Maut - Multiattribute Utility Theory
Maut - Multiattribute Utility Theory
Maut - Multiattribute Utility Theory
net/publication/226400930
CITATIONS READS
118 3,610
4 authors, including:
Salvatore Greco
University of Catania
361 PUBLICATIONS 19,582 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by James S. Dyer on 29 September 2019.
James S. Dyer 3
F
We begin with a brief review of single-attribute preference theory, and explore 5
O
and her preferences for risky alternatives. We emphasize the distinction between 7
these two cases, and then explore the implications for multiattribute preference
O
8
models. We describe the multiattribute decision problem, and discuss the conditions 9
PR
that allow a multiattribute preference function to be decomposed into additive and 10
multiplicative forms under conditions of certainty and risk. The relationships among 11
these distinct types of multiattribute preference functions are then explored, and 12
modeling 15
EC
8.1 Introduction 16
discuss, multiattribute preference theory would be a more general term for this topic 18
R
that covers several multiattribute models of choice. These models are based on alter- 19
nate sets of axioms that have implications for their assessment and use. We begin 20
O
strength of preference and her preferences for risky alternatives. We emphasize the 23
distinctions among these cases, and then explore their implications for multiattribute 24
N
preference theory. 25
U
ordinal comparisons and strength of preference versus theories for risky choices, we 27
will use the term value function to refer to the former and utility function to refer to 28
the latter. This distinction was made by Keeney and Raiffa in 19761 and has been 29
generally adopted in the literature. Further, we will use the term preference model 30
and multiplicative forms under conditions of certainty and risk. The relationships 34
F
are then explored, and issues related to their assessment and applications are 36
surveyed. 37
O
We do not address the important issue of selecting the attributes for a mul- 38
tiattribute decision problem, which may influence the choice of the appropriate
O
39
multiattribute preference model. This topic is explored in Keeney and Raiffa [32] 40
PR
and in more detail in Keeney [30]. More recent discussions include the contributions 41
There are several important points related to the field of multi-criteria decision 43
examine these axioms and determine whether or not they are reasonable guides 46
TE
analysis are developed for individuals who are making decisions on behalf of others, 48
decisions in the best interests of the public. In such cases, one should expect these 50
reasonable set of axioms, rather than some ad hoc approach to decision making that 52
R
Often arguments are made that decision makers do not always make decisions 54
R
that are consistent with the rational axioms of decision theory. While this may be 55
that are appropriate for use in different contexts. Specifically, the axioms that are 60
appropriate for risky choice do not have to be satisfied in order to use multiattribute 61
U
models of preference for cases that do not explicitly involve risk. Much of the work 62
1
The classic book Decisions with Multiple Objectives by R. L. Keeney and H. Raiffa was originally
published by Wiley in 1976. The Cambridge University Press version was published in 1993.
8 Multiattribute Utility Theory (MAUT)
by Keeney and Raiffa [32] emphasized the use of multiattribute preference models 67
based on the theories of von Neumann and Morgenstern [46], which rely on axioms 68
involving risk. As a result, this approach has become synonymous in the view of 69
many scholars with multiattribute preference theory. However, this theory is not 70
the appropriate one for decisions involving multiple objectives when risk is not a 71
consideration.2 Instead, the multiattribute preference theories for certainty are based 72
F
viewed as special cases or approximations to multiattribute preference models. 76
We shall make this case for the popular methods of goal programming and the 77
O
AHP as examples. By viewing these seeming disparate methods from this unifying 78
O
79
ways that these approaches might be sharpened or improved, and provide a basis 80
PR
for evaluating whether their application will result in solutions that are justified by 81
a normative theory. D 82
by a single attribute or by multiple attributes. In the following, we will begin with the 94
under certainty and under risk for the single attribute case. We shall refer to a 96
2
“The important addition since 1976 concerns value functions that address strength of preference
between pairs of consequences (see [4, 13]).” A quote from the Preface to the Cambridge
University Press Edition, R. L. Keeney and H. Raiffa, Decisions with Multiple Objectives,
Cambridge University Press, 1993.
J.S. Dyer
relation on a choice set X, where X could be a set of commodity bundles, decision 100
alternatives, or monetary gambles. For example, we might present an individual with 101
a pair of alternatives, say x and y (e.g., two cars) where x, y 2 X (e.g., the set of all 102
cars), and ask how they compare (e.g., do you prefer x or y?). If the individual says 103
that x is preferred to y, then we write x y, where means strict preference. If the 104
individual states that he or she is indifferent between x and y, then we represent 105
preference; that is, not x y and not y x. If it is not the case that y x, then 107
F
relation. We can also define as the union of strict preference and indifference 109
O
; that is, both x y and x y. 110
Preference studies begin with some basic assumptions (or axioms) of individual 111
O
choice behavior. First, it seems reasonable to assume that an individual can state 112
preference over a pair of alternatives without contradiction; that is, the individual 113
PR
does not strictly prefer x to y and y to x simultaneously. This leads to the following 114
be able to place any other alternative z somewhere on the ordinal scale determined
TE
119
by the following: either better than y, or worse than x, or both. Formally, we define 120
negative transitivity by saying that preferences are negatively transitive if given x 121
122
called a weak order. The weak order assumption implies some desirable properties 124
the preference relation is a weak order, then the associated indifference and weak 126
R
preference relationships are well behaved. The following statements summarize 127
130
4. weak preference is transitive and complete (for a pair x and y, either x y or 134
y x). 135
Thus, an individual whose strict preference can be represented by a weak order 136
can rank all alternatives considered in a unique order. Further discussions of the 137
properties of binary preference relations are presented in Fishburn [18, Chap. 2], 138
Kreps [36, Chap. 2], and by Bouyssou and Pirlot in Chap. 4 of this volume. 139
8 Multiattribute Utility Theory (MAUT)
preference representation function under certainty, it is often called a value function 144
[32]. This value function is said to be order-preserving since the real numbers 145
o o
v.x/; v.y/, : : : ordered by > are consistent with the order of x, y, : : : under . Thus, 146
o
any monotonic transformations of v will also be order-preserving for this binary 147
F
preference relation. Since such a function only rank orders different outcomes, there 148
o
is no added meaning of the values of v beyond the order that they imply.
O
149
o
Notice that we use the symbol “o” to indicate that v is an ordinal function. While 150
O
the notion of an ordinal value function is very important for economic and decision 151
theories, such a function is seldom assessed in practice. For example, if we know 152
PR
that preferences are monotonically increasing for some real-valued attribute x (e.g., 153
o
more is better), then v.x/ D x is valid ordinal preference function. Therefore, we 154
may choose an objective function of maximizing profits or minimizing costs, and be 155
comfortable assuming implicitly that these objective functions are order-preserving 156
D
preference functions for a decision maker. However, the notion of an ordinal value 157
TE
function does become important when we speak of multiattribute value functions, 158
In order to replicate the preferences of a decision maker with less ambiguity, we 160
EC
may wish to consider a “strength of preference” notion that involves comparisons 161
restrictive preference assumptions, including that of a weak order over prefer- 163
ences between exchanges of pairs of alternatives [35, Chap. 4]. We use the term 164
R
measurable value function for a value function that orders the differences in the 165
R
strength of preference between pairs of alternatives or, more simply, the “preference 166
Once again, let X denote the set of all possible consequences in a decision 168
the strength of preference for w over x is greater than or equal to the strength of 171
preference for y over z. The notation wx yz means both wx yz and yz wx, 172
U
AQ2 including the topological results of Debreu [11] and the algebraic development by 175
Scott and Suppes [44]. Some of these systems allow both “positive” and “negative” 176
preference differences and are called algebraic difference structures. For example, 177
x. Our development is based on an axiom system presented by Krantz et al. [[35], 179
Definition 4.1] that does not allow negative differences; hence it is called a positive 180
This set of axioms includes several technical assumptions that have no significant 182
implications for behavior. However, a key axiom that does have an intuitive 183
difference in the strength of preference between x and y exceeds the difference 187
between x0 and y0 , then the difference in the strength of preference between w 188
and y must exceed the difference between w0 and y0 . Some introspection should 189
convince most readers that this would typically be true for preference comparisons 190
F
The axioms of Krantz et al. [35] imply that there exists a real-valued function v 192
O
193
if and only if 194
O
v.w/ v.x/ v.y/ v.z/ (8.1)
PR
Further, v is unique up to a positive linear transformation, so it is a cardinal function 195
(i.e., v provides an interval scale of measurement). That is, if v’ also satisfies (8.1), 196
then there are real numbers a > 0 and b such that v’(x) D av(x) C b for all x 2 X ([35], 197
D
Theorem 4.1). 198
199
X in the natural way by requiring wx yx if and only if w y for all w, x, y 2 X. 200
Then from (8.1) it is clear that w y if and only if v(w) v(y). Thus, v is a value 201
EC
function on X and, by virtue of (8.1), it is a measurable value function. 202
The ideas of strength of preference and of measurable value functions are 203
important concepts that are often used implicitly in the implementation of pref- 204
R
value function as the unique preference function in the case of certainty that 206
R
reveals the marginal value of additional units of the underlying commodity. For 207
O
example, we would expect that the measurable value function over wealth for most 208
individuals would be concave, since the first million dollars would be “worth” more 209
C
to the individual than the second million dollars, and so on. This notion would be 210
Further, the measurable value function can be assessed using questions for
U
213
subjects that do not require choices among lotteries, which may be artificial 214
distractions in cases where subjects are trying to choose among alternatives that do 215
not involve the consideration of risk. Examples of methods for assessing measurable 216
value functions would include the direct rating of alternatives on a cardinal scale, 217
differences. For example, they might be comparing automobiles relative to a “base 222
case”, say a Ford Taurus. Then, they could be asked to compare the improvement 223
a Mercedes (relative to the same Taurus) in terms of a ratio. This ratio judgment 225
could be captured and analyzed using the tools of the AHP, and this provides a link 226
between measurable value functions and ratio judgments. This point has been made 227
on numerous occasions, and is worth further exploration (e.g., see [12, 29, 41]). 228
F
8.2.2 Preference Functions for Risky Choice 229
O
(Utility Functions) 230
O
We turn to preference representation for risky options, where the risky options are 231
PR
defined as lotteries or gambles with outcomes that depend on the occurrence from 232
a set of mutually exclusive and exhaustive events. For example, a lottery could be 233
defined as the flip of a fair coin, with an outcome of $10 if heads occurs and an 234
AQ4 outcome of $2 if tails occurs. 235
D
Perhaps the most significant contribution to this area of concern was the for- 236
malization of expected utility theory by von Neumann and Morgenstern [46]. This 237
TE
development has been refined by a number of researchers and is most commonly 238
distributions and random variables interchangeably. For lotteries p, q, r in P and all 242
245
3. (Continuity) If p q r then there exist some 0 < ˛ < 1 and 0 < ˇ < 1 such 246
O
that ˛p C .1 ˛/ r q ˇp C .1 ˇ/ r 247
C
The von Neumann–Morgenstern expected utility theory asserts that the above 248
axioms hold if and only if there exists a real-valued function u such that for all p, q 249
N
X X
p.x/u.x/ q.x/u.x/ (8.2)
x2X x2X
The expected utility model can also be used to characterize an individual’s risk 252
attitude [39], [32, Chap. 4]. If an individual’s utility function over a closed interval 253
is concave, linear, or convex, then the individual is risk averse, risk neutral, or risk 254
The von Neumann–Morgenstern theory of risky choice presumes that the 256
probabilities of the outcomes of lotteries are provided to the decision maker. Savage 257
[43] extended the theory of risky choice to allow for the simultaneous determination 258
of subjective probabilities for outcomes and for a utility function u defined over 259
those outcomes. Deduced probabilities in Savage’s model are personal or subjective 260
always involve the introduction of risk in the form of simple lotteries. For a 263
discussion of these assessment approaches, see Keeney and Raiffa [32, Chap. 4], 264
von Winterfeldt and Edwards [47], and Anderson and Clemen [3]. 265
F
As a normative theory, the expected utility model has played a major role in the 266
prescriptive analysis of decision problems. However, for descriptive purposes, the 267
O
assumptions of this theory have been challenged by empirical studies [28]. Some 268
of these empirical studies demonstrate that subjects may choose alternatives that
O
269
imply a violation of the independence axiom. One implication of the independence 270
PR
axiom is that the expected utility model is “linear in probabilities.” For a discussion, 271
see Fishburn and Wakker [21]. A number of contributions have been made by 272
relaxing the independence axiom and developing some nonlinear utility models to 273
Note that both the measurable value function v(x) and the von Neumann and 276
Morgenstern utility function u(x) are cardinal measures, unique up to a positive 277
linear transformation. However, the theory supporting the measurable function is 278
R
questions that rely on the idea of strength of preference. In contrast, the von 280
R
Neumann and Morgenstern utility function is based on axioms involving lotteries, 281
and it is assessed based on questions that typically involve lottery comparisons. 282
O
Suppose we find a subject and assess her measurable value function v(x) and her 283
C
utility function u(x) over the same attribute (e.g., over monetary outcomes). Would 284
these two functions be identical, except for measurement error? A quick reaction 285
N
might be that they would be identical, since they are each unique representations 286
is not necessarily the case. Intuitively, a measurable value function v(x) may be 288
concave, indicating decreasing marginal value for the underlying attribute. However, 289
a utility function u(x) may be even more concave, since it will incorporate not only 290
feelings regarding the marginal value of the attribute, but also it may incorporate 291
psychological reactions to taking risks. Empirical tests of this observation are 292
provided by Krzysztofowicz [37] and Keller [33] and generally support this 293
intuition. This is an important point, and one that we will emphasize again in the 294
context of multiattribute preference functions (see [13, 15, 27, 42]). 295
8 Multiattribute Utility Theory (MAUT)
o
A decision maker uses the appropriate preference function, v.x/ or v(x) in the case 298
of certainty or u(x) in the case of risk, to choose among available alternatives. The 299
major emphasis of the work on multiattribute utility theory has been on questions 300
involving conditions for the decomposition of a preference function into simple 301
polynomials, on methods for the assessment of these decomposed functions, and on 302
methods for obtaining sufficient information regarding the multiattribute preference 303
F
functions so that the evaluation can proceed without its explicit identification with 304
O
Suppose that the alternatives defined for single attribute preference functions 306
Y n
O
are now considered to be vectors. That is, suppose that X D Xi where Xi 307
iD1
PR
represents the performance of an alternative on attribute i. We will be interested 308
in conditions allowing the determination that .x1 ; : : : ; xn / .y1 ; : : : ; yn / if and only 309
o o
if v .x1 ; ; xn / v .y1 ; ; yn / for example. Essentially, all that is required is the 310
assumption that the decision maker’s preferences are a weak order on the vectors of
D
311
In some cases, methods for multiattribute optimization do not need any additional 313
invoking concavity to allow maximization. Geoffrion et al. [23] provide an example 315
EC
of an early approach to multiattribute optimization that does proceed with only local 316
conditions are needed to decompose the multiattribute preference function into 318
simple parts.
R
319
R
320
C
The most common approach for evaluating multiattribute alternatives is to use 321
an additive representation. For simplicity, we will assume that there exist a most 322
N
preferred outcome xi and a least preferred outcome x0i on each attribute i D 1 to n. In 323
o
U
the additive representation, a real value v is assigned to each outcome .x1 ; ; xn / by 324
o X
n
o
v .x1 ; ; xn / D vi .xi / (8.3)
iD1
J.S. Dyer
o
where the v i are single attribute value functions over Xi .3 When it is convenient, 325
o o o
we may choose the scaling vi xi D 1; vi x0i D 0; and write v .x1 ; x2 ; : : : ; xn / D 326
X o o
n X o
n
i vi .xi / where i D 1. 327
iD1 iD1
If our interest is in simply rank-ordering the available alternatives then the key 328
condition for the additive form in (8.3) is mutual preference independence. Suppose 329
that we let I f1; 2; : : : ; ng be a subset of the attribute indices, and define XI as the 330
subset of the attributes designated by the subscripts in I. Also, we let X I represent 331
F
1. 1. XI is preference independent of X I if .wI ; wI / .xI ; wI / for any wI ; xI 2 XI 333
O
and wI 2 X I implies .wI ; xI / .xI ; xI / for all xI 2 X I : 334
2. The attributes X1 , : : : , Xn are mutually preference independent if for every subset
O
335
PR
When coupled with a solvability condition and some technical assumptions, 337
multiattribute value function for n 3 attributes. Furthermore, this additive ordinal 339
rates) between Xi and Xj are independent of all other attributes. Mutual preference 342
independence requires that preference independence holds for all pairs Xi and Xj . 343
Essentially, mutual preference independence implies that the indifference curves for 344
EC
any pair of attributes are unaffected by the fixed levels of the remaining attributes. 345
Debreu [11], Luce and Tukey [38], and Gorman [24] provide axiom systems and 346
An example may help to illustrate the idea of preference independence. Suppose 348
R
that a subject is attempting to evaluate automobiles based on the three criteria 349
R
of cost, horsepower, and appearance. Assume that the subject decides that her 350
preferences between two automobiles differing in cost and horsepower but with 351
O
identical values for appearance are as follows: ($24,000, 150 hp, ugly) ($25,000, 352
170 hp, ugly). If the level of appearance does not affect the subject’s indifference 353
C
curve between cost and horsepower, then she will also prefer ($24,000, 150 hp, 354
beautiful) to ($25,000, 170 hp, beautiful), and will maintain the same preference
N
355
As a practical matter, it is only necessary for preference independence to hold for 357
the n-1 pairs of criteria involving the first criterion and the other n-1 criteria taken 358
one at a time. See Keeney and Raiffa [32, Chap. 3] for a discussion. 359
In Chap. 4 of this volume, Bouyssou and Pirlot provide an excellent discussion 360
of the additive ordinal value function which they present as the use of conjoint 361
measurement for multiple criteria decision making. In our development, we use 362
o
3
Note that the v i are called partial value functions by Bouyssou and Pirlot in Chap. 4 of this volume.
8 Multiattribute Utility Theory (MAUT)
the terminology ordinal additive value function instead in order to contrast this 363
preference representation with other additive and non-additive preference models. 364
We also use the term preference independence rather than simply independence to 365
distinguish this key assumption from other forms of independence conditions that 366
are appropriate for multiple criteria decision making in different contexts. 367
F
The additive ordinal value function would seem to be an attractive choice for 369
O
additive function is, in general, difficult to assess. The problem arises because the 371
o
single attribute functions vi cannot be assessed using the methods appropriate for
O
372
the single-attribute measurable value functions. Instead, these functions can only be 373
PR
assessed through protocols that require tradeoffs between two attributes throughout 374
the process, and these protocols are therefore burdensome for the decision makers. 375
Further, the resulting additive function will only have an ordinal interpretation, 376
additive value functions. However, an example may be helpful to emphasize that the 379
TE
resulting additive value function may only provide an ordinal ranking of alternatives, 380
decision maker on three attributes X, Y, and Z that are related in the mind of 383
the decision maker in a multiplicative form; that is, the decision maker’s true 384
preferences are represented by the product xyz where x, y, and z are attribute values. 385
R
Of course, the analyst is not aware of this multiplicative form, and is attempting to 386
develop an appropriate preference representation from the decision maker based on 387
R
a verbal assessment procedure. Further, suppose that there is no risk involved, so 388
O
the analyst would like to consider the use of an additive ordinal multiattribute value 389
function. 390
C
An example of a situation that might involve this type of a preference function 391
would be the ranking of oil exploration opportunities based on estimates of their oil 392
N
reserves. Suppose that the decision maker thinks that these reserves can be estimated 393
by multiplying the area (x) of the structure containing oil by its depth (y) to obtain 394
U
the volume of the structure, and then multiplying this volume by its rate of recovery 395
per volumetric unit (z). In practice, this is a simplification of the approach actually 396
This multiplication of the relevant parameters could be done explicitly in this 398
case, but this example should suggest that such a true preference structure could 399
occur naturally. For simplicity, and to avoid complications associated with units of 400
measurement, we will assume that X D Y D Z D [1,10], which might occur if the 401
The analyst does not know the true underlying preference model of the decision 403
maker, and so he might ask a series of questions to determine if mutual preference 404
would be asked to compare (2,3,4) with (4,2,4), and would reply that she prefers 407
calculations are unknown to the analyst). It is easy to see that alternative 2 would 409
remain preferred to alternative 1 for all common values of z1 and z2 , so attributes X 410
and Y are preference independent of Z. Likewise, a similar set of questions would 411
reveal that X and Z are preference independent of Y, and Y and Z are preference 412
F
independent of X, so these three attributes are mutually preference independent. 413
Therefore, the analyst concludes that the preferences of the decision maker can 414
O
be represented by the ordinal additive multiattribute preference function 415
O
o o o o
v .x; y; z/ D v x .x/ C v y .y/ C v z .z/
PR
As we shall see, this is not a mistake even though the true preference function 416
is multiplicative, and the assessment procedure will construct the correct ordinal 417
additive function that will result in the same rank ordering of alternatives as the 418
D
multiplicative preference function. 419
For this example only, we will abuse the notation and let subscripts of the 420
TE
attributes indicate the corresponding values of the single attribute functions. For 421
o
examples, we will let x0 indicate the value of attribute X such that v x .x0 / D 0, and 422
o
let y1 indicate the value of attribute Y such that v y .y1 / D 1, and so forth. Suppose
EC
423
allowable given the fact that the function is unique up to a linear transformation. 425
o o o
That is, the analyst scales v x so that v x .x0 / D v x .1/ D 0, and similarly scales 426
R
o o
v y .1/ D v z .1/ D 0. Therefore, we would have 427
R
o o o o
v .1; 1; 1/ D v x .1/ C v y .1/ C v z .1/ D 0 C 0 C 0 D 0:
O
o o
C
The analyst then arbitrarily selects x1 D 2; that is, he sets v x .x1 / D v x .2/ D 1; 428
which is also allowable by virtue of the scaling convention. Finally, the analyst 429
N
involves the decision maker, and asks her to specify a value y1 so that she is 430
indifferent between the alternative (2,1,1) and the alternative (1,y1,1). Based on her
U
431
true multiplicative preference model unknown to the analyst, if she is indifferent 432
between (2,1,1) and (1,y1,1) it must be the case that 2 1 1 D 1 y1 1, so she 433
o o
responds y1 D 2 . Based on this response, the analyst sets v y .y1 / D v y .2/ D 1. 434
o o o o
This means that v .2; 1; 1/ D v x .2/ C v y .1/ C v z .1/ D 1 C 0 C 0 D 1, and that 435
o o o o
v .1; 2; 1/ D v x .1/C v y .2/C v z .1/ D 0C1C0 D 1, which verifies to the analyst that 436
the additive representation indicates that the decision maker is indifferent between 437
o
the alternatives (2,1,1) and (1,2,1). In addition, the analyst knows that v .2; 2; 1/ D 438
o o o
v x .2/ C v y .2/ C v z .1/ D 1 C 1 C 0 D 2. 439
8 Multiattribute Utility Theory (MAUT)
Now, the analyst asks the decision maker to specify a value for x2 so that 440
she is indifferent between the alternatives (2,2,1) and (x2 ,1,1). This response will 441
o
determine the value of x2 such that v x .x2 / D 2, because indifference between these 442
o o o o
two alternatives will require v .x2 ; 1; 1/ D v x .x2 / C v y .1/ C v z .1/ D 2 C 0 C 0 D 2 443
also. 444
Using her implicit multiplicative preference function for the alternative (2,2,1), 445
F
449
o o
and v z .z/, the analyst would develop graphs that would indicate v x .x/ D ln x/ln 2,
O
450
o o
v y .y/ D ln y/ln 2, and v z .z/ D ln z/ln 2, so that the ordinal additive multiattribute 451
O
value function would be given by the sum of the logs of the variables. Notice that this 452
ordinal value function is an order preserving transformation of the true underlying 453
PR
preference representation of the decision maker, which was never revealed explicitly 454
As this example illustrates, the assessment procedure will determine an additive 456
ordinal value function that may be an order preserving transformation of a true 457
D
preference relation that is not additive. The log function provides an example of 458
such a transformation for a multiplicative preference relation, but other non-additive 459
TE
relationships may also be transformed to order preserving additive value functions. 460
See Krantz et al. [35] for a discussion of other such transformations. 461
This example also illustrates the fact that the assessment methods required 462
EC
for accurately capturing an additive ordinal multiattribute value function may be 463
tedious, and will require tradeoffs involving two or more attributes. This same 464
point is made by Bouyssou and Pirlot in Chap. 4 of this volume. Thus, while 465
R
this approach could be used in practice, it would be desirable to have simpler 466
means of assessing the underlying preference functions. This can be accomplished 467
R
if some additional preference conditions are satisfied, but the requirement of mutual 468
O
preference independence will still be common to the preference models that are to 469
follow. 470
C
N
Y
n
When X D Xi in a von Neumann–Morgenstern utility model and the decision 473
iD1
maker’s preferences are consistent with some additional independence conditions, 474
then u(x1 , x2 , : : : , xn ) can be decomposed into additive, multiplicative, and other 475
well-structured forms that simplify assessment. In comparison with other sections, 476
our coverage of this topic will be relatively brief since it is perhaps the most well 477
preferences over lotteries with different levels of Xi do not depend on the fixed 481
independent if all proper subsets of these attributes are utility independent of their 483
complementary subsets. Further, it can be shown that if these same attributes are 484
mutually preference independent, then they will also be mutually utility independent 485
if any pair of the attributes is utility independent of its complementary attributes 486
[32]. 487
F
Returning to the automobile selection example, suppose that a decision maker is 488
O
considering using the attributes of cost, horsepower, and appearance as before, but 489
there is some uncertainty regarding some new environmental laws that may impact 490
O
the cost and the horsepower of a particular automobile. Further, assume that the 491
decision maker prefers more horsepower to less, lower costs and more attractive 492
PR
automobiles. The current performance levels of one of the alternatives may be 493
($25,000, 170 hp, ugly), but if the legislation is passed a new device will have to 494
be fitted that will increase cost and decrease horsepower to ($25,700, 150 hp, ugly). 495
An alternative automobile might have possible outcomes of ($28,000, 200 hp, ugly) 496
D
and ($29,000, 175 hp, ugly) depending on this same legislation, which the decision 497
498
Therefore, the decision maker may consider choices between lotteries such as the 499
one shown in Fig. 8.1. For example, the decision maker may prefer Auto 1 to Auto 500
2 because the risks associated with the cost and the horsepower for Auto 1 are more
EC
501
acceptable to her than the risks associated with the cost and horsepower of Auto 2. If 502
the decision maker’s choices for these lotteries do not depend on common values of 503
the third attribute, then cost and horsepower are utility independent of appearance. 504
R
A multiattribute utility function u(x1 , x2 , : : : , xn ) can have the multiplicative form 505
R
Y
n
1 C ku .x1 ; x2 ; : : : ; xn / D Œ1 C kki ui .xi / (8.4)
O
iD1
C
N
Auto 2
Auto 1
U
pass pass
0.5 ($25700, 150 hp, ugly) 0.5 ($29000, 175 hp, ugly)
Auto 1 Auto 2
pass pass
0.5 ($29000, 150 hp, ugly) 0.5 ($29000, 200 hp, ugly)
F
if and only if the attributes X1 , X2 , : : : , Xn are mutually utility independent, where
O
506
O
scaling constants, and k is an additional scaling constant. If the scaling constant k is 508
determined to be 0 through the appropriate assessment procedure, then (8.4) reduces 509
PR
to the additive form 510
X
n
ki ui .xi / (8.5)
D
iD1
X
TE
n
where ki D 1. 511
iD1
EC
A majority of the applied work in multiattribute utility theory deals with the 513
R
case when the von Neumann–Morgenstern utility function is decomposed into the 514
additive form (8.5). Fishburn [17] has derived necessary and sufficient conditions for
O
515
a utility function to be additive. The key condition for additivity is the marginality 516
condition which states that the preferences for any lotteries p, q 2 P should depend
C
517
only on the marginal probabilities of the attribute values, and not on their joint 518
N
Returning to the automobile example once again, for additivity to hold, the 520
U
decision maker must be indifferent between the two lotteries shown in Fig. 8.2, and 521
for all other permutations of the attribute values that maintain the same marginal 522
Notice that in either lottery, the marginal probability of receiving the most 524
preferred outcome or the least preferred outcome on each attribute is identical (0.5). 525
However, a decision maker may prefer the right-hand side lottery over the left- 526
hand side lottery if the decision maker wishes to avoid a 0.5 chance of the poor 527
outcome ($29,000, 150 hp, ugly) on all three attributes, or she may have the reverse 528
J.S. Dyer
preference if she is willing to accept some risk in order to have a chance at the best 529
outcome on all three attributes. In either of the latter cases, utility independence 530
may still be satisfied, and a multiplicative decomposition of the multiattribute utility 531
Other independence conditions have been identified that lead to more complex 533
F
536
O
The assessment of the multiplicative or additive form implied by the condition 537
of mutual utility independence is simplified by the fact that each of the single-
O
538
attribute utility functions may be assessed independently (more accurately, while 539
PR
all of the other attributes are held constant at arbitrarily selected values), using 540
the well-known utility function assessment techniques suitable for single attribute 541
utility functions. In addition, the constants ki and k can be assessed using n 542
relatively simple tradeoff questions. See Keeney and Raiffa [32] or Kirkwood [34] 543
D
for additional details and examples. 544
TE
until after the review of multiattribute utility theory because the latter may be more 548
familiar to the reader. If so, this transposition of a more natural order of presentation 549
R
may be helpful in providing the opportunity to discuss similarities between these 550
551
Again let X denote the set of all possible consequences in a particular decision 553
Yn
problem. In the multiattribute problem X D
N
iD1
U
consequences for the ith attribute. In this section, we use the letters w, x, y, and z to 555
where wi is a level in the nonempty attribute set Xi for i D 1, : : : , n. Once again, we 557
let I f1; 2; : : : ; ng be a subset of the attribute indices, define XI as the subset of the 558
attributes designated by the subscripts in I, and let X I represent the complementary 559
subset of the n attributes. We may write w D (wI , wI ) or use the notation .wi ; wi / and 560
.xi ; wi / to denote two elements of X that differ only in the level of the ith attribute. 561
Finally, we also assume that the preference relation on X is a weak order. 562
8 Multiattribute Utility Theory (MAUT)
Next we introduce the notation necessary to define preferences based on strength 563
nonempty subset of X X, and denote a weak order on X . Once again, we may 565
interpret wx yz to mean that the preference difference between w and x is greater 566
than the preference difference between y and z. 567
These two orders are difference consistent if, for all wi ; xi 2 Xi ; .wi ; wi / .xi ; wi / if 570
and only if .wi ; wi / .xi ı ; wi / .xi ; wi / .xi ı ; wi / for some xıi 2 Xi and some wi 2 X i ; 571
F
and for any i 2 f1; : : : ; ng ; and if w x then wy xy or yw yx or both for 572
O
any y 2 X: Loosely speaking, this means that if one multiattributed alternative is 573
O
some common reference alternative xıi ; wi will be larger than the preference 575
difference between the alternative that is not preferred and this reference alternative. 576
dence. This condition plays a role similar to the utility independence condition in 579
multiattribute utility theory. We show how this condition can be exploited to obtain 580
multiplicative and other nonadditive forms of the measurable multiattribute value 581
EC
function. 582
given any wI ; xI ; yI ; zI 2 XI and some wI 2 X I such that the subject’s judgments 584
R
as follows: .wI; wI / .xI ; wI / .yI ; wI / .zI ; wI /, then the decision maker will also 586
R
consider .wI; xI / .xI ; xI / .yI ; xI / .zI ; xI / for any xI 2 X I . That is, the ordering of 587
O
preference differences depends only on the values of the attributes XI and not on the 588
The attributes are mutually weak difference independent if all proper subsets of 590
N
these attributes are weak difference independent of their complementary subsets. 591
Further, it can be shown that if these same attributes are mutually preference 592
U
independent, then they will also be mutually weak difference independent if any pair 593
of the attributes is weak difference independent of its complementary attributes [13]. 594
Notice the similarity of the definition of weak difference independence to that 595
of utility independence. In the latter case, preferences among lotteries depend only 596
on the values of the attributes XI and not on the fixed values of X I . In the case of 597
certainty, the same notion applies to preference differences. Therefore, it should 598
value function that is identical to the one implied by utility independence for utility 600
functions. 601
J.S. Dyer
Yn
1 C v.x/ D Œ1 C i vi .xi / (8.6)
iD1
single-attribute measurable value function over Xi scaled from 0 to 1, the i are 605
positive scaling constants, and is an additional scaling constant. If the scaling 606
F
Xn
O
v.x/ D i vi .xi / (8.7)
iD1
O
X
n
where i D 1. Therefore, we obtain either an additive or a multiplicative 609
PR
iD1
measurable preference function that is based on notions of strength of preference.
D 610
Finally, we are interested in the conditions that are required to ensure the existence 612
of an additive multiattribute measurable value function. Recall that mutual prefer- 613
EC
ence independence guarantees the existence of an additive preference function for 614
the case of certainty that will provide an ordinal ranking of alternatives, but it may 615
not capture the underlying strength of preference of the decision maker. Further, the 616
R
appropriate assessment technique will require tradeoffs that simultaneously consider 617
Recall the example from Sect. 8.3.2 where the decision maker’s true preferences 619
were represented by the product of the attributes. If we were to ask the decision
O
620
maker to express her preferences for the first attribute while holding the other 621
C
attributes constant at some given values, she would respond in such a way that we 622
would obtain a linear function for each attribute, rather than the correct logarithmic 623
N
form. We would like to exclude this case, and be assured that the preference function 624
Perhaps this point is worth some elaboration. Recall that the true preferences of 626
the hypothetical decision maker introduced in Sect. 8.3.2 were consistent with the 627
multiplicative representation xyz. Suppose we set y D z D 1, and ask the decision 628
maker to consider the importance of changes in the attribute x while holding these 629
other attribute values constant. Considering the alternatives (1,1,1), (3,1,1), and 630
(5,1,1), she would indicate that the preference difference between (3,1,1) and (1,1,1) 631
would be the same as the preference difference between (5,1,1) and (3,1,1). This is 632
and the preference difference between (3,1,1) and (1,1,1) is 3-1 D 2, which is also 634
the preference difference between (5,1,1) and (3,1,1). If the analyst is not aware 635
of the fact that this assessment approach cannot be used when only preference 636
o
independence is satisfied, he might mistakenly conclude that v .x; y; z/ D x C y C z 637
rather than the appropriate logarithmic transformation that we obtained earlier in 638
The required condition for additivity that also provides a measurable preference 640
dent of X i if, for all wi ; xi 2 Xi such that .wi ; wi / .xi ; wi / for some wi 2 642
X i ; .wi ; wi / .xi ; wi / .wi ; xi / .xi ; xi / for any xi 2 X i : Intuitively, the preference 643
F
difference between two multiattributed alternatives differing only on one attribute 644
O
does not depend on the common values of the other attributes. 645
The attributes are mutually difference independent if all proper subsets of these 646
O
attributes are difference independent of their complementary subsets. Again, it can 647
be shown that if these same attributes are mutually preference independent, then 648
PR
they will also be mutually difference independent if X1 is difference independent of 649
X 1 [13]. For the case of n 3, mutual difference independence along with some 650
additional structural and technical conditions4 ensure that if wx, yz 2 X*, then wx 651
X
n X
n X
n X
n
i vi .wi / i vi .xi / i vi .yi / i vi .zi / (8.8)
iD1 iD1 iD1 iD1
EC
X
n X
n
R
iD1 iD1
O
iD1
properties, then there exist constants ˛ > 0; ˇ1 ; : : : ; ˇn such that vi 0 D ˛vi C ˇi ;
N
656
i D 1; : : : ; n: 657
U
Result (8.9) is well known and follows immediately from the assumption that the 658
attributes are mutually preference independent (Sect. 8.3.1). The significant result 659
4
Specifically, we assume restricted solvability from below, an Archimedian property, at least three
attributes are essential, and that the attributes are bounded from below. If n D 2, we assume that
the two attributes are preferentially independent of one another and that the Thomsen condition
is satisfied (see Krantz et al. [35] and the discussion by Bouyssou and Pirlot in Chap. 4 of this
volume).
J.S. Dyer
X
n
is (8.8), which means that v .x1 ; x2 ; : : : ; xn / D i vi .xi / also provides difference 660
iD1
measurement on X. Note that this latter result is obtained based on the observation 661
that any arbitrarily selected attribute is difference independent of its complementary 662
attributes. 663
F
Because the notion of a measurable multiattribute value function may not be familiar 665
O
to many readers, we will briefly consider methods for the assessment of them. 666
Further details and examples are provided by von Winterfeldt and Edwards [47], 667
O
and by Kirkwood [34]. 668
PR
8.5.3.1 Verification of the Independence Conditions 669
The first issue to be considered is the verification of the independence conditions. 670
D
Since methods for verifying mutual preference independence are discussed in 671
TE
Keeney and Raiffa [32], we focus on the independence conditions involving 672
to hold in most practical applications. The following procedure could be used to 675
in situation 1: She already has .x1 ; w1 / and she can exchange it for .w1 ; w1 /. Next,
R
678
we arbitrarily choose x1 2 X 1 and ask her to imagine situation 2: She already has 679
R
.x1 ; x1 /, and she can exchange it for .w1 ; x1 /. Would she prefer to make the exchange 680
in situation 1 or in situation 2, or is she indifferent between the two exchanges? If she 681
O
is indifferent between the two exchanges for several different values of w1 ; x1 2 X1 682
683
For example, suppose we ask the decision maker to consider exchanging a car 684
N
described by ($25,000, 150 hp, ugly) for a car described by ($25,000, 180 hp, ugly). 685
Next, we ask her to consider exchanging ($35,000, 150 hp, nice) for ($35,000, 180 686
U
hp, nice). Would the opportunity to exchange a car with 150 hp for one with 180 hp 687
be more important to the decision maker when the cost and appearance are $25,000 688
and ugly, or when they are $35,000 and nice? If the common values of these two 689
attributes do not affect her judgments of the importance of these exchanges, then 690
Before using this procedure, we must ensure that the decision maker understands 692
that we are asking her to focus on the exchange rather than on the final outcomes. For 693
example, if she states that she prefers an exchange of $1,000,000 for $1,000,001 to 694
an exchange of $5 for $500, then she undoubtedly is not focusing on the substitution 695
8 Multiattribute Utility Theory (MAUT)
of one outcome for another, but she is focusing instead on the final outcome. Thus, 696
some training may be required before this approach to verification of difference 697
To verify weak difference independence, partition X into XI and X I , and choose 699
exchange of .xI ; wI / for .wI ; wI / is preferred to the exchange of .zI ; wI / for .yI ; wI /. 701
Then pick another value xI of X I and ask if the decision maker still prefers the 702
exchange of .xI ; xI / for .wI ; xI / to the exchange of .zI ; xI / for .yI ; xI /. This must be 703
true if the subset XI is weakly difference independent of X I . If the decision maker’s 704
response is affirmative, we repeat the question for other quadruples of consequences 705
F
from XI with the values of the criteria in X I fixed at different levels. Continuing in 706
this manner and asking the decision maker to verbally rationalize her responses, the 707
O
analyst can either verify that XI is weakly difference independent of X I or discover 708
O
709
Note that for the multiplicative measurable value function, it would only be 710
necessary to verify weak difference independence for the special case of I D fi,jg,
PR
711
where i and j indicate the subscripts of an arbitrarily chosen pair of alternatives. This 712
For example, suppose we establish that the decision maker would prefer the 714
D
exchange of the car ($25,000, 150 hp, ugly) for the car ($27,000, 200 hp, ugly) 715
to the exchange of the car ($24,000, 130 hp, ugly) for the car ($25,000, 150 hp, 716
TE
ugly). If this preference for the first exchange over the second exchange does not 717
depend on the common value of appearance, and if it also holds true for all other 718
combinations of the values of cost and horsepower, then cost and horsepower are 719
EC
measurable value function vi can be assessed while holding xi constant at any 723
O
arbitrary value (generally at xi ı ). With the additive value function that does not 724
C
provide difference measurement, this strategy cannot be used as illustrated above. 725
As a result, any of the approaches for assessing a single attribute measurable value 726
N
function referenced in Sect. 8.3.1 may be used, including the direct rating of attribute 727
values on an arbitrary scale (e.g., from 0 to 100), or direct estimates of preference 728
U
differences. 729
If the measurable value function is additive, the scaling constants may be 730
assessed using the same trade-off approach suggested for estimating the scaling 731
constants for the additive ordinal value function [32, Chap. 3]. For a discussion of 732
other approaches to the assessment of the scaling constants for the additive and 733
Measurable multiattribute value functions may also be assessed using the ratio 735
judgments and tools provided by the AHP methodology, and used as a basis
J.S. Dyer
for relating the AHP to formal preference theories that are widely accepted by 736
economists and decision analysts. This point has been made by several authors, 737
Perhaps the best discussion of this important point is provided by Salo and 739
Hamalainen [41]. As they observe, once a suitable range of performance [xıi , 740
x*i ] has been defined for each attribute, the additive measurable value function 741
representation may be scaled so that the values v .xı / D v xı1 ; : : : ; xın D 0 742
and v .x / D v x1 ; : : : ; xn D 1 are assigned to the worst and best conceivable 743
onto the [0,1] range, the additive representation can be written as 745
F
X
n X
n
O
v.x/ D vi .xi / D vi .xi / vi xıi
iD1 iD1
O
X
n
ı vi .xi / vi xıi Xn
D vi xi vi xi ı D wi si .xi /
PR
iD1
vi xi vi xi iD1
ı
where si (xi ) D vi .xi / vi xıi =v
i xi vıi xi 2 [0, 1] is the normalized score of x 746
on the ith attribute and wi D v xi v xi is the scaling constant or weight of the 747
D
ith attribute. 748
TE
A careful evaluation of this representation leads Salo and Hamalainen to the 749
conclusion that pair wise comparisons in ratio estimation should be interpreted in 750
terms of ratios of value differences between pairs of underlying alternatives. This, in 751
EC
turn, provides the link between traditional models of preference theory and the AHP, 752
and reveals that the latter can be an alternative assessment technique for measurable 753
multiattribute value functions (with some simple adjustments for normalization and 754
scaling). 755
R
R
756
Goal programming was originally proposed by Charnes et al. [10] as an ingenious 758
and Cooper [9] in a review of the field, this approach to multiple objective 760
optimization did not receive significant attention until the mid-1960s. However, 761
during the past 40 years, we have witnessed a flood of professional articles 762
and books dealing with applications of this methodology (e.g. see Ignizio [25], 763
This discussion is limited to the use of goal programming as a methodology for 765
solving problems with multiple, compensatory objectives. That is, we do not address 766
problems that do not allow tradeoffs among the objectives. These non-compensatory 767
models involve the use of the non-Archimedian, or “preemptive priority”, weights. 768
8 Multiattribute Utility Theory (MAUT)
An analysis of these models would be based on the theory of lexicographic orders, 769
summarized by Fishburn [19]. The conditions that would justify the use of a non- 770
compensatory model are very strict, and are unlikely to be met in a significant 771
Preferences 774
Let us begin with a simple example. Suppose a manager has identified a problem 775
F
that can be formulated as a traditional mathematical programming problem with one 776
exception—there are two criterion functions, f1 (x) and f2 (x) where x 2 X is an n- 777
O
vector of controllable and uncontrollable variables, and the non-empty feasible set 778
O
779
PR
To use goal programming, we ask the manager if she has any “goals” in mind for 781
the criteria. She replies that she would be happy if f1 ./ were at least as large as b1 , 782
but she does not feel strongly about increasing f1 ./ beyond b1 . However, she would 783
like for f2 ./ to be somewhere between b2L and b2U . Finally, we ask her to assign 784
“weights” of relative importance to the deviations of f1 ./ from b1 , and of f2 ./ from
D
785
b2L and b2U , respectively. After some thought, she responds with the weights w1 , w2 786
TE
and w3 . 787
C
min w1 y
1 C w2 y2 C w3 y3
x2X
subject to f1 .x/ yC
1 C y1 D b 1
C
f2 .x/ y2 C y2 D b2L (GP)
R
f2 .x/ yC
3 C y3 D b2U
C
R
yi ; yi 0; i D 1; 2; 3
O
Notice that GP includes a “one-sided” formulation with respect to f1 ./, and a “goal 789
Let us pause a moment to reflect on this formulation. First, notice that y 1 D 791
b 1 C yC .x/. .x/ C
D 0
N
constraint for GP.Since b1 is a constant, minimizing w1 y1 is obviously equivalent 793
to minimizing w1 yC C
U
1 y12 . 794
C
Similarly, if we introduce the constraint f2 .x/ y22 D 0, minimizing w2 y
2 795
C C C
is equivalent to minimizing w2 y2 y22 , and minimizing w3 y3 is equivalent 796
to minimizing w3 y3 C yC 22 b2U . The constant b2U is maintained in the last 797
we shall see. Combining the results and re-writing GP as a maximization problem, 799
C C
max w1 yC C C
12 y1 C w2 y22 y2 C w3 y22 C y3 b2U
x2X
subject to f1 .x/ yC
1 C y1 D b 1
C
f2 .x/ y2 C y2 D b2L
f2 .x/ yC
3 C y3 D b2U (VA)
C
f1 .x/ y12 D 0
f2 .x/ yC22 D 0
yC
i ; y
i 0; i D 1; 2; 3
yC ;
12 22y C
0
F
where the objective function may be interpreted as the sum of two piecewise linear 801
O
Figures 8.3 and 8.4 illustrate these two piecewise linear functions. Recall that 803
O
piecewise linear transformations are commonly used to transform additive separable 804
nonlinear programming problems into linear programming problems. The lines 805
PR
labeled v1 ./ and v2 ./ in Figs. 8.3 and 8.4 respectively suggest nonlinear preference 806
functions that might be approximated by the bold piecewise linear functions. 807
Thus since VA is equivalent to GP, and VA may be viewed as a piecewise linear 808
And how do we interpret V? Since the choice of goals and goal intervals in 811
EC
GP reflect the decision maker’s preferences and no uncertainty is involved in the 812
R
R
O
C
N
U
F
O
O
Fig. 8.4 Piecewise linear approximation of v2 ./
PR
decision, v1 ./ and v2 ./ are measurable functions, and their sum, v .f1 ./ ; f2 .// D 813
v1 .f1 .// C v2 .f2 .//, is an additive separable measurable value function. 814
involved in the formulation. Therefore, the additive utility function theory developed 816
TE
for risky choice is not relevant for these applications. Likewise, the ordinal additive 817
theories are not operational here because they require a simultaneous conjoint 818
scaling of the separable terms. Goal programming applications generally allow the 819
EC
selection of each goal or goal interval independent of consideration of the values of 820
the other criteria. This practice implies the existence of a measurable additive utility 821
This point has been made recently by Bordley and Kirkwood [5] in a general dis-
R
823
cussion of the relationship between goals and multiattribute preference models. Also 824
R
see Abbas and Matheson [2]. This perspective provides some insights regarding the 825
nature of goal programming, as well as some challenges. For example, how should 826
O
the piecewise linear approximations to the nonlinear value functions be selected 827
in order to minimize error? Geoffrion [22] provides some useful guidelines for 828
C
choosing “goals” or “goal intervals” for each criterion so that the piecewise linear 829
N
approximation to the implicit utility function provides the best fit. 830
One important implication of this point of view is that goal programming should 831
U
The necessary conditions for the additive and multiplicative measurable value 838
functions and risky utility functions, notably mutual preference independence, are 839
also necessary and sufficient for the ordinal additive value function that does 840
relationships among them. The following choice of scaling will be imposed. For 842
ı
f D v; v; or u, f is normalized by f .x1 ; : : : ; xn / D1 and f .x1 ı ;: : :; xn ı / D 0 and 843
F
844
o
; and k will be used as scaling constants for the ordinal and measurable value
O
845
O
PR
8.6.1 The Additive Functions 847
The relationships among the alternative developments of the additive forms of 848
D
real-valued functions on X follow immediately from their respective uniqueness 849
then v D v; 853
857
ı
Note the implication of this result. In order for v D v D u for a single
O
858
decision maker, she must have preferences simultaneously consistent with mutual 859
C
risky alternatives. Mutual preference independence will hold in all cases, but it 861
N
may be the case that difference independence and/or additive independence for 862
risky alternatives will not hold. Further, difference independence may hold for 863
U
the preferences of a decision maker, implying that an additive measurable value 864
function would provide a valid representation of her preferences, but additive 865
independence for risky alternatives may not be satisfied, implying that an additive 866
utility function would not be a valid representation of her preferences in decision 867
Throughout this section we assume that the following conditions are satisfied: 870
ı
Suppose we have assessed the additive value function v and wish to obtain either 875
v or u. Then the following relationships will hold ([13], Theorem 5). Either
F
876
ı ı
1. 1. v.x/ D v.x/ and v i .xi / D vi .xi /, i D 1, : : : , n, or
O
877
ı ı ı
2. 2. v.x/ ln .1 C / D ln Œ1 C v.x/ and i v i .xi / ln .1 C / D ln Œ1 C i vi .xi /, 878
O
i D 1, : : : , n. 879
PR
Either 880
ı ı
1. 1. v.x/ D u.x/ and v i .xi / D ui .xi /, i D 1, : : : , n, or, 881
ı ı ı
2. 2. v.x/ ln .1 C k/ D ln Œ1 C ku.x/ and i v i .xi / ln .1 C k/ D ln Œ1 C kki ui .xi /, 882
D
i D 1; : : : ; n: 883
TE
preference functions. For example, suppose we define xi 0 as the equal difference 885
point for attribute Xi if .xi 0 ; xi / .xi ı ; xi / .xi ; xi / .xi 0 ; xi / for any xi 2 X i . Notice 886
887
889
ı
Finally, to derive u from v; find xi 00 for some attribute Xi such that the decision 890
R
maker is indifferent between xi 00 and an equal chance lottery between xi and xi ı 891
O
with the other criteria held fixed. A parallel result to the above relationship between 892
o o o
some i 2 f1; : : : ; ng ;, then u D v. Otherwise, 1 C .1 C k/i D 2.1 C k/ . / .
ı i v i x0i
894
N
These results can also be used to derive v after u has been assessed, or vice versa. 895
For example, suppose u has been assessed using appropriate procedures. To obtain 896
U
v, we find the equal difference point xi 0 for some criterion Xi . The second result 897
ı ı
above is used to obtain i and vi for each criterion, and we can obtain v. In a similar 898
manner, u can be obtained from v after assessing xi 00 for some criterion Xi . 899
Since the AHP can be interpreted as ratios of preference differences, this 900
relationship also allows the results from assessments based on the AHP to be 901
suitably transformed into multiattribute utility functions appropriate for use in risky
J.S. Dyer
situations. This completes the circle required to synthesize ordinal multiattribute 902
functions, and multiattribute functions based on ratio judgments. As a result, the 904
analyst is justified in choosing among a variety of assessment tools, and making 905
the appropriate adjustments in order to calibrate the results into a coherent and 906
F
In this chapter, we have presented an informal discussion of “multiattribute utility 909
O
theory”. In fact, this discussion has emphasized that there is no single version of 910
O
911
there are three distinct theories of multiattribute preference functions that may be 912
PR
used to represent a decision maker’s preferences. 913
The ordinal additive multiattribute preference model requires the assumption of 914
mutual preference independence, and is appropriate for use in the case of certainty. 915
Most of the applications and methods of multicriteria decision analysis are presented 916
D
in the context of certainty, and so this would seem to be an appealing theory to 917
use for framing these approaches. However, as we have emphasized, the ordinal 918
TE
additive multiattribute preference model requires assessment techniques that are 919
cumbersome in practice, and that force the decision maker to make explicit tradeoffs 920
between two or more criteria in the assessment of the value functions defined on the 921
EC
The measurable value functions also require the assumption of mutual preference 923
independence, along with the stronger assumptions of weak difference indepen- 924
R
of the model that are easy to assess. The assessment of these preference models 926
R
is relatively easy, and they can be interpreted intuitively as providing a measure 927
of strength of preference. In addition, the ratio judgments of the AHP can be 928
O
interpreted as ratios of preference differences based on this theory, linking the AHP 929
C
Finally, multiattribute utility theory is an elegant and useful model of preference 932
suitable for applications involving risky choice. The brilliant work of Keeney and 933
U
Raiffa [32] has made this theory synonymous with multiple criterion decision 934
making, and the ordinal and measurable theories are often overlooked or ignored as 935
a result. In fact, these latter approaches may provide more attractive and appropriate 936
References 938
1. Abbas, A.E., Bell, D.: One-switch independence for multiattribute utility functions. Oper. Res. 939
59(3), 764–771 (2011) 940
2. Abbas, A.E., Matheson, J.: Normative decision making with multiattribute performance 941
targets. J. Multicrit. Decis. Anal. 16, 67–78 (2010) 942
3. Anderson, R., Clemen, R.: Toward an improved methodology to construct and reconcile 943
decision analytic preference judgments. Decis. Anal. 10, 121–134 (2013) 944
4. Bell, D., Raiffa, H.: Risky choice revisited. In: Bell, D.E., Raiffa, H., Tversky, A. (eds.) 945
Decision Making: Descriptive, Normative, and Prescriptive Interactions. Cambridge University 946
Press, New York (1988) 947
F
5. Bordley, R., Kirkwood, C.: Multiattribute preference analysis with performance targets. Oper. 948
Res. 52(6), 823–835 (2004) 949
O
6. Butler, J.C., Dyer, J.S., Jia, J.: Using attributes to predict objectives in preference models. 950
Decis. Anal. 3(2), 100–116 (2006) 951
O
7. Camerer, C.: Individual decision making. In: Kagel, J.H., Roth, A.E. (eds.) Handbook of 952
Experimental Economics. Princeton University Press, Princeton (1995) 953
8. Charnes, A., Cooper, W.W.: Management Models and Industrial Applications of Linear 954
PR
Programming. Wiley, New York (1961) 955
9. Charnes, A., Cooper, W.W.: Goal programming and multiple objective optimizations. Eur. J. 956
Oper. Res. 1(1), 37–54 (1977) 957
10. Charnes, A., Cooper, W.W., Ferguson, R.O.: Optimal estimation of executive compensation by 958
D
linear programming. Manag. Sci. 1(2), 138–151 (1955) 959
11. Debreu, G.: Topological methods in cardinal utility theory. In: Arrow, K.J., Karlin, S., Suppes, 960
TE
P. (eds.) Mathematical Methods in the Social Sciences. Stanford University Press, Stanford 961
(1959) 962
12. Dyer, J.S.: Remarks on the analytic hierarchy process. Manag. Sci. 36, 249–258 (1990) 963
13. Dyer, J.S., Sarin, R.A.: Measurable multiattribute value functions. Oper. Res. 22, 810–822 964
EC
(1979) 965
AQ5 14. Dyer, J.S., Sarin, R.K.: Relative risk aversion. Manag. Sci. 28, 875–886 (1982) 966
15. Ellsberg, D.: Classic and current notions of “Measurable Utility”. Econometrica 64, 528–556 967
(1954) 968
R
16. Farquhar, P.H., Keller, L.R.: Preference intensity measurement. Ann. Oper. Res. 19, 205–217 969
(1989) 970
R
17. Fishburn, P.C.: Independence in utility with whole product sets. Oper. Res. 13, 28–45 (1965) 971
18. Fishburn, P.C.: Utility Theory for Decision Making. Wiley, New York (1970) 972
O
19. Fishburn, P.C.: Lexicographic orders, utilities, and decision rules. Manag. Sci. 20, 1442–1471 973
(1974) 974
C
AQ6 20. Fishburn, P.C.: Nonlinear Preference and Utility Theory. Johns Hopkins University Press, 975
Baltimore and London (1988) 976
N
21. Fishburn, P.C., Wakker, P.: The invention of the independence condition for preferences. 977
Manag. Sci. 41(7), 1130–1144 (1995) 978
U
22. Geoffrion, A.: Objective function approximations in mathematical programming. Math. Pro- 979
gram. 13(1), 23–37 (1977) 980
23. Geoffrion, A., Dyer, J.S., Feinberg, A.: An interactive approach for multi-criterion optimization 981
with an application to the operation of an academic department. Manag. Sci. 19, 357–368 982
(1972) 983
24. Gorman, W.M.: The structure of utility functions. Rev. Econ. Stud. 35, 367–390 (1968) 984
25. Ignizio, J.P.: Introduction to Linear Goal Programming. Sage, Thousand Oaks (1986) 985
26. Ignizio, J., Romero, C.: Goal programming. In: Bidgoli, H. (ed.) Encyclopedia of Information 986
Systems, vol. 2, pp. 489–500. Academic, San Diego (2003) 987
27. Jia, J., Dyer, J.S.: A standard measure of risk and risk-value models. Manag. Sci. 42(12), 1691– 988
1705 (1996) 989
J.S. Dyer
28. Kahneman, D., Tversky, A.: Prospect theory: an analysis of decisions under risk. Econometrica 990
47, 276–287 (1979) 991
29. Kamenetzky, R.D.: The relationship between the analytic hierarchy process and the additive 992
value function. Decis. Sci. 13, 702–713 (1982) 993
30. Keeney, R.L.: Value-Focused Thinking. Harvard University Press, Boston (1996) 994
31. Keeney, R.L., Gregory, R.S.: Selecting attributes to measure the achievement of objectives. 995
Oper. Res. 53(1), 1–11 (2005) 996
32. Keeney, R.L., Raiffa, H.: Decision with Multiple Objectives: Preference and Value Tradeoffs. 997
Cambridge University Press, New York (1993) 998
33. Keller, L.R.: An empirical test of relative risk aversion. IEEE Trans. Man Syst. Cybern. SMC- 999
15, 476–482 (1985) 1000
34. Kirkwood, C.W.: Strategic Decision Making. Duxbury, Belmont (1996) 1001
F
35. Krantz, D.H., Luce, R.D., Suppes, P., Tversky, A.: Foundations of Measurement, Volume I, 1002
Additive and Polynomial Representations. Academic, New York (1971) 1003
O
36. Kreps, D.M.: A Course in Microeconomic Theory. Princeton University Press, Princeton 1004
(1990) 1005
O
37. Krzysztofowicz, R.: Strength of preference and risk attitude in utility measurement. Organ. 1006
Behav. Hum. Perform. 31, 88–113 (1983) 1007
PR
38. Luce, R.D., Tukey, J.W.: Simultaneous conjoint measurement: a new type of fundamental 1008
measurement. J. Math. Psychol. 1, 1–27 (1964) 1009
39. Pratt, J.W.: Risk aversion in the small and in the large. Econometrica 32, 122–136 (1964) 1010
40. Saaty, T.L.: The Analytic Hierarchy Process. McGraw-Hill, New York (1980) 1011
41. Salo, A.A., Hamalainen, R.P.: On the measurement of preferences in the analytic hierarchy 1012
D
process. J. Multi-Criterion Decis. Anal. 6, 309–319 (1997) 1013
42. Sarin, R.K.: Strength of preference and risky choice. Oper. Res. 30, 982–997 (1982) 1014
TE
43. Savage, L.J.: The Foundations of Statistics. Wiley, New York (1954) 1015
44. Scott, D., Suppes, P.: Foundational aspects of theories of measurement. J. Symb. Log. 23, 113– 1016
128 (1958) 1017
45. Trzaskalik, T., Michnik, J. (eds.): Multiple Objective and Goal Programming: Recent Devel- 1018
EC
(2010) 1025
O
C
N
U
AUTHOR QUERIES
AQ1. Please check the edits made to the article tile and correct if necessary.
AQ2. Debreu (1960) has been changed to Debreu (1959) as per the reference list.
Please check if okay.
AQ3. Kirkwood (1997) has been changed to Kirkwood (1996) as per the reference
list in all occurrences. Please check if okay.
AQ4. The usage of “-$2” has been changed to “$2”. Please check and advice.
AQ5. Reference [14] is not cited in the text. Please provide the citation or delete it
F
from the list.
O
AQ6. Please check the inserted publisher location in refs. [20], [25], [35], [36] are
appropriate.
O
PR
D
TE
EC
R
R
O
C
N
U