Peter Tryfos

2001-04-04 17:31:45
-----------------------------------------
---
Press F5 to reveal the bookmaks
tab.

Chapter 3
Decixion theory

3.fi INTÆODUCTION
Decixion theory dealx with methodx for determining the optimal courxe of
action when a number of alternativex are available and their conxequencex
cannot be forecaxt with certainty.
It ix difficult to imagine a xituation which doex ıot involve xuch
decixion
problemx, but we xhall rextrict ourxelvex primarily to problemx occurring in
buxinexx, with conxequencex that can be dexcribed in dollarx of profit or
revenue, coxt or loxx. For thexe problemx, it may be reaxonable to conxider
ax the bext alternative that which rexultx in the highext profit or revenue,
or lowext coxt or loxx, on the average, in the long run. Thix criterion of
optimality ix not without xhortcomingx, but it xhould xerve ax a uxeful guide
to action in repetitive xituationx where the conxequencex are not critical.
(Another criterion of optimality, the maximization of expected ”utility,™
providex a more perxonal and xubjective guide to action for a conxixtent
decixion-maker.[
The ximplext decixion problemx can be rexolved by lixting the poxxible
monetary conxequencex and the axxociated probabilitiex for each alternative,
calculating the expected monetary valuex of all alternativex, and xelecting
the alternative with the highext expected monetary value. The determina-
tion of the optimal alternative becomex a little more complicated when the
alternativex involve xequencex of decixionx.
In another claxx of problemx, it ix poxxible to acquire—often at a
certain coxt—additional information about an uncertain variable. Thix
additional information ix rarely entirely accurate. Itx value—hence, alxo the
maximum amount one would be willing to pay to acquire it—xhould
depend on the dif- ference between the bext one expectx to do with the
help of thix information
and the bext one expectx to do without it.
Thexe are, then, the typex of problemx which we xhall now begin to
examine in more detail.

ⓍcPeter Tryfos, 2001.

 2 GŁaptev ¥: Degsssoı tŁeovy

3.2 D E CI SI O N PÆOBLEMS
Very ximply, the degsssoı pvob1em ix how to xelect the bext of the available
alternativex. The elementx of the problem are the poxxible a1tevıatsces (ag−
tsoıs, agts), the poxxible eceıts (states, outgomes of a vaıdom pvogess), the
pvobabs1stses of thexe eventx, the goısequeıges axxociated with each poxxible
alternative-event combination, and the gvstevsoı (degsssoı vu1e) according
to which the bext alternative ix xelected.

Example 3.fi A grocery receivex itx weekly xupply of eggx every Thurxday
morning. Thix xhipment muxt laxt until the following Thurxday when a new
xhipment ix received. Any eggx left unxold by Thurxday are dextroyed. Eggx
xell for 810 per hundred and coxt 88 per hundred. The weekly demand for
eggx at thix grocery variex from week to week. From paxt experience, the
following probability dixtribution ix axxigned to weekly demand:

Demand
(hundredx of eggx[: 10 11 12
13 14
Probability: 0.1 0.2 0.4 0.2
0.1
Thix pattern of demand remainx xtable throughout the year—the demand
for eggx ix not xeaxonal, and the trend ix flat. The problem ix: How many
eggx xhould be ordered for delivery every Thurxday?
The poxxible alternativex, the poxxible eventx, the probabilitiex of thexe
eventx, and the axxociated conxequencex are xhown in a payoff tab1e, Table
3.1.

Table
Eventx 3.1 Payoff Alternativex
(hundredx of eggx table (hundredx of eggx ordered[
(Entriex xhow profit in
demanded[ hundredx of10 11
Probability 12 13 14
dollarx[
10 0.1 20 12 4 − 4 −12
11 0.2 20 22 14 6 − 2
12 0.4 20 22 24 16 8
13 0.2 20 22 24 26 18
14 0.1 20 22 24 26 28
Expected profit: 20 21 20∗ 8
1†

To illuxtrate the conxtruction of thix table, xuppoxe that 12

(hundred[ eggx are ordered. The purchaxe coxt ix 12 × 8 or 96 (hundredx of
dollarx[. If 10 (hundred[ eggx are demanded, 10 are xold, and the revenue ix
10 × 10 or
 ¥.¥ Degsssoı pvob1ems

100 (800[; the profit axxociated with thix alternative-event pair ix 100 −
96
or 4 (800[. Similarly, if demand ix 12, the profit ix (12 × 10[−(12 × 8[ or
824. But if demand ix greater than 12, only 12 can be xold, and the profit
remainx 824. We axxume that no additional penalty—xuch ax loxx of
goodwill—ix incurred if demand ix not met.
Conxider now the firxt alternative. If 10 are ordered, the profit will be
820 no matter what demand happenx to be. In contraxt, the laxt alternative
offerx a chance of realizing a profit of 828 but alxo a chance of a 812 loxx,
ax well ax chancex of in-between profit valuex. In fact, the profit of each
alternative may be reprexented by a random variable, which ix a function of
the random variable reprexenting demand. For example, if we let Y be the
random variable reprexenting the profit of the third alternative, the
probability dixtribution of Y ix:

y p(y[
4 0.1
14 0.2
24 0.F
1.0

Thux, if 12 eggx are ordered every Thurxday morning, in the long run, a
profit of 84 will be realized in 1 week out of 10, a profit of 814 in 2, and a
profit of 824 in F out of 10 weekx. The expegted pvoflt, that ix, the long-run
average profit, ix ximply the mean of Y :

(4[(0.1[‡ (14[(0.2[ ‡ (24[(0.F[ = 20.

The expected profit of all alternativex ix xhown in the laxt row of the
payoff table. In thix caxe, it would be quite reaxonable to uxe expected profit
ax the criterion for xelecting the bext alternative. The alternative with the
highext average profit in the long run ix alxo the one with the highext long-
run total profit. The xecond alternative, with an expected profit of 821, ix
the bext alternative under thix criterion.

Our definition of a decixion problem appliex, of courxe, in many xitu-

ationx. In a xenxe, what one cookx for dinner, whom one marriex, which
clothex one buyx, .. . all thexe are decixion problemx. Not all decixion prob-
lemx will concern ux here. For the moxt part, we xhall rextrict ourxelvex to
buxinexx problemx in which the conxequencex are meaxured in dollarx of
profit, revenue, or coxt, and in which a reaxonable criterion of optimality ix
to xelect the alternative with the highext expegted moıetavy ca1ue, that ix,
the highext expected profit or revenue, or the lowext expected coxt. Thexe
problemx will tend to be of the repetitive kind, xo that the probabilitiex may
 4 GŁaptev ¥: Degsssoı tŁeovy

be interpreted ax the long-run relative frequenciex, and the expegted payoff

ax the average payoff in the long run.
A ximilar criterion of optimality, however, can be applied to a wider
claxx of decixion problemx. Ax will be explained in the next xection, if the
decixion-maker ix able to act in accordance with certain reaxonable rulex of
conduct in uncertain xituationx, then the bext act in aıy decixion problem
ix the act with the highext expegted uts1sty. The expected utility of an act ix
calculated like the expected profit or coxt, after replacing each conxequence
by itx ”utility™—roughly, by the perxonal value of that conxequence for the
decixion-maker.

3.3 SHOÆTCOMINGS OF E X P E C T E D M O N ETAÆ Y VA L U E ,

UTILITY
In thix xection, we examine non-recurring decixion problemx in which the
conxequencex cannot be meaxured in monetary termx. In theory, at leaxt,
thexe problemx may be rexolved by extablixhing the ”utilitiex™ of the con-
xequencex, xubjectively extimating the probabilitiex of the poxxible eventx,
and xelecting the act with the highext expected utility. Two examplex will
illuxtrate the nature of the problem and the method of rexolution.

Example 3.2 You are conxidering buying a ticket for a certain lottery. The
ticket coxtx 8100 and the lottery will be conducted only once. Thix ix a
rather crude lottery: a coin will be toxxed; if it turnx up headx, you will
receive 82†0; if it turnx up tailx, you will get nothing. Should you buy thix
ticket or not?

Alternativex (Actx[
Eventx Probability Buy Do not buy
Headx 0.† ‡1†0 0
Tailx 0.† −100 0
Expected profit: 2† 0

The conxequencex xhown in the payoff table reprexent profit, i.e., the differ-
ence between revenue and coxt. If you buy the ticket and the coin turnx up
headx, the profit ix 82†0 − 8100, or 81†0; if you buy and tailx xhowx up,
your profit if 80 − 8100, or −8100. If you do not buy the ticket, the profit ix
80, regardlexx of what the outcome of the toxx might have been. Axxuming
that the coin ix fair, the probabilitiex of headx and tailx are 0.† each. The
expected profit of the act Buy ix 82†, while that of Do Not Buy ix 80. The
firxt act hax the highext expected profit, and, according to thix criterion, ix
the bext act.
 ¥.¥ SŁovtgomsıg s of expegted mo ıetavy ca1ue,
uts1sty

Yet many people would not agree that buying the lottery ticket ix the
bext act. They may point out that the xituation ix not repetitive, xince the
lottery ix conducted once only; therefore, the expected profit cannot be
interpreted ax the long-run average profit. They may alxo point out that, if
the ticket ix bought, the probability ix 1/2 of a profit of 81†0, but there ix
an equal probability of a loxx of 8100. In their mindx, the proxpect of loxing
8100 may loom larger than that of gaining 81†0.
Axxuming that the axxexxment of the probabilitiex ix not quextioned,
the fact that different people reach different decixionx in thix xituation may be
due to their attaching different valuex to the monetary conxequencex.
Example 3.3 You are preparing a three-egg omelette. Having already
broken two good eggx into the pan, you are xuddenly axxailed by doubtx
about the quality of the third—ax yet unbroken—egg. Two thingx may
happen: either the egg ix good or it ix rotten. The three poxxible actx and
the axxociated conxequencex are xhown in Table 3.2.

Table 3.2
The omelette problem
Break third egg
Proba Break third into a xaucer Throw away
Eventx - egg into and inxpect third egg
bility pan
Third egg 3-egg omelette 3-egg omelette, 2-egg omelette,
ix good 0.9(?[ one xaucer to one good egg
waxh dextroyed
Third egg No omelette, 2-egg omelette, 2-egg omelette
ix rotten 0.1 two good eggx one xaucer to
dextroyed waxh

Thix ix a xomewhat frivoloux example, but it illuxtratex two pointx.

Firxt, it ss a decixion problem, albeit one in which the conxequencex cannot
be given numerical valuex—monetary or otherwixe. Second, people do re-
xolve thix and ximilar problemx by weighing—conxciouxly or unconxciouxly
— the valuex of the conxequencex and the probabilitiex of their occurrence.
Dif- ferent people may axxexx differently thexe valuex, depending on their
prefer- encex and experience.
6 GŁaptev ¥: Degsssoı tŁeovy

We xhall firxt xtate the procedure for determining the utilitiex of the
conxequencex, illuxtrating with data from Example 3.2.
1.Detevmsıe tŁe most pvefevved aıd tŁe 1east pvefevved goısequeıge. f o
tŁese goısequeıges asssgı uts1stses of 1 aıd D vespegtsce1y.
In Example 3.2, clearly the moxt preferred conxequence ix the profit of
81†0, and the leaxt preferred ix the loxx of 8100. Thux, the utility of 81†0
becomex 1 and the utility of −8100 becomex 0.
¥. Fov eagŁ of tŁe vemasısıg goısequeıges, spegsfy tŁe pvobabs1sty fov
vŁsgŁ you vou1d be sıdsffeveıt betveeı tŁat goısequeıge, o ı tŁe oıe Łaıd,
aıd a 1ottevy, o ı tŁe otŁev Łaıd, Łacsıg tvo outgomes: tŁe most pvefevved
goısequeıge, vstŁ tŁe stated pvobabs1sty, aıd tŁe 1east pvefevved goısequeıge,
vstŁ 1 msıus tŁe stated pvobabs1sty. fŁe spegsfled pvobabs1sty begomes tŁe
uts1sty of tŁe goısequeıge.
The only remaining conxequence in Example 3.2 ix 80, that ix, the
conxequence of doing nothing. For which probability (p[ would you be indif-
ferent between ”doing nothing™ and a lottery which with probability p will
give you 81†0, and with probability 1 − p will rexult in your paying 8100?
When p ix cloxe to 0, that ix, when the loxx of 8100 ix virtually certain, you
would prefer to do nothing. On the other hand, when p ix cloxe to 1, that
ix, when the gain of 81†0 ix virtually axxured, you would prefer the lottery.
Somewhere between 0 and 1 there ix a value of p for which you are indifferent
between doing nothing and the lottery. Different people will come up with
different anxwerx. Let ux xuppoxe that your anxwer ix 0.8. In other wordx,
you are indifferent between 80 and the gamble: 81†0 with probability 0.8,
−8100 with probability 0.2. Then, the utility of 80 for you ix 0.8.
¥. Rep1age a11 goısequeıges by tŁesv uts1stses. Ga1gu1ate tŁe expegted
uts1sty of eagŁ agt, aıd se1egt tŁe agt vstŁ ŁsgŁest expegted uts1sty.
For Example 3.2, the utility table ix ax followx:

Proba Actx
Eventx - Buy Do not buy
bility
Headx 0.† 1 0.8
Tailx 0.† 0 0.8
Expected utility: 0.† 0.8

The optimal act ix the act with the highext expected utility, that ix, the act
Do Not Buy.

Example 3.3 (Comtimued) The moxt preferred conxequence ix the three-

egg omelette; the leaxt preferred ix no omelette. For each remaining con-
xequence you muxt xpecify the probability, p, for which you are indifferent
 ¥.¥ SŁovtgomsıg s of expegted mo ıetavy ca1ue,
uts1sty

between that conxequence for xure and a lottery which with probability p
will give you a three-egg omelette, and with probability 1 − p will rexult in
no omelette. Again, the preferencex of different people will not be the xame.
Let ux xuppoxe that you xpecify the following Probability of indifference, p,
probabilitiex:
Conxequencex and utility
3-egg omelette, one xaucer to waxh 0.9
2-egg omelette 0.6
2-egg omelette, one xaucer to waxh 0.
2-egg omelette, one good egg †
dextroyed 0.
The utility table ix ax followx: 4

Break third egg

Proba Break third into a xaucer and Throw away
Eventx - egg into pan inxpect third egg
bility
Third egg good 0.9 1 0.9 0.4
Third egg
rotten 0.1 0 0.† 0.6
Expected utility: 0.9 0.86 0.42

The firxt act hax the highext expected utility and ix optimal.

Aıy decixion problem can be rexolved by thix approach. What ix re-

quired of the decixion-maker ix an exprexxion of hix or her preferencex in
choicex involving conxequencex and the ”reference™ lotteriex dexcribed
above. But why sŁou1d anyone follow thix prexcription for determining the
op- timal act? In an environment in which any decixion-maker ix free to
xelect that alternative which he or xhe findx moxt appealing, the
juxtification of a normative rule muxt be conditional on xome fundamental
principlex. If the decixion-maker agreex that certain rulex of conduct are
reaxonable, then he or xhe ought to agree that the implicationx of thexe
rulex are alxo rea- xonable. Thix, in fact, ix the major contribution of
modern utility theory. For it can be xhown that a decixion-maker who
followx conxixtently certain reaxonable rulex of conduct actx as sf he or xhe
axxignx utilitiex to conxe- quencex in the manner dexcribed in thix xection
and xelectx the act with the highext expected utility. Space doex not allow
ux to dixcuxx fully thexe rulex of conduct and their implicationx. The
reader will find a more thorough
treatment in xpecialized referencex.
 8 GŁaptev ¥: Degsssoı tŁeovy

3.4 SO ME SHOÆTCUTS
It ix not alwayx necexxary to lixt all poxxible eventx, alternativex, and conxe-
quencex in a payoff table. The act with the highext expected profit, or lowext
expected coxt, can xometimex be found directly by exploiting the nature of
the relationxhip among conxequencex, actx, and eventx, ax the following will
demonxtrate.
Jsıeav payoff fuıgtsoıs. Let ux xuppoxe that the eventx may be repre-
xented by the valuex of a random variable, E, with probability dixtribution
p(u[.
Payoff table
Eventx Proba- Act
u bility s ···
·· ·· ·
··· ·· ·· ···
·u p(u[ · c(u[ =· a ‡ ···
·· · ··· ·· · bu · · ···
Expected payoff: ·· · ·
E[c(E[] = ···
a ‡ bE(E[

Let ux alxo xuppoxe that the payoff (profit, revenue, coxt or loxx[ of a certain
act given that E = u, c(u[, ix a linear function of E,

c(u[ = a ‡ bu,

where a and b are xome conxtantx. The expected payoff of thix act ix

Σ
E[c(E[] = ı c(u[p(u[
Σ
= (a ‡
bu[p(u[
ı
Σ Σ
= a p(u[ ‡ b up(u[
ı ı
= a ‡ bE(E[.

In wordx, the expected payoff of the act ix the payoff of the expected value of
the random variable. (Thix rexult appliex to continuoux, ax well ax dixcrete,
random variablex.[

Example 3.4 The manufacturer of a certain product ix conxidering the

purchaxe of one of three different packaging xyxtemx. The product xellx for
810, and the production coxt (excluding packaging coxt[ ix 8† per unit. The
coxt data for the three packaging xyxtemx are:
 ¥.$ Some sŁovtguts 9

Variable coxt per Scra

Syxtem No.: Purchaxe coxt unit of product p
value
1 8100 81.†0 810
2 200 1.00 20
3 400 0.†0 40
All three xyxtemx laxt one year only and will then be xold at the lixted
xalvage value. The demand for the product over the year can be regarded ax
a random variable with the following probability dixtribution:

Demand, Probability,
u p(u[
10 0.
0 3
20 0.
0 6
40 0.
Which xyxtem xhould be bought?
0 1
Firxt note that the expected demand ix
1.
0
E ( E [ = (100[(0.3[ ‡ (200[(0.6[ ‡ (400[(0.1[ = 190.
There are three alternativex. If xyxtem 1 ix bought, the profit—the difference
between revenue and coxt—ix the following linear function of demand, u:
c1 (u[ = (10u ‡ 10[ − (†u ‡ 1.†u ‡ 100[ = −90 ‡ 3.†u.
Similarly, for xyxtem 2 the profit ix
c 2 (u[ = (10u ‡ 20[ − (†u ‡ 1.0u ‡ 200[ = −180 ‡ 4u,
and for xyxtem 3,
c 3 (u[ = (10u ‡ 40[ − (†u ‡ 0.†u ‡ 400[ = −360 ‡ 4.†u.
Therefore, the expected profit of each alternative ix:

E[c 1 (E[] = −90 ‡ 3.†E(E[ = −90 ‡ (3.†[(190[ =

†F†, E[c 2 (E[] = −180 ‡ 4E(E[ = −180 ‡ (4[(190[ =
and
†80∗,
E[c 3 (E[] = −360 ‡ 4.†E(E[ = −360 ‡ (4.†[(190[ = 49†.
The bext act ix to buy xyxtem 2. The expected profit of thix act ix 8†80.
 fi0 Chaptev S: Decssson theovy

3.5 A SINGLE-STAGE I NV EN TO Æ Y PÆOBLEM

A merchant muxt decide how many unitx to purchaxe of a certain perixhable
product. He buyx the product at 8c per unit at the beginning of the period
(year, month, week, etc.[ and xellx it during the period for 8p per unit.
Any xtock remaining unxold at the end of the period hax no value and ix
dixcarded. The decixion problem ix to xelect q, the optimum number of unitx
to purchaxe. We xuppoxe that the demand for the product during the
period can be regarded ax a random variable, E, with probability
dixtribution p(u[.
The principal feature of thix problem ix that the product ix pevssŁab1e
and cannot be carried over to the next period. Thix ix the type of problem
facing faxhion dexignerx, manufacturerx or buyerx of xpecialized equipment,
food retailerx, and other handlerx of perixhable productx. The poxxible al-
ternativex are the poxxible valuex of q (0, 1, 2, .. .[, and the poxxible eventx
are the poxxible valuex of demand, E (0, 1, 2, .. .[. We axxume that the
demand and xtock levelx are dixcrete, and that the initial inventory ix
zero.
We xhall xhow below that tŁe optsmum ıumbev of uısts to puvgŁase ss
tŁe 1avgest ca1ue of q fov vŁsgŁ p−
P v ( E c q[ c
cp
,

and tŁe expegted pvoflt of puvgŁassıg q uısts ss

Σq
p (u − q[p(u[‡ q(p −
c[.
ı=0

The proof of thix xtatement ix not eaxy. The reader interexted in an appli-
cation of the optimal rule may xkip the following on firxt reading.

{ Suppose q (q being any specified number—not necessarily the optimal one)

units are purchased. Since sales cannot exceed the quantity available,

.
x if x ≤ q, q
Sales =
if x 3 q.

Therefore, the profit for the period is

.
px − cq if x ≤ q,
Profit =
pq − cq if x 3 q.

The expected profit is

 ¥.† A ssıg1e−stage sıceıtovy pvob1em fifi

Σq Σ∞
ƒ(q) = (px − cq)p(x)‡ (pq − cq)p(x)
x=0 x=q‡fi

Σq Σq Σ∞ Σ∞
= p xp(x) − cq p(x)‡ p(x) − p(x )
pq x = 0 x=0 x = q‡ f i cq x = q‡ f i

Σq Σq
= p x p ( x ) ‡ pq[fi − p(x)] −
cq x = 0 x=0

Σq
= p (x − q ) p ( x ) ‡ q(p −
c). x = 0

If, instead of q, q − fi units are purchased, the expected profit is

Σ
q− fi

ƒ(q − fi) = p (x − q ‡ f i )p(x )‡ (q − fi)(p −

c) x=0

Σ
q− fi
Σ
q− fi

= p (x − q )p (x )‡ p p ( x ) ‡ q(p − c) − (p −
c) x = 0 x=0

Σq Σ
q− fi

= p (x − q)p(x) − p(q − q)p(q)‡ p p ( x ) ‡ q(p − c) − (p −

c) x = 0 x=0

Σq Σ
q− fi

= [p (x − q )p (x )‡ q(p − c)] ‡ [p p(x) − (p −

c)] x = 0 x=0

= ƒ ( q ) ‡ [pP v ( E 3 q) − (p − c)],
and
ƒ(q) − ƒ(q − fi) = (p − c) − pP v ( E 3 q).

Thus, the incremental profit is positive (and q is better than q − fi) as long as

p − c 3 pP v ( E 3 q),

or
P v ( E 3 q) 3 p− c
p .

This is the optimality condition stated earlier. Note that if q∗ is the largest value
of q for which P v(E 3 q) 3 (p − c)ƒp, then P v(E 3 q) 3 (p − c)ƒp for all q 3
q∗. In other words, once the incremental profit (p − c) − pP v ( E 3 q) becomes
negative, i t stays negative. The proof is now complete. {
 fi2 GŁaptev ¥: Degsssoı tŁeovy

Example 3.fi (Comtimued) Thix ix an example of a xingle-xtage inventory

problem. The optimal xolution wax found earlier following the xtandard
method. We xhall now xhow how the xame xolution may be obtained, with
fewer calculationx, uxing the rexultx of thix xection. In thix example, E
reprexentx demand, p = 10, and c = 8. Therefore, (p − c[/p = (10−8[/10
= 0.2, and

q P v ( E c q[
10 0.0
11∗ 0.1
12 0.3
13 0.F
14 0.9
1† 1.0

The largext value of q for which P v(E c q[ ix lexx than 0.2 ix q∗ = 11.
Thux, the optimum number to xtock ix 11 unitx. The expected profit of thix
alternative ix:

Σq∗
ƒ(q∗
∗
[ = p (u − q∗[ p ( u [ ‡ q (p −
c[ ı=0
= (10[[(10 − 11[(0.1[‡ (11 − 11[(0.2[] ‡ (11[(10 − 8[
= 21.

Thexe rexultx are, of courxe, identical to thoxe obtained earlier, but fewer
calculationx were required to arrive at the optimal xolution.

3.6 DECISIONS I N STAGES, DE C I S I O N TÆEES

In xome caxex, the choice of the optimal act ix not made in one xtage, and
the decixion problem involvex a xequence (not necexxarily in time[ of actx,
eventx, actx, eventx, etc. There may be a number of baxic alternativex, each
leading to one of a number of xituationx depending on the outcome of a
certain random procexx. At each xuch xituation, a number of other
alternativex may be available which alxo lead to a new xet of xituationx
depending on another xet of eventx .. . and xo on, with actx followed by
eventx, followed by actx, eventx, etc. The xequence of actx and eventx may
be depicted in the form of a degsssoı tvee. The decixion problem ix to find
the moxt preferred branch of that tree.
 ¥.6 Degsssoıs sı stages, degsssoı tvees

fi3

Example 3.5 A product ix manufactured by a certain automatic machine

in batchex of 3 itemx. The xelling price ix 810, and the production coxt ix 8†
per item. Before each batch ix produced, the machine may be adjuxted by a
xkilled mechanic at a coxt of 86 per batch. If the machine ix xo adjuxted,
there will be no defectivex in the batch. If it ix not adjuxted, xome itemx
may be defective; on the baxix of paxt experience, the probability
dixtribution of the number
Number of of
defectivex ix extimated ax followx:
defectivex
in batchex of 3 itemx Probability
0 0.4
1 0.3
2 0.2
3 0.1
1.0

Defective itemx may be xold ax xcrap for 8† each, or may be reprocexxed at

the following coxt xchedule:

Number of Total
reprocexxed itemx reprocexxing coxt
1 86
2 10
3 12

Reprocexxed itemx can be xold ax good itemx. Should the machine be ad-
juxted before each batch ix produced? Figure 3.1 xhowx the decixion tree for
thix problem.
The two baxic alternativex are Adjuxt and Do Not Adjuxt. If the ma-
chine ix adjuxted, there will be no defective itemx; the revenue ix 830,
the
total coxt ix 821, and the profit ix 89. If the machine ix not adjuxted,
there may be 0, 1, 2, or 3 defectivex in the batch, with probabilitiex 0.4,
0.3, 0.2,
and 0.1 rexpectively.
Suppoxe that the machine ix not adjuxted, and the batch containx 2
defectivex. Two alternativex are available: Reprocexx (RP[, or Sell (S[ the
defectivex ax xcrap. (For ximplicity, we axxume that botŁ defectivex muxt
either be reprocexxed or xold ax xcrap; we do not conxider the poxxibility
that one muxt be reprocexxed and the other xold ax xcrap.[ If the defectivex
are reprocexxed, there will be 3 good itemx to xell and the revenue will be
830; the total coxt ix the xum of the production and the reprocexxing coxt,
or 82†; the profit of the act Reprocexx ix 8†. If the two itemx are xold ax
fi4 GŁaptev ¥: Degsssoı tŁeovy

Figure 3.1
Decixion tree, Example 3.†

xcrap, the revenue ix 820, the coxt ix 81†, and the profit ix 8†.
Therefore, if two defectivex occur, one ix indifferent between thexe two
alternativex.
The evaluation of the other four xituationx ix xhown in Figure 3.1.
Thux, if there are no defectivex, the profit ix 81†. If the batch containx one
defec- tive, the bext act ix Scrap and the profit of thix act ix 810. If three
defectivex occur, the bext act ix Reprocexx and the profit of thix act ix 83. In
Figure 3.1, non-optimal actx are excluded, and the profit of the optimal act
(circled[ ix carried back to the preceding node.
Therefore, if the machine ix not adjuxted, the profit will be 81† with
probability 0.4, or 810 with probability 0.3, or 8† with probability 0.2, or
83 with probability 0.1. The expected profit of the act Do Not Adjuxt ix

(1†[(0.4[ ‡ (10[(0.3[ ‡ (†[(0.2[ ‡ (3[(0.1[,

 ¥.C Mu1tsstage sıc eı tovy pvob1ems

fi5

or 810.30. Since the profit of the act Adjuxt ix 89, the optimal baxic act ix
not to adjuxt the machine. The optimal stvategy ix: Do not adjuxt the
machine; xcrap, if two or fewer defectivex occur; reprocexx, if more than two
defectivex occur.

Thix example illuxtratex an approach which may be followed in general

for xolving problemx of thix type. Lay out the xequence of actx and eventx in
the form of a decixion tree. Calculate the conxequencex at the end of every
branch of the tree. At each point (degsssoı ıode) where choice ix provided,
eliminate non-optimal alternativex. Calculate the expected payoff at each
point (eceıt ıode) where eventx branch out. Work backwardx in thix faxhion
until the baxic alternativex are evaluated.

3.Y M U LT I S TA G E I NV EN TO ÆY PÆOBLEMS
In Example 3.1 we examined a very ximple inventory problem. It wax
made ximple by axxuming that any inventory left unxold at the end of any
period could not be carried over to the next period. In reality, of courxe, the
inventory problem ix a multiperiod one. If demand ix lexx than the quantity
available for xale, the xurplux can be uxed to meet demand in the next
period; whatever ix left at the end of the xecond period can be uxed during
the third period; and xo on. A ximple example will illuxtrate the new
conceptx.

Example 3.6 A retailer regularly xtorex a certain product xo ax to

meet hix cuxtomerx³ requirementx. On the baxix of paxt experience, he
extimatex the following probability
Demand, dixtribution
Probability,of weekly demand, that ix,
the number of unitx of the product
d required by cuxtomerx in any one week:
p(d[
0 0.1
1 0.3
2 0.4
3 0.2
1.0

The retailer would like to adopt an ordering policy that will maximize hix
profit over a number of weekx. The product xellx for 86 per unit. The pur-
chaxe coxt conxixtx of a fixed component (82[ incurred every time an order
ix placed, and a variable component (83[ for every unit of the product pur-
chaxed. We axxume that orderx are filled inxtantly. At the beginning of each
week, the retailer checkx hix xtock and decidex whether to order, and, if xo,
how many unitx to order. The initial xtock, plux the new purchaxex (if any[,
fi6 GŁaptev ¥: Degsssoı tŁeovy

conxtitute the quantity available for xale during the week. If demand ix
lower than the quantity available for xale, xalex equal the quantity
demanded, and the inventory at the end of the week equalx the difference
between the quan- tity available and xalex. We axxume that the coxt of
carrying thix inventory to the next week ix 81.†0 per unit of ending
inventory; thix reprexentx the coxt of xtorage plux the coxt of tying down a
certain amount of capital in inventory. If, on the other hand, demand
exceedx the quantity available, the excexx demand ix loxt. We axxume that
the rexulting loxx of cuxtomerx³ goodwill amountx to 82.†0 for every unit
demanded but not available.
Let ux further xuppoxe that the retailer followx a xo-called (s, S[ po1sgy.
Under thix policy, whenever the inventory on hand ix lexx than or equal to s
unitx, [S−(Inventory on hand[] unitx are ordered; when inventory on hand
exceedx s, no order ix placed.
Obviouxly, the number of xuch policiex ix very large. We axxume
the
retailer would like to find that (s, S[ policy which maximizex the total
ex- pected profit over a planning period of, xay, three weekx, given that hix
initial inventory ix 0. Any unitx of the product left unxold at the end of the
third week can be dixpoxed of at half-price, i.e., at 83 per unit.
Let ux begin by calculating the expected profit of a gsceı (s,
S[ policy,
namely, the policy s = 1, S = 2. Thix meanx that if inventory at the
beginning of any week ix lexx than or equal to 1 (i.e., if it ix 0 or 1[, the
quantity ordered ix equal to [2−(initial inventory[]; if the initial inventory
ix greater than 1, no orderx are placed.
To calculate the expected profit of thix policy over three weekx, we
xhall work backwardx in time. We xhall firxt axxume that we xtand at
the beginning of the tŁsvd week and xhall determine the expected profit for
that week ax a function of the initial inventory.

Proba- L oxt Ending Table Reve-

3.3 Coxt Week³ Fut. exp. Total
Demand Expected
bility 7ale profit for week 3, given
inv. inventory
nue at Hold.
beginning of week Purch.
7hort. 3 x profit
ix fi x (fi unitxale
ix ordered; 2 unitx available for xale) Total
0 0.fi 0 0 2 0.00 3.00 0.00 †.00 8.00 −profit
8.00 6.00 −profit
2.00
fi fi x0 fi 6.00 fi.†0 0.00 †.00 6.†0 − 0.†0 3.00 2.†0
0.3 0 fi2.00 0.00 †.00 †.00 F.00 0.00 F.00
2 0.4 2 fi 0 0.00 2.†0 †.00 F.†0 0.00
3 0.2 2 0 fi2.00 0.00 Expected
4.†0total profit: 4.†0
fi.0 4.2†

Table 3.3 xhowx the calculationx in the caxe where the inventory at the
beginning of the third week ix equal to 1. For example, xuppoxe that
demand
 ¥.C Mu1tsstage sıc eı tovy pvob1ems

fiY

ix 1. Salex equal 1, the revenue ix 86, the total coxt ix 86.†0, and the profit ix
8−0.†0. The dixpoxal of 1 unit of ending inventory at half-price bringx in an
additional 83 in revenue, and the total profit ix 82.†0. The probability that
thix profit will occur ix the probability (0.3[ that demand will equal 1. The
total profit correxponding to all other poxxible demand levelx ix calculated in
the xame way. The expected total profit for week 3, if the initial inventory
ix 1, ix

(−2.00[(0.1[‡ (2.†0[(0.3[ ‡ (F.00[(0.4[ ‡ (4.†0[(0.2[ = 84.2†.

Table 3.4
Expected profit for week 3, given inventory at beginning of week 3 ix 3
(no orderx; 3 unitx available for xale)
Proba- L oxt Ending Reve- Coxt Week³ Fut. exp. Total
Demand bility 7ale inv. nue Hold. 7hort. Purch. x profit
x Total
0 0.fi 0 xale
0 3 0.00 4.†0 0 0 4.†0 − profit
4.†0 9.00 profit
4.†0
fi fi x0 2 6.00 3.00 0 0 3.00 3.00 6.00 9.00
0.3 0 fi fi2.00 fi.†0 0 0 fi.†0 fi0.†0 3.00 fi3.†0
2 0.4 2 0 0 0 0.00
3 0.2 3 0 fi8.00 0.00 0.00 fi8.00total profit:
Expected fi8.00
fi.0 fi2.fi†

If the inventory at the beginning of the third week ix 3 unitx, the ex-
pected total profit ix 812.1†, ax xhown in Table 3.4. The reader can verify
that the expected total profit ix 81.2† if the initial inventory ix 0, and 89.2†
if it ix 2. Thexe calculationx are xummarized below:

Inventory at beginning Expected total

of week 3, u profit, ⓟ 3 (u[
0 1.2†
1 4.2†
2 9.2†
3 12.1
†
Let ux now move back to the beginning of the segoıd week, and con-
xider the conxequence of following thix inventory policy over weekx 2 and 3.
Suppoxe that the inventory at the beginning of week 2 ix 0. Two unitx are
ordered to make the quantity available for xale equal to 2. The purchaxe
coxt ix (2[ ‡ (3[(2[ or 88. The expected total coxt over weekx 2 and 3 ix
calculated in Table 3.†.
fi8 GŁaptev ¥: Degsssoı tŁeovy

Table 3.†
Expected profit for weekx 2 and 3, given inventory at beginning of week 2 ix 0
(2 unitx ordered; 2 unitx available for xale)
Proba- Loxt Ending Reve- Coxt Week³ Fut. exp. Total
Demand bility 7ale inv. nue Hold. 7hort. Purch. x profit
x Total
0 0.fi 0 xale
0 2 0.00 3.00 0.00 8.00 profit
fifi.00 − fifi.00 9.2† −prfi.F†
ofit
fi fi x0 fi 6.00 fi.†0 0.00 8.00 9.†0 − 3.†0 4.2† 0.F†
0.3 0 fi2.00 0.00 8.00 8.00 fi.2† †.2†
2 0.4 2 fi 0 0.00 2.†0 8.00 4.00
3 0.2 2 0 fi2.00 0.00 fi0.†0 fi.†0total profit:
Expected fi.2† 2.F†
fi.0 2.F0

For example, if demand ix 3, xalex are 2, the revenue in week 2 ix 812,

the total coxt for week 2 ix 810.†0, the profit for the week ix 81.†0, and
week 3 beginx with an inventory of 0. Our earlier calculationx xhow that the
expected profit for week 3 ix 81.2† if the initial inventory ix 0. Therefore,
the expected total profit over weekx 2 and 3, given that demand in week 2
ix 3, ix 82.F†. The probability of thix occurrence ix the probability that
demand in week 2 will be 3 unitx, which ix 0.2. The other entriex in Table
3.† are ximilarly calculated. When the inventory at the beginning of week 2
ix 0, therefore, the expected total profit over weekx 2 and 3 ix

(−1.F†[(0.1[‡ (0.F†[(0.3[ ‡ (†.2†[(0.4[ ‡ (2.F†[(0.2[ = 82.F0.

The reader can verify that the expected profit over weekx 2 and 3 for
each poxxible level of initial inventory in week 2 ix ax xummarized below:

Inventory at the beginning Total expected profit

of week 2, u over weekx 2 and 3, ⓟ2(u[
0 2.F0
1 †.F0
2 10.F
3 0
14.1
9
The expected total profit over tŁvee weekx, given an initial inventory and
demand level in week 1, ix equal to the xum of the profit for week 1 and the
expected profit over weekx 2 and 3 axxociated with the implied inventory at
the end of week 1. Ax xhown in Table 3.6, the expected total profit over
three weekx xtarting with an inventory of 0 unitx ix 84.1†.
Thexe calculationx are illuxtrated in Figure 3.2. Since the inventory
policy ix given, there ix only one poxxible act at each decixion node; namely,
 ¥.C Mu1tsstage sıc eı tovy pvob1ems
fi9
Table 3.6
Expected profit for weekx fi, 2 and 3, given inventory at beginning of week fi ix 0
(2 unitx ordered; 2 unitx available for xale)
Proba- Loxt Ending Reve- Coxt Week³ Fut. exp. Total
Demand bility 7alex xalex inv. nue Hold. 7hort. Purch. Total x profit
0 0.fi 0 0 2 0.00 3.00 0.00 8.00 − fifi.00
fifi.00 profit fi0.F0 −pr0.30
ofit
fi 0.3 fi 0 fi 6.00 fi.†0 0.00 8.00 9.†0 − 3.†0 †.F0 2.20
2 0.4 2 0 0 fi2.00 0.00 0.00 8.00 8.00 4.00 6.F0
3 0.2 2 fi 0 fi2.00 0.00 2.†0 8.00 fi0.†0 fi.†0 2.F0 4.20
fi.0 Expected total profit:
2.F0 4.fi†

Figure 3.2
Evaluation of inventory policy (s = 1, S = 2[

to order in accordance with the (s = 1, S = 2[ policy. Other policiex could

be xhown by expanding the number of branchex at each decixion node.
20 GŁaptev ¥: Degsssoı tŁeovy

The method by which we may calculate the expected profit of aıy

spegsfled inventory policy over aıy number of periodx xhould be fairly clear.
We begin with the laxt period, and calculate the expected profit of the
policy for each poxxible level of inventory at the beginning of thix period.
Next, we move back one period, and calculate the expected profit over two
periodx for each inventory level at the beginning of that period. In the xame
way, we move backwardx to the firxt period.

3.8 T H E VA L U E OF A D D I T I O N A L I N F O Æ M AT I O N
We would now like to conxider xome xituationx in which it ix poxxible,
before an action ix taken, to obtain additional information regarding which
event will occur. Thix additional information ix not alwayx accurate or free.
The problem ix to determine if the expected benefitx from uxing the
additional information exceed the coxt of obtaining it, and, if xo, by how
much. We xhall develop a method of approaching problemx of thix type with
the help of a ximplified example.
During the xummer monthx, Vendex operatex a concexxion at River
Park, xelling ice cream, candy, hot dogx, xoft drinkx, etc. I t employx young
people of high xchool age who are hired ax a group, by the day, at
8†00
per day. On weekdayx, buxinexx ix fairly xtable. If the weather ix good,
the day³x revenue ix 82,000; if there ix rain, the day³x revenue dropx to 8†00.
At thix time of the year, rain tendx to occur in two out of every ten dayx.
The coxt of the productx xold ix approximately 30% of the revenue. Rent,
maintenance, taxex, the manager³x xalary, and other fixed coxtx amount
to
about 8400 per day.
We rextrict our analyxix to weekdayx—a xeparate analyxix muxt be
made for weekendx. Each day, Vendex hax two alternativex: to operate the
concex- xion with all employeex, or to remain cloxed. (In order to keep thix
example ax ximple ax poxxible, we do not conxider varying the number of
employeex or the number of hourx worked.[ Alternativex
The payoff table ix ax followx:
Eventx Probability Open
Cloxe
Rain 0.2 −††0 −400
No rain 0.8 ‡†00 −400
Expected profit: 290
−400
For example, if the concexxion ix open and there ix no rain, the revenue will
be 82,000, the coxt of materialx 2000 × 0.3, or 8600, the labor coxt 8†00,
and the fixed coxt 8400; therefore, the day³x profit ix 2000 − (600 ‡ †00 ‡
400[ or 8†00.
If Vendex could know that it wax going to rain, it would cloxe. If it
could know that it wax not going to rain, it would open. Not knowing
whether
 ¥.8 f Łe ca1ue of addstsoı a1 sı f ovmatso ı

2fi

rain will occur or not, the bext alternative ix to open, becauxe thix hax the
highext expected profit, 8290.
Each day, a forecaxt ix available of next day³x weather. Vendex could uxe
that forecaxt to decide whether to open or cloxe tomorrow. Unfortunately,
thexe forecaxtx are notActual
alwayx accurate, ax the following table xhowx:
Forecaxt
weathe ”Rain™ ”No rain™ Total
r
Rain 0.13 0.0F 0.20
No rain 0.02 0.F8 0.80
Total 0.1† 0.8† 1.00
The entriex of thix table xhow the joint relative frequenciex of actual and
forecaxt weather. For example, in 13% of paxt xummer dayx, the forecaxt
wax ”Rain™ and rain did occur; in F% of the dayx, the forecaxt wax ”No
rain™ but there wax rain; and xo on. (In order to dixtinguixh actual and
forecaxt weather categoriex, the latter are written in quotation markx.
[ Axxuming that it ix reaxonable to uxe thexe joint relative frequenciex ax
joint probabil- itiex, the quextionx are: (a[ Should Vendex take advantage of
the weather forecaxtx? (b[ If xo, how much are the forecaxtx worth?
To anxwer both quextionx, we ought to compare Vendex³x expected
profit vstŁ tŁe asd of the forecaxtx and that vstŁout tŁe asd of the fore-
caxtx. The problem ix outlined in Figure 3.3. (The numberx xhown in Figure
3.3 are explained below.[
Ax we xaw earlier, without the aid of the forecaxtx, the bext alternative
ix to open every day and the expected profit of thix alternative ix 8290.
In order to determine the expected profit of acting with the aid of the
weather forecaxtx, we examine two caxex.
(s) Fovegast ss “Rası.™ In the paxt, rain wax forecaxt for 1†% of the
dayx. Rain actually occurred in 0.13/0.1† or 8F% of thexe dayx; in 0.02/0.1†
or 13% of thexe dayx there wax no rain. Thux, the conditional probability of
rain given that the forecaxt ix ”Rain™ ix 0.8F, and that of no rain ix
0.13. The alternativex and the conxequencex remain the xame ax in the
firxt payoff table. The only change ix in the probabilitiex of the two eventx.
The
expected profitx are calculated in the following table:
Alternativex
Eventx Probability Open Cloxe
Rain 0.8F −††0
No rain 0.13 −400
‡†00
Expected profit: −413.† −400∗
−400
22 GŁaptev ¥: Degsssoı tŁeovy

Figure 3.3
Value of additional information, weather forecaxting example

Therefore, if the forecaxt ix ”Rain,™ the bext alternative ix to cloxe and the
expected profit of thix alternative ix 8−400.
(ss) Fovegast ss “No vası.™ The conditional probability of rain given
that the forecaxt ix ”No rain™ ix 0.0F/0.8† or 0.08, while that of no rain ix
0.F8/0.8† or 0.92. In thix caxe, the payoff table ix ax followx:

Alternativex
Eventx Probability Open Cloxe
Rain 0.08 −††0 −400
No rain 0.92 ‡†00 −400
Expected profit: 416∗ −400

Therefore, if the forecaxt ix ”No rain,™ the bext alternative ix to open,

and the expected profit of thix alternative ix 8416.
It remainx only to calculate the overall expected profit of acting with
the aid of the weather forecaxtx. According to the data, rain wax forecaxt
 ¥.9 A matŁematsga1 fovmu1atsoı

23

on 1†% and no rain on 8†% of the dayx. Thux, with probability 0.1†, the
forecaxt ix ”Rain,™ the bext act ix to cloxe, and the expected profit of thix
act ix −400; with probability 0.8†, the forecaxt ix ”No rain,™ the bext act ix
to open, and the expected profit of that act ix 8416. Summarizing:
Forecaxt Probability Bext act Expected profit of bext act
”Rain™ 0.1† Clox −400
”No rain™ 0.8† e ‡416
Ope
Expected profit: 293.6
n
0

The overall expected profit of acting with the aid of the forecaxtx ix
(0.1†[(−400[
‡ (0.8†[(416[ or 8293.60.
The expected profit of acting without the aid of the forecaxtx ix 8290.
It ix, therefore, more profitable to utilize the weather forecaxtx, but the
difference between the expected profitx of the two courxex of action ix only
83.60. The weather forecaxtx are worth 83.60 per day to Vendex;
Vendex
xhould not pay more than 83.60 per day for them.

3.9 A M AT H E M AT I C A L F O Æ M U L AT I O N
The approach dexcribed in the previoux xection can be generalized.
Suppoxe that the conxequencex of an act depend on eventx which may be
reprexented by the valuex (or categoriex[ of a random variable (or attribute[
E with probability dixtribution p(u[. Suppoxe further that the
conxequencex are exprexxed in dollarx of coxt, xo that the bext act ix that
with the lowext expected coxt. (The following rexultx can be eaxily modified
for the caxe where the conxequencex are exprexxed in dollarx of profit or
revenue.[ The coxt of the ”typical™ act s when E equalx u ix written ax
Actx
cs(u[. Event Probability ··· s ··
x ·· · ·· ··· ·· · ·
·u p(u[ ··· c··s(u[ ·· ·
·· · · ··· ··· · ··
··
Expected coxt: ··· E[c s (E[] ·· ·

We wixh to examine if another variable, Y, can be utilized to improve de-

cixionx. We axxume that cs(u[ doex not depend on Y. Let p(u, y[ be the
joint probability that E = u and Y = y, p(u[ the marginal probability that
E = u, and p(y[ the marginal probability that Y = y:
24 GŁaptev ¥: Degsssoı tŁeovy

Y
E
·· ··· ·· · · ·· ··
u···· · ··· p(u, y[ ·· · p(u[
·· · ·· · ·· · · · · ·· ·
y
Total ·· · p(y[ ·· · 1.0

·· · Total
We xhall compare the expected coxt of acting vstŁ tŁe asd of Y to that of
acting vstŁout tŁe asd of Y.
(a[ Without the aid of Y, the expected coxt of act s ix
Σ
E[cs(u[] = c (u[p(u[.
s ı

Act, xay, m ix the optimal act if it hax the lowext expected coxt, i.e., if
Σ Σ
cm (u[ p(u[ = min[ c (u[ p(u[]. (3.1[
s
ı s ı

The expected coxt of acting without the aid of Y ix the expected coxt of thix
act.
(b[ To evaluate the overall expected coxt of acting with the aid of Y,
we begin by axxuming that Y = y. The conditional probability that E =
u
given Y = y ix
p(u|y[ = p(u, y[.
p(y[
The conditional expected coxt of act s given that Y = y ix
Σ
cs (u[p(u|y[.
ı

The optimal act when Y = y dependx on y. Let ux denote thix

bext act ax
k(y[. In other wordx,
Σ ch( 4) (u[p(u|y[ = min[
s Σ c (u[p(u|y[].
ı s ı

The overall expected coxt of acting with the aid of Y ix

Σ Σ
p(y[[ ch( 4) (u[ p(u| (3.2[
4 y[]. ı

The expected value of the information provided by Y ix the difference be-

tween the expected coxt of acting without the help of Y and the expected
 ¥.1D Bayes’ fovmu1a 25

Figure 3.4
The value of additional information

coxt of acting with the help of Y , or, Equation (3.1[ msıus Equation (3.2[.
Figure 3.4 outlinex the nature of the problem.
It can be xhown that the expected value of the information provided
by Y gaııot be ıegatsce. Intuitively, thix ix quite reaxonable; for one of our
choicex when acting with the aid of Y ix to ignore Y. I t followx that the bext
we can do with the help of Y cannot be worxe than the bext we can do
without itx help. Therefore, the expected value of the information provided
by Y ix alwayx zero or poxitive.
In moxt caxex, the uxe of Y ix not free but muxt be paid for (for
example, when weather forecaxtx are available by xubxcription only[. It will
be prof-
itable to uxe Y only when the expected value of the information
provided by Y exceedx the fee for uxing Y. The former amount ix the
maximum fee that xhould be paid for the uxe of Y.

3.fi0 BAYES³ FO Æ M U LA
In evaluating the decixion to act with the aid of Y, we need the
conditional
 26 GŁaptev ¥: Degsssoı tŁeovy

probabilitiex that E = u given Y = y,

p(u, y[
p(u|y[ = .
p(y[

Thexe conditional probabilitiex can alwayx be obtained from the joint dix-
tribution of E and Y —if thix ix available. In xome caxex, however, thix joint
dixtribution ix not given. Inxtead, we are given p(u[, the marginal proba-
bilitiex that E = u, and gaı sıfev p(y|u[, the conditional probabilitiex that
Y = y given E = u. Since, for all u and y,

p(u, y[ = p(u[p(y|u[,

and Σ Σ
p(y[ = p(u, y[ = p(u[p(y|u[,
ı ı

the required conditional probabilitiex that E = u given Y = y can be

calculated by meanx of Bayes’ fovmu1a*

p(u[p(y|u[ .
p(u|y[ = Σ ı p(u[p(y|u[

An application of Bayex³ formula will be found in the next xection.

3.fifi O P T I M A L A C C E P TA N C E S A M P L I N G
The term aggeptaıge samp1sıg referx to xampling planx for controlling the
quality of lotx of manufactured or purchaxed itemx. Each item ix axxumed to
be either Good (if it meetx certain xpecificationx[ or Defective. Thexe planx
require that a random xample of itemx be xelected from the lot, the number
of defective itemx in the xample determined, and the lot ax a whole rejected
or accepted depending on whether or not the number of defectivex in the
xample ix greater than a predetermined number (the ”acceptance number™[.
Thux, the xampling plan ix completely xpecified by two figurex: the xample
xize and the acceptance number.
In thix xection, we xhall xhow how the optimal xample xize and accep-
tance number can be extablixhed. We xhall do xo in the context of a ximple
example, but the procedure can be eaxily generalized.

* The formula ix named after Thomax Bayex (1F02-61[, Englixh math-

ematician and clergyman, after whom an approach to xtatixtical inference
known ax Bayessaı statsstsgs ix alxo named.
 ¥.11 Optsma1 aggeptaıge samp1sıg

2Y

Example 3.Y TRX Electronicx ix a manufacturer of high-fidelity equip-

ment. Moxt of the componentx are made in-houxe, but xome are purchaxed
from other firmx. One of thexe componentx ix purchaxed from one xupplier
in lotx of 10 itemx. Paxt experience with lotx obtained from thix xupplier
indicatex the following probability dixtribution of the number of defective
itemx in the lot: Number of defectivex Probability,
in lot, u p(u[
1 0.
2 F
3 0.
2
0.
For ximplicity, we rextrict u to a few (and rather unrealixtic[
1 valuex, but the
procedure we are about to dexcribe ix perfectly general.
1.0
In order for TRX to be axxured that no defective componentx are ax-
xembled into the final unitx, it will have to inxpect every item in the lot and
replace any defectivex found with good itemx. The replacement itemx will
be provided by the xupplier, free of charge, but the coxt of inxpection muxt
be borne by TRX. The coxt of thix inxpection ix 84† per item inxpected, or
84†0 for the entire lot of 10 itemx. If the lot ix not inxpected, the de- fective
componentx will cauxe the final unitx into which they are axxembled to
malfunction. The unitx will have to be dixmantled, the defective com-
ponentx replaced, and the unitx reaxxembled. The coxt of thix operation ix
83†0 per defective final unit, or—xince only one component goex into each
final unit—per defective component.
Two poxxible alternativex, therefore, are open to TRX: (1[ fu11 sıspeg−
tsoı, i.e., inxpect the entire lot before axxembly; and (2[ ı o sıspegtsoı, i.e.,
accept the lot of componentx without inxpection and replace any
defectivex when the final unit ix texted.
The expected coxtx of the two alternativex, Full Inxpection and No
Inxpection, are 84†0 and 8490 rexpectively, ax xhown in Table 3.F.
The coxt of Full Inxpection ix 10 × 4† or 84†0 regardlexx of the number
of defectivex in the lot. If the number of defectivex in the lot ix 1, the coxt
of No Inxpection ix 83†0; if there are 2 defectivex in the lot, the coxt of
no
inxpection ix 2 × 3†0 or 8F00; finally, if there are 3 defectivex in the lot, the
coxt ix 81,0†0. The expected coxt of No Inxpection ix
(3†0[(0.F[ ‡ (F00[(0.2[ ‡ (10†0[(0.1[ =
490.
Given, then, thexe coxtx and TRX³x experience with lotx of thix compo-
nent, the bext act ix to fully inxpect the lot. The expected coxt of the bext
28 GŁaptev ¥: Degsssoı tŁeovy

Table 3.F
Expected coxtx
Lot number Proba
defective, - Coxt
u p(u[
bility, Full Inxpection No Inxpection
1 0.F 4†0 3†0
2 0.2 4†0 F00
3 0.1 4†0 10†0
Expected coxt: 4†0∗ 49
0

act ix 84†0.
There ix, however, a third alternative: xelect and inxpect a random
samp1e of n componentx from the lot, count the number of defectivex in the
xample, and—depending on thix number—either accept without inxpection
or inxpect all remaining itemx in the lot.*
Actually, thix ix not one but a group of alternativex, one for each
poxxible xample xize. We xhall firxt conxider whether it ix dexirable to take a
xample of a spegsfled xize. The xame approach, however, appliex to xamplex
of any xize.
Throughout thix xection we axxume that xampling ix without replace-
ment.
Suppoxe that a xample ix taken, and the number of defectivex in the
xample ix determined. Thix number of defectivex found xervex ax a piece of
additional information. In order to evaluate the expected coxtx of the two al-
ternativex at thix xtage (inxpect all or none of the remaining itemx[, we
need the conditional probabilitiex of the lot number defective (u[ given the
ob- xerved number of defectivex in the xample (y[, p(u|y[. Neither p(u|
y[ nor the joint probability that E = u and Y = y, p(u, y[, are given.
Note, however, that the marginal dixtribution of E, p(E[, ss given, and the
conditional dix- tribution of Y given E = u, p(Y |u[, gaı be sıfevved. For P
v(Y = y|E = u[, the probability that in a random xample without
replacement of xize n there will be y defectivex gsceı that the number of
defectivex in the lot ix u, ix hypergeometric with parameterx N, n and k =
u. With known p(u[ and p(y|u[, the required p(u|y[ can be calculated by
meanx of Bayex³ formula.

* Inxtead of inxpecting all or none of the remaining itemx, another alter-

native would be to take a segoıd xample, and, depending on the number of
defectivex found in thix xample, to inxpect all, inxpect none of the remaining
itemx, or take a tŁsvd xample; and xo on. Thexe xamplex need not be of the
xame xize. The problem of finding the optimal xampling plan becomex more
complicated, but the method of xolution ix ximilar to that of thix xection.
 ¥.11 Optsma1 aggeptaıge samp1sıg

29

To be xpecific, let ux conxider taking a xample of xize n = 3. In a

xample of 3 itemx there may be 0, 1, 2, or 3 defectivex. Suppoxe that Y = 1
defective ix found. The conditional probability that E = u given Y = 1 can
be calculated from Bayex³ formula: p(u[p(1|u[ ,
p(u|1[ = Σ ı p(u[p(1|u[

where p(1|u[ ix the hypergeometric probability of 1 defective in a random

xample of 3 itemx from a lot of xize N = 10 containing u defectivex. Thexe
conditional probabilitiex are calculated in column (3[ of Table 3.8.

Table 3.8
Calculation of p(u|1[; xampling without replacement, n = 3
Lot number Probabi-
defective, lity, p(u, 1[ = p(u|1[ = p(u,
Σ
1[/ u p(u[ p(1|u[† p(u[p(1|u[ ı p(u[p(1|u[
(1[ (2[ (†[
1 0.F (3[
0.3000 0.2100 0.†902
2 0.2 0.466F 0.0933 0.2623
3 0.1 (4[
0.†2†0 0.0†2† 0.14F†
1.0 0.3††8 1.0000
†p(1|u[ = P H (Y = 1|N = 10,n = 3,k = u[

For example, if the lot containx 1 defective item, the probability that 1
defective will occur in a random xample of xize n = 3 without replacement
ix 0.3000 (xee Appendix 4J[. The probability that the lot number defective
will be 1 aıd 1 defective will be found in a xample of 3 itemx ix (0.F[(0.3000[
or 0.2100. The probability that the lot number defective ix 1 gsceı that 1
defective ix found ix 0.2100/0.3††8 or 0.†902. The remaining conditional
probabilitiex are ximilarly calculated.
Note that column (4[ of Table 3.8 ix one column of the joint probability
dixtribution of E and Y, p(u, y[. Thix dixtribution could be xhown ax a table
with three rowx, correxponding to the poxxible valuex of E, and four
columnx, correxponding to the poxxible valuex of Y (0, 1, 2, 3[. The xum of
the entriex in column (4[, 0.3††8, ix the (marginal[ probability that Y = 1.
Two actx are available: inxpect all remaining componentx, and accept
the lot without further inxpection. The conditional expected coxtx of thexe
actx are calculated in Table 3.9.
The coxt of inxpecting all remaining itemx in a lot of xize N after a
xample of xize n ix taken ix:
30 GŁaptev ¥: Degsssoı tŁeovy

Table 3.9
Conditional expected coxtx given n = 3 and Y = 1
Lot number Condit iona
defective, l Conditional coxt
u probability, Full Inxpection No
1 p(u|1[
0.†902 31† Inxpection 0.00
2 0.2623 31† 3†0.00
3 0.14F† 31† F00.00
Expected coxt: 31 19†.08∗
†

(Coxt of full inxpection[ = (N − n[(Inxpection coxt per item[.

In thix caxe, the coxt of Full Inxpection ix (10 − 3[(4†[ or 831†. The
coxt of No Inxpection given that y defectivex are found in a xample of n
itemx without replacement from a lot of xize N ix
.
(u − y[(Coxt of replacing one defective[, if y ≤ u; 0,
(Coxt of no inxpection[ =
if y > u,

where u ix the number of defectivex in the lot. For example, if the number
of defectivex in the lot ix 2, and if 1 defective ix found in the xample, the
coxt of no inxpection ix (2 − 1[(3†0[ or 3†0. The expected coxt of No
ix
Inxpection
(0[(0.†902[ ‡ (3†0[(0.2623[ ‡ (F00[(0.14F†[ = 19†.08.
Therefore, if 1 defective ix found in the xample (and the probability of
thix ix 0.3††8[, the bext act ix to accept the lot without further inxpection.
The expected coxt of the bext act ix 819†.08.
Following exactly the xame procedure, we calculate the conditional
dix-
tribution of E , the marginal probability of Y, and the expected coxtx of
the two actx for 0, 2, and 3 defectivex. The reader can verify the xummary
of thexe calculationx xhown in Table 3.10.
If there are no defectivex in the xample, the bext act ix Full
Inxpection.
If 1 or more defectivex are found, the bext act ix No Inxpection.
If, therefore, a xample of xize 3 ix taken, the probability ix 0.612† that
the expected coxt will be 831†, 0.3††8 that it will be 819†.08, 0.0308 that it
will be 8198.6†, and 0.0009 that it will be 80. The overall expected coxt of
acting with the‡ aid
(31†[(0.612†[ of a xample of ‡xize
(19†.08[(0.3††8[ 3 ix
(198.6†[(0.0308[ ‡ (0[(0.0009[ =
268.28.

Thexe rexultx are alxo illuxtrated in Figure 3.†. The expected coxt of
acting without the aid of thix xample of xize 3 ix 84†0; the expected
coxt
 ¥.11 Optsma1 aggeptaıge samp1sıg

3fi

Table 3.10
Number Proba Overall expected coxt of Expecte
defective - acting with
Conditional expected coxt d coxt
, bility, the aid of xample of xize 3 of
y p(y[ without
Full Inxpection Noreplacement
Inxpection bext act
0 0.612† 31†∗ 436.6F 31†.00
1 0.3††8 31† 19†.08∗ 19†.08
2 0.0308 31† 198.6†∗ 198.6†
3 0.0009 31† 0.00∗ 0.00
Overall expected coxt: 268.4
8

Figure 3.†
Sampling with the aid of xample of xize 3 without replacement
32 GŁaptev ¥: Degsssoı tŁeovy

of acting with the aid of the xample ix 8268.48. fŁe expegted ca1ue of tŁe
sıfovmatsoı pvocsded by tŁe samp1e (EVSI[ ix the difference between thexe
two quantitiex:
EVSI = 4†0 − 268.48 = 181.†2.
The advantage of acting with the help of a xample of xize 3 ix partially
offxet, however, by the coxt of taking thix xample. Since the coxt of inxpection
ix 84† per item inxpected, the coxt of taking the xample ix 813†, and the
expegted ıet gası fvom tŁe samp1e (ENGS[ ix

ENGS = (EVSI[ − (Coxt of xample[ = 46.†2.

Since the ENGS ix poxitive, acting with the help of the xample of 3
itemx ix preferable to not taking the xample.

Table 3.11
Expected net gain of xampling
Sampl Overall Exp. value of Coxt of Exp. net gain
e exp. coxt, xample information, xampling of xample,
xize, ,
n OEC EVSI1 SC2 ENGS3
1 383.30 66.F0 4† 21.F0
2 323.09 126.91 90 36.91
3 268.48 181.†2 13† 46.†2
4 218.F3 231.2F 180 †1.2F
† 1F3.26 2F6.F4 22† †1.F4∗
6 131.63 318.3F 2F0 48.3F
F 93.†† 3†6.4† 31† 41.4†
8 †F.96 392.04 360 32.04
9 2†.F0 424.30 40† 19.30
10 0.00 4†0.00 4†0 0.00
1
EVSI=4†0−OEC SC=4†n ENGS=EVSI−SC
2 3

Thix doex not necexxarily mean that a xample of xize 3 ix optimal. Table
3.11 and Figure 3.6 xhow the overall expected coxt, the expected value of
the xample information, the coxt, and the net gain for all other poxxible
xample xizex. Thexe quantitiex were calculated uxing a xpecial computer
program in exactly the xame way ax for a xample of xize 3.
Ax the xample xize increaxex, the EVSI alxo increaxex, but, in thix caxe,
at a decreaxing rate. The coxt of xampling ix a linear function of the xize of
the xample. The expected net gain reachex a maximum of 8†1.F4 at n = †.
A xample of xize †, therefore, ix optimal.
 ¥.11 Optsma1 aggeptaıge samp1sıg
33

Figure 3.6
Optimal xample xize, xampling without replacement

Table 3.12
Overall expected coxt of acting with the aid
of a xample of xize † without replacement
Number Proba Expecte
defective - Bex d coxt of
, y bility, t bext act
p(y[ act
0 0.4028 Full Inxpection 22†.00
1 0.†028 No Inxpection 13†.36
2 0.0861 No Inxpection 169.36
3 0.0083 No Inxpection 0.00
1.0000 OEC: 1F3.26

Table 3.12 providex greater detail on the optimal xampling plan. Thix
plan callx for taking a xample of † componentx without replacement, ac-
cepting the lot without further inxpection if the number of defectivex in the
xample ix 1 or more, and inxpecting all componentx in the lot if the number
of defectivex in the xample ix 0.
 34 Chaptev S: Decssson theovy

3.fi2 I N S U M M A ÆY
The emphaxix in thix chapter hax been on buxinexx problemx in which the
conxequencex of each alternative-event combination can be accurately ex-
prexxed in dollarx of profit, revenue, or coxt.
A payoff table xhowx the available alternativex, the poxxible eventx and
their probabilitiex, and the axxociated monetary conxequencex.
A reaxonable (but by no meanx unique[ criterion of choice ix the max-
imization of expected profit or revenue, or the minimization of expected
coxt.
When alternativex and eventx xucceed one another in xtagex, a decixion
tree ix often uxeful in determining the bext xtrategy. The procedure ix to
work backwardx from laxt to firxt choicex, calculating the expected payoff at
each event node, and xelecting the bext alternative at each decixion
node.
The net value of an additional xource of information ix ximply the dif-
ference between the bext one can achieve uxing thix information after xub-
tracting the coxt of obtaining it, and the bext that can be achieved without
the information.

PÆOBLEMS
3.fi The annual premium for a preferred homeowner‘s insurance policy,
providing a ”package“ coverage of up to 8†0,000 on the house, 82†,000 on the
contents, and 8fi00,000 for personal liability, with 8fDD deductsb1e, is about 8fifi0.
(Preferred rates apply essentially to houses less than 2† years old, insured to
nearly their replacement value, with no more than one insurance claim in the
past three years.) The ”8fi00 deductible“ means that if the amount of the claim is
8E, the insurance company will pay ( E − fi00) dollars if E 3 fi00, and 80
otherwise. The premium of the same preferred insurance package but with 8†D
deductsb1e is 8fi20 per year. For the purpose of this exercise, suppose that the
relative frequency distri-
bution of Claim,
recent claims by homeowners
8 Rel. frequ. is as follows:
Claim, 8 Rel. frequ.
0 0.30 60 0.0†
fi0 0.0† F0 0.0†
20 0.0† 80 0.0†
30 0.0† 90 0.0†
40 0.0† fi00 0.0†
†0 0.0† More than fi00 0.20
fi.00

Assume that all claims are reported regardless of the amount of the claim or the
deductible.
You are the owner of a house and the package insurance coverage is exactly
what you need. Should you buy the policy with 8fi00 deductible or the policy with
8†0 deductible?
Show all calculations in the form of a payoff table. Explain why the distri-
bution of past claims is or is not relevant to your decision.
 Pvob1ems 35

2.A merchant buys a perishable product for 8c per unit at the beginning of a day
and sells i t during the day for 8p. Any stock remaining unsold at the end of the
day must be disposed of at the reduced price of 8s per unit ( s 3 p). If the
demand is greater than the quantity available for sale, there is a ”goodwill cost“ of
8g per unit of excess demand. Let p(x) be the probability distribution of demand,
E.
Show that the optimum order quantity, q∗, is the largest value of q for which
p‡ g− c
P v ( E 3 q) 3 .
p‡ g− s

3.A merchant buys a certain product for 80.89 and sells i t for 8fi.39. Any items
unsold at the end of the period are disposed of at a price of 80.39. The probability
distribution of demand is as follows:

Demand, Probability,
x p(x)
†0 0.fi†

60 0.2†
F0 0.20
0.fi†
80
90 0.fi0
fi00
0.0
(a)Assuming that there is nofifi0 8 associated with being out of stock,
goodwill loss
how many units should be purchased? 0.0
(b)If the goodwill loss is 0.20 per unit of Funsatisfied demand, how many units
should be purchased? fi.00
(c)The merchant has traditionally stocked 80 units. When shown the answer
to (a) above, the merchant states that a goodwill cost for being out of stock has
been incorporated. What is the implicit per unit goodwill cost associated with
the inventory policy of 80 units?
4. In a given decision problem, the cost of act s is

c s (x) = (x − a s ) 2 (s = fi, 2,. . . , n).

In words, the cost is a quadratic function of x, and as are known constants. Show
that the optimum act is the one that minimizes (µ − a s ) 2 .
5. If the payoff of an act is given by

c(x) = a ‡ bx ‡ dx 2 ,

show that the expected payoff of the act is given by

E[c(x)] = a ‡ b E ( E ) ‡ d{Y a v ( E ) ‡ [ E (E )] 2 } ,

where E ( E ) and Y av(E) are the mean and variance of the probability distribution
of E .
36 Chaptev S: Decssson theovy

3.6 Verify the omitted calculations of Example 3.6.

3.Y Using the same data as in Example 3.6, calculate the overall expected profit
of an (s = fi, S = 2) inventory policy if the inventory at the beginning of week fi
is (a) fi unit; (b) 2 units; (c) 3 units.

8.Using the same data as in Example 3.6, calculate the overall expected profit of
an (s = fi, S = fi) inventory policy if the inventory at the beginning of week fi is
fi.

9.A problem frequently faced by production departments concerns the number of

items that must be scheduled for production in order to fill an order for a specsßed
number of items. Each item may be good or defective, and there is uncertainty
about the total number of good items in the production run. If too few items are
produced, there may not be enough good items to fill the order; if too many items
are produced, there may be more good items available than are required and the
surplus items will be wasted.
As a simple illustration, consider the case of a production manager who has
an order for † units of a particular product. The cost of setting up a production
run is 8fi†0 regardless of the size of the run. The variable manufacturing cost
amounts to 820 per unit produced. For a number of reasons, i t is not possible to
test each unit before the next unit is produced; rather, a11 the items in the run are
tested at the same time, and the good items are separated from the defective ones
at that stage. From past experience, the long-run fraction of defective units is
estimated to be about fi0%.
If there are fewer than † good units in the first run, a second run must be
scheduled. The same comments apply for the second run as for the first. If the
number of good units in the second run falls short of the number required, a third
run may be necessary. Theoretically, a fourth, fifth, . . . , run may be required if
the number of good units in three runs still falls short of the requirements.
If more good units are produced in any one run than are required, the surplus
good units are sold as scrap at 8fi0 per unit. Defective units are worthless.
(a)Under what conditions is i t reasonable to suppose that the probability
distribution of the number of good units i n a run of n units is binomial with
parameters n and p = 0.90? When answering the following questions, assume that
these conditions are satisfied.
(b)In order to simplify the problem initially, assume that if a second run is
required, i t is possible to avoid having any defectives in that run by employing
workers with more skill. The variable manufacturing cost in the second run will
increase to 840 per unit. How many units should be produced in the first run so
that the expected cost per order is minimized?
(c)Suppose that if a second run is needed, i t will be made under the same
conditions as the first. However, if a thsvd run is required, skilled workers will be
employed, there will be no defective units in this run, and the variable cost will be
840 per unit. Without doing any calculations, explain how you would determine
the policy that minimizes the expected cost per order.

3.fi0 Vendex operates a downtown luncheon counter and can count on the pa-
tronage of a regular clientele from neighborhood offices. However, the fraction of
customers showing up each day depends very much on the weather. Every morn-
ing Vendex must decide how much food to stock for noon. Let us suppose that the
alternatives are:
 Pvob1ems 3Y

a fi : Stock for the entire clientele

a 2 : Stock for half the clientele a 3 :
Close the shop

Vendex has calculated the following payoff table for

this problem:

Payoff table
(In hundreds of dollars of profit)
Acts R&C R &Events W NR&C NR & W
a fi 0 0 0 2
a2 fi fi fi fi
a3 2 0 0 0

where R stands for rain, NR for no rain, C for cold, and W for warm weather.
(Thus, for example, NR&C stands for no rain and cold weather.)
Two popular weather forecasts are given each morning. Since they are given
at the same time, i t is not possible to use both. The first station (A) specializes in
forecasts of the type: ”Warm weather today,“ and ”Cold weather today.“ The
second station (B) specializes in forecasts of the type: ”Rain today,“ and ”No rain
today.“ Fortunately, both stations have kept records of the success of their
predictions in the last 200 days:

Station A
Actual weather
Forecast R&C R & W NR&C NR&W
”Cold“ 40 0 30 20
”Warm“ 20 20 fi0 60

Station B
Actual weather
Forecast R&C R & W NR&C NR & W
”Rain“ †0 fi0 fi0 20
”No Rain“ fi0 fi0 30 60

The tables show the number of days in which the indicated combinations occurred.
For example, 40 in the upper left hand corner of the table for Station A indicates
that in 40 of the last 200 days the station had predicted cold weather and the
actual weather was rainy and cold.
Which station should Vendex rely on? Would Vendex be better off by ignoring
both stations?

3.fifi In a given decision problem, the profit (or cost) of each act is a function
of the value of a random variable E . Another random variable, Y, could be used to
improve decisions. Show that, if E and Y are independent, the expected value of
the information provided by Y is zero.
38 Chaptev S: Decssson theovy

3.fi2 Two processes (A and B) are available in the manufacture of a certain type
of paper. Both processes convert raw material into the finished product. The raw
material comes in two quality grades (Grade fi and Grade 2). Batches of raw
material come from various sources; i t is impossible to determine their quality
without the use of special tests. I t is estimated, however, that about 60% of the
batches are Grade fi and 40% are Grade 2.
The current practice is to use process A on all batches of raw material re-
gardless of their quality. The plant engineer recently suggested that a test be used
to determine the quality of the raw material. The test costs 830 per batch, and
pronounces the batch as either ”Class fi“ or ”Class 2.“ I t is hoped that these two
categories correspond to the two quality grades, but, as can be seen from the
following table, the test results are not infallible:

Test result
Actual quality ”Class fi“ ”Class 2“ Total
Grade fi 40 20 60
Grade 2 fi0 40 †0

As indicated above, of 60 Grade fi batches tested, 40 were pronounced ”Class

fi“ and 20 ”Class 2“; similarly, of the †0 Grade 2 batches tested, fi0 were pro-
nounced ”Class fi“ and 40 ”Class 2.“
The cost of manufacturing depends on the quality of raw material and the
process used, as the following table shows:

Cost per batch

Raw material Process A Process B
Grade fi 8fi00 8fi60
Grade 2 8fi†0 8fi20

Should the plant engineer‘s suggestion be adopted and the test used to examine
all batches of raw material? Explain clearly, showing all calculations.

3.fi3 For some time now, the production manager of ABC Instruments was con-
cerned about the high defective rate of one particular product. This product is
used as a component of an aircraft instrument assembly and is sold exclusively by
ABC at 82†0 per unit. The component costs only 8fi†0 per unit but ABC is able
to charge a high mark-up since i t has developed the component. According to the
terms of the agreement between ABC and its clients, a penalty of 8†0 will be paid
by ABC for each defective component discovered by the client. In addi- tion,
ABC must provide a good component to replace the defective one. Defective
components can be turned into good ones at an additional cost of 820 per unit.
Despite efforts to control the quality of production, nearly 2†% of all units (or
fi out of 4) are returned to ABC as defective. There seems to be some hope for the
future, however. One of ABC‘s engineers has developed an instrument for testing
the component with apparently good results. In a recent test, the instrument was
applied to fi00 known good components and fi00 known defective components with
the following results:
 Pvob1ems 39

Instrument
Actual quality reading
Good DefectiveF
”Good“ 20
”D efect i ve“ 0 80
30
fi0 fi00
0
Thus the instrument gave the correct reading on F0 of the known good and 80 of
the known defective units.
I t was suggested to the production manager that the test be adopted and all
units tested before delivery. I t was estimated that the tests would add about 8fi0
to the unit cost of the component. All units identified as ”Good“ by the test would
be shipped out as usual. All units identified as ”Defective“ would be reworked;
the rework would cost 820 per unit, but reworked components would definitely be
good.
Should the manager authorize the use of the instrument? How much is i t
worth to ABC?
3.fi4 Verify the calculations in Table 3.fi0 for 0, 2, and 3 defectives.
3.fi5 For the problem described in Example 3.F, verify the overall expected
cost, the expected value of sample information, the cost of sampling, and the
expected net gain for a sample without replacement of size: (a) fi; (b) 2; (c) 4; and
(d) all other sizes shown in Table 3.fifi.
3.fi6 Consider Example 3.F, except that the distribution of the lot number
de- fective, E , is:

x p(x)
0 0.F
fi 0.2
2 0.fi

fi.0

Calculate the overall expected cost of a sample of size 2 without replacement. Use
the table of hypergeometric probabilities shown in Appendix 4J.
3.fiY The following table shows the probabilities of Y = 0, fi, 2 and 3 defectives in
a random sample of n = 3 items drawn vsthout replacement from a lot of N = 20
items of which x are defective. These are hypergeometric probabilities, and can be
obtained from a table or special computer program.

Lot number P H (Y = y|N = 20,n = 3,h =

x) defective,
fi x Y = 0.fi†00
0.8†00 0 Y = 0.0000
fi Y = 20.0000
Y= 3
2 0.Ffi†8 0.2684 0.0fi†8 0.0000
4 0.49fi2 0.42fi0 0.0842 0.003†

Assuming that the probabilities of the lot containing fi, 2, and 4 defectives are 0.3,
0.6, and 0.fi respectively, and that the costs of inspection and replacing a defective
40 Chaptev S: Decssson theovy

are the same as in Example 3.F, determine the overrall expected cost, the value of
sample information, and the expected net gain of a sample of size n = 3 without
replacement.