Wentzel Operations Research Mir 1983
Wentzel Operations Research Mir 1983
Wentzel Operations Research Mir 1983
WENTZEL
OPERATIONS
RESEARCH
A Methodological Approach
E. C. BeHmenb
HCCJlEflOBAHME
OnEPALlklPl
3aaaqn
npnHi^nnu
M eTOA OJIOrHH
HsAaiejibCTBo «HayKa»
MocKBa
E. S. Wentzel
OPERATIONS
RESEARCH
A M ethodological A pproach
Mir Publishers
Moscow
First published 1983
Revised from the 1980 Russian edition
Ha amjiuucKOM si3um
FOREWORD
LINEAR PROGRAMMING
E.S. Wentzel
C h ap ter 1
J* H x
6* 7 * 8 * </• >/0
17*16 ’ 15* ff.. \f1
18* 13*
20 *
0 Wf
Fig. 2.3-1
Fi an a l2 a 13 Cl
F2 a 2l a 22 a 23 C2
a 3l a 32 a 33 C3
F4 a 41 a 42 a 43 C4
C= (3.1-2)
i—1
Thus, we have obtained the objective equation and
it is linear. Next we write formulas for the constraints
of protein, carbohydrate, and fat content. Considering
that x1 units of food Fx contain alxxx units of protein,
x 2 units of food F2 contain a21x2 units of protein, and
so on, we arrive at three inequalities
^11^1 a2lX2 4“ ^31*^3 "1“ ^41^4 ^ ^1
#12^1 4“ @22X 2 4“ ^ 3 2 X 3 4“ ^42*^4 ^2 (3.1 “3)
#13^1 + a 23X 2 a33x 3 -j- d43x 4 b3
constraining the variations of decision variables xx, x 2,
x3l and x 4.
The problem reduces thereby to the following:
find nonnegative values of decision variables
#i, x 2, x 3, and x 4 such that satisfy the constraint
inequalities (3.1-3) and simultaneously minimize
the linear objective function of these variables:
4
C= 2 c ^ ^ m in
i= 1
What we have obtained is a typical problem of linear
programming. We postpone the discussion of its solu
tion methods until later to illustrate three more prob
lems of the kind.
5 -0 4 3 2
66 Ch. 3 Linear Programming
P r o d u c ts
R aw
[ m a te r ia ls
Pi P2 Pz
m1 A ll a 2l a 81
m2 a 12 a 22 a 82
m9 a lS a 28 a 88
a 14 a 24 a 84
C lo th
L oom
ty p e
Gi g2 Gz
1 all a 12 a 12
2 a 21 a 22 a 23
or in a concise form
3 2
L = % Cj dijx ij (3.1-13)
i= 1 i= l
This linear function of six arguments is to be maxi
mized, i.e. L =»• max.
In words, this linear program may be spelled thus:
find nonnegative values of variables xlly x12, . . ., x 23
such that satisfy the constraint inequalities (3.1-10),
(3.1-11), and equalities (3.1-12) and maximize the
linear function of these decision variables (3.1-13).
The total count of the problem constraints amounts to
eight: six inequalities and two equations.
A raw material allocation problem. A large corpo
ration has three production plants, Plf P 2l and P 3
in demand each, respetively for a±, a2, and a3 units
of a raw material. Five sources variously distanced
from the plants can supply this material at various
transportation costs per unit tabulated in Table 3.1-4.
The capacities of the suppliers are limited by the
available stocks so that the sources 5 lt S 2, S 3, S 4,
3.1 Linear Programming Problems 71
and S 5 can supply not more than 6 1? b2, b3l fc4, and b5
units, respectively. It is required to design a ship
ment schedule, i.e. to allocate the amounts sup
plied by each source to each plant, such that the de
mands are satisfied at least possible transportation
costs.
Table 3.1-4
S u p p lie r
r ia n t
Si S2 S3 S4 s5
00
#15 + X 25 + #35 ^ ^5
Zf = 2 S C ijX u ^ m in (3.1-17)
i= 1 j= l
xx — 2x2 + 3x 3 ^ 1 0
and maximize the linear function of these variables:
L = 4#! — x 2 + 2# 3 max (3.2-4)
We start with providing inequalities (3.2-3) with
zeroes on the right-hand side, namely,
3xx + 2x2 — x 3 — 4 0
(3.2-5)
—xx + 2 x 2 — 3 ^ 3 + 1 0 ^ 0
74 Ch. 3 Linear Programming
1 2 • • • ^mri
For shortness, we will designate this matrix by the
symbol [c*7].
3A The Transportation Problem 87
x m 1 x m 2 • • • x mn
x m l "1” x m 2 T“ • • • ~\~ x m n —
# 1 n “1” # 2 n ”f “ • • • x mn ^n
variables
k = mn — (jn -f- n — 1) = (m — 1) (n — 1)
As will be recalled, an optimal solution to an LP
problem is attained at a vertex of the feasibility area,
where at least k variables assume the value of zero.
For the case in question, an optimal shipment schedule
requires that at least (m — 1) (n — 1) shipments must
vanish, that is, no shipment is effected from respective
origins to destinations.
A shipment schedule will be termed feasible if it
satisfies the constraints (3.4-4) and (3.4-5), i.e. all
demands are satisfied and all supplies are exhausted.
A feasible shipment schedule will be referred to as
basic feasible schedule (solution) if it has at most
m -\- n — 1 positive shipments (the other shipments
are zeroes), otherwise it is degenerate. A schedule [xtj]
will be termed optimal if among all feasible sched
ules it results in the lowest total transportation
costs.
Owing to the special structure of the transportation
problem its solution requires no lengthy recursive
manipulations of the systems of equations. Instead,
the search for optimal solution is performed within
a table containing the data of the problem set up in
a certain order. The entries to the table are origins
and destinations, demands and supplies, transportation
costs, and, if allocated, shipments positioned in the
respective cells. A transportation table has m rows
and n columns. In the right-hand corner of each cell
we place the transportation cost ctj payed to transport
a unit of commodity from Ai to Bj. A shipment will
be placed in the center of the cell. The cell correspond
ing to the shipment from Ai to Bj will be shortly
denoted as (i, /) shipping route.
90 Ch. 3 Linear Programming
D estin ation
Sup
O rigin p ly
Bi b2 b3 Bi b5
13 7 14 7 5
30
11 1 6 8
12
I8 48
6 10 10 8 11
^8 20
14 8 10 10 15
A4 30
Table 3.4-2
D e s tin a t io n
Sup
O rig in ply
Bi Ba b3 b4 b6
13 7 14 7 5
Ai 30
18 12
11 8 12 6 8
A2 48
15 33
6 10 10 8 11
20
9 11
14 8 10 10 15
a4 30
4 26
D e
m and 18 27 42 15 26 128
D estin a tio n
Sup
O rigin p ly
Bi [B2 *3 £4 £5
13 7 14 7 5
30
18 12
11 8 12 6 8
48
15 22 11
6 10 10 8 11
^3 20
20
14 8 10 10 15
/14 3>
4 26
D e 27 15 26 128
mand 18 42
+ 15 x 8 -f 33 x 12 + 9 x 10 + 1 1 x 8 + 4 x
10 + 26 X 15 — 1442 whereas those in Table 3.4-3
amount to L2 = 18 X 13 + 12 X 7 + 15 X 8 +
22 X 12 + 11 X 6 + 20 x 10 + 4 x 10 + 26 x 15
= 1398.
Thus we have lowered the total shipping costs by
1442 — 1398 = 44 monetary units. The result, though,
may have been predicted with no resort to the total
shipping costs account. To prove, the algebraic sum
of shipping costs of the moving route cells taken with
a plus sign if the shipment of the cell increases and
with a minus sign if decreases—called the net cost—
amounts for the above example to 6 — 8 + 1 0 —
12 = —4. It means that in moving one unit through
a closed set of routes (the so-called evaluation method)
the shipping costs diminish by four. We have moved
1 1 units which means that the costs must have been
diminished by 11 X 4 = 44 monetary units, which is
obviously the case.
We now may conclude that essentially all we need to
optimize shipping schedule is to transfer shipments
round closed systems of routes having a negative net
cost.
Linear programming has the following statement
proved for a nondegenerate shipping schedule.
For each unfilled cell in a transportation table
there exists a unique closed set of routes with one
turning point in this cell and the others in filled
cells.
Therefore, looking for cost saving paths of negative
net cost we must evaluate those unused cells with low
shipping costs. If such a cell is available, it should be
included in an appropriately chosen closed set of
routes. The next step is to evaluate the net cost of
moving a unit through this system and if it turns out
3.4 The Transportation Problem 95
Destination
Origin Supply
B, B2 B3 4 Bs
t8 13 12 1_7_ Ik I7 k 30
A,
Az
11 22 12 / / k 1 8 48
©■--- ,
A3
6 k 10
1 k ! 11 20
20 I |
14 8 10 i 15 30
a4
-4 . 16
Required 18 27 4-2 15 26 128
(3-4-7)
i= 1 j- 1
When 2 i=i ai > 2 j=i bj (the supplies are in
excess of the requirements) all the demands can be
satisfied, butnot all the supplies will be exhausted.
Thissituation may be handled by introducing a
fictitious destination, S f, to absorb the difference
(slack)
m n
0 , - 2 bj (3-4-8)
j=i 5=1
What should then be the costs of shipment from
origins A t to the fictitious location? Naturally, they
should be assigned zero values since in reality nothing
will be shipped to B t. Hence, for all origins cn = 0.
We augment the transportation table with a dummy
column corresponding to destination B t and contain
ing zero transportation costs. Solution is then attemp
ted in the usual manner to result in an optimal trans
portation schedule
x n x i2 • • • x m x i t
*^21 ^"22 • • • X2n X2t
x ml x m 2 • • • x mn x m t
DYNAMIC PROGRAMMING
Fig. 4.1-1
— 10 ■ -10 —I—8 - B
1
11 1Z 10 10 11 11 13 15
-1 4 - - 10- - /2 - j — 14 —|-
H -
10 13 11 15 10 to 9 8
-1 3 - ■10 8 — |— 10—|— .
I 9
\ ~ w ~\ ~
11 15 12 14 15 10 9 11
H+
12 15-
14 11
- 11 -
10
- 1 3 — \— 1 6 — f - 12—
12 11 10
10
-+
12
*)— ;1 0 — — 1 5 — — 13- - 1 5 —— 1 2 - -to
rn 13 12 12 11 10 12 15
1 2 — 1— 1 0 - \ — 1 6 - \ — 1 3 —I-------
A x (East)
Fig. 4.2-2
that is, the first four steps must be to the north, next
two to the east, then one step to the north, and the five
remaining to the east. This completes the solution.
i = 4 i = 2
S
*6(S)TV6(S) x4(S) W4(S) x 3 ( S ) W s ( S ) *a(S) W 2(S) X
I(S) W ^ S )
set up from the upper left corner to the right and then
line hy line downwards. The decision at the 5th stage
is enforced: all the available funds are invested; the
decisions at all other stages have to be optimized.
The sequential backwards optimization from stage 5 to
stage 1 yields a complete list of sequential decisions and
the total optimal return W* for the whole operation,
amounting to 5.6 in this case. The last two columns
have only one row because the initial state of the
system is known exactly: S 0 = R = 10. The optimal
decisions at all stages are confined in frames. The
optimal sequence of investments reads thus: allocate
two units of 1 0 to enterprise 2 , five units to enter
prise 2, two units to enterprise 5, none to enterprise 4,
and a unit to enterprise 5. With this allocation the
total profit attains a maximum of 5.6.
To provide the reader with an insight to how the
table was set up, we perform a sample computation.
Suppose, for example, we are to optimize a decision
x 3 (7), that is, to decide on how many units we should
allocate at stage 5, where we have arrived with S = 7
units of resource available, and what maximum amount
we can earn at all the remaining stages including the
third. Assume all the stages following the third (i.e.
4th and 5th) have already been optimized and, hence,
the columns for these two stages in the table are filled.
Let us find x 3 (7) and W3 (7). To effect this, compose
an additional table (see Table 4.2-3). Its first column
lists all possible investments x at the third stage, which
cannot be more than S = 7. The second column lists
all what remains from S = 7 after this investment.
The third column presents the return at stage 3 from
the investment of amount x in enterprise 3 (filled with
reference to the column for f 3 (x) in Table 4.2-1). The
fourth column gives the optimal return at the remaining
4.2 Solving Dynamic Programming 127
7 0 1.8 0 1.8
6 1 1.7 1.0 2.7
5 2 1.6 1.3 2.9
4 3 1.4 1.6 3.0
3 4 1.2 2.3 3.5
El
1
5
6
1.1
0.6
2.5
2.6
M
3.2
0 7 0 2.7 2.7
Table 4.2-4
Item 1 2 3 4 5 6
Weight 4 7 11 12 16 20
Valuo 7 10 15 20 27 34
Table 4.2-5
z = 6 i =5 i == 4 i == 3 i == 2 i -= 1
s
xi wi xi w. xi xi xi
wl wt *« wl
0 0 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0 0
2 0 0 0 0 0 0 0 0 0 0
3 0 0 0 0 0 0 0 0 0 0
4 0 0 0 0 0 0 0 0 0 0
5 0 0 0 0 0 0 0 0 ' 0 0
6 0 0 0 0 0 0 0 0 0 0
7 0 0 0 0 0 0 0 0 1 10
8 0 0 0 0 0 0 0 0 1 10
9 0 0 0 0 0 0 0 0 1 10
10 0 0 0 0 0 0 0 0 1 10
11 0 0 0 0 0 0 1 15 0 15
12 0 0 0 0 1 20 0 20 0 20
13 0 0 0 0 1 2v) 0 20 0 20
14 0 0 0 0 1 20 0 20 0 20
15 0 0 0 0 1 20 0 20 0 20
16 0 0 1 27 0 27 0 27 0 27
17 0 0 1 27 0 27 0 27 0 27
18 0 0 1 27 0 27 0 27 0 27
19 0 0 1 27 0 27 0 27 1 30
20 1 34 0 34 0 34 0 34 0 34
21 1 34 0 34 0 34 0 34 0 34
22 1 34 0 34 0 34 0 34 0 34
23 1 34 0 34 0 34 1 35 1 37
24 1 34 0 34 0 34 1 35 1 37
25 1 34 0 34 0 34 1 35 1 37
26 1 34 0 34 0 34 1 35 1 37
27 1 34 0 34 0 34 1 42 1 44
28 1 34 0 34 1 47 0 47 0 47
29 1 34 0 34 1 47 0 47 0 47
30 1 34 0 34 1 47 0 47 0 47
31 1 34 0 34 1 47 1 49 0 49
32 1 34 0 34 1 54 0 54 0 54
33 1 34 0 34 1 54 0 54 0 54
34 1 34 0 34 1 54 0 54 0 54
35 1 34 0 34 1 54 | 0 54 1 57 0 57
4.3 General Form of DP Problem 131
MARKOV PROCESSES
Fig, 5.1-2
0 t
Fig. 5.2-2
Fig. 5.2-3
Fig. 5.2-4
2 M *) = 1 (5.3-1)
1
d pj dt = ^ 2 1 /* 2 (^ 1 2 ~t" ^13)7*1
d p j d t = ^ 3 1 Pi - ( - ^ 4 3 ^ 4 ^32^3
2 Pi = i (5.3-6)
2=1
Physically these probabilities are as follows. When t
is at infinity, the system arrives at a steady state con
dition in which the system may change its states at
random, but their probabilities do not depend on time
any longer. The equilibrium probability of state, S t,
may be thought of as the mean relative time the sys
tem occupies this state. To illustrate, if the system
has three states 5 1? S 2, and S 3, and their equilibrium
state probabilities are 0.2, 0.3, and 0.5, respectively,
this implies that in the balanced state the system on
the average spends two tenths of its time in three
tenths in S 2, and a half of the time in S 3.
A way to compute these equilibrium probabilities is
rather easy. If they are constant, their derivatives must
6 This condition is sufficient but not necessary for equilibrium
probabilities to exist.
160 Ch. 5 Markov trocess^s
^k-i,hPh-l — ^U,k-lPk ( 6 . 2- 2)
XU) YU)
Fig. 6.2-2
If the queue is settled and the system is in the steady
state, the mean number of arrivals per unit time (input)
is equal to the number of elements leaving it (output),
both the input and the output have the same rate X.
Let X (t) denote the number of customers arriving
at the system up to time t, and Y (t) the number of cus
tomers leaving it by the same instant t. Both functions
are random and vary in jumps of unit value at the
instants an arrival occurs, X (t), or leaves the system,
Y (t). The forms of both X (t) and Y (t) are shown in
Fig. 6.2-2; both lines vary in steps, the upper is
X (t), the lower is Y (t). Obviously, for any moment of
time t their difference Z (t) = X (t) — Y (t) is but
the number of customers in system. When the lines
X (t) and Y (t) merge, the system has no customers.
Consider now what happens in the long run, taking
a large period of time T (we can mentally extend the
graph far beyond the margins of the page) and comput
ing for it the expected number in the system. It will
be given by the integral of the function Z (t) over this
6.2 Birth and Death Process 173
i
In the quantity TX we can recognize the expected
number of arrivals for the period T. If we divide the
sum of all the times t t by the expected number of cus
tomers, we obtain the expected time of a customer in
system, W. Therefore
L = XW
174 Ch. 6 Queueing or Waiting Line Theory
or
W = L /l (6.2-12)
This result is known as Little's formula which states
for any distribution of arrival and service times, and
any service mechanism that
the expected system time is equal to the expected
number of customers in system divided by the arri
val rate.
In the same manner we could derive a second Lit
tle’s formula relating the expected time in queue, W q,
which is the expected time a customer spends waiting
in line, and the expected number in queue, L q, which is
expected number of customers waiting to be serviced:
W q = L q!\ (6.2-13)
To effect the derivation, it suffices to use as the bot
tom line in Fig. 6.2-2 the function U (t), the number
of customers that leave the queue rather than the sys
tem by time t (if a customer entering the system im
mediately goes to service missing the queue, we may
still assume the arrival joins the queue but waits in
it a zero time).
The Little formulas (6.2-12) and (6.2-13) are signi
ficant results of queueing theory. Unfortunately, most
texts on the subject ignore these formulas (proved in
the general form relatively recently ) . 2
Fig. 6.3-1
graph with the corresponding rates indicated at the
arcs. The transitions from S 0 to S x occur at the rate
X (as soon as a demand arrives the system moves from
S 0 to S x). Generally, the system’s transitions from each
lower (left on the digraph) state to the next (right) state
proceed at the arrival rate X (the upper arcs of the
digraph).
Consider now the rates at the lower arcs in the digraph.
Let the system be in S ly i.e. with one channel busy
6.3 Analysis of Simplest Queueing Models 177
G = 1 - ^ 1 om= 1 - - S j Po (6-3-7)
Multiplying Q by the arrival rate X yields the abso
lute throughput
4 = A£ = x ( l - i £ p 0) (6.3-8)
* = p ( 1 ~ T T Po) (6.3-10)
We recommend the reader to solve a self-check ex
ample. An exchange possesses three trunk lines (n = 3),
the arrival rate of calls put for connection is X = 1.5
(calls per minute); the mean time to serve a call is 2
(minutes). Assuming a Poisson distribution of arrival
and service times, find the steady-state probabilities
and the performance measures, A, Q, P\0$s, and k.
For the sake of comparison, the answer must be this:
p0 = 1/13, Pi = 3/13, p 2 = 9/26, p 3 = 9/26 ^ 0.346,
A & 0.981, Q ^ 0.654, Pi 0 Ss ^ 0.346, and k ^ 1.96.
These results indicate, by the way, that our exchange
is substantially overloaded: of three channels about
12*
180 Ch. 6 Queueing or Waiting Line Theory
two are busy in the mean, and about 35% of all in
coming calls leave unserved. We suggest the reader
evaluates how many channels (trunk lines) the exchange
should possess to supply not less than 80% of sub
scribers with a required line. What proportion of chan
nels will then be idle?
Here there are already some trends to optimization.
To illustrate, each channel has a cost per unit time. On
the other hand, each served call realizes a certain profit.
Multiplying this profit by the average number of calls
served per unit time, A, gives the mean profit rate from
the system. Naturally this profit would increase with
any new trunk line, but the maintenance costs of the
lines also increase. Which will outbalance, the rise
of profit or expenses? It depends on the conditions of
the operation, i.e. the monetary rate per served call
and the maintenance cost of a channel. W ith these
values specified, one can find an optimal number of
lines which would be most efficient from the economic
viewpoint. We leave these computations to the reader,
who is to invent and solve his own example. The inven
tion of problems generally gives more intellectually
than the solution of those posed by somebody.
(ii) Single channel queue of unlimited size. Single-
server queues are most often met in applicational areas.
The examples are a queue to a doctor, a line of foot
ball fans to a ticket seller, an on-line computer perform
ing the subscriber demands in the time sharing mode,
etc. In queueing theory these single-server models are
also of special importance (most analytical expressions
derived so far relate to these systems). Therefore we
shall attach an especial attention to this type of queue.
Let there be a single-server queue with no constraints
imposed either on the queue length or waiting time.
The customers arriving at a rate X are serviced at a
6.3 Analysis of Simplest Queueing Models 181
Fig. 6.3-2
l + P + P2 + . . . + P h+ . . . = T q ^
or for (6.3-11)
Po = 1 ~ P (6-3-12)
The equilibrium probabilities are then found with
the expressions
Pi = PP01 P 2 = P2Po» Ph = P Pof *
which, by virtue of (6.3-12) become
Pi = P (1 — p), P2 = p2 (1 — p),-.-»
pk — p* (1 — p), ... (6.3-13)
The probabilities p0, ..., p k are seen to form a
geometric series with the common ratio p. However
strange it might seem, the largest of them is p0, that
is the probability of idle channel. In other words,
whatever the load of the server, if it is capable of serv
ing the incoming arrivals (p < 1 ), the most proba
ble number of customers in the system will be zero.
184 Ch. 6 Queueing or Waiting Line Theory
L = S V ( l —p)
h=1
or with p (1 — p) factored out ahead of the sum
L= p(l-p) § V 1
ft=1
We may again exercise our ingenuity if we identify
that Apf t - 1 is but a derivative of ph with respect to p,
hence
L = p ( l - P) 2 - £ p h
h=1
Interchanging the places of the differential and summa
tion operators, we get
z' = p ( 1 - p ) ^ S p* (6.3-15)
h=1
6.3 Analysis of Simplest Queueing Models 185
or, finally,
=7— (6-3-20)
Referring to the Little formula (6.2-13) find the ex
pected time in the queue
P2 (6.3-21)
Wg
M l-P )
By this we have found all the sought performance
measures of the system.
As a self work we suggest that the reader should
solve the following example. A single-channel queueing
system is a railway sorting yard which handles a Poi
sson arrival of trains incoming with rate X = 2
(trains per hour). A gathering of a t rain (servicing) takes
a random time exponentially distributed with a mean
of 20 minutes. The yard has waiting room for only two
trains to be positioned on two body tracks; if both
of them have waiting trains, the arriving trains have to
await service on the approach lines. It is required to
find (for the yard operating at the equilibrium condi
tions):
L = mean number of trains bound to the yard,
W = mean time a train spends in the yard (on
the body tracks, on the approach lines,
and in service),
L q = mean number of trains waiting in queue
for gathering (no matter on what track),
W q = mean time a train waits in the queue.
Try also to evaluate:
La = mean number of trains awaiting at the
approach lines, and
6.3 Analysis of Simplest Queueing Models 187
Fig. 6.3-3
O pk Pn
Pi — “Jj” P0 ’ • • • * Ph = i P0 ’ • • • > Pn == Po
(6.3-22)
pn+1 pn+r
Pn+i=~ ^ \ Po' • • • ’ Pn+r~ i ^ i r Po’ •••
Now we will seek for the performance measures of
the system. The easiest to find is the average number
of busy channels k = Mp, = p (this is in general val-
6.8 Analysis of Simplest Queueing Models 180
O R /
Fig. 7.2-2
F (x )= j f(x )d x , (7.2-1)
—oo
F(r)
Fig. 7.3-1
Poi = ^ r - , Pu = ^ f - (* = 0, 1, 2, 3)
Bi b 2 Bn
At a ll a l2 am
a2 a 21 a 22 •• 02n
Table 8.2-2
By B* By Bh
Ay 3 4 5 2 3
A% 1 8 4 3 4
A„ 10 3 1 7 6
Ay 4 5 3 4 8
By By #8 By B5 ai
Ay 3 4 5 2 3 2
A2 1 8 4 3 4 1
A3 10 3 1 7 6 1
a4 4 5 3 4 8
0
Pi 10 8 5 7 8
Bt B, Ba Bt a-i
Ai 2 4 7 5 2
At 7 6 8 7 JT j
a3 5 3 4 1 1
8 7
P;
7 i6
In this example the minorant game value is equal to
the majorant one: a = p = 6. What is the import of
this? Minimax strategies of A and B will be stable. As
long as both of them follow these strategies, the payoff
is 6. Let us see what will occur if we (A) find out that
the opponent (B) abides by j52? Nothing whatsoever
will change* Just because any deviation from strategy
A 2 may only impair our situation. In exactly the same
way, information obtained by the opponent will not
make him abandon his strategy Z?2. The pair of ^42
and B 2 has the property of equilibrium (balanced pair),
and the payoff (in this case 6) for this pair is called the
saddle point of matrix1. An indication of a saddle point
1 The term ‘saddle point1 has been borrowed from geomet
ry where it means a point of the surface at which simultaneously
a minimum is achieved in one coordinate, and a maximum in
the other.
224 Ch. 8 Gamo Theory for Decision Making
15-0432
226 Ch. 8 Game Theory for Decision Making
Table 8.2-5
Bi B2 Bs CLi
2 —3 4
CO
1
A„ —3 4 —5
1 1
^3 4 —5 6
Vi |T j 4 6
lh B2 B9 B4 B,
Ai 4 7 2 3 4
a2 3 5 6 8 9
As 4 4 2 2 8
a4 3 6 1 2 4
A» 3 5 6 8 9
230 Ch. 8 Game Theory for Decision Making
Table 8.3-2
B i b 2 Bg B 4 B5
A x 4 7 2 3 4
a 2 3 5 6 8 9
Ag 4 4 2 2 8
Bt Bg Bt B s
A t 4 2 4 2
3 6 At
A* 3 6
4 2 Ag
Ag
... Bn
Ai «ll a 12 •M a in
A2 fl 21 a 22 ... a 27l
••• ... ... ••• • •.
am i a m2 . . . amn
B1 b2 B„
Al 7 2 9
a2 2 9 0
A» 9 0 11
ft i Bi B3 j Ai Aa A3 V V V*
b2
1 3 9 0 11 2 2 9 0 0 9 4 .5
2 2 11 9 11 2 4 18 0 4.5 9 6.75
3 2 13 18 11 3 13 18 11 3.67 6 4.84
4 2 15 27 11 3 22 18 22 2.75 5.50 4.13
5 1 22 29 20 3 31 18 33 4.00 6.60 5.30
6 3 31 29 31 2 33 27 33 4.84 5.50 5.17
7 1 38 31 40 2 35 36 33 4.43 5.14 4.79
8 2 40 40 40 2 37 45 33 5.00 5.61 5.30
9 2 42 49 40 3 46 45 44 4.45 5.11 4.78
10 1 49 51 49 1 53 47 53 4.90 5.30 5.10
11 3 58 51 60 2 55 56 53 4.64 5.09 4.87
12 2 60 60 60 2 57 65 53 5.00 5.41 5.20
13 2 62 69 60 3 66 65 64 4.61 5.07 4.84
14 1 69 71 69 1 73 67 73 4.93 5.21 5.07
15 3 78 71 80 2 75 76 73 4.74 5.06 4.90
8.3 Resolving Finite Games 237
The reader has thus got some insight into the theory
of noncooperative games and methods of solving ma
trix games.
A few words of caution about this theory and its
applicability are in order here. At the time of its con
ception the theory of games was looked to for the
selection of decisions in conflicting situations. The
hopes were justified but to a slight degree.
Above all, in actual practice strictly non-cooperative
games are a not very frequent occurrence—may be
only in real games (draughts, chess, cards, etc.). Save
for these artificial situations where one side seeks at
all costs to maximize, and the other to minimize
the payoff, such conflicts hardly can occur. It would
seem that warfare is a field ideally suited for game
theory. After all, here we encounter the most “fierce”
antagonisms and sharply conflicting interests 1 But
the conflict situations in this field appear to be but
rarely reducible to two-person zero-sum games. Non-
cooperative scheme is, as a rule, only applicable to
small-scale operations of marginal importance. For
example, side A is a group of aircraft attacking an
object, side B is the antiaircraft defence of the object,
with the first seeking to maximize the probability of
destroying the target, and the second to minimize.
Here the scheme of the two-person zero-sum game can
find its application. Now consider a more complex
example: two forces of units (types of tanks, aircraft,
ships) are engaged. Each side wishes to hit as many
hostile combat units as possible. In that case the sit
uation loses the purity of antagonism: it is no longer
reducible to a two-person zero-sum game. If the sides’
goals are not directly conflicting but just not coin
cident, the mathematical model becomes more involved:
here we are not satisfied just with a victory, which
8.3 Resolving Finite Games 230
n2 Nn
Ni N2 N3 N4
1 2 3 5
A2 7 4 4 5
^3 3 4 4 1
A4 7 4 2 2
rU — Pj a H• (8.4-1)
16*
244 Ch. 8 Game Theory tor Decision Making
N* N* Nt
At 1 4 5 9
a2 3 8 4 3
As 4 6 6 2
P/ 4 8 6 9
Table 8.4-4
N,
At 3 4 1 0
A* 1 0 2 6
a3 0 2 0 7
Ni N* N3
A, 20 30 15
A2 75 20 35
25 80 25
A4 85 5 45
Table 8.4-6
N* N 3 0C;
At 20 30 15 15
A; 75 20 35 20
^3 25 80 25 iM!
A4 85 5 45 5
^2 N 3 Yi
A\ 65 50 30 65
A 2 10 60 10 0
^3 60 0 20 |60|
a4 0 75 0 75
Table 8.4-8
Ni n 2 N3 at hi
At 20 30 15 15 30 21
a2 75 20 35 20 75 42
^3 25 80 25 25 80 E3
A4 85 5 45 5 85 37
N2 N3 a; ©i hi
At 19 30 41 49 0 49 31
a2 51 38 10 20 10 51 26
^3 73 18 81 11 11 81 El
By the Wald criterion the optimal strategy is A x,
by the Hurwicz criterion (with x = 0.6) A s. We will
now see what the Savage criterion will tell us. The
risk matrix with an additional column containing
row maxima y* is given in Table 8.4-10.
The minimum in the last column is 38, so that the
Savage criterion, just as the Hurwicz criterion, votes
for A v
252 Ch. 8 Game Theory for Decision Making
Table 8.4-10
N, n 2 N< Vi
Ai 54 8 40 0 54
a2
. 22 0 71 29 71
^3 0 20 0 38