 100
These numbers imply that provisions for .regional rcen:ation, 0, is only half as important as
conservation, 0 4 ; day recreation, 0 1 , is forty percent as impoitant as 0 4 , and so on. The
difference between achieving 0 1 and 0 2 is relatively small
:Next, the relative weights w1 are examined fo~ consistency with the established preference
statements; for e~tample, the above weights yield . . . . __- . . .
w1 + wl + w 3 = 110
which implies that the combination of 0 1, 0 2 , and OJ is preferable to 0 4 This is inconsistent

with the second preference statement established in the ordinal ranking.
After discovering this inconsistency, the planners reconsidered both the preference state-
ments and the assigned relative weights. Following extensive debate, the following set of
preference statements and relative wei$1Jts were determined:
04 >- 0 1 >- 0 2 >-OJ

o~.~ o, + ol + ol
0 1 >- 0 2 +OJ
and
111 1 = 5(), "Hil = 25, WJ = )0, -..4 = }{)()
which may be shown to contain no inconsistencies.
' .
DECISIONS WITH JfULTIPLE OBJECTIVES 81
E.scablishmenc of Alrernacites
Six alternative land use plan$ were proposed. Although each plan could be implemented.,
none of the plans can achieve most of the objectives.
a, : Full Range Recreation. including picnic areas. cabins, campsites, beaches. boats. etc.
The plan fully exploits the recreational potential of the area.
a 1 : Limired Recreation. including a hunting and fishing preserve with limited cabins and
campsites. This plan would retain much of the natural character of the area while providing .
limited service facilities for the users.
a 3 : Conservation Lockout. development of a state or national forest with a number of hiking
trails. but very limited service facilities.
a. : Agricuitural, setting aside the tillable land for agricultural use only and preserving the
remainder in its natural state.
a 5 : Agricultural and Extractive Industry, combining the fanning aspect of a 4 with timber and
other natural resource (if any) removed. '
a 6 : Residential, the development of a community of second homes or vacation homes. nlis
would require the full development of the area in a traditional sense.
Efficiency l()de."C
In this step, the relative probability of achieving the objectives with each alternative is estimated.
For example. the conservation lockout alternative (a1 ) will completely achieve conservation
(0 4 ); thus, is assigned p 34 ,; lOOYo--Thc other alternatives will result in proportionately lesser
degrees of conseryation. The relative degree of achie\'ing a given objective (e.g.. 0 1) with each
alternative is given by p11 as tabulated in Table E23S.
Table E2.35 Decision Tableau for Land Use
Relative
weights, w1 so 25 10 100
o, Oz 03 o. Overall
Regional Day Relative
Recreation Recreation Employment Conservation Utility
a, \
Full-range
recreation 1.00 1.00 o.so 0.10 90.0
al
Limited . !
recreation 0.70 0.80 0.25 0.80 137.5
a,
Conservation
lockout 0.05 0.10 . . 0.05 1.00 105.5
a.
Agriculture 0 0.10 0.25 0.50 . 55.0
a,
Agriculture
and extractive
industry 0 0 0.70 0 7.0
~
Residential 0 0.10 0.40 0.10 16.5
Overall RelatiVI! Utility

The overall relative utility for each alternative is then computed according to Eq. 2.51 and
~in the last column of Table ~35. For example, the overall relative utility for the
full-range-recreation alternative (a 1) is
. UJ =. 1.0 X 50+ 1.0 X 25 + 0.5 X 10 + 0.1 X 100

= 90.0 .
A comparison of the overall relative utilities for the various alternatives indicates that
alternative al, providing limited n:c:reation, scores the maximum numerical utility values.
Observe that although al is the optimal alternative, it fails to provide the maximum possible
contribution to any single objective. However, it doe$ provide the best balance in meeting all
objectives.
EXAMPLE 2.36 (e:xcer:ptedfrom Erskine tUuJ Shih, 1972)

.
Tnis is a case study of the water supply problem in' San Angelo. Texas. Plans were developed
to correct the prc;>blem of inadequate water Supply. In 1972, the wate~ supply in San Angelo is
completely dependent on rainfall which has proved to be often inadequate. Moreover, there
is a general lack of ,systematic resource planning and mlinagement.
Following a .prelin1inary study, 'the 'feasible alternatives available for the future water
resourcedevelopment are proposed as follows: . . . . . .
a; : ~o~_tinuati<in of the present water management policies and complete reli~nce on rain-
fall.
al : Continuation of the present water management policies and development of adequate
ground wate~ resources in nearby McCulloch County.
a 3 : Revision of the current water management policies to include rationing and rate adjust-
ment programs, the development of a short-range water sup;JIY to smooth rainfall fluctuations,
and the initiation of a waste water reuse program. . -.. --.
.. The major objectives to be achieved in this wat~r resource planning co~ist of:
~.;
0 1: Meeting future wat~ demand_.

(!;i: Improving water quality to meet minimum health rcqu~ements.
,0_3 : .Mini~i%ing annual cost.
.0:.: Increasing recreational benefits such as water s~rts 3.Jid la.Wr! improvements.
o,, Social acceptance, considering social preferences based on faeto!'S. _other .than.the above
:objectives.
Weighing ofObjecrive~
. In order to estimate the relative weights for the.four objectives, questionnaires were sent out to
experts in the fie1ds'of engineering and economic-development, asking them to a,ssign weights
indicating the relative importance of the stated-objectives. Using ordina) and cardinal rankings,
each respondent reported his or her assigned weights for the fiye objectives together with a
justification for such weight. These individual assessments were summarized and then circu-
lated among all the experts involved. Each individual could revise his assessments at this point.
The updated TCS!Jlt~ were compiled and _the ~can weight for ~~ objective computed and
.......
DECISIONS WITH MULTIPLE Oli.JECTIJIES 83
~ded as the consensus. After normalization, the relative weights representing the consensus
opinion of the group on the five objectives were as folloow:s'
IVt = 44.9
w1 = 23.8
w3 = 15.8
w4 =1l.l
w, = 4.4
This approach of subjective assessment is known as the Delphi Method (Dalkey, 1969, 1972).
Probability Assessments
In order to compute the probability of meeting fut)lre water demands, data on the annual
runoff from rainfall and projection on the demand for water at the end of the planning period
under each alternative are evaluated. From 30 years of observed runoff data for the Concho
Rivers. and assuming that only 80% of the recorded runoff would be available for consumption
because oi evaporation, the annual net water supply may be shown to belognonnally.distri-.
buted with a !)lean of32,900 acrc-ft and a coefficient of variation of0.42. The planning period
in this case iS-taken to be 30 years. At the year 2005 the projected overall water demand is
estimated to be 22,000 acre-ft. Based on data compiled by Freese. Nichols,.and Endress (1970,
1971), the estimated additional water supply under alternatives a1 and a 3 are shown in Table
E2J6a. Observe that the demand .in_ the year 2005 is well-covered by the additional water
supply alone under a1 ; whereas for aJo about 0.75 x 'J'J 000- 8503::: 8000 acre-ft of water
will still be needed from the rainfall. Hence, the probability of meeting the damand 'in the year
2005 is simply the probability that the net annual surface water available exceeds its demand
for each alternative. These probabilities are computed to be 0. 7865, 1.0, and 0.9995 for alterna-
tives a" a1 , and a 3 , respectively.
To evaluate each alternative with respect to the improvement of water quality, interviews
with local political .ligures, city water department officials, and other experts were conducted.
Based on the limited information and reports by Freer. Nichols and Endress (1970, 1971),
ratings for each alter:native with respect to the water quality were established, respectively, as
0.6, 0.9 and 0.8. These values indicate that the quality of surface water is judged inferior .to that
of underground water. The rating is equivalent, in a relative sense, to the probability of each
alternative achieving the objective Qf good water quality. Similarly, ratings lor additional
recreational benefits are established as 0.0, 0.91, and 0.77, respectively, for the three altema._tives.
The costs of the altemati ves in most cases were estimated by informed officials. Based oq an
annual interest rate of 6% over a 30-year period, the annual cost figures for alternatives a1 and
a 3 are summarized in Table _E2.36a. The additional annual overall costs for a" a1 , and aj arc
computed to be $0, $1,10[,774, and $581,581, respectively. In order to convert this to numerical
ratings, the alternative with the lowest annual cost is assigned a score of 1.0 for achieving the
minimum cost objective, whereas the alternative with the highest annual cost scores a rating
Table El.J&z Cost and Water Supply Data (fr<?m Freese, Nichols, and ~ndress, 1970, 1971)
AdHitional Source Additional Additional Contn"bution

Alternative of-Water . . Annual Cost t~ Water Supply
Ground water . $1,101,774 22,40() acre-ft/yr

ReviSed management . . so 25% of normal consumption
Short -range supply -- . S94;540 2903 acre-ft/yr
Waste Water reuse $487,041 5600 acre-ft/yr
Total for a 1 $581,581 8503 acre-ft/yr ;t- 25% reduction
in water consumption
Table E2.36b Decision Tableau for Water Resource Planning Example
Relative Weights 44.9 23.8 1!i.8 11.1 4.4
Objectives o. 02 ~3' o. o,
Meeting Overall
Future Water Anilllal Recreational Social Relative
Alternatives Demand . Quality_ Cost Benefit Acceptance Utility
at 0.7865 0.6 1.0 0.0 1.0 69.78

al 1.0 0.9 0.0 0.91 1.0 80.82
al . 0.9995 . . 0.8 0.472 0.77 0.0 79.92
of zero. The annual cost is assumed to vary linearly with a rating bet wed~ these two points.
Thus, a nting of 0.4721 for alternative a3 is obtained. :
Finally; with respect to the social acx;eptance objective, a 3 is the only alternative with a
questionable sOcial implication because It includes a waste water' reuse program; therefore,
the first two alternatives were rated 1.0 with the third 0.0
Table E236b summarizes the informaJior. for the decision analysis. The overall relative
ut!l!ty for eac~ alternl!tive is computed acc<irding to Eq. 251. For example, the overall relative
utility for a 1 lS
. . . . .
. .
. .
U 1 "= 0:7865 X 44.9 + 0.6- X 23.8 + 1.0 )< }5.8 + 1.0 X 4.4
= ' 69.78 .. .
The.alternative with the hlghe~ relative utility .vaJue is a1 This implies that the best course of
action is to continue the present management policies and pursue the complete development .
of the ground water field in McCulloch County.
The we(ghted objective decision analysis provides a simplified method in the

solution of multiattributed problems. The decision maker can directly assess the
weights and probabilities as appropriate. In other words, he has complete know-
ledge of the entire process. A more rigorous approach to decision analysis with
multiple objectives is presented in Sectiori 2.63;nowever.this methodology lacks
the brevity and directness of the weighted objective approach described above. In
.. applying the weighted objective decision analysis, extreme care is required in
identifying the objectives such that they do not overlap with each other. Further-
more, the limitations of the linearity assumption should be recognized.
2.6.2 Multiattribute Utility Approach

The previous sections show that if all the consequences In a decision model can be
conveniently described by a single . utility measure, the expected utility of an
alternative; say att is computed by
.. . E(Ui) ~ P;1 ~i1 1 '- L (2.52a)
J
.w here uii is the jth conseq~ence of the itli alternative; ana Pii is the corresponding
probability. Alternatively, if the riatural .outcome 'is governed by the value of a
continuous random variable X, the expected utility ofa1 given by is
E(U 1) = f'., u(x) f~ (x) dx (2.52b)
where f~x) i~ the-PDF or'X, correspo~di~g to.~temative.a1:_

: ... . :.. :. ... .
DECISIONS WITH MULTIPLE OB.iECTIYES 85
This computation of. the expected utility can be generalized to deciiion analysis
in which the effectiveness is measured by several attributes, such: as cost, time.
noise pollution. and others. It is obvious that each of these attributes will reqUire its
respective units of measurement. For example dollars, minutes, and CNR are
standard measures of cost, time, and noise level, respectively. The corresponding
utility function and the associated probability density function will, therefore, be
multidimensional. Hence.
E(U i) = f
~
J~u(x 1, :c2 , :, x,.)f~, ... x.(x 1 , , x,.) dx 1 dx.
.
(2.53)
.
where X 1 t.o X. are random variables describing values of then respective attributes
associated with each alternative. Determining these joint utility and density
functions requires the evaluation of the conditional utility and probability functions.
This task can be cumbersome, especially when the number or attributes is large.
Moreover, these functions may have to be developed entirely or largely on the basis
ofsubjective judgments; interviews of experts to furnish descriptions of u(x 1, , x.)
and !L ...C:.t. (x 1, . .. , x,.), for all ~ would generally be impractical. Appro
priate assumptions have been proposed (Ting, 1971; Raiffa, 1969; Fishburn, 1970; -
Keeney, 1972; Keeney and Raiffa, 1976) to simplify the determination of these
functions. Some of these assumptions are as-follows..
If the random variables x,s, i = 1 ton, are statistically

Statistical lndependtiiiCI!
independent, the joint density function - -
(2.54)
- .
where f~1(x1) is the marginal density function of X 1 corresponding to alternative a 1
Hence, the assessment or an n-dimensional function is simplified to that of n one-
dimensional functions: - -""' -
'
- f
- . : -
Preferential Independence Trade-offs between any two attributes are 'governed
by the unique indifference curve between these two attributes regardless of the
values of other attributes.
Utility Independence The relative utillty of X ; remams the same-regardleS.s ofth~

values of other X/ s. In other words, the utility of each of the :X/s can be separately
determined. -
The assumptions of preferential indepenqence and utility independence together
imply that the joint utility fun<:tion may be-expressed asa function of the marginal
(single attribute) utility functio~, n~_el~ -
(2.55)
After the marginal utility functions have been obtained. the function g may be
determined by scaling u!..xJ with respect to other utility functions such that they
are _consistent with each other. The procedure is described in the following.
86 DECISION .ANALYSIS
For an n-dimensional utility function whose variables are mutually preferential

and utility inqependent, the joint utility function u(xh x 2 , , x,.) satisfies the
folloMzl:g equation (Kenney and Raiffa, 1976)
-
~.
1. ;.. ku(x 1 , , x..) = TI [l + kk u.(xJ]

t~l
1 (256)
"" . ' ~
where k and k1 .(i = 1 ton) are constants to be evaluated; u.(xJ is the marginal
utility function that has been scaled such that u.(*xJ = 0 and u.(xt') = 1; x1 and
xt are defined as the values of x 1 that give the m,inimum and maximum values of
u.(xJ, respectively.- . ~.
It is obvious that the outcomes (*x 1, x.l, ... , *x..) and (xf~ xJ, . . , .x:) are,
respectively, the least and most desirable ones. To .set up the relative utility scale,
u(*x 1 ;*x2 , ... , x~) and u(xf, xj, . . , , x:) are assigned the reference values of
0 and 17 respectively. Then, from Eq; 2.56, the utility function with .x1 at the most
desirable state and all other attributes at the least desirable suites is given by
1 + ku(*x 1, . . . , xt, ... , xft)
= [1 ld- kk 1 u 1(x 1)] . . . [1 + kk1ut("xnJ .... [1 + kk"u,.(x.)]
= 1 .. : (1 + kki) .. 1 = 1 + kki .
Thus,
i=l, ... ,n (257)
Moreover, b~ substituting u(xf, ... ~ x:) into Eq. 25~, we obtain
1 + ku(xt, .. . , x:) = TI [1 + kk,u~xt)]

I I
or
n (1 -t'kk,)
ft
1+k = (2.58)
il
Hence, the constants k 1 and k can be det~rmined from "Eqs. 2.57 and 258 once the
utilit} value u(*xlt ... , xt, ... , x.) is known for all i. :
Consider the two-dimensional case; assign u(*xlt *x2) = 0 and u(xT, xJ) = 1.0.
The value of u(xt,x 2 ) can be determined from a pair of indifferent lotteries as
shown in Fig. 2.14.'Suppose the decision maker is ind-ifferent between alternatives
I and II at probability p 1 ; tlJ,en,
u(xf, xi)= p 1 u(x!, xJ) + 0 :"""' p 1 )u{*xlt !x2 )
. = P!
cxl,xzl
Fixure l.U Indifferent lotteries. .
.
DECISIONS WITH JIULTIPL OBJECTI'YES 87
Similarly, the value of,u(*x 1,x!) may also be determined by replacing (xf, 4!xz) .
with (*x~o xt) in alternative 1 above. Again, it may be shown that u(*x 1, xt) will.
be equal to another probability p 2 .
Hence, from Eq. 2.57,
and from Eq. 2.58,
or
Eq. 256, therefore, becomes

1 + ku(x~> x1) = [1 + kk 1u 1(x 1 )][1 + kk2uz(~2)]
or v
u(x 1, x 2 ) = k 1 u 1(x 1) + k 1 u 1(x 1 ) + kk 1k 2 u 1(x 1 )uz(x1 )
= .Ptilt(Xt) -+. P.2U1(x1) + (1 - P1 - Pl)ul(xt)ul(xl) (259)
This procedure may be extended to derive the joint utility function for the n-
dimensional case. It can be shown that the joint function will be a sum of various
products of marginal utility functions. Again, the constants k1 (i = 1 to n) may be
evaluated from n pairs of indifferent lotteries.
EXAMPLE 2.37
As the scheduled completion date of a construction project approaches, a contractor has to
decide whether he should .rent a set of modem equipment to speed up the construction.
With the present equipment, the job may_be c_ompleted within no S weeks (see Fig. E231a)
\
f,(t)::,... 3 f 1(x);,.... 2
rr .1 ... , 112
~ 3
I, Duration In Weeks x, Cost (105 $1
(a) (b)
r2 !tl:,.... 2.5.. fz(xl:,.... 3

<T. 1/2 "'I/~
2 4
I, Duration In Weeks x, Cost (105 $1
(c) (d)
Figur~ E2J7 Distributions of durations and costs of projects.

costing from SlOO to 300 thousand (Fig. E2..37b). However, if the contractor rents modem
equipment, the time of completion will range between 2 and 4 weeks (Fig. E2.37c); whereas the
cost will be incr~ to between $200 and 400 thousand (Fig. E237d).
After interviewing the contractor, a decision analyst determined that the following con-
ditions are appropriate:
(i) The cost and job duration for each alternative are statistically independent.
(ii) The contractor's preference on cost and job duration arc preferential- and utility-
independent.
(iii) The marginal utility functions on cost x and job duration t arc
uzlx) = -x; x in 10' dolla~ .-

. J
u,(t) ~ 21 (25 - r); tin weeks
:.
(iv) The contractor is indifferent in the following pairs of lotteries:
where, for example (1, 5) denotes the event that the job costs $100 thousand and requires
5 weeks.
(a) Assume that u(1, 2) = 1.0 and u(4, 5) = 0; determine the joint utility function u(x, t).
(b) Should the contractor rent the additiona~ equipment to speed up the job"?
Solution
(a) The marginal utility functions should be check~ first if they are scaled from zero to
one. The most and least desirable costs in this example arc $100 and 400 thousand . Since
uJO =-I
u.-(4) = .:..4
the given ux{x) should be modified. Applying a linear transiormation. define
u~x) = a + bu.-(x)
The conditions to'be satisfied are :
u~l) =a+ b(-;-1) ='- ~

u~(4) =a+ b(-4)";: 0 .
The sol_u~io~ to these simultaneous equations gives a =. ~- and b ;= ! :thus.
. . . .. 4 .X
u..(x) = 3-3 .
Similarly, thc .most and least de5irablc durations ofthe project arc. respectively, 2 and 5 weeks.
Since
u,(2) = ir(25 - 4) =J
u,(S) = -fr{25 - 25) = 0
DECISIONS WITH .\.1CLTIPLE OBJECT/YES 89
the given utility function is scaled properly. From the given pairs o(indilferent lotteries,
u(,-c. r) = u(l. 5) = 0.2u( 1, 2) + 0.8u(4. 5)
= 0.2
Similarly. u(.-c. r) = u(4, 2) = 0.5. Hence, substituting p 1 = 0.2 and p~ = 0.5 in Eq. 2.59. the
joint utility function is
u(x. r) = 0.2u.(.-c) + 0.5u,(c) + O.Ju,(.-c)u,(c)
= 0.2---(43 X)3 + -(25-
0.5
21
0.3 (4
1 +-
21 J
X)(25- r)
---
3
t)
..1.
~ 1.338 - 0.186.-c -0.04Jc1 + 0.005:cr:

(b) The expected utility of not renting additional equipment (alternative a 1 ) can be
determined from Eq. 253, in which_.-c and care statistically independeot.
E(U I)= r., r.., u(:c,t)/}, r{.x, c) dx de

(,
' ~J
1
f J.
I
(1.338 - 0.186x - 0.043c1 + 0.005xt1 ) f}(x) J:(c) dx dt
Integrating term by term, ~e obtain- . ..

E(U 1) ~ 1.338 - 0.186!1.n - 0.043(p} 1 + uh) + 0.005!1xt<P?1 + u?1)
= 1.338- 0.186(2)- 0.043(3 1 + 11 ) + 0.005(2X3 1 + 11 )
=0.636 .
Similarly, for the other alternative a 1 , that is, renting the modern equipment,
E(U 1 ) ~ 1.338 - 0.!86!1.n - 0.043(p}1 + u}J + 0.005.u.u<Ph + oi-1 )
= 1.338 - 0.186(3) - 0.043(251 + 0.51 ) + O.OOS(3X251 + 0.51 )
=0.598
Since E( U 1) > E(U1 ), the contractor should not rent the additional equipment.
EXAMPLE 2.38
Two alternative designs of an offshore platform are proposed. The equivalent annual con-
struction cost or designs A and B are $5 and $6 million, respectively. Waves are the major
loading on the platform. The distribution of the annual maximum wave height H is lognormal
with a median of 20m and c.o. v. of 20 ~o- Design A will suffer some damage if H exceeds 26m;
moreover, if H exceeds 30m, it will collapse: On the other hand, design B will be damaged when
Table E2.38 Statistics of ConsequenCes

Design A Design B
Monetary loss Casualty Monetary toss Casualty
Mean c.o.v. Mean c.o.v. Mean c.o.v. Mean c.o.v.
Collapse 50 0.3 . 10 0.5 55 0.4 5 -o.5

Damage 10 0.3 1' 0.5 15 0.4 o- 0
In addition to initial coastruaion costs.

Con!!!!Ul!TICM
Monetary Loss Casualties
Q0214
A li
(5:1, 15) (10,5)
tl5,3) (I,Q5)
. 5 0
(61,22) (5,2.5)
(21,6) 0
6 0
Figur~ E2J8 Decision tree for platform design.
H exceeds 28m and will collapse when H exceeds 32m. The respective consequences associated
with lhe damage and collapse states of the two proposed designs are sliilllDarized in Table
E238. - - . , . :: . ., .
The joint utility functio~ of x and n has.bcen established as
. .
u(x, n) = 1 - 0.004x - 0.005n 2
wherex is the. toUI monetary cost (in s n)illion), and n is the nu~ber of casualties. Determine
the optimal design ~temativc.
The decision tree is shown in Fig. E238, where the consequences X and N for each path arc
described in terms of the resPective means and standard deviations ()l, u). The probability of
collapse (C) or damage (D) or intact (I) associatc:O with each proposed design can be computed
from the given statistical distribution on wave heights. For example, for platform design A,
P(D) =:o P(26 ~ H.< 30)
= 4>~ 30 - In 20) _ 41 {In 26 - In 20)

\- 0.2 \- 0.2
= 0.0737
The expected utility of platform design A if a collapse docs occu~ _i~
E(U,. Iq = E[1 - 0.004X Ac - 0.005N~c]
= 1 - 0.004E(X ,.c) - O.OQ5E(N~c)
= 1 - 0.004(55) .:.... O.OOS(W + 5 2)
= 0.155
where X Ac and N AC are the random variables describing the monetary loss and casualty
3.l'SOcialed with path AC. Similarly,
. .. E(U,.-fD) ~ I.:._ -~.004(15)- 0.005(1 2 + 0,52 )
. = 0.934
.....
E(U~Il) = 1- 0.004(5)- 0.005(0)
= 0.98()
Hence, the expected utility value associated with design A is
E(U~) = 0.155(0.02!4) + 0.934(0.0737)_+ 9.980(0.9049)
=0.959
The expected utility of design B can be similarly calculated, yielding
E(U 8 ) = 0.970
According to the above analysis, platform desi~ B is preferred.
' ~ ..
DECISIONS WITH MULTIPLE OBJECT/YES 91
EXAMPLE ].39 (ucerptedfrom de Neufrille and Keeney, 1972)
In 1970, the Mexican govemmem was faced with the problem of. deciding how the airport
facilities of Mexico City could be developed to insure adequate service for the region for the
next 30 years.
The decision required the following:
(1} The location of the airport relative to the city.

(2) The extent of the services.
(3) The timing of the several stages in the development of the airpot:t.
The Alternative Sites

Because of severe environmental constraints, there arc only two sites suitable for a large,
international airport in the Mexico City metropolitan area. One is the Texcoco site of the
existing airport. The site is close to the city; however, it has been surrounded on three sides by
mixed residential and cotnmercial developments, and thus, further expansion is severely
limited. This site is also on top of a thick alluvium that has caused considerable differential
settlement problems in the past The second site is in Zumpango. The Zumpango site is farther
from the city, has more room for expansion, and is on higher and firmer ground. Problems of_
di:lferential settlement, such :U those experienccci at the Texcoco site! are
not expected here. .
Objectives and Measures of Effectiveness

The interest groups in the planning of the airpOrt consist of the government as operator,
the users of the air facilities, and the interested public as IW~ULSers. To evaluate the potential
consequences of the di:lferent alternatives, it is necessary to study the impacts of the airport
sites on these interest groups. Six major objectives were identified as appropriate, as follows:
( 1) .Minimize total construction and maintenance costs.

(2) Provide adequate capacity to meet the projected air traffic demands.
(3) Minimize aa:ess time to the airport. .
(4) Maximize the safety of the system (airplane, airport, and its surroundings}.
\
(S) Minimize social disruption caused by the construction and operation of new airport
facilities.
( 6} Minimize the effects of noise pollution caused by air traffic.
. .
Although there iS some overlap among the di:lfcrent interest ~ups. the first two objectives
are primarily of concern to the government as operator, the third is importa.nt to the users.
and the last two objectives are primarily nonusers' concerns.
The measures of effectiveness, X~o associated with each of these six objectives wc:ie dc1ined
as follows: . '
X 1 = Total cost in millions of pesOs. .
X z '"' Hourly capacity in terms 'or the number of aircraft operations per hour.
X3 = Averagc aa:ess time in minutes, weighted by the number of travellers from each zone of
Mexico City.
X 4 = Expected number of people seriously inju'red or killed per aircraft accidcn~

X, = Number of people displaced by airPort expansion.
X 6 = Number of people subjected to a high noise level. specified to be <::. 90 CNR.
..,
" 1.0
.E.....
.... . .
~
.
~
]
....
.....
u
Q5
u
<[
~
:a
.
fi.
0
0 40 so eo 100
. x3 ""'. Ave~09i-~~-Time In ~inute;
Flgur~ 239 Probability distribution for aVt:rage access time.
Probability Distributions -
Because of inherent randomness and uncertainties in prediction, the six measures of effective-
ness cited-above cannot be accurately known for each altemative.lt is, therefore, necessary to
dmve the joint PDF for the six measures of effectiveness, denoted u / t . ..:.:r.
(.xh . . , .x6)
for alternative a1
A first step toward simplifying the usessment of the joint density function is to examine
if the X,'s are statistically independent The size of the airport and levels of services..would
obviously affect both construction costs and the amount of noise generated; thus X 1 and X 6
arc likely to be correlated. For simplicity, however, the X,'s were assumed. to be mutually
statistically independent in the study. The development of the marginal density functions for
each alternative was performed by the staff of the MinistJy_of Public Works, the Depannient
of Airports, and the Center for Computation and Statistics.'The-fractile method discussed
in Section 2..5.3 wu used. In most cues, dat2 were available to validate and.C:onfirm the judg-
mental probability measures. Hence, the derived densitY functions represent the judgment of a
group of people familiar with the problem and the impacts of the different alternatives. As
an example, Fig. E239 shows the CDF of X 3 if the Texcoco sit~. is s_eJected.
Utility Functions
To simplify the determination of ~e multiattribute utility function u(.x, ... ,.x 6 ), preferential
and utility independence were assumed. ,Discussions with the people responsible for the final
decision show that tb~ . assumptions ap~ reasonable.
The six marginal utility functions u.(.xJ, i = '1 io6, were first formulated. As an example of
these results, the utility function for mean. access time is presenled in Fig. 2.8 of Section 2..4.5.
The procedure summarized in Eqs. 2.56 through 2.58 is then used to rombine these marginal
utility functions to obtain the multiattribute ~~nctions u(.x,. .. : .x6).
.. ~ ..
Change of Effectiveness with Time
Since a duration of 30 years is involved, the measure of effectiveness associated with each
objective for each alternative may not be constant throughout this period, or the achievement
of the objective may occur at different points in time. It was necessary to determine a common
'
DECISIONS WITH .\IIULTIPLE OBJECTIVES 93
basis for comparison. For .the monetary cost in the first objective this was done using a. dis-
counted present value method. For all but one of the other measures of effectiveness. average
values over the years seemed appropriate. Specifically, the average access time, the expected
number of casualties per accident in the 30-year period, the total number of people displaced..
and the average of people subjected to high noise level were used, respectively, for X 3 , x .. ,
X 5 , X 6 The capacity, X 1 , was divided into three subattributes: (X1 ) 75 , (X1 ) 15 , (X 1 ) 95 ,
each denoting the capacity at 1975. 1985, 1995, respectively.
Results
The expected utilities fer each of the feasible alternatives were obtained using a computer
program developed for this purpose. The results show that two general types of developmental
strategies appeared to be superior to all others; these were:
(1) ~Phased Development at Zumpango," that is, a gradual development of this site over
the next 30 years, with some activities remaining at Tcxcoco.
(2) "All at Zumpango,'' implying that all airport activities should eventually be moved
to Zumpango. by 1985 at the latest.
EXAMPL/i.. 2.40 (excerpted from Kuney and Wood, 1977),
The multiattribute utility approach has been used for planning the water resources develop-
ment of the Tisza River Basin. coverjng an area of 130,000 km 1 , of which 30,000 km: is. in
Hungary. as shown in Fig. E240. The -main river runs from the Carpathian mountains and
after flowing through the flat Hungarian Plain, joins the Danube River in Yugoslavia. Five
alternative system plans for the Tisza River basin were proposed for further consideration,
which may be described as follows:
.
IOOkm
SCALE
Figure E2.4(} Schematic map of the existing water resources system.

System 1
A multipurpose canal-reservoir system to transfer water between the Tisz.a and Danube river.
The water would ~ transferred all year round from the Danube by a gravity canal jn the
flatland area and by a pumped canal-reservoir system in the BOrZsiiny-Cserbat mountams.
System.2
A pumped reservoir sy~tcm is developed in the northeastern pan Ortb~ regiori,.wbicb ~-sUpplied
from the Tisza River. In this system, the natural supply of water is available only 4 to 5 months
per year. It would usc all of the natural supply of this part of the Tisza River basin but not all
of the available storage capacity.
System :3 . . .
A reservoir system will be developed on the ftatland part of the .region, consisting of 2 to 4
m-dcep rCsc:rvoirs, using water from the Tisza River. Only a limited volume of 5.5 km 3 may be
used with such a system, and its capacity to regulate is limjt~ to only -I0~3jy.
Sysiem 4
Develop a mountai~ ~scrvoir system in the upper Tisza River basin, which involves building
rescrvoirsin 'tbe upper'Tisza River basin outside of Hungary. For this system to function
properly, extensive and long-term international cooperation would be necessary.
System 5
a
Involves !!round water storage system, mainly in the flatland part to the cast of the region.
The system would be a conjunctive groundwater system utilizing the Tisza River water for
both supply and groundwater exchange:The maximum yield of tbe aquifer is a limiting factor.
Objectives and Attributes

The basic purpt.>se of the water resources planning is to develop, over a 55-year period, the
natural supply of water resources by comprehensive runQf!' regulation, in terms of both quantity
and quality over space and time. The plans call for signifi.Ca.nt development with growth in
irrigated agricu!Lure, industry, and hydroelectric power generation. as well as an implementa-
tion of new technologies. The Hungarian National Water Authority planners have established
the following goals:
(i) To satisfy the demand for the quantity and quality of water.
(ii). To provide protection at least against the ~year flood.
(iii) To-develop the efficient use and reuse~ water.
(iv) To be able to drain an area of 15.000 km2 in 10 days on the average.
(v) To minimize the resources needed for the implementation and operation of the system;
the resources include water, energy, land and foi'Cl!t, capital and manpower.
(vi) To minimize environmental impact caused by the 'cxinstruction of the reservoir,
canal networks, and groundwater pumping iacilitics.
(vii) To provide flexibility in the systcm.to accommodate the future development of the
region and to cope with the uncertainties _inherent in forecasting, econ.omics, and technologies.
. .. .
Twelve attributes have been identifiee_for evaluating the alternative systems, namely:
X 1 = Cost in. 109 forintsjyr
X~ = Water shortage in percent of demand

XJ = Water quality in a scale of 1 to 100
DECISIONS WITH MULTIPLE OBJECTIVES 95
X 4 = Energy reuse factor in terms of the ratio of energy produced to the energy used
Xs = Recreation in a scale of 1 to 100
X6 = Flood protection in terms of the mean recurrence interval
X 1 = Land and forest use in thousand hectares
X 8 = Social impact in a scale of 1 to 100
X 9 =Environmental impact in a scale of 1 to 100
X 10 = International cooperation in a scale of 1 to 100
X 11 = Development possibility in a scale of 1 to 100
X 12 = Flexibility in a scale of 1 to 100
Utility Functions
The assumptions of preferential independence and utility independence are verified as follows.
For example, it was found that (.x 1 = 80, .x1 = 60) was indifferent to (x 1 = 110, x 1 =- 30)
regardless of the fixed levels of other attributes .x 3 to x 11 The same condition was observed for
other pairs (.x" x 1 ) that were found to be indifferent. Thus, it seemed appropriate to assume that
X 1 , X z is ~ferentially. independent of {X 3 , , X 11 }. In. a similar manner each pair ef
attributes was verified to be preferentially independent of the other 10.
In establishing the utility function of X" the levels of all the eleven other attributes were
first fixed; the certainty equivalent of a 50-50 chance lottery with x 1 = 80 and x 1 = 110 was
then obtained as 98. It was found thanhe ccr:tainty equivalent oi 98 did not change when only
the levels of the other attributes x 2 to .x 11 were varied. Next, the certainty equivalent for the
50-50 chance lottery with x 1 = 80 and .x 1 = 98 was assessed to be 91 and this also did not
depend on the levels of the attributes x 1 through .x 11 Hence, one felt justified in assuming X 1
is utility-independent of {X1 , . , X 11}. The analyses of the other attributes established that
they too were utility-independent.
According to Eq. 2.57, the joint utility function u(x 1, x 1 , ,.x 12) is given by
11
1 + ku(x1, ... .Xu) = n [1 + lck,ut(xJ]
11
where u~x 1) is the normalized marginal utility function of X 1; k and k, i = 1 to 12 arc scaling
constants. Table E240a summarizes .this information obtained from extensive interview
procedures.
\ .
Table E2.40a Summary of Information for Evaluating Joint Utility Function
Worst .. Best ;
Attribute X, Value *x1 Vaiue xf Utility Function u,(x,) k,
XI 110 80 1.784 - 0.088e0017.. 0.150 .

xl 60 0 1.611 - 0.6lle001 6x 0.243
X) 0 100 . 1.198 - 1.198e-~01h 0.189
x.. ' 0 1 -0.784 + 0.784e0u... 0.090
x, . 0 100 -0.476 + 0.476e0011 ... 0.132
x6 40 500 -1.08 - 1.354e-:o.oo56... 0.200
Xr 100 50 2- 0.02x7 0.090
x, 0 100 1.198 - l.l98e-o.ou... 0.165
x, 0 .100 t784- 1.784e-:- 0 ~ 111.... ' 0.132
X,o 0 100 1.191! _ 1.198e-0.01bto 0.189
XII 0 100 1.198- i.198e-o.ou.... 0.034
X11 0 100 0.01.-.: 11 .0.034
k = -0.715
96 DECISION .ANALYSIS
Evaluation of Utility of Alternatives

The levels of attributes associated with each alternative system are assessed and summarized
in Table E1.40b. The uncertainties l!S50ciated with these attribute levels are assumed to be
negligible: By substituting these attnoute levels into the joint utility function, the utility
corresponding to each of the five proposed systems are determin:rl as:
. u, = 0.832
U 2 = 0.821
U 3 = 0.503
v~ = o.648
u, = 0.521 .~
indicatjng that systems 1 and 2 ~re superior to the others. Hence. systems .1 and 2 should be
subject to further investigation. Nevertheless, based on the current utility values, system 1
should be prefen:ed to syst~ 2 In order to determine how much syStem I is better than system
2, the cost of system I in :I"able E240b'.may be increased with all other attribute levels re-
maining fixed, until the utility of syslem I equals the current utility value of system 2; namely,
0:821. This occUrred at .x 1 = 104.2 x 109 forinujyr; so S)'t!tem 1 is better than system 2 by
(104.2 - 99.6) X 109 0.. 4.6 X 109 forints/yr. .. \ .
Table E2.40b. Attribute Levels for Alternative Systems
Alternative Systems
. - Attribute 2 3 4 5
x, 99.6 85.7 101.1 95.1 101.8

X! 4 19 50 50 so
XJ 80 60 20 80 40
x. 0.7 0.5 o.eJ .... 0.1 0.01
x~ 80 60 40 20 -- . 20 ..
x. 100 200 67 100 500
X1 90 80 80 60 70
Xs 80 so 60 40 40
.x9 80 60 20 60 40
X,o 80 60 40 20 40
Xu 80 60 40 20 40
X:: 80 80 20 40 .. 20
PROBLEMS
2.1 Figure P.2: rshows a network ofroadslinkingfivc towns: A, B, C;D, and E. The respec-
.. tive fravel times between the towns ate independently normallydistributed with means
and standard deviations shown in parentheses in Fig. P2: 1. For example, the travel time
between A and C is N(S. 2) in hours. A bus company is .planning a route going from
A to E v.ith intermediate stops at the other three towns. The sequence of service is not
imponant. Assume that the time spent at each station is negligible. Determine the best
route if th~; objective is to:
(a) Minimize the expected travel time, .
(b) Maximize the probability of completing the entire route from A toE within 2S
hours.
. (c) ~~t ~(b). fo.r 35 hou;s.
If one of the roads was to be improved so trui.t the mean and standard deviation of the
travel time for that road is reduced by 10~.. which road should the bus company recom-
mend if the objective is (a}? Repeat this if the objective is (c). :.
2.2 A highway intersection will be closed for the ~anitary department to improve the se..,;er
line crossing the intersection. The project will require at least 10 days but no more than
30 days, and will most likely talce 16 days. A service-station owner on the corner of the
intersection estimates that his loss as a function of the project duration is as follows:
,. Project Duration T (days) Loss
T ~ 15 negligible
.L5 .< T ~ 20 S4000
20 < r ~- 25-- - S10,000
25<T~30 $20.000
Assume that the distribution of the project timeT is triangular; determine the expected
loss to the service station owner arising from the closing of the intersection.
2.3 The construction of a causeway is proposed over a long stretch of swampy areas. The
causeway is a system of girders resting on piers. As the span length of the girder increases,
the construction cost per foot of girder decreases, whereas the cost of pier construction
increases .substantially. For the following costs,
Cost per pier = 2x1
. 2000
. Cost per ft of girder = r:
. vx
where :c is the span length of each girder, det~rmine the optimal span length.
2.4 The performance function of an engineering system is given by
G. = XL + Xl +Xl- aD
. ~ . . . . ~t -;.,. ... '
where the X ,"s arc independent normal variates N(O, 2}, N(2. 1} and N(20, 3}. respectively; .
Dis the present design system capacity; and a: is an additional safety factor that could be
imposed to achieve a h1ghcr reliability. Failure of the system is denoted by the event
(G > O).AssumethatD = 30;thecostofthepresentdesignisC0 ;thecostofasystemwith
safety factor a: is .<XC.; and the cost in the event of system failure is C,.. Derive an ex-
pression for the optimum value of a: in terms of C,/C~. such that the total expected cost
is minimizccl Determine:~,.. for c,;c. _,;, .101 and 106
d - -: ... : d . . 'dg(x) dg(x)
Hint: - {C!I[g(x)]} = - {C!I[g(x)]}- = t/>[g(x)J-
. . dx . dg(x) .. dx dx
where!)( ) and t/J{ ) a.n: the CDF and PDF of the standard normal variate.
2.5 Construct the new decision tree for Example 2.4 if the following alternatives are also
included :
(i) A field test of the embankment system can be constructed. The ainount of seepage
recorded during the test period is classified into three categories: low, medium,
and high. If low seepage is indicated. the actual How will be Q 1 with 95% prob-
--ability; whereas, if high seepage is indicated, the actual How will be Q1 with 80%
.. 98 DECISION ANALYSIS
probability. If medium seepage is indicated, Q1 and Q2 will be equally likely.

Assume that the field test is equally likely to indicate any one of the three seepage
level&. The total cost of the field test program is 3 units.
(ii) A. perfect test is available, which will cost 10 units, that will give perfect informa-
tion on the actual flow.
2.6 A housing project is proposa;! for low-income citizens. From feasibility studies, three
alternative sizes of the housing project were suggested for further study, namely:
" 1: 1200 units
a2 : 1500 units
a3 : 1000 units in the first phase with an optional seco.,d phase of 500 units
The actual dei:nand for housing, however, is uncertain. From the TeSU)ts of a survey
'. study, the probabilities of the various levels of demarid were
estimated as follows:
Demand ~- . . Probability ..
Less than 1000 units 0.2
1000-1100 0.1
1100-1200 0.2
1200-1300 0.3 \
1300-1400 0.1
1400-1500 . 0.05
>1500 0.05
After a consideration of the socio-economic factors, the following monetary conse-
quences are inferred, in terms of S thousand.
(i) Foultematives a 1 and a 2 , if th~ demand does not exceed the number of housing
uni,!~" provided, the expected losses are 100 and 200, respectively; otherwise, if
theaemand is not met, the expected losses are 400 in both cases.
(ii) FoPalternative a 3 , if the demand does not exceed the capacity provided in the
firif'phase, the expected loss is 100; if the demand exceeds 1000 units but is less
tharr 1500 units, the expected loss is 200 associated with the inconvenience
caused by delay; whereas, ifthe demand exceeds 1500, the expected loss is 400.
Draw the decision tree and determine the optimal a}tcniativebased on the maximum
EMV criterion.
2.7 A contractor is faced With the opportunily of bidding on two projects, A and B. His
alternatives are to submit bids for both projects, or for either project, or none at all.
Based on 'his experience and considering the expected. number of bidders for the jobs,
he estimates that his chances of winning the.bids for A and B are, respectively, 0.2 and 0.1.
The costs for preparing the bids will be $1000 for the two jobs, and S600 for one job.
The estimated profit on j.;b A will be $4000, whereas that for job B will be $6000. U he
fails to win any bid, the contractor will lose S400; this includes the cost of employees and
equipment left idle. Assume that the probabilities of winning projects A and B are
. statistically independent. Construct the decision tree and .detennine the best alternative
based on ihe maximum EMV criterion.
2;8 In order to protect an S8 million dam. fr~m {)Vertopping in the event of extreme floods,
two alternative spillway designs. are proposed; systein A costing.S3.5million and system
B costi ngS3 million. The spillway system may fail.because of exceptionally large flood
inflow during the expected life ofthedan] and/or incorrect assessment of the spillway
capacity. If the spillway systein fails to perform satisfactorily, the dam could be de-
strayed; however, because of the location of the dam, a failure will not cause other losses
except the replacement co~t of tl]e pam and the spill~!.-ay system.
With the available information, the lifetime failure probabilities for spillway systems A
and B arc estimated to be 0.15 and 0.3, respectively~
The uncertainty in the capacity, of the proposed spillway systems may be reduced by
performing a model test at a cosi of $0.3 million to. examine the performance under
.. ~
PROBLE.\IS 99- -
large Rood condition. In the judgment of the engineer, there is a 0.7 probability that the
test will indicate that system B is preferable, and a 0.3 probability that system .4 is
preferable. If the test indicates a preference for system A, system B may still be built
but the failure probability will be increased to 0.6. whereas system A will have a failure
probability of 0.2. On the other hand, if the test indicates that system B is preferred, its
failure probability will be 0.1, whereas system A will have a corresponding failure
probability of 0.05.
(a) Draw a decision tree and include all of the relevant alternatives, outcomes,
probabilities, and monetary consequences.
(b) Detennine whether a model test of the dam and spillway should be conducted
- before building the actual spillway. Assume that decisions will be based on EMV.
2.9 A contractor is deciding between submitting a bid for a highway project or a dam
project. Because of time and manpower limitations, he cannot alford to submit bids for
both projects. The estimated bid for the highway job is $1.5 million; whereas, that for
the dam is $3.0 million. The cost for preparing a bid-is about 2% ofthe bid price. Suppose
the contractor estimated his gross profit is 10% of the bid price. After a careful study of
the potential bidders for the two jol, he estimates from his experience that there is a
70% chance of winning the highway job, but only a 50-50 chance of winning a bid on the
dam r)roject. If the contractor does not acquire any job, be will incur an additional loss
of $50,000 due to the idling of machinery and permanent employees.
(a) Construct the decision tree.
(b) Determine the-Ocst altl:(l!&tive based on the following decision criteria:
(i) mini-max (minimizC ina.ximum loss);
(ii) maxi-max (maximize maximum gain);
(iii) muimum expected mon~ value.
2.10 The following data are taken from a study of Whitman, et aL (1975) in which earth-
quake damages are evaluated.
(i) For buildings designed according to the Uniform Building Code Zone 3 re-
quirement, the damago: probabilities after an earthquake of given Modified
Mercalli Intensity are as follows:
Modified Mercalli Intensity
Damage State V . VI VII VIII IX X

\
NoneO 1.0 0.57 0.25 0 0 0
Light L o 0.43 . 0.50 0.25 -0 0
ModerateM 0 -0 o.2S O.S3 0.20 0.
Heavy If . 0 o 0 . 0.21 0.52 0
TotalT 0 0 0 O.Ql 0.23 0.80
Collapse c 0 0- 0 0 0.05 0.20
(ii) . The _losses,. expru.sed as the puc~ntage of the replacement cosr of the building,
for_ ea~ damage state are tabulated below: . . '
Damage
Repair Incident Human
Damage State Cost Cost Costs .. - -
NoneO < 0 0 0
Light L 0.3 0.3 0
- ModerateM S 0.4 --- 0.6
HeavyH 30 2 7
Total T 100 3 30
Collapse c 100 3 600
100 DECISION .ANALYSIS_
The total loss for each damage state is the sum of the damage repair cost, incident
rosl, and human cost.
(iii) For a building founded on soft soil in a given region, the probability that the
largest annual earthquake will be of MM intensity V; VI, VII, and VIII, re-
. - spectively, are estimated to be 0.0&, 0.0155, 0.0043, and 0.0002, respectively.
: _: ..-. (a) Determine the ..annual expected total loss from the earthquake of a building
designed acx:Ording to the UBC Zone 3 requirement. --~~- -
.(bj . For a~building whose expected service life is 50 years. determine the present value
of the total lifetime expected earthquake loss. Assume a net discount rate of 2%
per annum, accounting fQJt-fut~re interest rate and inflation rate. :
(c) . - ~ alternative design is to satisfy the UBC Zone 2 requirement whose damage
_ probabilities ai-e given below., _ _ c ,._. _ .. _.._ ._ . :~ :
. "'-"~
. Modified Mercalii Intensity

'-
Damage St;ue -- v. VI
-. VII VIII IX "'"'"'- " X
0 . 1.0 0.47 0.20 o 0 0

. -- \ L 0 .. 0.53 0.50 OJ 0 0
M 0 0 0.29 . 0.53 0 0
H 0 0 0.01 0.31 0 0
T 0 0 0 0.05 0.8 0.6
c 0 0 0 0.01 0.2 0.4 .
Furtherniore, suppose for a given type of building, the initial cost for the UBC
Zon~ 3 requirement is I % higher than that for the UBC Zone. 2 requirement.
Which code will minimize the total expected cost (initial construction plus
expeCted earthquake losses) for this type of building with an expected service life
of seyears~ . .
(d) Would the decision in part (c) change if"The-~te is more seismiCally active? The
corresponding probabilities for the annual largest earthquake are 0.3, OJ, 0.02,
and 0.0001 for MM intensity VI. VI!, VIII. and IX. respectively.
2.11 In Example 2.5, suppose the reliability of the Tharp and LOuisiana models are expressed
by the con~~tional probabilities given below:
Conditional Probabilities or P(/:C H',IH .. H ,). etc.
Tharp ' Louisiana

H~H", H~Li- L~1;J; L~L1 ' H~lf1 H~L1 L:,Hi- : L~L",
H,.I:IF- Q-:4 . 0.3 0.2 0.1 0.25 0.25 0.25 0.25

H,.-LF 0.2 0:5 0.2 o.r . 0.1'5" 0.35 o.~s 0.25
L:,.HF 0.2 0.2 o:5 '0.'1 0.2 . 0.2 0.4 0.2
L,.Lr 0.15 0.15 0.25 0.45 0.1 O.l 0.3 0.5
The Louisia~a "model wil_l cost $1 o:ooo l~ss io perform than the Tharp model. Determine
the optimal decision. Assume that the prediction models apply only to the improved road
scheme, whereas the .probabilities associat~ w~th the _prtSCf!t, (no improvement) road
~. . ....
PROBLEMS JOJ
design have bceti well established from records of past pcrfonnance as listed in Fig.
E2.5.
2.12 An engineer is driving his car to attend an imPortant meeting in anearby town. The
meeting is scheduled to begin in another 90 minutes. He judges that if he stays within
the legal speed limit. his travel time w.ill be uniformly distributed over 2 0.6 hours. .
However. if he decides to overspecd. say by x mph, his traveltime will decrease and be
uniformly distributed over (2 - :t/20) 0.6 hours. In this case, however. he would be
running the risk of being ticketed by a traffic officer; this probability is given by x/25
for 0 ~ x ~ 25. Suppose that the loss a.sso<.:iated with the delay in attending the meeting
is equal to the amount of the delay (in hours). If the engineer is ticketed. 0.3 hour needs
to be added to his travel time; in addition, the loss resulting from the traffic ticket is
x/30, which depends also on the overspecd. Assume there is no other loss. Formulate the
decision problem and determine his optimal ovcrspeed.
2.13 An engineer is selecting the pavement material for the landing strip of a new airport.
There arc two kinds of pavement materials available: namely, good quality material (G)
that costs $0.4 million and regular quality material (R) costing S0.28 million. Both of
these materials will last for 10 years with proper maintenance. The annual maintenance
cost will depend on the quality of the pavement material as well as the volume of traffic.
..
as follows:
Materia~ Quality _.. Traffic Annual Maintenance Cost

Good --- - High 50.02 million
Good Low 0.01 million
Regular High 0.04 million
Regular . Low 0.02 million
In the planner's judgment. there is a 30% probability that in the next 10 years the air
traffic volume will be high. Assume that the decision will be based on the maximum
EMV criterion and the interest rate is 6% per annum.
(a) Without any additional information, which pavement material should the
engineer select? .
(b) A research group is willing to.conduct a study to predict the traffic volume for a
price of 520,000. However, the prediction from this research group is not per-
fectly reliable. If the actual future traffic is high (H), the group will predict high
traffic (H.) with probability 0.9 and low traffic (L.) with probabilitj.0 .1. that is,
P(H.IH) = 0.9 and P(L.IH) = 0.1. Similarly, P(L.IL) = 0.9 and P(H. IL) = 0.1.
Should such a study be performed before the selection of the pave:ncnt material?
(c) What is the value of perfect information on the actual volume of future traffic? . .
2.14 The present w~tc treatment facility at a Midwestern city ~onsists of a trick~g lilt~
operating at a cost of SS.40 per capita per year. The filter is packed with particles so that
a large surface area is provided for the oxidation and purification process as the waste
water is sprinkled from the top arid allowed to percolate through the particles. The
efficiency of this device:is equally likely to be between 60 and 80~.,. There is. how~.
public pressure for meeting higher stream standards, Based on the prevailing technology
and after a feasibility stud_y, the following alternative waste treatment processes wen:
suggested: .. ' -
.A: Enlarging the existing trickling filter at iui additional cost of$210 per capita/year.
8: The output from the existing trickling filter will be treated chemically and the
wastes allowed. tQ precipitate slowly into a lagoon and removed. The cost of this
scheme depends partly on th.e cost of acquiring the lagoon. Based on the judgment
of the planner, the additional costs for this scheme will be either $4.10 or S5.60
t.
per capitajyear with probabilities of i and respectively.
The efficiency of scheme A will depend on the workmanship of the expansion of the
trickling filter; assume that it is 70% likely that the workmanship will be good. The
102 DECISION .4NA.LYSIS
'.
probabilities zwociatcd with the levels of efficiency of each alternate scheme are sum-
marized as follows:
Probabilities
": ...__. _-. Improvement Scheme A
Efficiency Present Good . Bad . 'lmprovement

Level(%) : DeVice Workmanship Workrilanship Scheme B
60-70' 0.5 . . . . 0.2 . 0.3 0..1

70-80 0.5 0.7 O.T 0..2
80-90 0 OJ 0 o.s
90-100 0 0 0 o..2
The classification of Water quality and its estimated loss from social and other con-
., s'?crati0~ arc tabulated~ foDow~:
Estimated
Efficiency Monetary Loss
(% Rc:moval of Pollutants) Water Qwi.lity per Capita/Year
<70 poor SIO
,70-90 good S2
.. 90-100 excellent so ,
Suppose you were consulted by the planner ofthc pollution control board to perform a
decision analysis on whether or not to improve the waste treatment.
(a) Draw the . decision tree and include all the relevant alternatives, outcomes,
probabilities, and monetary consequences. .
(b) Dct~nc the best alternative assuming EMV decision criterion.
(c) - Sup~c that alternative A is chosen. The plaooer is given the 'option whereby
the quality of the workmanship may be tested by a week's trial performance.
Because of the limited trial time, it C3ll only becstimated that if the workmanship
is good, the performance will be satisfactory with a probability-of 80%, whereas
if the workmanship is bad, the performance will be satisfactory with a probability
of only'10 %. Suppose that the job will be accepted if there arc satisfactory per-
formances during the trial week; other:_wise, the contractor will have to lower the
cost by an amount equivalent to Sl.OO per capita/year due to poor workmanship. .
Draw the. decision tree .for this alternative including all the outcomes and their
probabilities. At most, how much money (equivalent to dollars per capitajyear)
should the pllmner pay for pursuing this option.
.2.15 A .c~ntractor has to decide. whether h!= should order new ;nachincry.for assembling
precast st.ructural elements. Each of the n!=w machines is subject to breakdown, repair,
. and. maintenance. During each breakdown, the repair time is estimated to be 5 days.
Assume that't)le tinie between:bre8lcdowns for each machine is T; which is-exponentially
. 9istributed wi~h mean ). regardless of the number of preVious' breakdowns. -T he con-
tract!>r estimAtes that his :long-term benefit will be afuriction of.the number of.pieces of
.equipment that are opcratingon an average day, as s)lown in the. following table:
. . ; . . . . . . . . . ...
Number of Operating Machines Benefits
(onan Average Day) (in thousand $)
0 0
1 120
2 180
PROBLE.liS 103
After studying the 'recommendation-of the manufacturer, he assumes.that A. is either 15 .

or 25 days with relative likelihoods of 1 to J.
(a) If the contractor has the option of ordering 0, 1, or 2 pieces of equipment, which
alternative should he choose if the cost of the first piece of equipment is 560,000,
and any additional equipment may be acquired at a 5% discount? ,
(b) If the contractor is given a free trial period, say 20 days, on a piece of the equip-
ment, would the additional information from the 2o:day trial period influence
his decision or improve his CApccted utility from (a)? Assume that there are only
two possible outcomes in the trial period: the equipment will or will not break-
down.
(c) What is the value of perfect information in this case?
2.16 Two alternative designs, namely A and B, are considered in a project. There arc two
modes of failure associated with each design. Let IX.c 1, IX.cz be the performance indicators
of design A against the two modes of failure. Similarly IXa~o IXaz are the performance
indicators of design B against the corresponding two modes of failure. The performance
indicators are all statistically independent random variables as follows:
IX.cl - N(l.S, 0.3) IX.cl - N(1 .6, 0.4)
,, lXIII - N(l.6, 0.4) IXn - N(1.4, 0.28)
If t:ftc performance indicator is less than 1, the failure of the corresponding mode will
occur; that'is, failure,; {IX< 1}. Suppose the initial costs of design A and design Bare,
respectively, 1.0 and 0.8 S million. Moreover, failure in modes r and 2 will incur ad-
ditional costs of S5 million and S2 million, respectively. Assume that a design can be
subject to both modes of failure and all costs are additive.
(a) Construct the decision tree and determine the optimal design based on mininuun
exprcced coca/ cost criterion.
(b) Suppose design B can be load-tested with respect to its first mode of failure. The
survival of a specific load test will imply that IX11 will be at least 0.9, and all the .
other IX's will remain unchanged. On the other hand. if design B do.:s not survive
the load test, it will be removed from further consideration; assume $0.5 million is
lost and design A will be adopted. Determine the value of this load tcsL
(c) In part (a), if. the failure costs associated with mode.2 depend on the actual values
of IX,.1 and IXaz such that
C(IX,u) = 5(1 ~ IX,.z) + 5(1. - IX.cz.}z = Sod1 - + -10;
1?cr.u
C{IX!f>)= 5(1- IXazl + 5(1- IX81 ) 1
= 5a~z- 151X.n + 10;
In other words, as ;,,.z
or 1Xez is-significantly below unity, larger damages will be
cxpected compared to the case in :which 2A 1 or <182 is close to unity.
Construct-the decision trec.and determine the expected total cost assOciated with
. design A .. Numerical integration may be needed..
2.17 .In a project. involving tunnel construction tiu:ough rock presented in Einstein et al
(1978), an .~gineer may select one of three construction strategies: St. S2. or SJ. The
. cost of construction-depends on the strategy adopted and the geology or the tunnel.
which could be classified into.threc conditions: namely go!)d (G), fair (F), and poor (P).
The constru~ion ~st is given by the following cost matrix, in terms of dollars per foot
of construction.
_... !unri~ Geology

G F P
Construction 500 1500

- - strategy 600 . 1000
600 soo
Furthermore, the prior probabilities of the tunnel geology conditions at a proposed site
. are 0.2, 0.6 and 0.2 for G, F, and P, respectively.
(a) Uno further exploration is conducted. which construction strategy should be
adopted at the proposed site? .
(b} The engineer could perform soil exploration programs to gather additional
information about the tunnelgcology. The reliability of one such program is
summarized by the following reliability matrix: - . .
ExJJoration Re~ult
EG ,, . . E,.
.Geologic G o.6 . Q.2 Q.2
state F 0.3 0.6 0.1
p . 0.2 . .. 03 Q.5
For ex~ple, if.the actual gWI~gic staie is good (G), the exploration will indicaie
.. good geology (EG) with probability 0.6, that is, P(EG IG) -= 0.6. Draw the decision
tree including this 5oil exploration progniin as an altcmativC. Determine the
value (in terms of dollars per foot of construction) of this exploration program.
(c) Suppose a more elaborate soil exploration program is available with the following
reliability matrix:
Exploration Result
EG E, E.,.
Geologic G 0.8 0.1 0.1

state F O.IS 0.8 o.os
p 0.1 0.2 0.7
~-
How much more is this prOg$<lm worth than the one in part (b)?
(d) What"is the maximum value (in terms of dollars per foot of construction) that the
engirleer should spend iri verifying the actual geologic condition of the site~
?.F
2.11! A geotechnical engineer. has to decide between .t wo.foundation designs for a given site,
which is believed to contain a lens of soft clay. The size of the lens could be discretized
into large L (width 20 to 30ft). medium M (width 10 to 20ft), and small S (width 0 to
10ft). Design A basically assumes that there is indeed a lar!e lens of soft clay, wherea~
design B assumes that the size of the soft clay lens, if any, is insignificant. The expected
cost, which ncludes initial cost and expected failure costs, for each design and lens size
combination is given by the following cost ~atrix (in uni~ of S milhon):
Size of Soft Clay l :->s

L M S
Design . A
B
I
.
4.0
4.8
3.8
3.8
3.6
. 3.0
Suppose the engineer believes that' the soft clay lens Is equally likely among L. M, and
S based on preliminary information. .
(a) Which is the better design based on the maximum EMV Criterion?
(b) Suppose four borings spaced at 60 ft could be driven: at the site to verify the
existence of soft clay .lens. .Probabilistic analyses show that the probabilities of
such soil exploration programs encountering a weak spot are 0.99. 0.44. and 0.02
for lens sizes L. M. and S. respectively. Draw the decision tree and determine if
such borings should be driven before adopting the foundation design. Assume
that the boring program wi!l cost $0.1 ~"il!io~ .
PROBLE.WS 105
(c) The geotechnical tngineer could perform a more extensive soil exploration
program instead. How much, at most, should the geOtechnical engineer spend for
verifying the size of soft _clay lens? ' _.
2.19 Consider Problem 2.4 again and -
(a) -Suppose X 1 represents the model error, whose uncertainty level could be reduced
by half, that is, to N(O, 1}, by additional research. What is the maximum amount
that should be spent on such research? Assume CFIC. = lOJ and express your
answer in terms of c. -
(b) What is the value of perfect information on X 1 ? AssumeCF/C. = lOJ and express
your answer in terms of c.
2.20 For most engineering variables, such as material strengths, the penalty associated with
underestimating the actual value of the variable, say 8, is generally less than the loss
associated with overe:: imation. Suppose the loss function is given by
O -{C 1(8- _8); _. e-s.o
8
g( ' _) - C 1 (8 - 9); - 8 > 9
where 8 is the estimated value; C 1 > C 1 and both are constants fn the respective linear
loss functions. Determine the optimal Bayes' estimator.
2.21 For a medium-size town (population 50,000 to 100,000) the daily per capita demand for
water may be modeled by .a normal distribution N(p.. IT). Data on water demand (per
capita per day) coli~~ for four such towns are as follows :
120. 1 io; 128, 112 (in gallons/day, gpd)
obtaining: sample mean = 117.5 gpd; sample variance = 71.7 gpd 1 . Assume that a is
known, with IT= 8.47 gpd.
(a) Data on the water demand for larger towns are also available. _Based on this
information, it is believed that J.1 will be 120 12 gpd with 95% probability.
Determine the distribution of J.1 based on the combined information.
(b) Additional data on water consumption from other towns of similar size can be
obtained at a cost of approximately S2100 per town. Assume that the error in the
prediction of Jl will introduce an expected loss of (in dollars)
L = 2000<7;1
where u;z is the posterior variance of JL Determine the optimal number of ad-
ditional towns that should be surveyed for data on water consumption. Determine
the corresponding expected total loss.
2.22 (a) Determine the optimum sample size in the estimation of the mean value J.l if the
loss function is
L<n. p., J.L) = c(p - fll 2 + b(p - P.> + k:;.
where fl is the estimate chosen for JL k is the unit sample cost, and n is the sample
size, under the following conditions;
(i) There is no prior information except that obtained from the sample obser-
vations. - - .
(ii) The prior distribution of J.1 is N(J.I', o'). .
Assume that the basic random variable is normally distributed whose standard
.. derivation ri is known. -
(b)- Repeat part (a) under condition (ii) if t_h e loss function Is
L(n, P., J.L) = c(p - {l) 1 + bji + kn
2.23 A foundation engineer has to determine the design pressure for a rigid foundation
in order to perform a trade-off study between construction cost, soil exploration cost,
and the penalty for excessive settlement. The compression index X is measured from
soil samples during soil testin~ Let X be a normal random variable with mean J.l and
standard deviation a. The parameter Jl depends on the soil encountered, which will be
estimated from laboratory measurements; whereas, a may be assumed to be a known
constant Sen_lement of the foundation, Y, hua mean and variance given by
P.r ~ bqp.,
a}= 0 2q2u;
wherea and b are foundation dimensions; q is the design pressure; p., and a, arc,
respectively, the mean and standard deviation of /L. Suppose the total load of the struc-
ture is Q (per unit foundation area). Because of excavation to depth d, the net design
pressure of the foundation is .
where y is the density ofthe soil excavated. 11Je following overall cost function is assumed:
.L = C1 + C 2 d + C3 Y2 + C 4 n + c,
where
C 1 = Costof construction, excluding foundation cost.
C2 = Additional cost per meter of excavation depth.
C3 = ~ss per (metO: of settlement)\
C4 = Unit cost for sampling and testing.
C5 = Cost of setting up the testing program.
. n = Number of tests for determining compression index.
Suppose a prior distribution of p. is available and is normal N(Jl', a'). Decisions arc to be
bued on minimizing the expected overall cost. .
(a) Draw the decision tree including the sampling alternative and the selection of a
design pressure.
(b) Determine q.pe for given samplemean x and sample size n.
(c) For the case of no sampling, determine q""' and the corresponding expected
overall cost. .
2.24 A home-owner is building a house on the outskirts of a town and is considering the
alternatives between drilling his own well on a one-acre property and subscribing to the
city water. It costs $500 to drill a well and $2()(). to complete the necessary facilities
if water is found. in adequate quantity. The probability of striking a water-bearing
stratum is estimated to be 0.4 in the vicinity. However, if the first well is dry, it may be
assumed that any other well on the property will yield the same result. The city requires
an immediate payment of $500 to make a connection. The present worth of water
payments to the city is estimated at $1000, whereas, the present worth of operation and
maintenance on a successful well is only $300. Answer the following questions, and draw
the corresponding decision trees wherever appropriate.
(a) What should the engineer do if he is an EMV decision maker?
(b) An ultrasonic device to detect the presence of water can be rented. Although
the device will indicate "wet" if there is a water-bearingstratum, it may fail to
indicate "dry" in the absence of"a water-bearing stMtum. In this latter case, the
device will read ~wet" 30% of. the time. How much should the engineer pay to
rent the ultrasonic device to obtain additional information on the presence of
water beneath his property?
(c) At most, how much should perfect information be worth (i.e. certain knowledge
about whether or not a water-bearing stratum is present)?
(d) Suppose the probability of striking a: water-bearing stratum in the vicinity can
only be estimated within the range 0.4 0.1. Will the engineer still insist on his
optinlal decisions previously obtained in (a)? Justify. ; :
PROBLEMS 107
(c) Would the decision in part (a) change if the uti)ity function of money for the home-
owner is
X :50
where xis the total present monetary value in doll~? Is. the utility function risk-
aversive or risk-affinitivc?
2.25 A contractor may order ready-mix concrete from plant A or plant Bat S820 and S900,
respectively. Although the concrete mix from plant A is cheaper, it is of poorer quality;
the chance that it may not pass the strength test is lOY.. The probability that concrete
mix from plant B will fail the strength test is only 5%. In the.event that a batch ofconcrete
fails the strength test, the contractor will lose an additional $2000, in the form of penalty
and delay.
(a) Draw the decision tree.
(b) Because of'kecn competition, plant A offers an option to the contractor. The
plaBt will guarantee that the concrete mix has the required quality, but the cost
will be 20% higher.
Plant B is also offering an option for an additioMI cost of 6% on the price of the concrete.
Instead of placing a guarantee on its products; plant B will replace any concrete mix
that had been proved to be under-strength. However, because of a delay in the project,
the contr:l.ctor will still lose $540 (but not S2000) in addition to the cost of material
from plant B under !tlis alternate plan. Draw the new decision tree including these
options. --- ..
(c) Determine the best alternative ii the decision is based on expected monetary
value (EMV). .
(d) If the decision is based on expected utility value (EUV), . determine the best
alternative for each of the following utilitY functions.
(i} u(d) = 02d where d = monetary value
1
(ii) u(d) = -0.000ld ; forti < 0
2.26. The utility function of mone!ary value -for the contractor:.in Problem 2.9 is given as
shown in Fig. P226.
u (d)
1.0 .
\
. Q8
0.6
Q4
. Q2
0 ~~~~--~--~o--.------~a~,--
- ----~a~2------
d (~lliion ti
FigllTe P2.26 Utility function of money.
Determine the. best aitcrMtiv~ for the contraCtor on tlie basis ofthc maximum expected
utility value criterion. .. .
2.17 If the utility function for monc~ vaiue. is .exjio~cntial, -tha~ is, u(.x) = 1.- e-o.,..,
determine the expected utility of a project whose monetary return is normally distri-
buted with a mean ofSlO,OOOandc.o.v. of30%- Observe that xis in units ofS thousand.
. 108 DECISION ANALYSIS
2.28 Two modes of

public transportation are available between two cities 150 miles apart
A passenger could either ride the train or take a plane. The uavcl time for each transpor-
tation mode is randomly distributed according to the weather condition as follows:
Travel time (hours)
Mode Weather Mean Standard Deviation

Train '' Any 3 0.5
Plane Good 1 0.2
Bad 6 - 2.5
During the winter months, the probability of bad weather on a given day oftravcl is 0.2.
Suppose only travel time is considered in the choice of a specific mode of transportation.
Furthermore, because of limited seat capacities in both modes of transportation, a
passenger would have to decide on his travel plan at least one week ahead.
(a) Should train travel be pianned if the decision criterion is to minimize expected
travel time? .
(b) Suppose the.utility\!unction for time is
u(r) = -10r2
Which mode should be chosen if one wants to maximize expected utility?
(c) What is the.value of perfect information on the weather condition on the day of
travel. in terms of time units? Assume the same criterion as in (b).
2.29 A manufactur~r is deciding between two alternatives:
A : Devoting the entire firm to the production of a conventional product,
whose volume of sale units is described by a normal variate N(lOO, 10);
the profit is $70 for each unit sold.
B: Devoting the entire firm to the production ofa new product, whose volume
of sale will depend on public acceptance;-lfthe pro9uct is weli accepted,
the manufacturer is 90 ~-~confident that he will net a pro.fit of$10,000, with
the remaining 10% chance of zero profit. Whereas, if the product does not
meet public app_roval the manufacturer is cenain to make zero profit.
With h1s five years of experience in the manufacturing business, the manufacturer
believes that the probability of the new product meeting public acocptance is 80%-
(a) What should.be the decision on the baSis ofEMV analysis?
(b) Suppose the utility function ofmoney'is
u(y) = {IO,~r
where r is the profit in dollars. What should he decide on the. basis of EUV
analysis?
(c) Suppose a market research firm is offering to conduct a market survey, guaran-
teeing a reliable indication of the public reaction to the new product, at a price of
$1600. Should the manufacturer obtain such information? Assume EMV analysis.
(d) The manufacturer doubts the validity ofthe assumption of SO% public acceptance
probability for the new product. ' -
(i) What should, the value of this probability be before.he changes his mind
about the original decision made in part (a)?
(ii) What should the value of this probability be before the manufacturer
re\'erscs his choice of part (c)?
PROBLEMS 109
2.30 The traffic in the existing highway belween towns A and B is approaching the full
capacity of the highway. Th~ local government is considering alternatives to expand the
traffic capacity between the two towns, for the next ten yean. The attcrnatives considered
arc:
(I) Expand the Existing Highway. The existing highway is a 4-lane road with a
capacity of 50 cars per minute. .Expanding it to a &-lane r<i3d will provide an
additional capacity of 30 cars per minute. The cost rcquin:d for tbis expansion
is estimated to be 53 million for construction and S3 million for displacing and
relocating the existing business and buildings adjacent to the existing roadway.
(II) Construct a New Highway. A new highway is proposed that will go through a
presently sparsely populated area such that negligible cost will be incurred in
displacing the existing buildings to make way for the new road. It is expected
that the new highway will increase the land value in the adjacent area, which is
estimated to bring 52 million of revenue to the local government. The proposed
highway will provide a capacity of 50 cars per minute, at a construction cost of
SlOmillion.
(III) Rapid-Transit System. This alternative is a response to reduce the amount of
air pollution caused by autos, which is becoming a problem in this region. The
cost of constructing the rapid-transit system is estimated to be. 58 million. In
addition, the loss resulting from the displacement of some of the existing
structures will be S1 million.
The local govcrnme.nt will subsidize the operation of the transit system at a total cost of
$2 million. As far as thcnumber .Qf potential oassengcrs arc concerned; it is expected that
50% of the future traffic dcmani:! between. the two towns would be handled by the
transit system. The desired objectives for each alternative are as follows:
(0 1): To meet future traffic demand. This is measured by the probability that an
alternative is able to handle the future peak hour traffic demand.
(0 2 ): To minimize total cost. This is measured by the following utility function:
U = e-O.b X >0
where x is the total cost in S million associated with c:~ch alternative.
(0 3 ): To minimue air pollution. The possible effects on. air pollution from each
alternative have been studied. The ratings statingchow well each alternative
can achieve this objecti\e hav~ been subjectively dctermin~ as 0.0, 0.5, :md 1.0
for alternatives I. II. III, respectively. In other words, expanding the present
roadway will increase the amount of air pollution the most, whereas the transit
system will result in the maxiihum reduction of air pollution.
It is assumed that the peak hour traffic demand in the next ten years is a normal random
variable with a mean of80 cars per minute and a coefficient of variation of20~~-
(a) Suppose the relative importance of achieving objectives 0" 0 1 , and 0 3 is 5:1:3.
Construct the decision tableau and determine the best alternative using the
weighted objectiue decision approach.
(b) What additional alternatives and other relevant objectives do you think shoiild be
included in the decision analysis?
2.31 In a multiattnbute utility analysis, ~uppose.the following joint utility function has been
established:
u(.x..Y>==' 2+x+xy+xlr~
(i) Determine the expected utility of an alternative in which X and Yare statistically
independent normal random variables N(2, 1) and N(l, 0.5). resPectively.
(ii) Suppose one of the following improvements can be made to part (i): .
I
II Increasing Jlr by 100%
whereas crx and cr, remain the same as before. Which improvement alternative
is better?
2.32 A foundation engineer is deciding between two types of foundation schemes for a
proposed building. The first sc;hcme is called flotation involving extensive excavation
and will cost $200,000 if normal conditions exist. However, there is a-40% probability
thai an excessive heave and drainage problem may occur; in such eVents an additional
cost of $50,000 will be required. The second alternative is a rigid base foundation
at shallow depth, costing SISO,OOO. There is a thick layer of normally consolidated clay
beneath the ground surface; .consequently, settlement may be encountered with the
second alternative. Depending on whether the clay has high or low compressibility,
the settlement is expected to be as follows:
Compressibility Setrlemeni Probability

High 2in. 0.4
4in. 0.6
Low 2 in. 0.8
4in. .. 0.2
The engineer believes that the clay is equally likely to have high or low compressibility.
Assume that the settlement is negligible with the flotation scheme. From a study of the
utility of moneyand settlement, the following joint utility function is assumed to be
appropriate:
u(d, s) = .:...o.Id- s2
where
d = monetary value in thousand dollars
s = settlement in inches
(a)Draw the decision tree and determine the best foundation scheme with the present
information.
(b) Suppose a test exists that will yield perfect information about the compressibility
ofthe clay at the site. At most how much should be invested for such informations'!
(c) Suppose there is an indirect test that will furnish the engineer with additional
information on the compressibility of the.;lay stratum. If the compressibility is
actually high, the probability that the test will
indicate high compressibility is
90%; whereas if the actual compressibility is low, this probiibility is only 20%.
Draw the decision tree for this alternative. Assume that the indirect test will
indicate either high or low compressibility only. Should the engineer proceed
with the indirect test if it costs SSOOO'!
2.33 The desirability of a proposed design is evaluated in terms of both cost x (in units of
million dollars) and reliability r. Suppose the following marginal utility functions are
,given ;
u(r) = 4r1 ; 0.5 :!f. r :5. I
u(x) = e-a; 1:5..x:S2
where the alX>ve ranges of rand x cover the possible values of rand x associated with all
the designs proposed. Moreover, the decision maker is found to be indifferent in the
following pairs oflotteries
1.0 ~10.5,2)
(1,2) :9
(1,1)
~(0.5,2)
1.0
(0.5,1) .
{1,1)
PROBLEMS lJJ
where ( !, 2} denotes the event that the reliability of a design is I and the design costs
S2 million. Assurrie that rand x are statistically independent; also assume preferential-
and utility-independence.
(a} Determine the joint utility function u(r, .'<). .
(b) Suppose design A has a reliability of 0.99 and its cost is uniformly distributed
between 1.5 and 20 million dollars. Determine its expected utility.
(c) Suppose the reliability of another design B has a mean of O.S and a C.o.v. of 0.1,
whereas its cost is uniformly distributed between 1.0 and 1.5 million dollars.
Determine its expected utility. Which design is better?
(d) What should the reliability of design A be before the decision maker is indifferent
between designs A and B?
2.34 Consider a contractor evaluating various new business markets; the attributes influenc-
ing his decisions include the markup (x 1) in percent return. contract size (x 1) in millions
of dollars, and the resn~latory environment (x 3) that is defined by a subjective scale.
Suppose the _m arginal utility functions of each of these :mributes have been established
(sec lbbs and Crandall, 1982) as follows:
u 1(x 1) = 0.8 In x 1 - 1.11;. 4 .S: x 1 .s; 14
u1 (x 1 ) = x1 - O.lxf - 1.48; 1.8 .s; x 1 .s; 4.55
and the joint utility (unction for the contractor is given as

u(x 1, x~, x 3 ) ~"tiJ5u 1 (x 1 } + 0.25u 2(x 1)
+ 0.25u 3(x 3 } +0.0525u 1(X 1)u 2(x2)
. + 0.0525u 1(x 1)u 3(x 3) + 0.0375u 2 (xi)u 3 (x 3 )
+ 0.007u 1(x 1)u 1(x1)u 3(x3)
(a) For a construction project whose attributes have the following statistical informa-
tion:
Artribute Mean Standard Deviation

In XI 1.66 0.4
xl 3.1 0.3
x3 -~ 225 1.0
\
and X 3 follows a gamma distribution, .determine the expected :ttility of the
project. ,
(b) Suppose two other alternatives, namely A:z and A 3 , are available to the contractor.
For A 1 , the mean_and standard deviation of In X 1 are !.66 and 0.2, the statistics
of X 2 and X 3 arc the same as those for A 1 ;whereas for A 3 ,the mean and standard
deviation of In X 1 are 217 and 0.23, the mean of X 3 is 4.0, and all other statistics
of X 1 and X 3 arc the same as those for the previous alternative .4. 1 Determine
which among the three alternatives is optimal. '
3. arkov, Queueing, and
Availab~lity . Afode.ls
3.1 THE MARKOV CHAIN
3.1.1 Introduction
The state of a system invariably changes with respect to some parameter, for
example. time or space. The transition from one state to another as a function of
the parameter. or its corresponding transition probability, may generally depend on
the prior states. J:Iowever, if the transition probability depends only on the current
StC1-te. the process of .change may be modeled wi.th the M arkot' process. If the state
space is. a cOuntable or finite set. the process is called a Markov chain. Moreover,
if change can occur only at discrete points of the parameter, for example, at
discrete instants of time, . the process is a discrete parameier M ark01: chain. If the
transition probability is independent of the state of the system, the proce~s reduces
to the Poisson process (see Chapter 3, Vol. I).
As an example. after an earthquake a structure may be damaged to various
degrees. If the damages are not repaired, the structure could be subject to even
further damage during the next earthquake. It is obvious that a structure is more
likely to sustain heavy damage in the next earthquake if it had prior damage from
.earlier earthquakes. Conceptually, the damage may be classified into several
damage stares, such as no damage. light damage, moderate damage, heavy damage,
-and total collap~e. The structure either remains-in the. same state or moves to
another state after each earthquake according to the transitional probabilities
between states. These probabilities, therefore, will determine.the likelihood of the
structure being in a particular damage state after a number of earthquake occur-
rences or at some future time. Of interest here is perhaps the average number of
earthquake occurrences that will bring a structure to collapse, or the average time
a structure could remain undamaged from earthquakes.
3.1.2 The Basic Model

Consider a system with m possible states, namely 1, i ... ., ~.and changes in state
can occur only. at discretized values of the parameter; for example. at times
r 1 , t 1 ,. :. , t" . Let X n + 1 denote the state of the system at r~ 1 In general, the prob-
ability of a future state of the system may depend oil its entire history; that is, its
conditional probability is
(3.1)
112
THE MARKOY CHA.ll' 113
where X 0 = :c0 , . , X.= x. represent all previous states of the system. If the-
future state is governed solely by the present state of the system, that is, the con-
ditional probability, Eq. 3.1, is
P(X.+ 1 = iiX 0 = x 0 , X.= x.) = P(X.+ 1 = iiX. = x.) (3.2)
then the process is a Markov chain. For a discrete parameter Markov chain. the
transitional probability from state i at time i,. to statej at timer. may be denoted by
p1 1{m, n) = P(X. = jl X,.= i); n>m (3.3)
The Markov chain is homogeneous if p1.J.m, n) depen<;l.s only on the difference
t. - t,.; in this case, we define
p,.J.k) = P(Xt = jiXo = i) = P(Xu. = jiX. = i) s 2!:: 0

. a:S the k-step transition probability function. Physically, "this represents the con-
ditional probability that a homogeneous Markov chain will go from state i to
state j after k time stages. This probability can be determined from the one-step
transition probabilities, namely p 1"'{1) or simply Pl.i between all pairs of states in
the system. These transition probabilities can be summarized in a matrix .for a
system with m states, called the transition probability matrix .
P=[~:;: :::: : ::,::] (3.4)
P.... l P... 2 p,.,

As the states of a system are mutually exclusive and collectively exhaustive after
each transition, the probabilities in each row add up to 1.0. For a homogeneous
discrete Markov chain. the probabilities of the initial states are the only other
information needed to define the mod.el behavior at any future time.
3.1.3 St3te Probabilities \

The probabilities of the respective initial states of a system may be denoted by a
row matrix
P(O) = [pt(O), P2(0) .. p.,(O)]
where p,(O) is the probability that the system is initially at state i. In the sPecial case
for which the initial state of the system is known, for example, a~ state i, then
p1(0) = 1.0 and all other elements in the row matrix P(O)" are zero. After one tran-
sition, the probability that the system is in state j is given by the theorem of total
probability as
pp) = P(X 1 = J) = Li P(X 0 ~ i)P(X1. ~ jiX

.. .
0 = i)
Hence,
(3.5)
114 MARKOV, QUEUEING, AND AJIAILABILITY MODELS
In matrix notation, the single stage state probabilities become

P(1) = P(O)P (3.6)
which is also a row matrix.
. .Similarly, the probability that the system is in state j after two transitions is
given by . .
p_/..2) = LP(X 1 = k)P(Xz =jiX 1 = k)
k . .
= L Pt(l)Pt.j (3.7)
t
or in matrix notation
P(2) = P(l)P = P(O)PP = P(O)Pz (3.8)
Therefore, by induction, it can be shown that the n~stage state probability matrix is
given by
P(n) = P(n - 1)P = P(n - 2)PP = = P(O)P" (3.9)
Clr.apmo.n-KolmogoroP Equations The state probabilities after n stages of tran-

sition can be determined alternatively by considering the fact that when the process
goes from state i to state j in n time stages, the process must be in some inter-
mediate state k after exactly r stages (r < n). Hence, on.the basis of the theorem of
total probability, the transition from state ito statej inn stages is given by:
Pi_j(n) = P(X. = jiX 0 = i)
=I P(X. = j!Xo = i, X.= k)J>(X. = kiX 0 = i)
t
O.~.r_< n (3.10)
Equation 3.10 is known as the Chaprnan-Kolmogoror Equation. It is valid between

any two states i andj, and any r < n. In the case in which n = 2, and r = 1, Eq. 3.10
becomes
Pl../..2) = .L Pt.P)P!,.I:(l)
t
=It Pi.thj (3.11)
which is the product of the ith row by the jth column. of the one-stage transition
. probability matrix.
EXAMPLE 3.1
The tn~cnsition of wet and dry days in a town is modeled as a homogeneous Markov chain with
the following transition probabili~y matrix
Dry
bi-y[o.s
Wet 0.5
THE MARKOV CHAIN 115
For example, the probability of going from a dry day to a wet day is 0.2 Let the dry and wet
days be denoted as states 1 and 2. respectively.
(a) lf it is dry today, the probability that it will be dry 2 days from now may be computed
by first determining P(2) using Eq. 3.8.
P(2) = P(O)P1
= [1 0][0.8 0.2][0,8 0.2]

0.5 0.5 0.5 0.5
= [0.8 0.2][0.8 0.2]

0.5 0.5
= [0.74 0.26]
Hence. it is 74% probable that the day after tomorr0w will be dry. Alternatively. Eq. 3."11 may
be applied to yield
p,,,(2) = L Pt.lPk.i
l
= Pt.tPt.t + Pt.1P1.1
= 0.8 X 0.8 .+ 0.2 X 0.5 = 0.74
Similarly for five days frot:t__now.
P(S)',;, -P(O)P'
= [1.0 olo:8 0.2]'

Lo.s 0.5
= [0.715 0.285]
Hence, the probability of dry weather 5 days from now is 0.715.
(b) However, if today is wet instead of dry, then the corresponding probabilities are:
P(2) = [0 - 1][0.8 0.2r

0.5 0.5
= [0.65 0.35] \ :
P(5) = [0 1][0.8 0.2r
0.5 0.5
- [0.713 0.287]
In this case the probability that 2 days from now it will be dry is less than that in part (a);
however. the probability that it will be dry 5 days from now is practically the same as that in
part (a). Table E3.1 summarizes the state probabilities at subsequent time stages. Row I lis~
Table E3.! Successive State Probabilities with Two Different Initial States
2 3 4 s. 6 7
Pt.t(n) 0.8 0.74 0.722 0.717 0.715 0.714 0.714

Pz.t(n) 0.5 0.65 0.695 . 0.709 0.713 0.714 0.714
the results starting with a dry day, whereas row 2 indicates the results starting with a wet day.
Observe that when the number oi stages becomes large, the state probabilities appear to be- .
come independent of the initial states.
116 MARKO I', QUEUEING, A/I"D A I'A/LABILITY MODELS
EXAMPLE 3.2
The amount of stored water in a reservoir may be idealized into three states, namely: full,
half-full, and empty. Because of the probabilistic nature of the inflowing water to the reservoir
as well as the outflow from the resen'oir to meet uncertain demands for water, the amount of
water storage may shift from one state to another between seasons. Suppose the transition
probabilities from one state to another are as indicated in Fig. E3.2. Denote empty, half-full,
04
0.1
Figure 3.2 Transition probabilities between states of reservoir storage.
and full storage with-the states I, 2, and 3, respectively. The corresponding transition proba-
bility matrix is as follows:
0.4 0.5
p = 0.3 0.3 0.1]
0.4
[
0.1 0.7 0.2
Suppose two weeks before the end of a season, it is predicted that there is 80% probability
the reservoir will be full at the beginning of the next season: if it is not full. the reservoir will
equally likely be empty or half-full. What is the probability that the reservoir will be full at the
end of the next season?
The initial state probabilities are: P(O) = (0.1 0:-1- ..0.8]; therefore. after I season, the
state probabilities will become -
0.4 0.5
0.1]
P( I) = (0.1 0.1 (1.8] 0.3
[
0.3 0.4
0.1 0.7 0.2
= (0.15 0.64 0.21]
The probability that the reservoir will be full at the end of the following season is only 21 %.
EXAMPLE 3..3
The record of hurri~ane occurrences at Mustang Island. Texas (1818-1970) is shown in Fig.
E3.3.
A homo1.1eneous Markov chain is assumed to describe the number of hurricane occurrences
in a year. Suppose states 0, I. 2 denote. respectively, the occurrence of 0. 1, and 2 hurricanes
in a year.
(a) Determine the transition. probability matrix.
In estimating the transitional probabilities. we have io examine all pairs of consecutive

years. From the record of observations, there an; IS~ pai.rs.of consecutive years. o.u t of which 122
are preceded by years with no hurricane. Amo~g these 1:!2. pairs. fJ7 were followed byanother
THE ;\-fARKOY CHAIN 117
!
~ 3
.,..
~ Hurricane! (Class H)
21 z
~
~
u
~
'0
.!
~
z
1900 19~
Time (years)
Figure E3J Hurricane occurrences on. Mustang Island, Texas (1818-1970) (from Russell
and SchuCI.ler, 1971).
year of no hurricane; 23 were followed by a year.with one hurricane and the remaining 2 by a .
year with two hurricanes. On the basis of these statistics. the following.probabilities.may be
estimated:
Po.o = tn = 0.795i; -Po:; =tn = 0.1885; PQ.l = rh = 0.0164
Similarly,
Pl.O = H = 0.8148; Pl. I ""1-1 = 0.1182; Pl.l = -b = 0.0370
P!.o = i = 0.6667; P1.1 = t = 0.3333; p 1 , 1 = ~ = 0.0000
Hence, the single-stage transition. probability matrix is
0.7951 0.1885 0.0164]
p = [ 0.8148 . 0.1482 0.0370
0.6667 0.3333 0.0000
(b) If the above model was used to predict hurriclme frequencies in the beginning of 1971;
what is tbe probability that there will be one hurricane in 1973 on Mustang Island?
The state probabilities after 3 transitions are required. With the initial state as 1 (i.e., there
was I hurricane in 1970), the state probability matrix after 3 years is
P(3) = P(O)P3
[0.7951 0.1885 0.0164r
= [0 0] 0.8148 . 0.1482 o:o370
0.6667 0.3333 0.0000
~ [0.7964 . 0.1836 0.02~]
Henoe, the probability of one hurriCUle in 1973 is about 18%

(c) Suppose the engineer in (b) would also like to estimate the probability that there will.
be hurricanes in 1971,1972,and 1973,butnotin 1974.Since both states I and 2 imply occurrences
of hurricanes, considerable simplification in computation can be achieved in this case: by
combining these two states. Redefinethe year without a hurricane as state 0 and the year with
hurrii:ane(s) as state L Hence, the transition probabilities are
Po.o = Nr,. 0.7951 Po. 1 == -ftr = 0.2049
P1.0 = ~ = 0.8000 P1.1 = -k = 0.2000
11/J MAHJIU V, l,lUJ:.UJ:.INI,, AI\U A I'AILAISILIJ r MVUI:.~
The probability that each of the years 1971, 1972, and 1973 would be rubjectecfto at least one
hurricane (but not in 1974) at Mustang Island, is
P= Pf.1P!.o
= (0.2)3 (0.8)
= 0.0064
3.1.4 Steady State Probabilities

In Example 3.1, we saw that the state probabilities starting with two different
initial states approach one another as the number of transition stages increases.
In fact, for a class of Markov chain possessing the ergodic property, the state
probabilities will converge to a set of steady-state probabilities p*, which are
independent of the initial states. Therefore, at steady-state condition,
P(n + 1) = P(n) = P* .(3.12).
Also from Eq. 3.9,

P(n + 1) = P(n)P (3.13)
Hence,
For a Markov chain with m states, this matrix equation represents a set of simul-
taneous equations as follows :
P. P1."']
[p! . . . p:J = [p! p!] .;
[ P.... P......
or ----.
(3.14)
P! = PJ.,.PT .+ P2 ... P! + + P... ,p!
where pr is the_probability in state i when a steady-state condition is achieved.
Adding both sides of Eq. 3.1 4, we obtain the identity
PT + .. . + p! = PT + ... + p;,
Hence, Eq. 3.14 contains one degre:e of freedom. The required constraint to obtain
Pr,pJ, . ~ . , p! is
PT + P! + ... + P!.= 1.0 .. (3.15)
EXAMPLE 3.4
For the town in Example 3.1, what is the probability that the day one year from now will
be dry?
...
THE MARKOV CHAIN H9
According to Eq. 3.9, the probability matrix after 365 transitions is
P(365) = P(O)P36 ,
This obviously requires considerable computations. Alternatively, it is reasonable to assume
that the state probabilities will have reached steady-state. Thus, it is sufficient to compute p
instead. From the given transition probability matrix, the first equation inEq. 3.14 yields
P! = 0.8p! + 0.5p!
whereas Eq. 3.-15 gives
P! + P! = 1.0
The solution of these two simultaneous equations yields
p! = 0.714
P! = 0.286
Therefore, the probability that it will be a dry day one year from today is 0.714. This result
may be o~ed also in Table E3.1, which shows that the state probabilities have reached
steady-state by the sixth day.
EXAMPLE 3.5
. ... ..
The bus in a mass-transit system is operating on a continuous route with intermediate stops.
Suppose the arrival of a bus at an intermediate stop is classified into three states, namely:
!-Early arrival.
2-0n-time arrival.
3-Late arrival.
The state of arrival at an intermediate stop may be assumed to depend on the state of arrival
;~.t the previous stop; governed by the transitional probability matrix as foUows:
0.5 0.4
0.1]
p = 0.1 0.6 0.3
[
0.1 0.2 . 0.7
)ver a long period of operation, what fraction of the intermediate stops can be ex~ to .
1e late?
From the solution to the following set of simultaneous cquatioruJ
PT = O.Sp! + O.lp! + 0.1p!
P! .= 0.4pt + 0.6pf + 0.2p!
p! + pf + p; = 1.0
e obtain the steady-state probabilities pf = 0.167, p! = 0.389,--and p; = 0.444. Therefore,
e can expect the bus to be late at 44.4% of its intermediate stops.
,... "..:.. ..
XAMPLE 3.6
Jndreds of prestressed tendons are evenly spaced along a circular circumferential concrete
:II of a nuclear reactor containment structure. After a period of operation, these tendons may
:orne corroded and/or subject to loss of prestress. The occurrences of these pbenomenll
JJO MARKOV, fJUEUEJNG, A!\"D JI'YJIILJIBJLITY MODELS
between adjao:nt tendons may not be independent. An inspection of a sequence of 10 adjacent

tendons in a structure revealed the following:
Loss of
Tendon Corroded Prestress
1 no no
2 no no
3 .. yes no
4 yes -yes
5 no yes
6 no no
7 no no
8 DO no
9 no yes
10 DO yes
Assume a Markov model for occurrences of corroded tendons and .also tendons with Joss of
prestress. What percentage of the tendons are expcctc:O to be corrod~ in this structure?
The transition probability matrix that governs the occurrence of corroded tendons is first
determined from the given infonnation. Observe- that out of 7 tendqns (1, 2, 5, 6. 7, 8, and 9)
that are not corroded, 6 of them'are followed by another noncorroded tendon; tendon number
2 is tbc oitry one out of the seven noncorrodcd tendons that is not followed by another nco-
corroded tendon. Suppose states I and 2 denote noncorroded and corroded tendons, re-
- sptttively. The transitional probabilities p 1 , 1 and p 1, 1 arc estimated to be~ and.;, respectively.
Similarly, for the two corroded tendons (3 and 4), only one is followed by a corroded tendon.
Hence, p1 , 1 and p 1 , 1 ar~ both 1. The transitional probability matrix is
The steady-state probabilities arc tncn evaluated f~o~ the follo.wing simultaneous equations:
. ..
pj = ~P! + !P!
P'r + p! = 1
yielding p! = 0.222 In other words, 22% of the tendons arc expected to be corroded in the
structure.
Altcmativcly, it appears that the probability of finding a corroded tendon could have been
estimated simply by the fact that only two out of the ten tcndons.observed were corroded.
This would yield 0.2 as the fraction of corroded tendons in the structure; this latter result, of
course, is based on the assumption of random sampling. In a case in which dependent behavior
is exhibited among adjacent tendons, observing ten tendons in a sequence would not consti-
tute a set of independent sample selected from the entire population of tendons in the structure.
Therefore, estimates based on the Markov model should give ~>c<ttcr results.
EXAMPLE 3.7
Suppose that sand and clay layers arc the two basic soil materials that constitute the stratum
at a !riven site. Assume that the layers may be idealized as horizontal. Consider: a l~foot layer
as the unit distance of transition, that is, within each 1-foot layer (e.g., 0-1' or 2-3', etc.), it is
either all sand or all clay. An example of a soil profile is shown in Fig. E3.?. Suppose that a
homogeneous Markov chain could be used to model the transit!on between clay and sand
THE .'riARKOJI CHA/i\i Ill
o~~~~~~~~~~
z Clay '
4 .. -Sand . . .. . ..
6t:::=:=~l~~
:
1
-.-~/;i;A:hnf~{>~':,~~~/);\;
Clay
IZ~------~-----------
H'
Figuu 3.7 A soil profile.
for each foot interv'll over a depth of 10 feet, and the transition probability matrix has been
estimated based on ~>?ring logs as follows:
Clay Sand
ctayro.8 o.2]
SandL0.5 0.5
(a) If the surface layer (0-l') is clay, .. . .

(i) What ~.the probability that sand is foun.d in 2 < H < ..3?.in 4 < H '< 5?
(ii) Suppose thar sand is found in 2 < H < 3. \\'hat is the probability that sand is also
found in 4 < H < 5? .... - - ..
(iii) Suppose a layer of at least 2 feet of s~d i~ required for proper foundation support
within the tirst 4 feet of the soil stratum. What is the probability that such conditions will be
met for the site?
(iv) What is the probability that 18 < H < 20 consists of sand only?
(b) What is the expected cumulative thickness of sand over a 1o-foot stratum?
Solution
(aXi) The probability that sand is found in {2 < H < 3} can be determined from the 2-
stage probability vector, "::ith the initial state being clay.
02 1
\ P(2) =. [1 . oJ[o:s
. 0.5
] = [0.74
0.5
o.26J
Similarly;
P(4) = [1
o:s
OJ [0.5 ~:~r = [0.7~ 0.18)
Hence,
P(S4 -s) = 0.28
(ii) .If sand is found in (2 <; H < 3), this information supersedes that .for (0 < H < 1);
the initial state will be sand and the probability that sand is found in (4 < H <. 5) is determined
from the 2-stage probability vector.
P(2) = (0 !)[~:: . ~~r = [0.65 0.35)

122 MARXOJ', QUEUEING, AND AJ'AILABILITY MODELS
Hence,
P(S4-5) = 0.35
(iii) The combinations of stages for the first four feet of soil that will provide at least 2 feet
of sand may be identified as fo!lows:
C- ~ - S ._ S/C . .
C-C-S-S
where C and S denote clay and sand, respectively. Hence, the probability of proper foundation
support is
P(C- S- S) + P(C- C- S- S) = PcsPS5 + PccPcsPss
= 0.2 X 0.5 + 0.8 X 0.2 X 0.5
= 0.18
(iv) Since the depths of(l8 < H <J9) arc far from the sUrface, the states may be assumed
to be independent of the initial state; the probabilities, therefore, arc expected to approach the
steady-state probabilities. The solution to the following simultaneous equations
. \ .
0.8pt + 0.5p~ = Pt and Pt + P~ = 1
yields p~ =~-Hence,
P(Su-lo) = P~ Pss =~ x 0.5 =-+
(b) Since the steady-state probability of finding sand in a 1-foot layer is~ from (iv) above,
the fraction of sand over a stratum is expected to also be~- The expected cumulative thickness
of sand over a 20-foot stratum is thus' x 20 = 5. 7 feet.
- 3.1.5 First Pas5age Probabilities

In addition to the n-stage and steady-state probabilities, another quantity of
interest associated with a Markov chain is the. p_umber of transitions before a
system goes from state i to state j for the first time. For instance, in Example 3. 7
the foundation engineer may be interested in knowing if sand would be encountered
within the first ten feet of pile driven through the stratum, whereas a contractor may
_oe interested in the expected length of a wei spell that could seriously hamper the
progress of a construction job. The probabilities of these events may be computed
fro:p1 the first passage probabilities. Define fi,){n) as the probability that a system
starting from state i will be in state jfor the .first time after n transitions.
Consider first a two-state Markov chain (with states 1 and 2) as shown in Fig. 3.1
wi~h the corresponding tr~nsition probabilities. The first-passage probabilities may
be evaluated as follows:
/ 1, 1(1) = P1,1
f 1, 1(n) = P1,2P'2~lP2.1
f 1. 2(n)- ,;..-J.P1,2
- 111.1
(3.16)
f2.1 (n) = Pi~l P2.1
f2.2(l) = P2,2
f2.2(n) = P2.1P~~lP1.2 n ~1
Pz,l
Figure 3.1 A two-state Markov chain.
For example, the event that the system starting at state 1 will return to state 1 for the
first time after n transitions (for n ~ 2) requires that the system starting at state 1
switches to state 2 in the first transition, remains in state 2 during the next n - 2
transitions, and then swi.::hes back to state 1 at the last transition."
When more than two states are involved in a system, the direct determination of
the first passage probabilities Ji.;{n) using the procedure illustrated above could
become unmanageable. Alternatively, the first passage probabilities may be
evaluated from the n-stage probabilities through a step-by-step procedure as
follows:
Ji;j{1) 7" ~~. P> = PI,/
J;,J{2) = Pi. ;{2) -.:. Ji:J.l )pJ, j
(3.17)
fi./..3) = Pi.P> - fi.f..1)pd2)- AJ{2)pj,j
fi.J..n) = P1.f,.n) - fi.P)PJ,J..n- 1)- - h.;{n- 1)PJ.J

It is obvious that the first passage probability in one transition (n = 1) is the same
as the one-step transitional probability. For n = 2, the two-stage transitional
probability p1,J..2) includes the probability of visiting ~tate j immediately after the
first transition and remaining in state j during the second transition. HenCe.. the
probability of this even't shoulci be subtracted from p1.;{2) to obtain the first passage
\ probability in exact!y two transitions. The rest of Eq~ 3.17 is also based on simjlar
reasoning. Through a recursive procedure and applications of Eq. 3.17 the first
passage probabilities between any two states and at any stage. can be computed
accordingly. _
Once these first pas_sage probabiliti~ -have been -determined, they may be Used
to evaluate various probabilities and quantities of interest, including the followin~:
(1) The probability that a system, initially at state i, will visit statej (at least once)
within m transitions is
q1,J{m) =
..
L f 1,J{n) (3.18)
I
The event of interest is the same as Uie event that the first passage from i to j
occurs within m transitions. - . . . . . .. .
(2) The probability that a system starting at state fwill eventually reach statej is
thus given by
co
q1,f..oo) = L ft.;{n) = !, 1 (3.19)
I
J1d M.4RKOY, fJUEUEING, AND .4V.4JLABIUT1' MODELS
(3) The probability that a system starting in state i will eventually return to state
i is
. ao
q1, 1(oo) = L fi.1(n) = fi. 1 (3.20)
.. -1 .
- - .
(4) The mean recurrence time for state iis the weighted average period from the
time the system leaves state i until it subsequently returns to state i for the
first time. Hence,
00
t 1 = E(recurrence time)= L nfi.t_n) (3.21)

nl
Furthermore, the expected length of occupation in state i is the weighted average

: period until the system first leaves state i. Hence the expected occupation time is
\
It = I npi,l
I
1
: .
1
(1 -Pi,;)= - --
- 1 - PI, I
(3.22)
For illustrations, consider the two-state Markov chain m Fig. 3.1. Applying
Eqs. 3.18 and 3.16, we obtain
ql,l(m) =
..
L J1,1(n)
..
=P1 , 1 + "' .. -2P2,
L..Pt. 2P2. l 1
~l
1- P~.% 1
=P1.1 +PJ,2P2;1 l .
- P2.-2
Since 1 - P2.z =Pl.~> and P1, 1 + P1, 2 = 1, it can be shown that
q1,1(m) = 1- Pt,2Pi.2 1 (3.23)
Similarly,
. q1, 2(m) = L ft.z(n)

.: I
..
"' . -1
= L..])l , t P1.2
. 1-pj l.
= Pt. z .
1- Pt , t
= 1- Pi. t (3.24)
ql. 1(m) = .1 - Pi.2 (3.25)
q2.2(m) = i - Pi. tPi':1
1
. . (3.26)
-
THE MARKOV CHAIN ]25
For the average ocCupation time, Eq. 3.22 yields

1
11=--- (3.27)
1- Pt.t Pt.z
Similarly:
(3.28)
From Eq. 3.21, the average recurrence time

..,
ti. = L .n/1, j(n)
1
..,
= Pt,t + ~ np~,zti:~bl.t
-2
= Pt . t +
.. 1
PuPz.t [ (l _ P:z.z)i
.
+
1 J
1 _ Pz.z
= 1 + Pt.z (3.29)
Pz.t
Similarly,
t:~;-= 1 +Pz.t (3.30)
Pt.Z
\
3.1.6 Recurrent, Transient, and Absorbing States
If fi.1_in Eq. 3.20 is equal to unity, state i is known as a recurrent state:. th;~.t is. a
system starting at state i will surely return to state i eventually. On the other hand,
if fi. 1 < 1~ it implies that a system starting at state i will have a finite probability
of never returning; in such a case, state i is then called a transient (or nonrecu"ent)
state. In the case of a two-state Markov chain, / 1 1 and [ 1 . 2 can be determined by
substituting m. = ::t::> in Eqs. 3.23 and 3.26. If the one-step transitional probabilities,
namely p1, 1 < 1.0 (i = 1, 2;j = 1, 2), then / 1 , 1 = fz. 1 = 1.0, indicating that states
I and 2 are recurrent states: On the other hand. if p 1: 1 is 1.0 and p 1 , 1 > 0, then
fz.z = qz,:z( oo} = 1 - p1 , 1pi: 1 = 1 - p1 , 1 < 1.0, indicating that state 2 may
. 11ever bereached again. State 2, therefore. is an.example of a transient state... .... ..
When there are many states in a Markov chain, the calculation of the first
passage probabilities using Eq. 3.17 could be cumbersome. If all the states in the
chain are recurrent, the mean first passage time from one state to another, or the
mean recurrence time of any state, may be computed without these first passage
probabilities; such alternative methods are presented later in this section.
J:J6 MARKOJI, QUEUEING, AND AYAILABIUTY MODELS
EXAMPLE 3.8
For the probability matrix given in Example 3.1, determine the following:
(a) Suppose it is wet today. What is the probability that it will be wet again for the first
time the day after tomorrow?
. (b) Suppose it is wet today. What is the probability that there will be a dry day within the
next .five days?
(c) What is the expected length of a dry spell1 A wet spell?
(d) If it is wet today, what is the expected length of time (m days) until the next wet day1
Solution
Let states 1 and 2 denote dry and wet days, respectively.
(a) From Eq. 3.16, the probability that it will be wet for the first timetwo days from now is
/1.1(2) = P1.1P1,1
""0.5 X 0.2
= 0.1
(b) From Eq. 3.25, the probability there will be a dry day within the next five days is given
by
ql,1(5)"" 1 - Pt1
= I - (0.5) 5
= 0.96875
(c) From Eqs. 3.27 and 3.28, the average lengths of dry and wet spells are, respectively:
I I
11 = - =--= Sdays
P1,1 0.2
and
I I
12 = - =--= 2 .da)'$
P2,1 0.5 -
(d) From Eq. 3.30, the expected recurrence time of wet days is
t1
P1 1
-= I + - = I
. 111.2
+0.5
- -= 3.5 da}'s
. 0.2
EXAMPLE 3.9
A quality control program is set up to inspect the quality of aggregates that are produced
"c4ily froma rock core. Suppose the daily aggregate quality is divided. into two classes, namely,
satisfactory (S) and unsatisfactory ( U). It is found that the probability of a given quality of
aggregates in a given day depends on the quality of the product for the previous day, in ac-
cordance with the following trai)sition probability matrix:
s . u
sr{).9 .. . O.IJ
VL0.4 0.6
(a) If the aggregate is of satisfactory quality t~ay, what is the probability that it will
become unsatisfactory within the next week?
(b) What Is the average period that satisfactory quality aggregates will be produced?
That unsatisfactory quality aggregates will be produced? . ..
TH MARXOY CH:UN 127_
Solution
Let states 1 and 2 denote sati~factory and unsatisfactory aggregate quality, respectively,
(a) From Eq. 3.24,

q,,z(7) = 1- PL
= 1 - (0.9)' ' ..
= 0.522
(b) The expected period of satisfactory quality aggregate is given by Eq. 3.27 :l.S
1 .
11 = - = 10days
p,;z
where3S from Eq. 3.28. the expected period of unsatisfactory quality aggregate is
1 .
11 = - = 2.5 days
Pz.1
EXAMPLE 3.10
Consider Example 3.7. again; the expected depth of excavation (from the surface) until sand is
encountered is given by .....
1 1 .
l c = - = - =' Sfeet
Pes 0.2
whereas the expected thickness of a sand layer is
1 1
Is = - =- = 2 feet
Psc 0.5
EXAMPLE JJJ
For the nuclear reactor containment structure in Example 3.6, the size (number of tendons) of
an average group of corroded tendons may be detennined as the expected occupation time of
the corroded state; thus, \
1
. /1 = - = 2
4::: .. , . P.z. . .
Suppose an inspection pl8D calls for the. i!l4~ion o(agi-oup of tendons in sequerice. If
the first tendon inspected is free of conosiori and loss of prestress, what is the average nWI1ber
of tendons inspc!:ted before a defective ~endon (corroded and/or loss of pre5uess) will be dis-
co~? ~ .
In this case, the probability matrix iD Example 3.6 will noi be applicable. A new two-state
transition probability matrix must be constructed with state 1 representing defective tendon
(corroded andfor loss of prestress) and state 2 representing nondefective tendon; thus. with
the data summarized in Example 3.6
With the first tendon inspected being n 0 ndefectivc, the average number oitendons inspected
until a defective tendon is found is equivalent to determining the average group size of non-
defective tendons, which is given by
1 1
lz = - = - = 2.S
Pz. 1 f
JlB MARXOJI, QUEUEING, AND AVAILABILITY MODELS
Absorbing States A state j in. the Markov chain is an absorbing scare if its one-step
transition prob.ability is pj,J = 1.0. An absorbing state, therefore, is defined as a
state that cannot reach any other states in the system except itself.
Consider the deflection at mid-span of a wooden beam subject to time-varying
load~. If the performance of the wooden beam at a particular time.is described by the
states (level) of deflection, the beam system will switch from one state to another
as the load on the beam varies. However, there is a maximum limit to which the
beam may deflect, beyond which tlie beam may collapse, and the system will not
be able to move back to its previous state again..The state corresponding to the
maximum deflection limit in this case is an absorbing state. Another example is
the bankruptcy state of a construction firm. Once banlcruptcy is declared, no other
financial states of the firm can be recovered, as the firm ceases operation.
If an absorbing state exists in a Markov chain, the system will eventually reach
and be trapped in that absorbing state. A Markov chain may contain more than
one absorbing state. In this case, 'only one of the absorbing states will ever be
visited. Of interest is the probability that a system will be absorbed in a particular
absorbing state b for a given initial statej. This absorption probability is equal to
the probability of ever getting to state b from state j, namely fJ.b Applying the
theorem of total probability, we obtain
:. jj,b = P(ever getting to biX0= j)
r
= I P(ever getting to bfX 0 = j, X 1 = i)P(X 1 = iiX 0 =f)

11
r ..
= I P(ever getting to b Iinitial state at i) Pi. 1

il
r
j = 1, ... , r
= I
I I
fr.bPj.i j -::1 absorbing states-
(3.31)
where r is the number of states in the system. When the initial state is b, it is already
absorbed in b; hence, clearly fb b = 1.0; whereas when the initial state is at some
other absorbing state, say a, th~n it will .never be absorbed at b. Hence, J.,b = 0.
For all other initial states, a set of simultaneous equations may be written from
Eq. 3.31, whose solution is the absorption probability, jj,b The computation of
absorption probabilities are demonstrated in Examples 3.14 and 3.15.
Another quantity of interest is the mean time before absorption if the system
starts from a given initial state. Let m1 be the mean time of absorption if the initial
state is j; then,
m1 = E[ absorption time IX 0 = J]
,
= LE[absorption timeiX 0 =j, X 1 = l]P(X 1 = iiXo =j)
i=l
= I {1 + E[absorption timejinitial state-l]}p;,;

ial
j = 1, ... , r
= 1 + 'f.m;PJ.i (3.32)
i =l 'j -::1 absorbing states
THE MARKOV CHA/ili 129 .
Equation 3.32 represents a set of simultaneous linear equations whereby the mean
absorption time m/s can be determined in. terms of the transition probabilities.
Observe that if state i is ail absorbing state. the system is already in an absorbing
state; in such a case, m; = 0.
EXAMPLE 3.12
A pile is to bedriveri into a soil stratum until rock is reached. Consider each foot of pile-driving'
as a unit distance of transition between states, consisting of sand (state l), clay (state 2), and
rock (state J). In other words. it is implicitly assumed that within each !-foot layer, the soil
is either sand or clay or rock. Although the pile may penetrate sand and clay layers, pile-<lriving
will stop once rock is encountered; state 3, therefore~ _is an absorbin~ state. The Markov chain
model is shown in Fig. E3.12. Let m 1 and m1 denote thi:: mean depths of pile-driving until rock
is encountered if the suifaa: of the soil stratum starts with sand and clay, respectively.
Applying Eq. 3.32, we have .
m1 =I+ Pt.lml +. P.~~zmz
mz =1+ Pz.1m1 .+ Pz.zmz
Solving these two simultaneous equations, we obtain
m,= --~..~-~-~~-~P~z~.2~+~P~~-~~---
(1- Pt.1XI - Pz.z)- Pt.zPz.:t
m _ 1-pl,t+Pz.t
2
- (1- Pt.1Xl- Pz.z)- P1.1P2.1
Suppose
p 1, 1 = 0.2; Pt. 1 = 0.74; Pt.l = 0.06; Pz. 1 = 0.1; Pz.z- 0.88; Pz.l = 0.02
Then,
1 - .0.88 + 0.74
m1 = = 39.1
~1 - 0.2Xl "': 0.88) - 0.74 X 0.1
- 1 - 0.2 + 0.1
m =
l l1 - 0.2Xl - 0.88)- 0.74 X 0.1
= 40.9
Therefore, a pile starting penetration in Sand will take 39.1 ft, on the average; before hiiting
rock, whereas a pile starting penetration in clay V::!ll .require ari average of 40.9 fL
Mean First Passage Time for Recurrent States For a Markov chain in which all
the states are recurrent, the mean first passage time from one state .to another can
be conveniently computed by applying the cohcept of absorbing states.
-:: . .,...... -
Figure E3.12 Markov chain with an absorbing state.

130 MARKOV, QUEUEING, AND AVAILABILITY MODELS
Let mi.~ denote the mean first passage time from state j to k. If a new Markov
chain is constructed from the original Markov chain by artificially making k an
absorbing state, the mean absorption time mi in the new Markov chain is equivalent
to the mean first passage time mi.~ of the original Markov chain, a:nd Eq. 3.32 may
be used to compute the required probability. The.
construction of the new Markov
chain requires the definition of a new transition probability matrix P' = [pi.i] as
follows:
Pi,J = 1 ifi=_k,j=k
Pi.J~ 0 if i = k,j =F k (3.33)
Pi.J =Pl.} ifi ~ k
The resulting Markov chain consists of a single.absorbing.state ;k with all other
states becoming nonrecurrent. The behavior of the new Markov chain before
absorption is the same as that ~of the original. Markov chain before state k is
visited for the first time.
EXAMPLE 3.13
For the nuclear reactor coniainmcni stnicture in Example 3.6, suppose the first tendon in-
spected is free of Corrosion and has no loss of prestress. Determine the average number of
tendons inspected befor:e a corro4ed tendop (with no loss oi prestress) is discovered.
Four states may be identified to characterize each tendon; namely:
State 1: Not corroded, no Joss of prestress.
State 2: Not corroded, Joss of prestress.
State 3: Corroded, no Joss of prestress.
State 4: Corroded, loss of prestress.
BaSed on the data for the "ten consecutive tendons inspected, a transjtion probability matrix
can be established as follows:
! !
j
0
0 0]0 .
-~
0
0
Starting with a tendon in state 1, we see that the mean number of tendons inspected before a
corroded tendon with no loss of prestress is discovered is equivalent to determining the mean
first passage time from state 1 to state 3. From Fig. E3.13a it may be observed that all four
states of the system are recurrent. Hence. if state 3 is an absorbing_sta,te (see Fig. E3.13b),.the
mean absorption time for the modified Markov chain gives the mean
first passage time for the
original chain. The modified transition probability matrix is
~]
!
.
P' =[!
. 0 . 0
0
!
f .:
Applying Eq. 3.32. we obtain the following set of slmuitaneousequations:

m1 = I + !m 1 + !m2
m2 = I + !m1 +!rri2
. THE MARKOV CHAIN 131- .
(0)
Figuu EJ.JJ Transitional pr51babilities.
which yields m 1 = 7 and m2 = 9. Since the. initial state is 1, it will require ~n the averagethe
. inspection of seven tendons before a corroded tendon with no loss of prestn:ss is discovered.
Mean RecUTrence Time ofRecurrent States The mean recurrence time associated
with state j in a Markov chain may be computed by Eq. 3.21. However, si.mpli.fic:1- -.
tion is possible if all the: states in the Markav chain are recurrent states. It requires
first determining the ste:idy-state..probabilities p!, ... ,pf. Observe that the mean
recurrence time will be inversely proportional to the corresponding steady-state
probability (see Parzen, 1962). Hence,
1
tj=- j = l, ... ,k (3.34)
Pt
As a verification, consider a two-state Markov chain. Solving Eqs. 3.14 and 3.15
gives the steady-state probabilities.
_ Pz.1
P1 - and Pz* = P1,1
Pt,l + Pz.1 Pt.z + Pz.t
\
Hence, Eq. 3.34 yields
- . 1 Pt 1
.. . . t1 -
Pt
* = 1 +-'-
Pz,t
.- - .
.. -
' which is the same as Eq. 3.29..
'.
EXAMPLE 3J4
The daily BOD concentration of a siream has been measured over a long period of time and
may be divided into four levels, namely: negligible, low, medium, and high. By representing.
these four concentration levels as states 1, 2, 3, and 4, respectively, a homogeneous Markov
chain is c_onstructed ~.the f~liowing ~tion probability matrix: . . . . , --~;..
. [0.5 0.4 0.1

0.1 0]
~5
0.5 0.2
P = 0.25 0.6 0.1
. 0 0.2 0.4 0.4
132 MARKOV, QUEUEING, AND AVAILABILITY MODELS
If the BOD concentration in a stream is high, it implies a polluted state.
(a) Detenni_ne the steady-state probabilities.

(b) Suppose the stream starts out with negligible BOD concentration. How long will it
take, onthe average, to reach a polluted state? Repeat this with the stream starting out with a
low BOD concentration. ..
_(c) How long will it take, on the average, for a pollutCd nream to become one with negligible
or low BOD concentration?
(d) For the stream starting with medium BOD concentration, what is the probability
that it will become one with negligible BOD concentration (excellent water quality) before
becoming polluted (i.e~ state 4).
(e) Determine the mean recurrence time of the polluted state, that is, the mean time until
the stream becomes polluted again.
Solution
(a) If we apply Eqs. 3.14 and 3.15, the set of four simultaneous equations required for the
solution of the steady-state p~obabilities are: . _
P1 = 0.5pj + 0:2P~ + 0.05p1
p~ = 0.4pj + 0.5p~ + 0.25p~ + 0.2p:
p! = OJpt+ O.lpj + 0.4p! \
pj+p~+P1+p!=l.O
..
The solution to these equations yields pj = 0.1788, p~ = 0.3613, P1 = 0.3426, p!. = 0.1173.
-. (b) The first passage time to state 4 is required; hence, state 4 is made into an absorbing
state in the modified Markov chain._The_corresponding transition probabilities bec~me
0.~]
0.5 0.4 0.1
0.2 0.5 0.2
P' = [ ~.05 0.25 0.6 0.1
0 0 1.0
- With Eq. 3.32, the mean absorption times 111 1 (corr-e5ponding to the initial state 1) and m 2
(corresponding to initial state 2)can be determined from the sol?tion to theJollowing equations:
111 1 =I+ 0.5111 1 .+ 0.4111 2 + 0.11111
111 2 = I + 0.2m 1 + 0.5111 2 + 0.2m 3
111 3 = I + 0.05111 1 + 0.25111 2 + 0.6111 3
yielding m1 - 14.8. 111 2 = 12.9, and 111 3 = 12.4. Therefore, it takes, on the average, approxi-
mately 1:5 days for a stream with negligible BOD concentration to become polluted; whereas
it takes an average of about 13 days for a stream with low BOD concentration to become
polluted. .
(c) In this case. both states I and 2 must be converted to absorbing states in the new Markov
chain. The transition probability matrix becomes
0 o
p - [i., 0.25
0.2
0
0.6
0.4
~J
0 ]"
0.4
The set ofequations to be solved is
m3 = I + 0.6111 3 + 0.1111 4
The solution yields ml. = 3.5 and m4 = 4. Hence, the mean time for a polluted stream to become
one containing negligible or low BOD concentration is 4 days. .
l
(d) To obtain the desired probabilities requires convet:ting states 1 and 4 into absorbing
states. Hence, the transition probability matrix becomes
1 0 0 ... 0
02 0.5 0.2 0.1
P' = [ ~.05 0.15 0.6 0.1
0 o 1
Applying Eq. 3.31and observing that/1. 1 = I and / 4, 1 = 0, we obtain
/1.1 = Pz. 1 + Pz.:z.fz.1 + Pz.Jfl.l

fl. I = Pl. t + PJ.zfz.1. + Pl,l/3. 1
or
/:z.,l = Q.2 + 0.5/:z., 1 + 0.2{,, 1
fl. I= 0.05 + 0.15/:z.. i + 0.6/],l
The solution yields / 2 1 ~ 0.6 and / 3 1 = 0.5. By
"Similar analyses, we obtain./3 4 = 0.5.
(Observe that / 3 , 1 + f 3 . .. ... 1.0 as required.) Therefore, a stream starting with mediwn BOD
concentration is equally likely to turn into one with excellent water quality or become polluted.
(e) Applying Eq. 3.34, we obtain
I l .
t4 = -=--=8.53
p:
0.1173
Therefore, on the average, it will take 8.53 days for a polluted stream to become polluted again.
EXAMPLE 3.15
The amount of storage in a reservoir may be discretized into 0, I, 2. or 3 units of water. Empty
and full reservoirs com:spond to states 0 and 3 units. respectively. Suppose the amount of
storage between two sua:essi ve months is governed by a homogeneous Markov chain with the
following transition probability matrix :
0.4
~.1]
0.3 0.3
0.1 0.4 0.4
. - P= [ g 0.2 0.2 0.6 .

0 02 0.8
If the reservoir contains lunit of storage in a give~~ month,
(a) What is the mean time until the reservoir will be full?
(b) What is the probability t~at ihe reServoir will ~empty _before it ever beco~es full?
Solution
(a) Make state 3 an absorbing state, with. the corresponding transition probability matrix
_as follows:
., ...,.:..,:;
.. . .. .. : - - -
. : :-: .r~ ::t:::2. ~ .
[~ ~']
0.3 0.3 ...~~,.,....
0.1 0.4 0.4
P' = 0 0.2 0.2 0.6
0 0 0 1.0
..... v'. \lt.I<OIJLIJTU, Jt/WU A JfAlLAJJ/UTY MODELS
Applying Eq. 3.32 for j = 0, we obtain

3
mo ,. 1 + I 171(Po.l = 1 + 0.4mo + 0.3ml + 0.3m;z
10
. Similarly,
- ...:.;...., l
. ~ ~ - - ..
m1 = 1 +I m1p1 J- 1 + O.lm0 + 0.4m 1 + 0.4m2
10
and
3
m2 = 1. +. Lm p
lO
1 2 ,1 = I + 0.2m 1 + 0.2m2
After simplifiCation, these become
0.6mo- 0.3m 1 - 0.3m2 = 1

-0.1m0 + 0.6m 1 - 0.4m2 = 1
. -D.2m1 + 0.8m1 = 1
\
whose solution yields m 1 = 3.95. In other words, with a present storage of 1 unit, we can expect
almost 4 months before the reservoir becomes full.
{b) In this case, both states 0 and 3 must be converted into absorbing states. The cor-
responding transition probability matrix becomes
~-1]
0 0
0.4 0.4
0.2 0.2 0.6
~-
. -~.
0 . 0 1
Applying Eq. 3.31 for j = 1 and b = 0, we have

2 - ---.
ft. o =

I fi.oPt,l = fo.oPt.o + !t,oPt. l + J2: oPi~2
10
= I x 0.1 + / 1, 0 x 0.4 + /2 , 0 X 0.4
Similarly,
l
f 2. 0 = 1L0 / ;. oPl.l = fo . oP>.o + f . oPz. +. f:. oP::
= 1 X 0 + f t.O X 0.2 + / 2,0 X 0.2
which can be reduced to the following:
0,6ft.o .- 0.4f2.0 = 0.1
and
-0.2!1 0 + 0.8./2. 0 = 0
The solutions aref1. 0 = 0.2 and/2 0 ,. O.OS.ln other words, therefore. for a reservoir with an
initial storage of I unit, there is a 20 % probability that it will become empty before it ever
becomes full.
THE .'rfriR.KOV CHAIN 1J5
EXAMPLE 3.16 ..
The degree of damage to an existing building following an earthquake of MM scale ~ V may
be classified into several damage states as follows (Whitman et al. 1975):
State Damage Le~:ei
1 No damage
2 Light damage
3 Moderate damage
4 Heavy damage
5 Total damage or collapse
Suppose the additional damage that may be incurred in a subsequent earthquake depends on
the initial state of the structure prior to the next earthquake, as defined by the following transi-
tion probability matrix.
~']
0.15 0.045 0.005
o
["' 0.6 0.3 0.07
P= 0 o. 0.5 ~
0.4 .0.1 .
0 0 0 0.3 0.7.
i) 0 0 . 0 1 ..
If building damages are ~ot repaired, .
(a) Determine the probabilities of the building being in the various damage states after
three earthquakes.
(b) On the average, how many earthquakes can a building withstand before it incurs at
least moderate damage {i.e.. state 3)?
Solution
(a) For a building that is initially in an undamaged state (state 1). the state probability
matrix following three earthquakes is .
[~
0.15 0.045 0.005
0.6 0.3 0,07 0.03
0 ]'
P=[l 0 0 0 0] 0 0 0.5 0.4 0.1
-. . 0 -0 0 0.3 0.7
0 0 0 0 1
= [0.512 0.222 0.144 0.069 0.053]
(b) To determine the mean number of earthquakes necessary to inflict at least moderate
. damage to a building, states 3, 4~arid 5 may be combined into a single absorbing state 3';
the corresponding transitional probability matrix becomes
_:_~_[r -~~5]..
0.15
0.6
.. . ~-
-- o .. ...---......-
:.. --
The mean time of transition from state 1 to state 3 can be evaluated by applying Eq. 3.32
m1 = 1 + O.Sml + 0.15m1
m2 = 1 + 0.6m 1
.- .
136 MARKO~, QUEUEING, AND ll~AILABIUTY MODELS
from which m 1 = 6.88 and ml - 2.5. In other words, it will require, on the average, seven
earthquakes before the building suffers at least moderate damage. On the other hand, if the
building already had light damage from earlier quakes, a mean number of 2.5 additional
ear:thquakes will be required ~ore the bui~ding suffers at least moderate d:unage.
w "
.......
3.1.7 Random Time Between:Stages
The time required for each stage of transition in a Markov chain model may be a
random variable itself. ln such, a case, the number of transitions in a given time
period is not deterministic. Therefore, the probability that the system is in state i
at time r, namely p,(r), will depend on both the number of stages n "that have taken
place and the state probabilities after n stages. Applying the theorem of total
probability, we obtain
. "' .
p,(r) = I p,(n)pN(n; r) (3.35)
0
\~(,here p,(n) is the n-stage probability of the system at state i, calculated according
to. Section 3.1.3, and PN(n; t) is the probability mass function of the number of
sta2es N within time r:
Theev ent that the number of staies N exceeds n in timet is the same as the event
that the total time taken for the (n +
1) stages of transition is Jess than or equal tor.
Hence, the cumulative distribution function of N
F J..n) = 1 - P(N > n) = 1 - P(T1 + T2 + .. + T,.;. 1 ~ t) (3.36)
where 7;, i = 1 ton + 1, is the time taken for the ith stage of transition.
LetSo+ 1 denotethesumofn + 1 transition times, namely T1 + T2 + .. + T,+ 1
Equation 3.36 becomes
F 1\.(n) = 1 -:- F s ,(i) (3.37)
where Fs._,(t) is the CJ?F of s.+ I The ~~F

pJ..n; t) = F N(n)- F N(n - 1)
= F5 .(t)- Fs ,(t) (3.38)
Suppose the transition times . ~s are independent and ...identically normally
distributed as N(p., a). Then, s. will also be normally distributed as N(nJ.L. Jna).
Hence,
( . ).:-. _ "'(t
p 1,,n,t - np.) - "'[t
.... --,-
. . ...;na
.... - (n + )p.J
. ,Jn+J.a
1 (3.39)
On the other hand. if the transition times are exponentially distributed, as in the
case of the Poisson process for stage transition occurrences; PN(n; r) in Eq. 3.35
could be calculated simply as
(3.40)
_ where ,. is the mean rate of transition.

THE JIARXOV CH.~IN 137
EXAMPLE 3.17
For the earthquake damage in Example 3.16, suppose severe earthquakes Occur at the site
according to a Poisson process with a mean occurrence rate of once in !en years.
(a) Determine the probability that a building will suffer at least moderate damage over a
period of 20 years. ,
(b) Repeat part (a) if the building is restored to its origi~al condition after each earthquake.
(c) How would the answer to part (a) change if the time between severe earthquake
occurrences is normally distributed with a mean of 10 years and a standard deviation of
J10 years?
Solution
(a) Tne probability that the building will suffer at least moderate damage (D 2: 3) over a
period of 20 years depends on the number of earthquakes that may occur during this period.
Applying Eqs. 3.35 and 3.40. we obtain
., e-12"
P(D ?:: 3) = I P(D ?:: 31n) - -
-o . n.1 .
where P(D 2: 3:n) is the probability of the building in state 3' after n earthquakes; this is the
n-stagc probability of state 3', which can be calculated from the modifi.:ri transition probability -
matrix of pan (b)in Example 3.16. The following table summarizes the requi.::d calculations:
. .
e- 12"
n P(N = n) = - , -. P(D 2: 3jn) P(D 2: 31n)P(N = n)
n.
0 0.1353 0 0
I 0.2707 0.050 0.0135
2 0.2707 0. 150 0.0406
3 0.1804 0.266 0.0480
4 0.0902 0.380 0.0343
5 0.0361 0.485 0.0175
6 0.0120 -0:576 0.0069
7 0.0034 0.654 0.0022
8 0.0009 0,7,19 0.0006
9 0.0002 0.773 . 0.0002
10 0.00004 0.817 0.0000
0.1638
Therefore, there is a 16.4% probability that the building will suffer at least moderate damage
duringa 20-year period. . .
(b) If the building is careful!y inspected after each earthquake and then completely re-
stored to its initial condition before the occurrence of the next earthquake, the probability that
it will suffer at least moderate damage in the next eanhquake is simply Pu = 0.05. Then,
over a 20-year period. the probability of a building suffering at least moderate damage is -
P(D?:: J) = 1.- P(D ~ 3l = , - - f P(D ~ 3ln)P(N = n)

C)
-1z-
r:.. r
~
= (1 = o.osr~
..-o . . , n!
e-1
= 1 - e-zxo.9J
= 0.0952
138 MARKOY, QUEUEING, Al\'D AYAILABILITY MODELS
which is Jess thall the probability of part (a). Obviously, therefore, inspection and repair of a
building after each earthquake will increase its safety.
(c) The probabilities of 11 earthquakes in 20 years are reevaluated according to F.q. 3.39
!orr '"' 20_years;-~----~-- . _
. P(N = .--:-
n) =
(20.=-io~)-- -(2o_._:1{x"+ 1)) - .... -
~ - --c==--=,..;..
~
. Jn.fiO Jn+'l.fiO
whose results arc summarized as follows:
II P(N=n) P(D e! 3ln) P(D e! 3jn)P(N = n)
0 0.000785 0 0
1 0.4992 0.050 0.0250
2 0.466 0.150 0.0699
3 0.0332 0.266 0.0088
4 0.000774 0.380 0.0003
5 0.000011 0.485 0.0000
0.1040
Therefore, there is only 10.4% probability that the building will suffer at least moderate
damage during a 20-year period, which is less than the probability of 16.4% obtained for the
Poisson occurrence model in part (a).
3.1.8 Continuous Parameter Homogeneous Markov Chain

ln the previous sections, the transition between states has been assumed to occur
in discrete stages; for example, from year to year, from earthquake to eanhquake,
or from bus stop to bus stop. ln some physical systems, the transitions may occur
at any value of the parameter, such as at any time. For instance, the queue length
at a toll station may change at any time. In a continuous parameter Markov
chain, the time of transition is describe.d by a continuous variable with {X(t), t. ~ 0}
denoting the state of the system at time r. For a homogeneous Markov chain, the
transition probability of the system from state j to state k in a time interval t is
. given by the conditional probability
Pi, 1(t) = P[X(t + s) = kiX(s) = J1 = P[X(t);, kiX(O) = j]
Instead of the one-step transition probability as in the discrete case, the likelihood
of changes between states is expressed in terms of the transition intensity. The
transition intensity is simply the rate of changing from one state to another pe1
unit time. Let q1 , 1 be the intensity of transition from state k to j, whereas q1 is the
intensity of passage out of state k. Hence, in a small time interval !it, the probabilit~
of a system going from state k to statej ~s ~pp(Oximately given by
Pk.Plt) ::< qk,j. (At); . k #- j (3.41G
and the probability that a sys~e~ remains at statek is

: : .. ..
(3.41.
TilE MARKOV CIIAJN 139
The Chapman-Kolomogorov equations are also applicable to the .c ontinuous

parameter Markov process. In a Markov chain with m states, the transition
probability between states i andj in timet is
p1,Jt) = P[X(t) = jiX(O) = i]
=
..
L P[X(t) = jlX(O) = i, X(r) = k]P[X(r) = klX(O) = i]
ta.l
= L"' PtoJ{t - r)p,,t(r)

tl
where 0 < r < t. Similarly, the probability that the process will be at state j at
time (t + 61) is
pj(t
.. .
+ At) = L Pt(t)Pt,J{61) (3.42)
kl
For small At Eqs. 3.4la and 3.4lb may be substituted

.... .into. Eq. 3.42 to yield
pJ{t + At) .= L Pt(t)qt,J{61) + pJ{tXl - q}. At) --
li-J .
...____ .
or
pJ{t + At) - pf_t) _ ~ ( )q , )q
'-'-'-------'-''--- L... Pt t t,J- Pil.t J
At . t-J
Taking the limit as 61-+ 0, we have
j = l, ... ,m (3.43)
Equation 3.43 provides a set of differential equations that can be used to determine
the state probability at a future time i for any given probabilities of the initial state,
provided that the transition intensities betweenail pairs of states in the system are
known. '
On the other hand, when tis very large, the steady-state condition on the state
probabilities may have been achieved, as in the. discrete case
. of Section 3.1.4. Thus,
dpf_t)' =0
. dt
and
. pf_t) = iJ!"
for j = 1, ... , m. Equation 3.43 then becomes.
j = l, ... ,m . (3.44)
Again with Eq. 3.15, the steady-state probabilities of a continuous parameter ~-

Markov process can be determined.
As an example. for a two-state Markov chain, substituting/= .tin Eq. 3.44 we
have
140 MARKOJ', QUEVEINC, AND A IIA/LABILITY MODELS
. Since q1 is the intensity of passage out of state 1, it will be equal to q 1 , 2 in this

case. Furthermore, since
P! + pJ =1
the steady-state probabilities Can be shown to be
(3.45)
and
P! = ql,l (3.46)
ql,l+ql,l
Section 3.2 describes an application of the continuous parameter Markov chain

to queueing problems in engineering. Availability problems, presented in Section
3.3, further discuss applications of the continuous parameter Markov chain with
emphasis on the transient beha~or of a system.
3.1 QUEUEING MODElS
3.2.1 Introduction
A class of physical problems involves the arrivals of"calling units" that must be
served by "service units." Depending on the arrival rate versus the service rate, a
queue of calling units may be formed. The study of the queue length as a function
of the arrival and service rates requires queueing models.
Queueing models have been applied to predict traffic flow characteristics
(Haight, 1963), in the study of the efficiency ot"operations in construction projects
(Shaffer; 1967); the determination of optimal constructioncrew size (Griffis, 1968),
.and in the study of waste treatment processes (Thatcher, 1968). A common
feature in these problems is a waiting line, or queue, that may consist of vehicles,
delivery trucks, shovels, construction workers, or solid waste loads.
.A typical example of a queueing problem is the toll booth operation at a turnpike
exiL The arrivals of vehicles constitute a random pro~ess, and the service time for
each vehicle is a random variable. The system may have"been designed such that
the average service capacity of the toll booth can adequately handle the -long-tern:
average traffic demand;.however, the instantaneous arrival rate of vehicles at th1
toil booth 'may exceed the service capacity'ofthe"booth resulting, temporarily, ir
a queue of vehicles. Through appropriate -queueing models of the toll booth opera
tion, information such as the. expected length of the waiting line and the meru
waiting time for a vehicle may be derived.
3.2.2 Poisson Process Arriv;ds and Departures

A basic queueing system may be schematical!y represented as in Fig. 3.2, with ;
service units that can operate simultaneously. The length of the queue is define
as the total number of calling units in the system including those being served. I
QUEUEING MODELS 141
------~------.,
// , ol 7z
o o~ o o I- .
Service
eo1r'-"Q--Io O Units
Unt \
.,_ '. i:
1
........... _ Qj~
0 ----~-----'
Figure 3.1 A queueing system.
general, the mean rates of arrival and service will depend on the queue length.
Suppose l. is the mean arrival rate of the calling units, J.nd J.l. is the corresponding
mean service rate of the system when the queue length is ;1. Consider a system with
queue length nat timet+ 6t. During an incremental time interval, !:l.t, between
t and t + ru, one of the following events must occur:
(i) Exactly one arrivaL

(ii) Exactly one dep_arture.
(iii) No arrival or departure: -
(iv) Multiple arrivals and/or departures.
If we assume that the arrivals and departures are respectively Poisson processes,
the probability of multiple arrivals or departures will be of higher order in 6t ;
hence, as !:l.t -. 0, the probability of event (iv) will be negligible relative to those of
the first three events. In other words, the queue Icugth at time t can only .be one of
three mutually exclusive possibilities, namely: n - 1, n, and n + 1. Thus, by virtue
of the theorem of total probability, the probability that the queue length X is n at
time (t t ru) is given by
P[X(t + !:J.t) 9 n] = .P[X(t .+ !:J.t) -~ niX(t),; n- l]P[X(t) = n - 1]
+P[X(t + 6t) = niX(t) = n]P[X(c) = n]
+ P[X(t + ~t) = n!X(t) = n + l]P[X(t) = n + 1]
The conditional probabilities correspond, respectively, to the events of exactly one
arrival at X = n - 1, no arrival or departure at X = 11, and exactly one departure
at X = n + 1, within the. time interval lll.. In terms of transitional probabilities
and with the notation p.(r) = P[X(t) = n], the above equation becomes
p.(t + 6t) = P .:.t,.(lll.)p.-t(t) + P(lll.)p.(t) + Po+l,.(lll.)p,..:_l(t)
in which the transition probabilities can be shown to be
P.-t(M) = A..-tlll.
P(&) = 1 - i,& :_/.t,.!:J.i .
Po+t.~(!:J.t) = J.'..+tru
Hence,
UJ MARKOf', QUEUEING, AND AVAILABILITY MODELS
Rearranging the above yields
. .:. =-
Therefore, as & -+ 0, the following differential equation is obtained:
for n = 1, 2, . . . (3.47)
- .
When n = 0, the state X_ 1 for the system at time t is an impossible event; hence,
p_ 1 (t) = 0. Moreover, at state X 0 , that is, queue length equals zero, the service
rate is obviously zero. Hence, .
(3.48) .
Equations 3.47 and 3.48 provide a system of differential equations for determining
p~(t). The solution is mathematically involved. Fonunately, for most practical
engineering problems, the queueing system may attain a steady-state condition in .
a relatively shon time. For such a long-term system performance, the steady-state
probabilities may be adequate for engineering planning purposes. At the steady"
state condition,
- dp.(t) = 0,. . forn=0, 1, 2, ... (3.49}

.dt
and the steady-state probabilities p's will be independent of time. Thus, Eqs. 3.47
and 3.48 become
..
0 = ).,.-IP:-1 + Jl..+ tP:+ I - (}., + Jl.)p: ; forn > 0 (3.50)
for.n =0 (3.51)
The recursive solutions of Eqs. 3.50 and 3.51 would yield

P1 =-Po
lo
J.ll

P2 = <A1 + Jlt)PT - -.-
AoPo =--Po
lolt
Jl2 .. . Jl2 J.ltlh
-1
,
_Ao " '"w-lp*- 10 p*
n ).1
P,.- J.lt . JL. o - - . - o (3.52)
n Jlt
11
In addition, of course,
..
QUEUEING MODELS 143
or
-I
Po+
"' nAt * 1.0
* "(" -.-Po=
~
i~ 0
nlli
=1
i:al
from which
p~ = ---~.-..,..,...1- (3.53)
~ n~
1+ I i:O
=I n Jli
. i=l
The steady-state probabilities p: can be determined from Eqs. 3.52 and 3.53 for
given values of~ and f.l.il i = 1, 2, ... , n. The:values of 11 and lli generaily depend
on the number of available servers and calling units. The expected _queue length is,
therefore, given by
"' nlJ*_
L = ,.- . (3.54)
~ . n
. 111 .
Queueing 1l1odel with One Server In the simplest case when there is only one
server with an unlimited number of calling units, both the arrival rate and service
rate will be constant regardless of the queue length (except for the obvious case
llo = 0); that is,
lli = }l; i ~ 1
and
1i =1; . i ~0
Then,
...
\
and
For the case A/fl..< 1.0, Eq. 3.53 becomes
1
- .,.....,.-"'"=7-
1+31!__
1 - J.ff.l.
. . 1
=1--
11
]44 MARKOV. QUEUE/1\'(i, .~SD AVAILABILITY MODELS
Substituting this into Eq. 3.52, we obtain
p: = (;)"(t-~); n = 0, l, ... ,and~ < 1.0 (3.55)
The corrcspo~ding queue length, Eq. 3.54, is
).
for- < 1.0 (3.56)
J1
As the: rate of arrivals approaches the rate at which the queue is served, that is,
). -+ J1, the _queue would become very long (i.e. L-+ oo).
EXAMPLE 3.18
During rush hours, the mean rate of arriving planes at an airpon is estimated to be 40 per hour.
The airpon is equipped with two traffic controllers who can independently process the land-
ings. Each controiler takes approximately two minutes. on the average, to process one landing.
For a plane arriving at this airpon during rush hours, what is the chance that it will be im-
mediately processed for landing? ..
The mean arrival rate will not be affected by the iiumber Qf planes in the system. Jf there
are two qr less planes in the queueing system, the mean process rate will be proportional to the
number of planes in the system. whereas if there are more than two planes in the system, both
ofthc;controlle~ will be busy and the process rate is limited to 60pcr ho~r. Thus,
i-o = ). = ... = ..t., '= J. = 40
Jlo = 0; fj 1 = Jl = 30; J12 = JIJ ==fl.,= 2p = 60
Substituting these into Eq. 3.53. we have
= [~ + G) 1 _ :J.f2p)r
2.p - ).
= 2p +X
= 0.2
and from Eq. 3.5:!
;.., 40 .
Pr =~ P~ =
Ill - 30
X
.
0.2 =; 0.167

QUEUE/lVG .~IODELS J.IS
The event that the arriving plane will be immediately processed for loading is equivalent to the
r event that there is zero or one plane in the queue. Hence, the pertinent probability is
pt + pf = 0.20 + 0.267 = 0.467
The expected number of planes ahead of an arriving plane may be computed, using Eq. 3.54, as
2
L = I -JJAp~ + 2 -JJA(A)
- p~ + 3 -A(A
2jJ
- ) p~ +
JJ 2JJ
A 1
=;Po.(l - -1)2
2JJ
40
= 3Q X0.2 X-:(-~---:_,..,..)~l
= 2.403
that is, on the average, the pilot of an arriving plane can expect 2 or 3 planes ahead of him in
the queue.
EXAMPLE 3:19 (Excerpted from Grijjis,J968)

---
An earth excavation operation using a single shovel is schematically represented in Fig. E3.19.
A typical cycle of the excavation operation consists of :
(i) Truck is loaded by the shovel at the production unit.

(ii) Travel to the fill area.
(iii) Dump the truck's load.
(iv) Travel back to the production uniL
(v) Wait in the queue until it is the truck's tum to be reloaded.
The contractor would like to estimate the: expected production per unit time for a single shovel
and a fleet of five trucks. First, this requires determining the fraction of time the shovel is idle
in the present mode of operation. This is p~. which is the steady-state probabilitNhat no trucks
are in the queue. The efficiency of the shovel is (I - p~). and the expected production Y (or
amount of excavation in cu yd) in a time .interval Tis
Y = (1 - p~)TJJC
where .u is the rate of shoveling in terms of the number of truck loads and C is the capacity of
each truck load in cu yd.
Oueue -ot
Trucks
....----000
Direction of
Travel of
Loaded Trucks
Fill Areo
Figure 3.19 Schematic diagram of earth excavation operation..

146 M-4RXOII, QUEUEING, A.'\'D tll'ti/Lt18/LITY MODELS
In the present case, of cou~e, the number of trucks in the queue ranges from 0 to S. Hence,
the mean rate ofarrivals of trucks jn the queue depends on the number of trucks not in the
queuing system. Assume that the mean arrival rate is proportional to the number of trucks
not in the queueing system; namely,
A. = (S _: n)l; '! "= 0,.1, -.- .., 5
where A. is the mean ai'rival rate of a particular truck. Since there is only one shovel, the mean
. rate of departure for loaded trucks would be constant, equal to 11 when one or more trucks are
present in the queue; otherwise, the shovel is not operating and the mean rate of service is zero.
Hence, .
JJa. = O, 111 = 111 = 113 - P4 = p, = J1.
With these values of A, and p.., Eq. 3.53 yields
P3 = [~ +'5G) + 54Gr + _543Gr + 5-4-_32Gr + 543 21Grrl

Suppose the engineer estimates l = 7.5 arrivals per hour, and 11 = 30 truck loads per hour.
The steady-state probability that no trucks are at the shovel. therefore. is
1
. - 0199 .
Po = 1 + 1.25 + 1.25 + 0.9375 + 0.4688 + 0.117l .-
which implies that the shovel is idle about 20~~ of the time. The expected production per hour
for. the present excavation operati~n (with 5 trucks, each with a capacity of 8 cu yd) is
. .
Y = (1 - 0.199) X 1 X 30 X 8 = 192 CU yd
The con.tract'br may also be concerned with whether the present operation is optimal;
that is, should he order additional trucks or transfer some of the trucks to other operations.
Intuitively, as the truck fieet expands, the shovel is less likely to become idle, and thus, the
efficiency of the shovel is increased. However, this also means an increased cost for hiring
additional trucks and drivers. Suppose from past cost records.. the cont~r determines that
each truck with a driver costs SJ4.80 per bout, whereas the power shovel, operator. and oiler
cost $35.60 per hour. Then, for a fteet of k trucks, the.exp~ted total cost per unit production is
E[TQ.k)] = Total cost per hour

Expected production per hour
. 14.8k + 35.6.
Y(k)
Since
..t,- (k - i).l.; i - 0, l , ... , k
= 0; i>k
and
ll;=jl.; i = '1, 2,". ,k .
the steady-state probability p~ associated with a fleet o! k trucks can be derived from Eq. 3.53 as
k(k-l) (k-n+l)l"]- 1
p~(k) = [
I+ Il
.
E I Jl.
= l ~ I~)]-
(k- n)! \j;
I
0
Hence, the expected production per hour with a fleet of .k trucks is

Y(k) = [1 - p~(k)] X I X 30 X - ~ .
QvEUEitvG .uoons Ul
Table E3.19 Summary of Calculations for

Expected Total Cost
Fleet Size
k p~(k) Y(k) E(TC(k)]
I 0.800 48.0 1.05 .

2 0.615 92.5 0.705
3 0.451 132.0 0.599
4 0.311 . 165.0 0.568
5 0.199 192.0 0.567
6 0.117 2JU 0.581
7 0.063 224.5 0.618
Tab!e E3.19 summaiizes the calculations of the expected total cost for various fleet sizes;
the results are basedon constant values of ..1. and JL It may be observed that the present fleet of
5 trucks is, in fact, the optimal fleet size for the excavation operation.
EXAMPLE 3.20
A municipal agency is studying the operation of a waste-treatment proceSs. The so.lid waste
anives in truck toads at an averag~ _!'ate of 10 loads per hour. Because or' the random size of
each truck load and the randomness in the treatment process, the processing time for each
truck load is assumed to be exponentially distributed with a m.:an of 12 minutes. The !Otal cost
for operating and maintaining a prcx:essing unit is $80 per hour. On the other hand, any backlog
of the solid waste will cause inconvenience and pollution problems. involving a loss estimated
to be $100 per truck load per hour on the queue (excluding those loads. being processed).
Determine the optimal number of prcx:essing units such that the expected total cost. and loss
will be minimum.
Solution
The expected total cost is given by
T;, 80m +lOON
where m is the number of solid waste treatment processing units and N is the net expected
\ number of loaded trucks waiting to be prcx:essed (this will depend on m).
For a treatment system with m processing units, the mean service rate per hour is
J.l.. = fiJl = Sn; forn s; m
= mJl = Sm; for n ~ m
If we assume that the arrivals of trucks follow a homogeneous Poisson process, ..1.,. ;= ..1. = 10
per hour for alln. Therefore. from Eq. 3.53
.[
p~ = 1 +
..
L ~+
..t . ..
L
.., G).
- I
1 1
. -~
_, .
nIJ.I ...... m.rrt"
l
ll .
l
. .
1
for in ?' 2
which can be determined for given m.. Subsequently, p: can be determined using Eq. 3.52
148 MARKOY, QUEUEING, AND AYA.ILABILJTY MODELS
For the case m "" 3,
p~ = [1 +2 + 2 + ~ + ~ (~r/(~)rl = 0.111
10
p! = - p~ = 0.222
5
10
p~ = - P! = 0.222
10
10
pJ = Pt = 0.148
15
10 0099.
P = P3 =
15
. . .
10
p! = p! = 0.066
15
When there arc only three or fewer trucks in the queue, none of these trucks will need to wait;
hence, no pollution cost will be incurred. If there arc more than three trucks_in the queue, all
the trucks except the first three will give rise to pollution. Hence,
N = L"' p~(i - 3) = 0.099 )( 1_+ 0.066 )( 2 + ... = 0.9

I
Jbe total expected cost iS

T = 80 x 3 + 100 x 0.9
- S330 per hour
.:r..
Table E3.20 summarizes the numerical calculations for other values of m. When m = 1
Table E3.20 Summary of Calculations for Ex-

pected Total Cost
m p~ N BOrn lOON T
I a:: 80 ' :X:. 00

2 . co 160 X oc
'
3 0.111 0.9 240 90 S330
4 0.130 0.18 3ZO 18 S338
5 0.134 0.27 400 27 $427
and 2 the incoming rate exceeds the service rate; thus, the queue is expected to become in-
finitely long. From the total costs shown in Table E3.20. the minimum-cost number of treat-
ment units would be three. with an expected total cost of$330 per hour.
EXAMPLE 3..21
In a given region, storms occur as a Poisson process with a mean arrival rate). = 0.1 per day.
For each occurrence. the duration of the storm follows an exponential distribution with a
mean of I day. Assume that the next storm will not occur umil the previous storm has ended.
Determine the probability of stormy weather in the region in a given day.
QUEUEING MODELS 149 .
The problem may be formulated as a queueing problem with Poisson arrivals and exponen-
tial service time, in which the queue length is limited to l or 0. Hence.
Ao = 0.1. JJt = l
Substituting these into Eq. 3.53. we obtain
l 111 I
Po=--=---=-
Ao JJt + Ao 1.1
1 +-
JJt
Therefore, on a given day, the probability of stormy weather in the region is
l
pf = 1 - - = 0.09
1.1
3.2.3 Poisson Arrivals and Arbitrarily Distributed Service Time

The results derived in Section 3.2.2 are valid only if the arrivals of the calling units
follow a Poisson process and the distribution of the service time is exponential.
The assumption of exponential service time may be reasonable for routine opera-
tions involving 'il single-stage l!l which the service time is generally short, with only
occasional longer service time:-as in the case of a toll booth where changing big-
bills may occasionally be necessary. However, if the service consists of several
operations, the total service time would no.t be exponentially distributed even
though the individual operation times are exponentially distributed. In general.,
therefore, the service time may have to be some other distribution, although the
arrivals may reasonably be Poisson.
Consider the instances at which departures from a service system ,take place,
which will be referred to as departure points. The corresponding queue length (not
including the one just departed) is defi_ned as the state of the system at that departure
point. If the system is.at state i at a preceding departure point. the event that it will
be at state j at the present departure point is equivalent to the event that (j - i + l)
arrivals have occurred during the last service period. For a Poisson arrival process
with mean rate .t, the probability of n arrivals within a service duration x is
(3.57)
Since the service time is random, this. probability becomes
(3.58)
. . .
. . where fx(x) is the_PDF of the service time. Hence, the transition probability
between states i andj for tWo successive departure-pointS is . ~--. . . . : : :-- ~ '
. .
j-i+l?:O, i>O . (3.59)
For the case iri which the queue length is zero at the previous departure point,
that is, i =._O, the transition to statej will involvej additional arrivals (immediately
]50 MARKOY, QUEUEING, Al'oD AYAILABILITY MODELS
after the next arriyal), during the period required to process the next arrival. Hence,
Po,J = kJ (3.60)
and p1, 1 = 0 in all other cases, that is,j- i + 1 < 0 and i > 0.
.. _F~om Eq. 3.14, the one-stage transition probabilities are related to the steady-
state probabilities as
ao
pj = I Pi,)Pt ; j = 0;1, 2, ... (3.61)
10
Substituting Eqs. 3.59 and 3.60 into Eq. 3,61, we obtain

ao
P1 = I Pi.1Pr + P~ k1
11
J+l
= IkJ-HtP1+p~kJ; j = 0, 1, ... (3.62)
11
or
\
p~ = p~ko + pTko
P! = p~kl + PTkt + p;~o
. .
p: = p~k. + Ptk. + p~k-t + + P:+ tko (3.63)
~The steady-state probabilities are 0btained as follows .
: Observe that the moment-generating function (seep. 96, Vol. I) of the steady-state
.:Probabilities in Eq. 3.62 is .
c
Gp(s) == I"" pjtti
jO
QJ j+ 1 ' ...
= I I kJ-i+ tPf1 + I. p~k1 rr

JOil )0
ao ao
I r kJ-I+tPTei + pf2_ kl1
~
=
il Jl-1 )0
~ ao
= L L k,pt"e"(I+I-Jl + p~G.(s) .
f l 10
where I = j - i + 1 and G.(s) is the moment-generating function of n, the number

of arrivals during a service period, whose probabilities are k .as defined in Eq. 3.58.
Hence,
Gp(s) = e- fpte' 1 fk,e' 1 + p6G.(s)

il . 10
Therefore,

Ang A. H-S, Probability Concepts in Engineering Planning and Design, 1984

Uploaded by

Copyright:

Available Formats

Ang A. H-S, Probability Concepts in Engineering Planning and Design, 1984

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ang A. H-S, Probability Concepts in Engineering Planning and Design, 1984

Uploaded by

Copyright:

Available Formats

What are some of the advanced probability and statistical tools developed in the book?

What are some of the advanced probability and statistical tools developed in the book?

What are the two advanced undergraduate or graduate level courses that the material is suitable for?

What are the two advanced undergraduate or graduate level courses that the material is suitable for?

Probability Concepts

ALFREDO H-S. ANG

JOHN WILEY & SONS f

Copyright l!:l 1984 by John Wiley & Sons. Inc.

Printed in the Republic of Sinppore

A. H-S. Ang AND W. H. Tang

1.1 Aims and Scope of Volume II

1.2 Essence and Emphasis I

2.2 The Decision Model 10

2.3 Applications in Sampling and Estimation 47

2.4 Elementary Concepts of Utility Theory 56

2.5 Assessment of Probability Values 74

2.6 Decisions with Multiple Objectives 77

Problems for Chapter 2 96

3. . Markov, Queueing, ~nd Availability Models 112

3.1 The Markm Chain 112

3.3 Availability Problems 156

Problems for Chapter 3 . 178

4. Statistics of Extremes 186

4.1 Introduction 186

4.2 Probability Distribution of Extremes 187

4.2..2 Asymptotic Distributions 194

4.3 The Three Asymptotic Forms 206

4.4 Extremal Probability Papers 235

4.5 Exceedance Probability 246

4.6 Estimation of Extremal Parameters 255

Problems for Chapter 4 268

5. Mont~ Carlo Simulation 274

\ 5.1 Introduction 274

5.2 Generation of Random Numbers 278

5.3 Variance Reduction Techniques 292

5.4 Application Examples ' 301

Problems for Chapter 5 326

6. Reliability and Reliability-Based Design 333

6.1 Reliability of Engineered Systems 333

6.2 Analysis and Assessment of Reliability 334

6.3 Modeling and Analysis:ofUncertainty 383

6.4 hobability-Based Design Criteria 412

Prob_lems for Chapter 6 < . 434

7. Systems Reliability 448

7.1 Introduction 448

7.2 Multiple Failure Modes 448

7.3 Redundant and Nonredundant Systems 471

7.4 Fault Tree, Event Tree Analyses 485

7.5 Approximate Methods 504

Problems for Chapter 1 517

Appendix A-Tables 531

Table A."I Values of the Gamma Function 532

Table A.2 CDF of the Standard Extremal VariateS 533

Table A.S Relations Among Parameters of the T)i:,e III Asymptotic

Appendix B-TransforTTUJtion of Nonir.orrnal Variates

B.l The Rosenb_latt Transformation 540

1.1 AIMS AND SCOPE OF VOLUME II

1.2 ESSENCE AND EMPHASIS

1.2.1. Decision Analysis (Chapter 2)