200 IEEE TRANSACTIONS ON SYSTEMS SCIENCE AND CY13ERNETICS, VOL. SSC-4, NO.

3, SEPTEMLIER 1968

A Tutorial Introduction to Decision Theory

D. WARNER NORTH

Abstract-Decision theory provides a rational framework for not obscure the fact that they represent the limiting case
choosing between alternative courses of action when the conse- of perfect predictions. It is often tempting to assume
quences resulting from this choice are imperfectly known. Two
streams of thought serve as the foundations: utility theory and the perfect predictions, but in so doing we may be eliminatinig
inductive use of probability theory. the most important features of the problem.' We should
The intent of this paper is to provide a tutorial introduction to this like to include in the analysis not just the predictions
increasingly important area of systems science. The foundations are themselves, but also a measure of the confidence we have
developed on an axiomatic basis, and a simple example, the "anni- in these predictions. A formal theory of decision making
versary problem," is used to illustrate decision theory. The concept
of the value of information is developed and demonstrated. At times must take uncertainty as its departure point and regard
mathematical rigor has been subordinated to provide a clear and precise knowledge of outcomes as a limiting special case.
readily accessible exposition of the fundamental assumptions and Before we begin our exposition, we will clarify our point
concepts of decision theory. A sampling of the many elegant and of view. We shall take the enginieering rather than the
rigorous treatments of decision theory is provided among the purely sc ientific viewpoint. We are not observing the way
references.
people make decisions; rather we are participanits in the
decision-making process. Our concerin is in actually making
a decision, i.e., making a choice between alternative ways
INTRODUCTION
of allocating resources. We must assume that at least two
THE NECESSITY of makinig decisions in the face of distinct alternatives exist (or else there is 11o element of
uncertainty is an integral part of our lives. We must choice and, consequently, no problem). Alternatives are
act without knowing the consequeinces that will result distinct only if they result in different (uncertain) rewards
from the action. This uncomfortable situation is partic- or penalties for the decision maker; once the decisioni has
ularly acute for the systems engineer or manager who must been made and the uncertainty resolved, the resource
make far-reaching decisions oIn complex issues in a rapidly allocation can be changed only by incurring some penalty.
changing technological environment. Uncertainty appears What can we expect of a general theory for decision
as the dominant consideration in many systems problems making under uncertainty? It should provide a framework
as well as in decisions that we face in our personal lives. in which all available information is used to deduce which
To deal with these problems oii a rational basis, we must of the decision alternatives is "best" according to the
develop a theoretical structure for decisiotn making that decision maker's preferences. But choosing an alternative
includes uncertainity. that is consistent with these preferences and present
Confronting uncertainty is Inot easy-. We naturally try knowledge does not guarantee that we will choose the
to avoid it; sometimes we even pretend it does not exist. alternative that by hindsight turns out to be most profit-
Our primitive ancestors sought to avoid it by consulting able.
soothsayers and oracles who would "reveal" the uncertain We might distiniguish between a good decision and a
future. The methods have changed: astrology and the good outcome. We are all familiar with situations in which
reading of sheep entrails are somewhat out of fashion to- careful management and extensive planning produced
day, but predictions of the future still abound. Much poor results, while a disorganized and badly managed
current scientific effort goes into forecasting future eco- competitor achieved spectacular success. As an extreme
nomic and technological developments. If these predictions example, place yourself in the position of the company
are assumed to be completely accurate, the uncertainty in president who has discovered that a valuable and trusted
many systems decisions is eliminated. The outcome result- subordinate whose past judgment had proved unfailingly,
ing from a possible course of action may then be presumed accurate actually based his decisions upon the advice of a
to be known. Decision making becomes an optimization gypsy fortune teller. Would you promote this man or
problem, and techniques such as mathematical program- fire him? The answer, of course, is to fire him and hire the
ming may be used to obtain a solution. Such problems gypsy as a consultant. The availability of such a clair-
may be quite difficult to solve, but this difficulty should voyant to provide perfect information would make deci--
sion theory unnecessary. But we should not confuse the
Manuscript received M\ay 8, 1968. An earlier version of this paper two. Decision theory is not a substitute for the fortune
was presented at the IEEE Systems Science and Cybernetics Con- teller. It is rather a procedure that takes account of all
ference, Washington, D.C., October 17, 1966. This research was sup-
ported in part by the Graduate Cooperative Fellowship Program of available information to give us the best possible logical
the National Science Foundation at Stanford University, Stanford,
Calif.
The author is with the Systems Sciences Area, Stanford Research I For further discussion of
this point, see Howard [10] and Klein
Institute, MTenlo Park, Calif. 94025. and Meckling [141.
NORTH: INTRODUCTION TO DECISION THEORY
201

decision. It will minimize the consequences of getting an

unfavorable outcome, but we cannot expect our theory to
POSSIBLE OUTCOMES shield us from all "bad luck." The best protection we have
IT IS YOUR
ANNIVERSARY
IT IS NOT YOUR
ANNIVERSARY
against a bad outcome is a good decision.
DECISION
ALTERNATIVES Decision theory may be regarded as a formalization of
common sense. Mathematics provides an unambiguous
language in which a decision problem may be represented.
BUY FLOWERS
There are two dimensions to this representation that will
presently be described: value, by means of utility theory,
and information, by means of probability theory. In this
representation, the large and complex problems of systems
analysis become conceptually equivalent to simple prob-
lems in our daily life that we solve by "common sense."
We will use such a problem as an example.
DO NOT You are driving home from work in the evening when
BUY FLOWERS you suddenly recall that your wedding anniversary comes
about this time of year. In fact, it seems quite probable
(but not certain) that it is today. You can still stop by the
florist shop and buy a dozen roses for your wife, or you
may go home empty-handed and hope the anniversary
Fig. 1. Anniversary problem payoff matrix. date lies somewhere in the future (Fig. 1). If you buy the
roses and it is your anniversary, your wife is pleased at
what a thoughtful husband you are and your household is
the very epitome of domestic bliss. But if it is not your
anniversary, you are poorer by the price of the roses and
your wife may wonder whether you are trying to make
amends for some transgressioil she does not know about. If
you do not buy the roses, you will be in the clear if it is not
your anniversary; but if it is, you may expect a temper
tantrum from your wife and a two-week sentence to thedog-
house. What do you do?
We shall develop the general tools for solving decision
problems and then return to this simple example. The
reader might consider how he would solve this problem by
BUY FLOWERS
''common sense" and then compare his reasoning with the
formal solution which we shall develop later (Fig. 2).
THE MACHINERY OF DECISION MAKING
//j Utility Theory
The first stage in setting up a structure for decision
making is to assign numerical values to the possible out-
DO NOT
comes. This task falls within the area covered by the
BUY FLOWERS modern theory of utility. There are a number of ways of
developing the subject; the path we shall follow is that of
Luce and Raiffa [16 ].2
The first and perhaps the biggest assumption to be
made is that any two possible outcomes resulting from a
X DECISION POINT
decision can be compared. Given any two possible out-
comes or prizes, you can say which you prefer. In some
RESOLUTION OF
o
UNCERTAINTY cases you might say that they were equally desirable or
undesirable, and therefore you are indifferent. For ex-
ample, you might prefer a week's vacation in Florida to a
season ticket to the symphony. The point is not that the
STATUS QUO vacation costs more than the symphony tickets, but rather
Fig. 2. Diagram of anniversary decision.
2 The classical reference on modern utility theory is von Neumann
and Morgenstern [22]. A recent survey of the literature on utility
theory has been made by Fishburn [5].
2029 IEEE TR \NSA\CTIONS ON SYST'EMS SCIENCE AND CYBERNETICS, SEPTEMBER 196S

that you prefer the vacationi. If you were offered the vaca-
tion or the symphony tickets on a nonnegotiable basis,
you would choose the vacation.
A reasonable extension of the existence of y-our prefer-
ence among outcomes is that the preferenice be transitive;
if you prefer A to B and B to C, then it follo(ws that Xyou
prefer A to C.3
pKI
The second assumption, originated by von NXeumann
arid Morgenstern [22], forms the core of modern utility-
theory: you can assign preferences in the same mariner to
lotteries involving prizes as you can to the prizes them-
selves. Let us define what we meani by a lottery. Imaginie a Fig. 3. A lottery.
poitnter that spins in the center of a circle divided inito
two regions, as showni in Fig. 3. If you spin the pointer and A
it lands in region I, you get prize A; if it lands in region P
II, you get prize B. We shall assume that the poiniter is
spun in such a way that, when it stops, it is equally likely
to be pointing in any given direction. The fraction of the I -P
circumference of the circle in region I will be denoted P, B
and that in region II as 1 - P. Then from the assumption Fig. 4. Lottery diagram.
that all directionis are equally likelv, the probability that
the lottery gives you prize A is P, and the probability that
you get prize B is 1 - P. We shall denote such a lottery as
(P,A;1 - P,B) and represent it, by- F'ig. 4.
Now suppose you are asked to state your preferences for
prize A, prize B, and a lottery of the above type. Let us
assume that y-ou prefer prize A to prize B. Then it would
seem natural for you to prefer prize A to the lottery, A
A
(P,A;1 - P,B), between prize A. arid prize B, arid to P
C(I- P)' B
prefer this lottery betweein prize A arid prize B to prize B IP B
for all probabilities P between 0 and 1. You would rather (I-P) (I-P)
pIP
have the preferred prize A than the lottery, and you would C C
rather have the lottery than the inferior prize B. Further- COMPOUND LOTTERY EQUIVALENT SIMPLE LOTTERY
more, it seems natural that, given a choice betweein two Fig. ,. "No fun in gamblitig."
lotteries involving prizes A and B, you would choose the
lottery with the higher probability of getting the preferred
prize A, i.e., you prefer lottery (P,A;1 - P,B) to (P',A; different, prizes and perhaps a differenit divisioin of the
1-P',B) if anid only if P is greater thai P'. circle (Fig. 5). If you spini the second pointer you will
The final assumptions for a theory of utility are not receive prize B or prize C, depending on where this pointer
quite so natural and have been the subject, of much dis- lands. The assumption is that subdividing region II into
cussion. Nonetheless, they seem to be the most reasonable two parts whose proportions correspond to the proba-
basis for logical decision makiing. The third assumption is bilities P' arid 1 - P' of the second lottery creates ain
that there is no intrinsic reward in lotteries, that is, "nio equivalent simple lottery in which all of the prizes are
fun in gambling." Let us consider a compound lottery, outcomes. According to this third assumptioii, you can
a lottery in which at least, one of the prizes is riot an out- decompose a compound lottery by multiplying the proba-
come but another lottery among outcomes. For example, bilit) of the lottery prize in the first lottery by the proba-
eonisider the lottery (P,A;1 - P,(P',B;1 - P',C)). If the bilities of the inidividual prizes in the second lottery; you
pointer of Fig. 3 lands in region I, you get prize A; if it should be indifferenit betweein (P,A;1 -IP,(P',B;1 - ',
lands in region II, you receive another lottery that has C)) anid (P,A;P' - PP',B;1 - P- P' + PP',C). In
other words, your preferences are riot affected by the way-
in which the uncertainty is resolved bit by bit, or all at
I
Suppose not: you would be at least as happy with C as with A. once. There is no value in the lotterv itself; it does riot
Then if a little man in a shabby overcoat came up and offered youl matter whether you spin the pointer oince or twice.
C instead of A, you would presumably accept. Now you have C;
and since you prefer B to C, you would presumably pay a sum of Fourth, we make a continuity assumption. Considei
money to get B instead. Once you had B, you prefer A; so you would three prizes, A, B, arid C. You prefer A to C, and C to B
pay the man in the shabby overcoat some more money to get A.
Butt now you are back where you started, with A, and the little mani (and, as we have pointed out, you will therefore prefer A
in the shabby overcoat walks away counting your money. Given that to B). We shall assert that there must, exist some proba-
youi accept a standard of value such as money, transitivity prevents
youi from becoming a "money ptump." bility P so that yrou are indifferetnt, to receiving prize C or
NORTH: INTRODUCTION TO DECISION THEORY 203

the lottery (P,A;1 - P,B) between A and B. C is called 3) (P,A;1 - P,(P',B;1 - P',C)) (P,A;P' - PP',B;
-

the certain equivalent of the lottery (P,A;1 - P,B), 1 - P - P' + PP',C), i.e., there is "nIo fun in gambling."
an11d on the strength of our "no fun in gambling" assump- 4) If A > C > B, there exists a P with 0 < P < 1 so that
tioit, we assume that interchangiing C and the lottery
(P,A;1 - P,B) as prizes in some compound lottery does C,(P,A;I1- P,B)
iiot change your evaluation of the latter lottery. We have i.e., it makes no difference to the decision maker whether C
not assumed that, given a lottery (P,A;1 - P,B), there or the lottery (P,A;1 - P,B) is offered to him as a prize.
exists a Prize C intermediate in value between A and B so Under these assumptions, there is a concise mathe-
that you are indifferent between C and (P,A;1 - P,B). matical representation possible for preferences: a utility
Inistead we have assumed the existence of the probability function u( ) that assigns a number to each lottery or
P. Given prize A preferred to prize C preferred to prize B, prize. This utility function has the following properties:
for some P between 0 and 1, there exists a lottery (P,A;
1 - P,B) such that you are indifferenit between this lottery u(A) > u(B) if and only if A > B (1)
and Prize C. Let us regard the circle in Fig. 3 as a "pie"
to be cut into two pieces, region I (obtain prize A) and if C (P,A;1 -P,B),
region II (obtain prize B). The assumption is that the then u(C) = P u(A) + (1 - P) u(B) (2)
"pie" can be divided so that you are indifferent as to i.e., the utility of a lottery is the mathematical expectation
whether you receive the lottery or intermediate prize C. of the utility of the prizes. It is this "expected value"
Is this continuity assumption reasonable? Take the property that makes a utility function useful because it
following extreme case: allows complicated lotteries to be evaluated quite easily.
A = receive $1; It is important to realize that all the utility function does
B = death; is provide a means of consistently describing the decision
C = receive nothing (status quo). maker's preferences through a scale of real numbers,
providing these preferences are consistent with the previ-
It seems obvious that most of us would agree A is pre- ously mentioned assumptions 1) through 4). The utility
ferred to C, and C is preferred to B; but is there a proba-function is no more than a means to logical deduction
bilitv P such that we would risk death for the possibility of
based on given preferences. The preferences come first and
gaining $1? Recall that the probability P can be arbi- the utility function is only a convenieint means of describ-
trarily close to 0 or 1. Obviously, we would not engage in ing them. We can apply the utility conicept to almost any
such a lottery with, say, P = 0.9, i.e., a 1-in-10 chance of
sort of prizes or outcomes, from battlefield casualties or
death. But suppose P = 1 - 1 X 10-50, i.e., the proba- achievements in space to preferences for Wheaties or
bility of death as opposed to $1 is not 0.1 but 10 -50. ThePost Toasties. All that is necessary is that the decision
latter is considerably less than the probability of being maker have unambiguous preferences and be willing to
stru(ck onI the head by a meteor in the course of going outaccept the basic assumptions.
to pick up a $1 bill that someone has dropped on your door- In many practical situations, however, outcomes are in
step. MAost of us would not hesitate to pick up the bill. terms of dollars and cents. What does the utility concept
Even in this extreme case where death is a prize, we con- mean here? For an example, let us suppose you were
clude the assumption is reasonable. offered the following lottery: a coin will be flipped, and if
We can summarize the assumptions we have made into you guess the outcome correctly, you gain $100. If you
the followiing axioms. guess incorrectly, you get nothing. We shall assume you
A, B, C are prizes or outcomes resulting from a decision.
feel that the coin has an equal probability of coming up
Notation: heads or tails; it corresponds to the "lottery" which we
have defined in terms of a pointer with P = 1/2. How
means "is preferred to;" much would you pay for such a lottery? A common answer
A > B means A is preferred to B; to this academic question is "up to $50," the average or-
means "is indifferent to;" expected value of the outcomes. When real money is in-
A B means the decision maker is indifferent be- volved, however, the same people tend to bid considerably
tween A and B. lower; the average bid is about $20.4 A group of Stanford
Utility Axioms: University graduate students was actually confronted with
a $100 pile of bills and a 1964 silver quarter to flip. The
1) Preferences can be established between prizes and average of the sealed bids for this game was slightly under
lotteries in an unambiguous fashion. These preferences are $20, and only 4 out of 46 ventured to bid as high as $40.
transitive, i.e., (The high bidder, at $45.61, lost and the proceeds were
A>B, B>C impliesA>C used for a class party.) These results are quite typical;
A B, B-C impliesA-C. in fact, professional engineers and maniagers are, if anv-
2) If A > B, then (P,A;1 - P,B) > (P',A;1 - P',B) if 4Based on unpublished data obtained by Prof. R. A. Howard of
and only if P > P'. Stanford University, Stanford, Calif.
204 IEEE TRANSACTIONS ON SYSTEMS SCIENCE AND CYBERNETICS, SEPTEMBER 1968

thing, more conservative in their bids than the less 1.00

affluent students.
The lesson to be learned here is that, by and large, most
people seem to be averse to risk in gambles involving what
is to them substantial loss. They are willing to equate the 0.75
value of a lottery to a sure payoff or certain equivalent
substantially less than the expected value of the outcomes.
Similarly, most of us are willing to pay more than the -' 0.50
expected loss to get out of an unfavorable lottery. This
fact forms the basis of the insurance industry.
If you are very wealthy and you are confronted with a
small lottery, you might well be indifferent to the risk. 0.25
An unfavorable outcome would not deplete your resources,
and you might reason that you will make up your losses in
future lotteries; the "law of averages" will come to your
rescue. You then evaluate the lottery at the expected value 0
of the prizes. For example, the (1/2, $0; 1/2, $100) lottery 0 20 40 60 80 100
would be worth 1/2($0) + 1/2($100) = $50 to you. Your MONEY- dollars

utility function is then a straight line, and we say you are Fig. 6. Utility cuirve for money: $0 to $100.
an "expected value" decision maker. For lotteries involv-
ing small prizes, most individuals and corporations are Even without such a monletary represeintationi, we caii
expected value decision makers. We might regard this as a always construct a utility function oin a fiinite set of out-
consequeiice to the fact that any arbitrary utility curve comes by using the expected value property (2). Let us
for money looks like a straight line if we look at a small choose two outcomes, one of which is preferred to the
enough section of it. Only when the prizes are substantial other. If we set the utilities arbitrarily at 1 for the preferred
in relation to our resources does the curvature become outcome and 0 for the other, we can use the expected value
evident. Then an unfavorable outcome really hurts. For property (2) of the utility function to determine the
these lotteries most of us become quite risk averse, and utility of the other prizes. This procedure will always wor-k
expected value decision making does niot accurately reflect so long as our preferences obey the axioms, but it may be
our true preferences. unwieldy in practice because we are asking the decision
Let us now describe one way you might construct your maker to assess simultaneously his values in the absenice of
owin utility curve for money, say, in the amounts of $0 to uncertainty and his preference among risks. The value of
$100, in addition to your present assets. The utility func- some outcome is accessible onily by reference to a lottery
tion is arbitrary as to choice of zero point and of scale involving the two "reference" outcoomes. For example,
factor; changing these factors does not lead to a change in the reference outcomes in the aniniversary problem might
the evaluation of lotteries using properties (1) and (2). be "domestic bliss" = 1 and "doghouse" = 0. We could
Therefore, we can take the utility of $0 as 0 and the utility then determine the utility of "status quo" as 0.91 since the
of $100 as 1. Now determine the minimum amount you husband is indifferent between the outcome "status quo"
would accept in place of the lottery of flipping a coin to and a lottery in which the chances are 10 to 1 of "domestic
determine whether you receive $0 or $100. Let us say your bliss" as opposed to the "doghouse." Similarly, we might
answer is $27. Now determine the certain equivalent of discover that a utility of 0.667 should be assigned to "sus-
the lotteries (1/2, $0; 1/2, $27), and (1/2, $27; 1/2, $100), picious wife and $6 wasted on roses," siince our friend is ilndif-
and so forth. We might arrive at a curve like that shown ferent between this evenituality and a lottery in which the
in Fig. 6. probabilities are 0.333 of "doghouse" and 0.667 of "do-
We have simply used the expected value property (2) to mestic bliss." Of course, to be consistenit with the axioms,
construct a utility curve. This same curve, however, our friend must be indifferent between "suspicious wife,
allows us to use the same expected utility theorem to etc. ," and a 0.73 probability of "status quo" and a 0.27
evaluate new lotteries; for example, (1/2, $30; 1/2 $80). probability of "doghouse." If the example included
From Fig. 6, u($30) = 0.54, u($80) = 0.91, and therefore additional outcomes as well, he might find it quite difficult
1/2 u($30) + 1/2 u($80) = u(x) > x = $49. If you are to express his preferences among the lotteries in a maiinner
going to be consistent with the preferences you expressed consistent with the axioms. It may be advisable to proceed
in developing the utility curve, you will be indifferent in two stages; first, a numerical determination of value ill a
between $49 and this lottery. -Moreover, this amouint risk-free situation, and then an adjustment to this scale
could have been determined from your utility curve by a to include preference toward risk.
subordinate or perhaps a computer program. You could Equivalent to our first assumption, the existence of
send your agent to make decisions on lotteries by using transitive preferences, is the existence of some scale of
your utility curve, and he would make them to reflect your value by which outcomes may be rankied; A is preferred to
preference for amounts in the range $0 to $100. B if and only if A is higher in value than B. The numericCal
NORTH: INTRODUCTION TO DECISION THEORY 205

structure we give to this value is not important since a morning or it will rain this afternoon," should have the
monotonic transformation to a new scale preserves the same value as "It will rain today."
ranking of outcomes that corresponds to the original These assumptions are equivalent to the assertion
preferences. No matter what scale of value we use, we can that there is a function P that gives values between 0 and 1
construct a utility function on it by usiIng the expected to events ("the statement is true" is an event) and that
value property (2), so long as our four assumptions hold. We obeys the following probability axioms.5
may as well use a standard of value that is reasonably Let E and F be events or outcomes that could result
intuitive, and in most situations money is a convenient from a decision:
standard of economic value. We can then find a monetary
equivalenit for each outcome by determining the point at 1) P(E) > 0 for any event E;
which the decision maker is indifferent between receiving 2) P(E) = 1, if E is certain to occur;
the outcome and receiving (or paying out) this amount of 3) P(E or F) = P(E) + P(F) if E and F are mutually
moliey. In addition to conceptual simplicity, this pro- exclusive events (i.e., only one of them can occur).
cedure makes it easy to evaluate new outcomes by pro- E or F means the event that either E or F occurs. We
viding an intuitive scale of values. Such a scale will be- are in luck. Our axioms are identical to the axioms that
come necessary later on if we are to consider the value of form the modern basis of the theory of probability. Thus
resolving uncertainty. we may use the whole machinery of probability theory for
We will return to the anniversary decision and demon- inductive reasoning.
strate how this two-step value determination procedure Where do we obtain the values P(E) that we will
may be applied. But first let us describe how we shall assign to the uncertainty of the event E? We get them from
quantify uncertainty. our own minds. They reflect our best judgment on the
basis of all the information that is presently available to us.
The Inductive Use of Probability Theory The use of probability theory as a tool of inductive reason-
We now wish to leave the problem of the evaluation of ing goes back to the beginnings of probability theory.
outcomes resulting from a decision and turn our attention In Napoleon's time, Laplace wrote the following as a part
to a means of encoding the information we have as to of his introduction to A Philosophical Essay on Proba-
which outcome is likely to occur. Let us look at the limiting bilities ([15], p. 1):
case where a decision results in a certain outcome. We
might represent an outcome, or an event, which is certain Strictly speaking it may even be said that nearly all our
to occur by 1, and an event which cannot occur by 0. knowledge is problematical; and in the small numbers of
A certain event, together with another certain event, is things which we are able to know with certainty, even in
the mathematical sciences themselves, the principal means
certain to occur; but a certain event, together with an for ascertaining truth-induction and analogy-are them-
impossible event, is certain not to occur. Mlost engineers selves based on probabilities ....
would recognize the aforementioned as simple Boolean
equations: 1 1 = 1, 1 .0 = 0. Boolean algebra allows us Unfortunately, in the years following Laplace, his writ-
to make complex calculations with statements that may ings were misinterpreted and fell into disfavor. A definition
take on only the logical values "true" and "false." The of probability based on frequency came into vogue, and the
whole field of digital computers is, of course, based on this pendulum is only now beginning to swing back. A great
branch of mathematics. many modern probabilists look on the probability assigned
But how do we handle the logical "maybe?" Take the to an event as the limiting fraction of the number of times
statement, "It will rain this afternoon." We cannot now an event occurred in a large number of independent
assign this statement a logical value of true or false, but repeated trials. We shall not enter into a discussion of the
we certainly have some feelings on the matter, and we general merits of this viewpoint on probability theory.
may even have to make a decision based on the truth of Suffice it to say that the situation is a rare one in which
the statement, such as whether to go to the beach. Ideally, you can observe a great many independent identical trials
we would like to generalize the inductive logic of Boolean in order to assign a probability. In fact, in decision theory
algebra to include uncertainty. We would like to be able to we are often interested in events that will occur just once,
assign to a statement or an event a value that is a measure For us, a probability assessment is made on the basis of a
of its uncertainty. This value would lie in the range from 0 state of mind; it is not a property of physical objects to
to 1. A value of 1 indicates that the statement is true or be measured like length, weight, or temperature. When we
that the event is certain to occur; a value of 0 indicates assign the probability of 0.5 to a coin coming up heads, or
that the statement is false or that the event cannot occur. equal probabilities to all possible orientations of a pointer,
WVe might add two obvious assumptions. We want the we may be reasoning on the basis of the symmetry of the
value assignments to be unambiguous, and we want the
value assignments to be independent of any assumptions I Axioms 1) and 2) are obvious, and 3) results from the assumption
that have not been explicitly introduced. In particular, the of invariance to the form of data presentation (the last sentence in
value of the statement should depend on its content, not the preceding paragraph). Formal developments may be found in
Cox [3], Jaynes [121, or Jeffreys [131. A joint axiomatization of both
on the way it is presented. For example, "It will rain this probability and utility theory has been developed by Savage [20].
206 IEEE TRANSACTIONS ON SYSTEMS SCIENCE AND CYBERNETICS, SEPTEMBER 1968

physical object. There is nO reason to suppose that one value, or the average of the random variable over its
side of the coin\iill be favored over the other. But the probability distributioni, is
physical symmetry of the coin does not lead immediately
to a probability assiginmenit of 0.5 for heads. For example,
consider a coin that is placed on a drum head. The drum (xjS) = X{.xs}. (4)
head is struck, atnd the coin bounces into the air. Will it
land heads up half of the time? We might expect that the One special state of iniformation will be used over anid
probability of heads would depend on which side of the over again, so we shall need a special name for it. This is
coin was up iniitially, how hard the drum was hit, anid so the iinformation that we niow possess on the basis of our
forth. The probability of heads is not a physical parameter prior knowledge and experience, before we have done aIny
of the coin; we have to specify the flipping system as wvell. special experimenting or sampling to reduce our unicer-
But if we kiiew exactly how the coin were to be flipped, we tainty. The probability distribution that we assigni to
could calculate from the laws of mechanics whether it values of an uncertaii (quantity on the basis of this prior
would land heads or tails. Probability enters as a means of state of information (denoted g) will be referred to as the
describiing our feelings about the likelihood of heads when "prior distribution" or simply the "prior."
our kinowledge of the flipping system is not exact. We must Nowv let us consider a problem. Most of us take as
conclude that the probability assignment depenids on our axiomatic the assignment of 0.5 to the probability of heads
present state of kniowledge. on the flip of a coin. Suppose we flip thumbtacks. If the
The most importanit consequence of this assertion is that thumbtack lands with the head up and poinit dowil, we
probabilities are subject to change as our iniformation shall deniote the outcome of the flip as "heads." If it laIlds
improves. In fact, it even makes sense to talk about with the head down and the point up, we shall denote the
probabilities of probabilities. A few years ago we might outcome as "tails." The question which we must answer
have assigned the value 0.5 to the probability that the is, "What is p, the probability of heads in flipping a
surface of the moon is covered by a thick layer of dust. thumbtack?" We will assume that both thumbtack anid
At the time, we might have said, "We are 90 percent means of flipping are sufficiently standardized so that wve
certain that our probability assignment after the first may expect that all flips are independent and have the
successful Surveyor probe will be less than 0.01 or greater same probability for coming up heads. (Formally, the
than 0.99. We expect that our uncertainty about the com- flips are Bernoulli trials.) Then the long-run fractioin of
positioIn of the mooin's surface will be largely resolved." heads may be expected to approach p, a well-definied
Let us conclude our discussion of probability theory number that at the moment we do not know.
with an example that will introduce the means by which Let us assign a probability distribution to this uncertain
probability distributions are modified to include new in- parameter p. We are all familiar with thumbtacks; we have
formation: Bayes' rule. We shall also introduce a useful no doubt dropped a few on the floor. Perhaps we have some
notationi. We have stressed that all of our probability experience with spilled carpet tacks, or coin flipping, or
assigniments are going to reflect a state of information in the physics of falling bodies that we believe is relevanit.
the mind of the decision maker, and our notation shall We want to encode all of this prior information inito the
indicate this state of information explicitly. form of a probability distribution on p.
Let A be an evenit, and let x be a quantity about which This task is accomplished by using the cumulative dis-
we are uncertain; e.g., x is a random variable. The values tribution functioin, {p < polg}, the probability that the
that x may assume may be discrete (i.e., heads or tails) parameter p will be less than or equal to some specific
or continuous (i.e., the time an electronic component will value of the parameter po. It may be convenienit to use
ruin before it fails). We shall denote by {A|S} the proba- the complementary cumulative
bility assigned to the event A on the basis of a state of
information S, and by {xjS} the probability that the lp > polg} = 1 - lp < Pol8}
random variable assumes the value x, i.e., the probability and ask questions such as, "What is the probability that p
mass function for a discrete random variable or the proba- is greater than Po = 0.5?"
bility density fuinetion for a continuous random variable, To make the situation easier to visualize, let us introduce
given a state of informatioin S. If there is confusion be- Sam, the neighborhood bookie. We shall suppose that we
tween the random variable and its value, we shall write are forced to do business with Sam. For some value Po
{x = x01S}, where x denotes the random variable and x0 between 0 and 1, Sam offers us two packages:
the value. We shall assume the random variable takes on Package 1: If measurement of the long runi fractioin of
some value, so the probabilities must sum to 1: heads p shows that the quantity is less than or equal to po,
then Sam pays us $1. If p > po, then we pay Sam $1.
f {xIS} = 1. (3) Package 2: We divide a circle into two regions (as showni
in Fig. 3). Region I is defined by a fraction P of the circum-
f is a generalized summation operator represeniting ference of the circle, and the remainder of the circle con-
summation over all discrete values or integration over all stitutes region II. Now a pointer is spun in such a way
contiinuous values of the random variable. The expected that when it stops, it is equally likely to be pointing in anyv
NORTH: INTRODUCTION TO DECISION THEORY 207

given direction. If the pointer stops in region I, Sam pays 1.0 T 1.01-

us $1; if it lands in region II, we pay Sam $1. z\ :VALUES OBTAINED THROUGH
SAM'S INTERROGATION
Sam lets us choose the fraction P in Package 2, but then
he chooses which package we are to receive. Depending on
the value of po, these packages may be more or less attrac- ti3
tive to us, but it is the relative rather than the absolute 8P 0.5
VI
value of the two packages that is of interest. If we set P 00
tco be large, we might expect that Sam will choose package 1<F PROBABILITY
THAT LONG-RUN
1, whereas if P is small enough, Sam will certainly choose 0o
00 FRACTION OF HEADS_
FOR THUMBTACK
package 2. Sam wishes (just as we do) to have the package < Po L
with the higher probability of winning $1. (Recall this is
our secoind utility axiom.) We shall assume Sam has the 0 0.5 1.0
same information about thumbtacks that we do, so his pO
probability assiginments will be the same as ours. The Fig. 7. Cumulative distribution functioni for thumbtack flipping.
assumptioni [utility axiom 4) ] is that givein po, we can find a
P such that Packages 1 and 2 represeint equivalent lot- 2.0
teries, so P = {p < pol8}F6 The approach is similar to the
well-known method of dividinig an extra dessert between
two small boys: let one divide and the other choose. The
first is motivated to make the division as even as possible - 1.0K I
so that he will be indifferent as to which half he receives.
Suppose Sam starts at a value po = 0.5. We might
reason that since nails always fall oIn the side (heads), and a 0 -
II II

thumbtack is intermediate betweeni a coin and a nail 0 0.5 1.0

PARAMETER p= LONG-RUN FRACTION of HEADS for THUMB TACK
heads is the more likely orientation; but we are not too
sure; we have seen a lot of thumbtacks come up tails. Fig. 8. Prior probability density function.
After some thought, we decide that we are indifferent about
which package we get if the fraetion P is 0.3, so {p < i.e., as the product of the probability we assign to the
0.518} = 0.30. experimental outcome E times the probability we would
Sam takes other values besides 0.5, skipping around in a assign to the value of p after we knew the experimental
random fashion, i.e., 0.3, 0.9, 0.1, 0.45, 0.8, 0.6, etc. The outcome E in addition to our prior information; or
curve that results from the initerrogation might look like
that shown in Fig. 7. By his method of randomly skipping {p,EI&} =
{Ejp,j} {pjg} (6)
around, Sam has eliminated aniy bias in our true feelings i.e., the product of the probability of that experimental
that resulted from an uncoinscious desire to give answers outcome if we knew that p were the probability of getting
consistent with previous points. In this fashion, Sam has heads times our prior probability assessment that p acttu-
helped us to establish our prior distribution on the param- ally takes on that value.
eter p. We may derive a probability density function by We assumed that probabilities were unambiguous, so
takiing the derivative of the cumulative distribution func- we equate these two expressions. Providing {E18} 0,
tion (Fig. 8): {pJ&} = (dldpo) {p < pol}. i.e., the experimental outcome is not impossible, we obtaiin
Now supposing we are allowed to flip the thumbtack the posterior (after the experiment) probability distribu-
20 times and we obtain 5 heads aind 15 tails. How do we tion on p
take account of this new data iii assigning a proba-
bility distribution based on the inew state of information,
which we denote as 8, E: our prior experience g plus E, the
lplEII J{Elp,6}&j~{p8}
E (7)
20-flip experiment? We will use oine of the oldest (1763)
results of probability theory, Bayes' rule. Consider the This expressioni is the well-known Bayes' rule.
prior probability that p will take on a specific value and {E g} is the "pre-posterior" probability of the outcome
the 20-flip experiment E will have a certain specific out- E. It does not depend on p, so it becomes a normalizing
come (for example, p = 0.43; E = 5 heads, 15 tails). Now factor for the posterior probability distribution. { Ejp,6} is
we can write this joint probability in two ways: the probability of the outcome E if we knew the value p
for the probability of heads. This probability is a function
{p,E18} = {pJE,&} {EZ8} (5) of p, usually referred to as the "likelihood function." We
notice since p must take on some value, the expectation of
the likelihood function over the values of p gives the pre-
6 We have equated the subjective probability that summarized our posterior probability of the experimental outcome:
information about thumbtacks to the more intuitive notion of
probability based on symmetry (in Package 2). Such a two-step
approach to probability theory has been discussed theoretically by
Anscombe and Aumann [1].
{ El&4 -= { p ,& { P (8&)
208 IEEE TRANSACTIONS ON SYSTEMS SCIENCE AND CYBERNETICS, SEPTEMBER 1968

ing his anniversary: the odds are 4 to 1 that it is not his

0.2 anniversary. Now let us look at the two lotteries that
represent his decision alternatives. If he buys the flowers,
he has a 0.2 probability of "domestic bliss" and an 0.8
0.1 probability of "suspicious wife." The expected utility of
the lottery is 0.2(1.0) + 0.8(0.667) = 0.734 = u($50).
I II
On the other hand, if he does not buy the flowers, he has ail
0
..j

0 0.5 I.0 0.8 chance of "status quo" anid a 0.2 chance of "doghouse."
PARAMETER p The expected utility of this alternative is 0.8(0.91) +
Fig. 9. Likelihood function for 5 heads in 20 trials. 0.2(0) = 0.728 = u($49). The first alternative has a
slightly higher value to him so he should buy the flowers.
5.0
On the basis of his values, his risk preference, and his
judgment about the uicertainty, buyiing the flowers is his
4.0 best alternative. If he were an expected value decision
U)
3.00 maker, the first lottery would be worth 0.2($100) +
2.0
0.8($42) = $53.60 and the second 0.2(0) + 0.S($80) =
$64. In this case he should not buy the flowers.
1.0 The foregoing example is, of course, very trivial, but
.0 conceptually aniy decision problem is exactly the same.
0 0.5
PARAMETER p
I.0
There is only one additioinal feature that we may typically
Fig. 10. Posterior probability density functioin.
expect: in general, decision problems mav involve a
sequence of decisiolis. First, a decision is made and theni ain
uncertain outcome is observed; after which another de-
For the specific case we are treating, the likelihood func- cision is made, and an outcome observed, etc. For example,
tion is the familiar result from elementary probability the decision to develop a new product might go as follows.
theory for successes in n Bernoulli trials xhen the
r A decision is made as to whether or not a product should
probability of a success is p: be developed. If the decision is affirmative, an uncertaill
research and development cost will be inicurred. At this
{ElpI8} = !(n- pr)! r(l - p)n-r. (9) point, a decision is made as to whether to go inito produc-
tion. The production cost is uncertaini. After the produc-
This function is graphed for r = 5 heads in n = 20 trials in tion cost is known, a sale price is set. Finally, the uincertain
Fig. 9. M\Iultiplying it by the prior jpj&}
(Fig. 8) and
normalizing by dividing by {E[8} gives us the posterior
sales volume determines the profit or loss onI the product.
We can handle this problem in the same wax as the
distribution {pjE,&} (Fig. 10). In this way, Bayes' rule anniversary problem: assign values to the fiinal outcomes,
gives us a general means of revising our probability assess- and probabilities to the various uneertaini outcomes that
ments to take account of new information.7 will result from the adoption of a decisiotn alternative.
We can represent the problem as a decision tree (Fig. 11),
SOLUTION OF DECISION PROBLEMS and the solution is coniceptually easy. Start at the final
Now that we have the proper tools, utility theory and outcome, sales volume (the ends of the tree). Go in to the
probability theory, we return to the anniversary decision first decisioin, the sales price (the last to be made chrono-
problem. We ask the husband, our decision maker, to logically). Compute the utility of the decision alternatives,
assign monetary values to the four possible outcomes. and choose the one with the highest value. This value
He does so as follows: becomes the utility of the chaice outcome leadiing to that
decision (e.g., production cost). The correspondilng
Domestic bliss (flowers anniversary): $100 certain equivalent in dollars reflects the expected utility
Doghouse (no flowers, anniversary): $ 0 of reaching that point ill the tree. II1 this fashioi, xve xork
Status quo (no flowers, anniversary): $ So
no
backwards to the start of the tree, findiing the best decision
Suspicious wife (flowers, no anniversary): $ 42. alternatives and their values at each step.
(For example, he is indifferent between "status quo" aind MIany decision problems enicountered inl actual practiC e
"doghouse" provided in the latter case he receives $80.) are extremely complex, anid a decision tree approach may
His preference for risk is reflected by the utility function of not always be appropriate. If all quatntities concerned in
Fig. 6, and he decides that a probability assessment of 0.2 the problem were conisidered uncertain (with prior dis-
sums up his uncertainty about the possibility of today be- tributions), the problem might be computationially ill-
tractable. It is often advisable to solve the model de-
terministically as a first approximation. We approximate
For certain sampling processes having special statistical proper-
ties, assumption of a prior probability distribution from a particular all uncertain quantities with a single best estimate anid
family of functions leads to a simple form for Bayes' rule. An ex- then examine the decisioni; i.e., if research and develop-
tensive development of this idea of "conjugate distributions" has
been accomplished by Raiffa and Schlaifer [19]. ment costs, production costs, and sales volume took the
NORTH: INTRODUCTION TO DECISION THEORY 209

PRODUCTION
Ilow will the secretary's information change his assess-
COST ment of the probability that today is his anniversary?
If she says, "No, it is not your anniversary," he may be
sure that it is not; but if she says "Yes, it is," she could be
joking. We can compute the new assessment of the proba-
bility from Bayes' rule. This new probability is equal to the
probability 0.2 that she says yes and it really is his anni-
DO NOT
versary, divided by his prior estimate, 0.2 + 0.5 X 0.8
DEVELOP = 0.6, that she will say yes regardless of the date of his
PRODUCT
anniversary. Hence the probability assignment revised to
X<DECISION POINTS include the secretary's yes answer is 0.333.
UNCERTAIN OUTCOMES- What is the value of this new alternative to our friend?
"CHANCE POINTS"
If his secretary says no (probability 0.4), he may return
home empty-handed and be assured of "status quo." On
Fig. 11. Product development decision tree. the other hand, if she says yes (probability 0.6), he will
buy the flowers. In either case, he has incurred a cost of
values we consider most likely, would it then be advisable $10 which must be subtracted from the values of the out-
to develop the product? This deterministic phase will comes. Calling the secretary then has a utility of
usually give us some insight into the decision. AMoreover, 0.4 u($70) + 0.6 [0.333 u($90) + 0.667 u($32) ]
we can perform a sensitivity analvsis by varying quantities
= 0.344 + 0.416 = 0.760 = u($53.50).
that we believe are uncertain to determine how they
affect the decision. The decision may be quite insensitive Since this value of $53.50 exceeds the value of $50 for his
to some quantities, and these quantities may be treated as previous best alternative (buy flowers), our friend should
certain (uncertainty is neglected if it appears not to call his secretary. If the husband were an expected value
affect the decision). On the other hand, if a variation that decision maker, the alternative of calling the secretary
lies within the range of uncertainty of a factor causes a would have a value of
major shift in the decision (i.e., from "develop the prod-
uct" to "do not develop the product"), we shall certainly 0.4 ($70) + 0.6 [0.333 ($90) + 0.667 ($32)] = $58.80
wish to encode our feelings about the uncertainty of that which is less than the value of $64 for the "do not buy
quantity by a prior distribution.8 flowers" alternative; in this case our friend should not call
his secretary. It is evident that in this example preference
THE VALUE OF RESOLVING UNCERTAINTIES toward risk is very important in determining the decision
There is a class of alternatives usually available to the maker's best course of action.
decision maker that we have not yet mentioned: activities In the complex decision problems normally encountered
that allow him to gather more information to diminish the in practice, there are usually several alternative options
uncertainties before he makes the decision. We have al- available for diminishing the uncertainty associated with
ready seen how new information may be incorporated into the unklnown factors. In theory, the expected gain for each
probability assessments through Bayes' rule, and we noted type of sampling could be computed and compared with
that we can assign a probability distribution to the results the cost of sampling as we have just done in the simple
of the information gathering by means of the pre-posterior anniversary example. But these calculations can be quite
probability distribution. Typical information-gathering involved as a rule, and there may be a great many alterna-
activities might include market surveys, pilot studies, tive ways of gathering information. Often the relevant
prototype construction, test marketing, or consulting with questions are, first, "Should we sample at all?" and then,
experts. These activities invariably cost the decision maker "What kind of sampling is best for us?"
time and resources; he must pay a price for resolving It is often useful to look at the limiting case of complete
uncertainty. resolution of uncertainty, which we call perfect informa-
Let us return to the husband with the anniversary tion. We can imagine that a gypsy fortune teller who
problem. Suppose he has the option of calling his secretary. always makes correct predictions is, in fact, available to us.
If it is his anniversary, his secretary will certainly tell him. The value of perfect information is the amount that we are
But if it is not, she may decide to play a trick and tell him willing to pay her to tell us exactly what the uncertain
that today is his anniversary. He assigns probability 0.5 to quantity will turn out to be. Note that her answer may be
such practical joking. In anv event, the secretary will of little value to us-we may be planning to take the best
spread the word around the office and our friend will get decision alternative already. On the other hand, her perfect
some good natured heckling, which he views as having a information may be quite valuable; it may allow us to
value of minus $10. avoid an unfavorable outcome. We are going to have to pay
her before we hear her information; our payment will
8 The decision analysis procedture has beeii described in detail by
reflect what we expect the information to be on the basis
Howard [8]. of our prior probability assessment.
210 IEEE TRANS ACTIONS ON SYSTEMS SCIENCE AND CYBERNETICS, SEPTEMBER 1968

In the husbanid's anniiversary problem, perfect iniforma- ACKNOWLEDGMENT

tion might correspond to a secretary who is certain to tell The author is deeply indebted to Prof. R. A. Howard for
him if today is his anniversary. If he could act on this valuable guidance and advice. Much of the viewpoiint
information, he would buy flowers if it were his anini- presented in this paper was acquired as a result of two
versary and would not buy flowers otherwise. Since he years of assisting Prof. Howard in his course in Decision
feels that there is a 0.2 chance the secretary will tell him Analysis at Stanford University, Stanford, Calif.
that it is his anniversary, the expected utility of the out-
comes if he bases his decision on perfect informatiotn is REFERENCES
0.2 u($100 - b) + 0.8 u($80 - b) where b is the amount he [1] F. J. Anscombe and R. J. Aumann, "A definition of subjective
must pay to get the information. By setting this expression probability," Ann. Math. Statist., vol. 34, pp. 1909-205, 1963.
[2] T. Bayes, "An essay toward solving a problem in the doctrine
equal to 0.734, the expected utility of his best alternative of chances," Phil. Trans. Roy. Soc. (London), vol. 33, pp. 370-
based oIn prior information, we can solve for b = $33.50. 418, 1763.
The husband should consider for more detailed ainalysis [31 R. T. Cox, The Algebra of Probable Inference. Baltimore, Md.:
The John Hopkins Press, 1961.
only those opportunities for resolving his uncertainty that [4] P. C. Fishburn, Decision and Value Theory. New York: Wiley,
1964.
"cost" him $33.50 or less. If he were an expected value [5] , "Utility theory," ilManagement Sci., vol. 14, pp. 333- 378,
decision maker, perfect information would be of less value January 1968.
[6] , "On the prospects of a useful unified theory of value for
to him; he would be willing to pay a maximum of only $20 engineering," IEEE Trans. Systemns Science and Cybernetics,
for it. vol. SSC-2, pp. 27-35, August 1966.
[7] R. A. Howard, "Bayesian decision models for system engineer-
ing," IEEE Trans. Systems Science and Cybernetics, vol.
SUMMARY SSC-1, pp. 36-40, November 1965.
[8] a) , "Decision analysis: applied decision theory," Proc. 4th
Decision theory is a way of formalizing commoni sense. Internat'l Conf. Operational Res. (Boston, Mass., 1966).
b) -," The foundationbs of decision analysis" this issue, pp.
The decision maker analyzes the possible outcomes re- 211-219.
sulting from his available alternatives in two dimensions: [9] ec, "Information value theory," IEEE Trans. Systems Sci-
ence and Cybernetics, vol. SSC-2, pp. 22-26, August 1966.
value (by means of utility theory) and probability of [101 ,"The science of decision-making" (unpublished).
occurrence. He thent chooses the alternative that he expects [11] , "Value of information lotteries," IEEE Trans. Systems
Science and Cybernetics, vol. SSC-3, pp. 54-60, June 1967.
to have the highest value. He cannot guarantee that the [12] E. T. Jaynes, "Probability theory in science and engineering,"
outcome will be as good as he might hope for, but he has Field Research Laboratory, Socony Mobil Oil Co., Inic., Dallas,
Tex., 1939.
made the best decision he can, based on his preferernces and [13] H. Jeffreys, Theory of Probability, 3rd ed. Oxford: Clarendon
available knowledge. Inferenice using Bayes' rule allows Press, 1961.
[14] B. Klein and W. Meckling, "Applications of operations research
the decision maker to evaluate information gatheriing to development decision.s," Operations Res., vol. 6, pp. 332-363,
activities that will reduce his uncertainty. .May-June 1958.
[15] P. S. Laplace, Essai Philosophique sur les Probabilih's (Paris,
Decision theory gives no magical formulas for correct 1814), translation. New York: Dover, 1931.
decisions. In fact, it forces the decision maker to rely more [16] R. 1). Luce and H. Raiffa, Games and Decisions: Introduction an(d
Critical Survey. New York: Wiley, 1957.
strongly than ever oni his own preferences and judgments. [17] J. W. Pratt, H. Raiffa, arid R. Schlaifer, "The foundations of
But it does give him a logical framework in which to work, decision under uncertainty: an elementary exposition," J. Am.
Statist. Assoc., vol. 59, pp. 333-375, June 1964.
a framework that is adaptable in principle to all decisioin [18] , Introduction to Statistical Decision Theory. New York:
problems, from the simplest to the most complex. As MlcGraw-Hill, 1965.
[19] H. Raiff a and It. Schlaifer, Applied Statistical Decision Theory.
modern society continues to grow in size and complexity, Boston, Mass.: Graduate School of Business, Harvard Uni-
such a framework for decision making will become more versity, 1961.
L. J. Savage, The Foundations of Statistics. New York: Wiley,
[20]
and more necessary. 1934.
[21] R. Schlaifer, Probability and Statisties for Business Decisions.
New York: McGraw-Hill, 1959.
I Additional discussion regarding the value of information in [22] J. von Neumann and 0. Morgenstern, Theory of Games and
decision theory is available from many sources, most notably Economic Behavior, 2nd ed. Princeton, N.J.: Princeton Uni-
Howard [8b], [9], [11] and Raiffa and Schlaifer [19]. versity Press, 1947.