Games Theory

Preliminaries
1. Matrix games with zero sum
1.1. Games with saddle point
1.2. Games without saddle point

2. Two persons games with arbitrary sum (bimatrix games)

2.1. Arbitrary sum non-cooperative games
2.2. Arbitrary sum many person cooperative games
2.1.1. Characteristic function of arbitrary sum many person cooperative games
2.1.2. Shapley Value of arbitrary sum many person cooperative games

Preliminaries
EXAMPLES
1) Prisoner's Dilemma

Two prisoners are each suspected of a serious crime. They are held in separate cells,
unable to communicate with each other or the outside world. Each is invited to confess
and implicate the other prisoner. Each is aware of the following consequences:

If one prisoner alone confesses then he obtains complete immunity (utility level 3)
and the other prisoner gets life imprisonment (utility level 0).

If both confess both get a substantial sentence, but less than life (utility level 1).

If neither confesses there is enough evidence to convict them both of some minor

Deny

(2, 2)

(0, 3)

Confess

Player 1

violation for which they are fined (utility level 2).

(3, 0)

(1, 1)

Deny

Confess

Player 2
If there were some way of each guaranteeing to the other that he would not confess then
they could secure for themselves the outcome (2, 2). But, under the circumstances, no
such enforceable guarantee is possible. Each has an incentive to confess immediately, for
fear of being implicated by the other. The solution is (1, 1).

2) Battle of the Sexes

A couple wants to decide on an evening's entertainment. He prefers to go to the West End

(there's car race); she wants to go to the East End (there's a new play). If they go as a
couple each person gets utility level 2 if it is his/her preferred activity and 1 otherwise.
However, for each person the evening would be ruined if the partner were not there to

West

(2, 1)

(0, 0)

East

Player 1

share it (utility level 0).

(0, 0)

(1, 2)

West

East

Player 2
3) Managing retention groups
Consider a group (Group I) of n1=100 individuals. Each of them is exposed to a possible
loss of 1, with a probability q1=0.1. Assume these persons decide to form a risk retention
group, a small insurance company to cover them against that risk. The premium charged
will be such that the ruin probability of the group is less than 0.001. By considering that
the risks are independent, and using the normal approximation of the binomial
distribution, the group must have the total funds equal to:

P1 n1q1 3 n1q1(1 q1) =10+9=19

Hence each person in Group I will pay, in addition to the net premium of 0.1, a safety
loading of 0.09.

Another group - Group II - consists of n2=100 persons exposed to a loss of 1 with a

probability q2=0.2. If they form their own retention group under the same conditions, the
total premium will be

P2 n 2q 2 3 n 2q 2 (1 q 2 ) =20+12=32.
Therefore, each person in Group II will pay the net premium of 0.2 and a safety loading
of 0.12.
Suppose now that the two groups decide to join and form one single company. In order to
ensure that the ruin probability shall be less than 0.001, the new company must have
funds amounting to:

P12 n1q1 n 2q 2 3 n1q1(1 q1) n 2q 2 (1 q 2 ) =10+20+15=45.

One can see from this example that it is to the advantage of the two groups to form one
single company. Total payment of premium will then be P12=45, while it will be
P1+P2=51 if each group forms its own company. The open question is how this advantage
shall be divided between the two groups. The classical actuarial argument is that each
group shall be charged its fair premium. However, this principle has meaning only as
far as the net premium is concerned, it does not say anything about how the safety
loading should be divided between the two groups. The conventional method would be to
divide the safety loading proportionally between the two groups, that is, to let them pay
total premium of 15 and 30, respectively. The fairness of this rule is certainly open to
question, since it gives Group I most of the gain resulted from the formation of a single
company.
4) Risk exchange between two insurers
Insurance company C1 owns a portfolio of risks, with a mean claim amount of 5 and a
variance of 4. Company C2 portfolio has a mean of 10 and a variance of 8. The two
companies decide to explore the possibility to conclude a risk exchange agreement.
Assume only linear risk exchanges are considered. Denote by X1 and X2 the claim
amounts before the exchange, and by Y1 and Y2 the claim amounts after exchange. Then
the most general form of a linear risk exchange is:
Y1=(1-)X1+X2+K

Y2=X1+(1-)X2-K
where , [0, 1], and K is a fixed (positive or negative) monetary amount. If K=5-10,
then E(Y1)=E(X1)=5 and E(Y2)=E(X2)=10. In this case the exchange does not modify the
expected claims, and only the variances need to be analysed. Assuming the
independence:
Var(Y1)=4(1-)2+82
Var(Y2)=42+8(1-)2
If, for instance, =0.2 and =0.3, then Var(Y1)=3.28<4 and Var(Y2)=4.08<8. Hence it is
possible to improve the situations of both partners. Can we define optimal values of
and ?

Observation

The last two examples have several elements in common:

- participants have some benefits to share;
- the opportunity to divide benefits results from cooperation of all participants or a subgroup of participants;
- individuals are free to engage in negotiations, bargaining, and coalition formation;
- participants have conflicting objectives; each wants to secure the largest part of the
benefits for himself.
COOPERATIVE GAME THEORY analyses those situations where participants
objectives are partially cooperative and partially conflicting.

DEFINITIONS

We call game the situation in which some elements (called players) make a decision (or
movement) from a set of alternatives independently and successively, having order and
fixed conditions given by a set of rules.
The decisions can be:
- determinist when the choice of alternative is unique;
- random- when the choice of alternative is determined by a random process (for
example the cast of a die).
Taking into account the quantity of the available information that every player has, the
games can be classified as follows:
- with complete information (chess);
- with partial information (bridge).
The set of rules that uniquely define the determined decisions according to the state of the
game is called strategy.
If one of the players has m alternatives and the game is over on the account of the choice
among the set of alternatives then we say that the player has m pure strategies. When the
game is repeated and the players choose certain pure strategies with some frequencies
then we say that a mixed strategy is used.
If the number of pure strategies is finite the game is said to be finite, otherwise is called
infinite.
The strategy that leads to a maximum gain is called optimal.
For every i 1,n (n is the number of players) is denoted by pi the gain of the player Pi at
n

the end of the game. If

p i =0 then the game is said to be with zero sum, otherwise is

i 1

called nonzero sum game.

1. Matrix games with zero sum

1.1. Games with saddle point
Case study
Consider two companies C1 and C2 that produce the same good G as follows:
- C1 produces four kinds of G, namely X1, X2, X3,X4;
- C2

produces three kinds of G: Y1, Y2, Y3.

A marketing research shows that if C1 launches X1 then it gains 30% of the market when
C2 answers with Y1, 30% when C2 launches Y2 and 70% if C2 answer is Y3 and so on.
The data are given in the next table and a certain value existing in the table means the
gain of C1 and the loose of C2 for the corresponding pair of strategies.

C2

Y1

Y2

Y3

X1

30

30

70

X2

70

40

50

X3

40

20

60

X4

60

30

40

C1

The problem can be stated as: what kind of product must C1 launch in order to gain the
maximum percentage of the market under the assumption that C1 and C2 are equally
competent and act prudently?
Answer
The following reasoning can be made:

If C1 chooses X1 then it can gain at least 30% of the market and maximum 70%. As a
response C2 will choose at the next movement one of the strategies Y1 or Y2, so C1
will gain the minimum percentage of the market on the first line 30%.

If C1 chooses X2 then it gains maximum 70% and minimum 40%.

If C1 chooses X3 then it gains maximum 60% and minimum 20%.

If C1 chooses X4 then it gains maximum 60% and minimum 30%.

The strategy of C1 is the one that should maximize the minimum gain (which is sure) and
not the one that offers a maximum gain which can be cancelled by an appropriate answer
of C2.
Therefore C1 computes in this way:
1=min{30, 30, 70}=30
2=min{70, 40, 50}=40
3=min{40, 20, 60}=20
4=min{60, 30, 40}=30
and takes =max{1, 2, 3, 4}=40 and launches X2 (the maximin strategy the
maximizing of the minimum gain).
The company C2 makes the following reasoning:
- if it replies with Y1 for the firs decision of C1 then loses at most 70% and at lest 30%;
- if C2 chooses Y2 its lose is maximum 40% and minimum 20%;
- if C2 takes decision Y3 the maximum lose is 70% and the minimum one 40%.
So C2 computes the values:
1=max{30, 70, 40, 60}=70
2=max{30, 40, 20, 30}=40
3=max{70, 50, 60, 40}=70
and calculate = min{1, 2, 3}=40, so C2 takes Y2 (the minimax strategy the
minimizing of the maximum loss).

C2

Y1

Y2

Y3

i= min a ij

X1

30

30

70

30

X2

70

40

50

40*

X3

40

20

60

20

X4

60

30

40

30

70

40*

70

C1

j= max a ij
i 1,4

j1,3

Definition
A matrix game is a triplet (X, Y, A) where X={X1, X2, , Xm} is the set of P1 player
strategies, Y={Y1, Y2, , Yn} the set of strategies corresponding to P2 and

A= a ij
the payoffs (gains) matrix, that for a matrix game with zero sum aij
i 1,m, j1,n
means the P1 gain when P1 chooses Xi and P2 answer with Yj.
As was described in the previous example the player P1 acts following the maximin
principle (the maximizing of the minimum gain):

i min a ij ,i 1, m

notation
max i max min a ij v which is called the inferior value of the
i 1,m
i 1,m j1,n

j1,n

game.
The player P2 computes according to the minimax principle (the minimizing of the
maximum loss):

j max a ij , j 1, n

notation
min j min max a ij v which is called the superior value of the
j1,n
j1,n i 1,m

i 1,m

game.
The strategy that assures to P1 the gain v is called maximin strategy and the one that
assures to P2 the gain v is aid to be the minimax strategy.
In the previous example the inferior and superior values of the game and their
corresponding strategies are:

v max min a ij 40 X 2
i 1,4 j1,3

v min max a ij 40 Y2
j1,n i 1,m

If v = v =v, then v is called the value of the game. The element aij for which this value is
obtained is a saddle point and the game is said to be a game with saddle point. The
strategies Xi, Yj corresponding to the value aij are called optimal strategies.

The reducing of the game strategies using the domination relationships

Example
Consider the matrix game (X, Y, A) given in the table

P2

B1

B2

B3

B4

B5

A1

-1

A2

-2

-1

A3

-1

-2

A4

-6

P1

We note that A1 strategy has all the gains, respectively, greater than those associated to
A2. Therefore, one says that A1 dominates the strategy A2 (or, equivalently A2 is
dominated by A1). As a consequence the line of A2 can be removed from the table
because the player P1 will never choose the strategy A2.
With respect to the player P2 the losses corresponding to B1 and B3 strategies are,
respectively, greater than those of B4 so B1 and B3 are dominated by B4. Hence, the
strategies B1 and B3 can be deleted from the table.
Finally, the table becomes

P2

B2

B4

B5

A1

-1

A3

-1

-2

A4

-6

P1

Conclusion
From the gain matrix of a game one can remove:
- the rows that have all the elements, respectively, less that those of another row;
- the columns that have all the elements, respectively, greater that those of another
column.
This is called the principle of strategies domination.
Exercises
Consider the matrix game:
a) Find the dominated strategies of the given game.
b) Find the inferior and superior values of the given game and solve the game.
Y1

Y2

Y3

Y1

Y2

Y3

X1

-2

X1

-2

X2

X2

X3

X3

-2

Y1

Y2

Y3

X1

-2

X2

Y1

Y2

X1

-1

X2

X3

-2

Y1

Y2

Y1

Y2

Y3

X1

-1

X1

-2

X2

X2

X3

Y1

Y2

Y3

X1

-2

X2

1.2. Games without saddle point

Consider G=(X, Y, A) a matrix game with zero sum without saddle point, where

X={X1, X2, , Xm}, Y={Y1, Y2, , Yn}, A= a ij

and
i 1,m, j1,n

v max min a ij < v min max a ij .

j1,n i 1,m
i 1,m j1,n

In this case, the average game =(M, N, v) is associated to G as follows:

m

M={(x1, x2, , xm)=xRm xi0, i 1,m , x i 1 } the set of all mixed strategies
i 1

for the player P1 where xi=P(Xi) is the probability of choosing the strategy Xi,

i 1,m ;
n

N={(y1, y2, , yn)=yRn yj0, y j 1 } the set of all mixed strategies for the
j1

player P2 where yj=P(Yj) is the probability of choosing the strategy Yj, j 1,n ;
m n

v: MNR, v(x, y) =

a ijx i y j is the average gain (payoff) function.

i 1 j1

Example
Consider the following zero-sum matrix game without saddle point.
P2

Y1

Y2

Y3

X1

-2

-2

X2

-3

-3

X3

-4

-4

P1

v =-2
v =2

Assume that P1 and P2 move for N=100 times according to this game.
As an example let us consider that P1 plays X1 for 25 times, X2 for 50 times and X3 for 25
times which corresponds to the mixed strategy x=(0.25, 0.5, 0.25). The total gain of P1 is
computed as follows:

25 0 50 5 25 2
0.25 0 0.5 5 0.25 2 =3 if P2 plays the pure strategy Y1
100

(equivalent to the mixed strategy y=(1, 0, 0));

25 (2) 50 4 25 5
0.25 2 0.5 4 0.25 5 =2.75 if P2 plays the pure
100

strategy Y2 (equivalent to the mixed strategy y=(0, 1, 0));

25 2 50 3 25 4
100

0.25 2 0.5 3 0.25 4 =-2 if P2 plays the pure

strategy Y3 (equivalent to the mixed strategy y=(0, 0, 1)).

Next, lets assume that the player P1 chooses the mixed strategy x=(x1, x2, x3), where
x1, x2, x30 and

xi 1.
i 1

If P2 answers with Y1 (that is y=(1, 0, 0)) then the average gain will be
v(x, y)=5x2+2x3.
If the response is Y2 (equivalent to y=(0, 1, 0)) then
v(x, y)=-2x1+4x2+5x3.
For Y3 (corresponding to y=(0, 0, 1)) the average payoff is computed as
v(x, y)=2x1-3x2-4x3.
In the general case, if P2 uses a mixed strategy y=(y1, y2, y3) then:
3 3

v(x, y)=y1(5x2+2x3)+y2(-2x1+4x2+5x3)+y3(2x1-3x2-4x3)=

a ijx i y j .
i 1 j1

Conclusion: v(x, y) show how much gains player P1 and loses P2 if the mixed strategies
x=(x1, x2, x3) and y=(y1, y2, y3) are used.

Definition
Consider the average game =(M, N, v) associated to the matrix game with zero sum
without saddle point G=(X, Y, A). We define:
- the inferior value of the game by:

v max min v x, y
xM yN

- the superior value of the game by:

v min max v x, y
yN xM

- (x0, y0) is a saddle point for v(x, y) if:

v(x, y0) v(x0, y0) v(x0, y), xM, yN
and in this case v(x0, y0) is called the value of the game and x0, y0 are called the optimal
strategies of player P1 and P2, respectively.
Using the previous definitions one can prove the following results:

Theorem 1
If v and v exist, then v v .
Theorem 2
If v and v exist, then (x0, y0) is a saddle point for v(x, y) v = v = v(x0, y0).
Theorem 3
Consider the matrix game with zero sum without saddle point G=(X, Y, A) and
=(M, N, v) the associated average game. Then (x0, y0)MN and v and v such
that v = v = v(x0, y0).
Theorem 4
If the notations v = v G and v = v G , then

v Gv =v v G.

The solution of matrix games with zero sum and without saddle point.

I) Games 22

Consider G=(X, Y, A) a zero sum matrix game 22 without saddle point with

a
a
A= 11 12 a ij
. The optimal strategies x0=(x1, x2), y0=(y1, y2) and the value
i, j1,2
a 21 a 22

of the game v=v(x0, y0) are determined by the following relations:

1
JA1J T

, x0

JA 1
JA1J T

, yT
0

A 1J T
JA 1J T

(*)

where J=(1 1).

Examples

2 1
1) Find the solution of the matrix game given by the matrix A=
.
0 3
Answer

P2

Y1

Y2

X1

X2

P1

v G =1
2

v G =2

One can compute v G max min a ij 1 , v G min max a ij 2 , so the game is

i 1,2 j1,2
j1,2 i 1,2

without saddle point.

Next, we use the formulas (*):
1 3 1
2 0
3 1 -1
1
det(A)=6, AT=
A*=
, then
, A =
, A*=
6 0 2
det(A)
0 2
1 3

 2 2
1 3 11 1
JA1J T = 11
11 ,
6 0 2 1 6
2 3
v=v(x0, y0)=

so

the

value

of

the

game

is

3
.
2

The optimal strategies are:

x0

JA 1
JA 1J T

v JA 1

3
1 3 1 3 1 1 3 1
11

2
60 2 2 2 6 4 4

1 1
2
A J
1 1
1 T 3 1 3 11 3 3
yT

A
J

so y 0

0
.
2 6 0 2 1 2 1 1
2 2
JA 1J T

3 2
1 T

2) The same problem for the games given by the matrixes:

2 4
1 1
5 1
a) A=
.
; b) A=
; c) A=
8 6
6 0
3 4
Answers: a) v=

3
3 1
1 7
, x 0 , , y0 ,
4
4 4
8 8

b) v=1, x0 0.7, 0.3 , y0 0.5, 0.5 .

II) Games 2n and m2

Consider G=(X, Y, A) a zero sum matrix game 2n without saddle point with

a12 ....a1n
a
A= 11
a ij i 1,2; j1,n . The case m2 is studied in the same way as this
a 21 a 22 ....a 2n

one by swapping the players.

By notation, the optimal mixed strategy of P1 is x0=(x1, x2). We search the optimal
strategy x0 with the property that v(x, y) is maximum, that is v=v(x0, y0) - the value of the
game. Using the definition we have:
v= v(x0, y0) v(x0, y), yN v v(x0, fj), j= 1, n (**)
where f1=(1, 0, 0,,0, 0), f2=(0, 1, 0, , 0, 0),, fn=(0, 0, 0, , 0, 1) stand for pure
strategies of P2.
Hence the conditions (**) yield:

a11x1 a 21x 2 v
a x a x v
12 1 22 2
...........................

a1n x1 a 2n x 2 v
x1 x 2 1

x1, x 2 0
and the associated linear programming problem can be written:
[max] f(x1, x2, v)=v

a11x1 a 21x 2 v 0
a x a x v 0
12 1 22 2
...........................

a1n x1 a 2n x 2 v 0
x1 x 2 1

x1, x 2 0, v R
Remark
The previous problem can be solved by graphical method.

Examples
1) Find the optimal strategy of P1 and the value of the game if the matrix attached to the

2 4 4
game has the form: A=
.
8

6
16

The linear programming problem for this problem is:

[max] f(x1, x2, v)=v
2x1 8x 2 v 0
4x 6x v 0
2
1

4x

16x

1
2 v 0.
x x 1
2
1
x1, x 2 0, v R

The variable x2 can be expressed in terms of x1 as x2=1-x1 so the problem can be

rewritten:

[max] f(x1, 1-x1,v)=v

10x1 v 8
10x v 6
1

20x1 v 16
x1 [0,1], v R

By graphical method one can obtain the optimal solution v=1, x1=0.7 and x2=0.3, so
x0=(0.7, 0.3).
2) Find the optimal strategy of P2 and the value of the game if the matrix attached to the
6 2

5 1
game has the form: A= 3 4 .

1 5
2 2

As it is easy to observe, the strategy X5 is dominated by X3 (or X2) so X5 is removed from

6 2

5 1

the game. Consequently, the game become 42 type with the matrix A=
. The
3 4

1 5
linear programming problem associated is:
[min] f(y1, y2, v)=v

6y1 2y 2 v
5y y v
2
1
3y1 4y 2 v
.

y1 5y 2 v
y1 y 2 1

y1, y 2 0, v R
By replacing y1=1-y2 the model becomes:

[min] f(1-y2, y2, v)=v

6 8y 2 v
8y 2 v 6
5 4y v
4y v 5
2

2
y 2 v 3
.
3 y 2 v
6y 1 v
6y v 1
2
2
y 2 [0,1], v R y 2 [0,1], v R
2
3
By graphical method one can obtain the optimal solution v=3.4, y 2 , y1
5
5

2. Two persons games with arbitrary sum (bimatrix games)

The two persons games with arbitrary sum are extensions of zero sum games we
have studied before. In this case the players can cooperate in order to maximize their
gains which is not possible in zero sum games. So, there are two types of arbitrary sum
games:

non-cooperative games - when the players make decisions independently;

cooperative games where the players can correlate their strategies such that the
gains are maximized.

2.1. Arbitrary sum non-cooperative games

Consider two players P1 and P2 and assume the following set of pure strategies:

A= {A1, A2, , Am} for P1 and

B= {B1, B2, , Bn} for Pn.

The triplet G={A, B; f1, f2} is called arbitrary sum game of two persons given in the
normal form if A and B contain the pure strategies of the players and f1, f2:ABR are
payoff functions of the P1and P2, respectively.
For a pair of strategies (Ai, Bj) we denote f1(Ai, Bj)=aij and f2(Ai, Bj)=bij, i= 1,m , j= 1,n .
Hence, the most convenient way for representing a game of this type is the matrix of
gains as follows:

P2

B1

Bj

Bn

A1

(a11, b11)

(a1j, b1j)

(a1n, b1n)

Ai

(ai1, bi1)

(aij, bij)

(ain, bin)

Am

(am1, bm1)

(amj, bmj)

(amn, bmn)

P1

As the elements of the previous table are pairs of real numbers that can be determined by
the merging of two mn matrixes associated to the players P1 and P2, then the considered
arbitrary sum game of two persons is called bimatrix game.
Is worth mentioning that in this definition there is imposed no condition about the sum
f1(Ai, Bj)+ f2(Ai, Bj), i= 1,m , j= 1,n .

Examples
1) Consider the game G={A, B; f1, f2} where A= {A1, A2}, B = {B1, B2}, f1(A1, B1)=3,
f1(A2, B1)=1, f1(A1, B2)=1, f1(A2, B2)=0, f2(A1, B1)=0, f2(A2, B1)=4, f2(A1, B2)=2,
f2(A2, B2)=1 having the matrix form

P2

B1

B2

A1

(3, 0)

(1, 2)

A2

(1, 4)

(0, 1)

P1

One can observe that the player P1 prefers to play strategy A1 because for any answer of
P2 the corresponding payoffs of A1 are greater than those of A2 (f1(A1, X)>f1(A2, X),
XB). Equivalently, the strategy A1 dominates A2. Regarding the pure strategies of P2,
one can remark that there is no dominated strategy among B1 and B2. Then, by players
rationality hypothesis, the game is reduced by removing the dominated strategy A2, so it
becomes

P2
P1
A1

B1

B2

(3, 0)

(1, 2)

Finally, the player P2 will choose the pure strategy B2 in order to maximize his gain.
Therefore, the solution of the game is the pair of strategies (A1, B2).

2) Prisoner's Dilemma

If one prisoner alone confesses and implicates the other prisoner then he obtains complete
immunity (utility level 3) and the other prisoner gets life imprisonment (utility level 0).
If both confess then both get a substantial sentence, but less than life (utility level 1).
If neither confesses there is enough evidence to convict them both of some minor

Player 1

violation for which they are fined (utility level 2).

A1
Deny

(2, 2)

(0, 3)

A2
Confess

(3, 0)

(1, 1)

B1
Deny

B2
Confess

Player 2
As in the previous example, the optimal pair of strategies can be found - (A2, B2).

Player 1

3) Battle of the Sexes

A1
West

(2, 1)

(0, 0)

A2
East

(0, 0)

(1, 2)

B1
West

B2
East

Player 2
In this case is easy to observe that the solution can not be obtained by the dominated
strategies removing method there are no dominated strategies.
We assume, as in the theory of a rational decision-maker, that each player chooses
the best available action. In a game, the best action for any given player depends, in
general, on the other players' actions. So, when choosing an action a player must have in
mind the actions the other players will choose. That is, any player must form a belief
about the other players' actions.
It is helpful to think of the following idealized circumstances. For each player in
the game there is a population of many decision-makers who may, on any occasion, take
that player's role. In each play of the game, players are selected randomly, one from each
population. Thus each player engages in the game repeatedly, against ever-varying
opponents. Players experience leads to beliefs about the actions of typical opponents,
not any specific set of opponents.
As an example, think of the interaction between buyers and sellers. Buyers and
sellers repeatedly interact, but to a first approximation many of the pairings may be
modeled as random. In many cases a buyer transacts only once with any given seller, or
interacts repeatedly but anonymously (when the seller is a large store, for example).
In summary, the solution theory we study has two components. First, each player
chooses his action according to the model of rational choice, given his belief about the
other players' actions. Second, every player's belief about the other players' actions is
correct.

Definition
Consider a bimatrix game in the normal form G={A, B; f1, f2}. A pair of strategies
(A*, B*) represents a Nash equilibrium point for G if the following relations hold:
-

f1(A*, B*) f1(Ai, B*), AiA;

f2(A*, B*) f1(A*, Bj), BjB.

The previous mentioned conditions mean:

- max f1 Ai ,B* is reached at A* and
Ai A

- max f 2 A*,B j is obtained for B*.

B jB

As (A*, B*) A B then the Nash equilibrium is in pure strategies.

Remark
The notion of Nash equilibrium point is a generalization of saddle point defined at the
zero-sum matrix games.
Examples
1) Consider the bimatrix game G given by the table:

P2

B1

B2

B3

B4

A1

(4, 0)

(3, 1)

(0, 4)

(1, 1)

A2

(0, 1)

(2, 0)

(3, 3)

(5, 2)

A3

(2, 4)

(1, 3)

(1, 2)

(0, 2)

P1

The determination of the Nash Equilibrium points is approached as follows: for each
player and each associated strategy one finds the optimal answer of the other player at the
considered strategy. In order to point out this selection, the maximal gain or gains on the
line/column is highlighted.
For instance, one can observe that if P1 would play the strategy A1 then the optimal
answer of P2 is B3, so the gain 4 is highlighted and the pair (0, 4) becomes (0, 4). In the
same way, for the strategies A2 and A3 of P1 we select the optimal answers B3 and B1,
respectively and fill in the table the values (3, 3) and (2, 4), respectively.

P2

B1

B2

B3

B4

A1

(4, 0)

(3, 1)

(0, 4)

(1, 1)

A2

(0, 1)

(2, 0)

(3, 3)

(5, 2)

A3

(2, 4)

(1, 3)

(1, 2)

(0, 2)

P1

Next, are studied the optimal answers of P1 for every possible strategy chosen by P2.
Therefore the pairs (4, 0) (on the first column), (3, 1) (on the second column), (3, 3) (on
the third column) and (5, 2) (on the last column) are selected and the gains of P1 are
marked. Hence the table becomes:
P2

B1

B2

B3

B4

A1

(4, 0)

(3, 1)

(0, 4)

(1, 1)

A2

(0, 1)

(2, 0)

(3, 3)

(5, 2)

A3

(2, 4)

(1, 3)

(1, 2)

(0, 2)

P1

Finally, the Nash equilibrium points are associated with the pairs where both components
are highlighted (if such pairs exist), that is in the presented case the pair (3, 3). The Nash
solution of the game will be the pair of pure strategies (A2, B3).

Player 1

2) Battle of the Sexes

A1
West

(2, 1)

(0, 0)

A2
East

(0, 0)

(1, 2)

B1
West

B2
East

Player 2

By applying the same reasoning as described before, the highlighted pairs are (2, 1) and
(1, 2):

Player 1

A1
West

(2, 1)

(0, 0)

A2
East

(0, 0)

(1, 2)

B1
West

B2
East

Player 2
The game has two Nash equilibrium points: (A1, B1) and (A2, B2).

Player 1

3) A game without Nash equilibrium point in pure strategies

A1

(2, 1)

(0, 2)

A2

(1, 2)

(3, 0)

B1

B2

Player 2

Player 1

If we try to solve the problem as in the previous case then we get the highlighted payoffs:

A1

(2, 1)

(0, 2)

A2

(1, 2)

(3, 0)

B1

B2

Player 2

Hence the game has no Nash equilibrium point.

The following properties present the relation between Nash equilibrium points and
dominated strategies.

Properties
Consider the bimatrix game in normal form G={A, B; f1, f2}.
1) If (S1*, S2*) represents a Nash equilibrium point of G, then it is not modified by
procedure of the dominated strategies removing.
2) If after the removing of dominated strategies of G remains only one strategy
(S1*, S2*), then this is the unique Nash solution of the game.
3) Any Nash equilibrium point contains, for each player, payoffs greater than the ones
corresponding to the MAXIMIN strategy.
Consequence: In a bimatrix game the MAXIMIN strategy is not involved in the method
for finding the optimal solution.

What is the solution if the game has no Nash equilibrium point?

Answer: The average game has always a Nash equilibrium point.

Definition
Consider the bimatrix game with the previous notations G={A, B; f1, f2}. The average
game is defined by:
m

X={x=(x1, x2, , xm)Rm xi0, i 1,m , x i 1 } is the set of all mixed strategies
i 1

for the player P1 over the set of pure strategies A= {A1, A2, , Am};
n

Y={y=(y1, y2, , yn)Rn yj0, y j 1 } the set of all mixed strategies defined on
j1

B= {B1, B2, , Bn} the pure strategies associated to the player P2;

m n

1: XYR, 1(x, y)= a ijx i y j is the average gain of P1 for mixed strategies
i 1 j1

x=(x1, x2, , xm) and y=(y1, y2, , yn).

m n

2: XYR, 2(x, y)= bijx i y j is the average gain of P2.

i 1 j1

A pair of mixed strategies (x*, y*) XY is a Nash equilibrium point for the game if:
1(x*, y*)1(x, y*), xX x* is the best answer of P1 to the strategy y* of P2
2(x*, y*)2(x*, y), yY y* is the best answer of P2 to the strategy x* of P1.

NASH THEOREM
For any two players game with arbitrary sum in normal form G={A, B; f1, f2} there is at
least one Nash equilibrium point in pure or mixed strategies.

2.2. Arbitrary sum many person cooperative games

2.1.1. Characteristic function of arbitrary sum many person cooperative games

In many-person cooperative games there are no restrictions on the agreements that

may be reached among the players. In addition, is assumed that all payoffs are measured
in the same units and that there is a transferable utility which allows side payments to
be made among the players. Side payments may be used as inducements for some players
to use certain mutually beneficial strategies.
Thus, there will be a tendency for players, whose objectives in the game are close, to
form alliances or coalitions. The structure given to the game by coalition formation is
conveniently studied by reducing the game to a form in which coalitions play a central
role.

Definition
Let n 2 denote the number of players in the game, numbered from 1 to n, and let N
denote the set of players, N ={1, 2, . . . , n}. A coalition, S, is defined to be a subset of N,
S N, and the set of all coalitions is denoted by 2N. By convention, we also speak of the
empty set, , as a coalition, the empty coalition. The set N is also a coalition, called the
grand coalition.
If there are just two players, n = 2, then there are four coalition, {, {1}, {2}, N}. If there
are 3 players, there are 8 coalitions, {, {1}, {2}, {3}, {1, 2}, {1, 3}, {2, 3}, N}. For n
players, the set of coalitions, 2N, has 2n elements.

Definition.
The coalitional form of an n-person game is given by the pair (N, v), where
N = {1, 2, . . . , n} is the set of players and v is a real-valued function, called the
characteristic function of the game, defined on the set 2N, of all coalitions and satisfying
(i) v() = 0
(ii) (superadditivity) if S and T are disjoint coalitions (S T = ), then
v(S) + v(T) v(S T).

Compared to the normal (strategic) or extensive forms of n-person games, this is a very
simple definition. Naturally, much detail is lost. The quantity v(S) is a real number for
each coalition S N, which may be considered as the value, or worth, or power, of
coalition S when its members act together as a unit. Condition (i) says that the empty set
has value zero, and (ii) says that the value of two disjoint coalitions is at least as great
when they work together as when they work apart.

Relation to the strategic form of the game

The

strategic

form

of

an

n-person

game

is

given

by

the

2n-tuple,

(X1,X2, . . . , Xn, u1, u2, . . . , un), where

(1) for i = 1, . . . , n, Xi is the set of pure strategies of Player i, and
(2) for i = 1, . . . , n, ui(x1, . . . , xn) is the payoff function to Player i, if Player1 uses
x1X1, Player 2 uses x2X2, . . ., and Player n uses xn Xn.
Transforming a game from strategic form to coalitional form entails specifying the value,
v(S), for each coalition S 2N. The usual way to assign a characteristic function to a
strategic form game is to define v(S) for each S 2N as the value of the 2-person zerosum game obtained when the coalition S acts as one player and the complementary
coalition, S = N S, acts as the other player, and where the payoff to S is the sum of the
payoffs to the players in S: ui x1 ,...,xn . Thus
iS

v(S)=Val( ui x1 ,...,xn )
iS

where the players in S jointly choose the xi for i S, and the players in S choose the xi
for i S.

Example
Three person game P1, P2 and P3 with two pure strategies X1={A1, A2}, X2={B1, B2},
X3={C1, C2} and payoffs vectors given in the tables:

If P1 chooses A1, then

P3

C1

C2

B1

(0, 3, 1)

(2, 1, 1)

B2

(4, 2, 3)

(1, 0, 0)

C1

C2

B1

(1, 0, 0)

(1, 1, 1)

B2

(0, 0, 1)

(0, 1, 1)

P2

If P1 chooses A2, then

P3
P2

The associated game in coalitional form the finding of v.

2N={, N, {1}, {2}, {3}, {1, 2}, {1, 3}, {2, 3}}

v()=0;

v(N)=the largest sum in the previous eight cells=4+2+3=9 (corresponding to the

strategy (A1, B2, C1));

v({1}) is computed using the payoff matrix for the winnings of P1 against (P2, P3)

(P2, P3)

P1

(B1, C1)

(B1, C2)

(B2, C1)

(B2, C2)

A1

A2

The strategy (B2, C1) is dominated by (B2, C2) and (B1, C2) by (B1, C1) , so they are
removed from the game. Therefore, the game becomes:

(P2, P3)

P1

(B1, C1)

(B2, C2)

A1

A2

0 1 1
Hence, the value of this game is v({1})=val
= .
1 0 2

v({2}) is computed using a similar reasoning

(P1, P3)

P2

(A1, C1)

(A1, C2)

(A2, C1)

(A2, C2)

B1

B2

This is a game with saddle point, having the value v({2})=0;

for v({3}) we have

(P1, P2)

P3

(A1, B1)

(A1, B2)

(A2, B1)

(A2, B2)

C1

C2

3 0 3
So, v({3})= val
= .
0 1 4

for v({1, 3}) we have to consider the winnings of (P1, P3) against P2 as in the table:
P2

(P1, P3)

B1

B2

(A1, C1)

(A1, C2)

(A2, C1)

(A2, C2)

As the strategy (A1, C2) dominates both (A2, C1) and (A2, C2) the game becomes
P2

(P1, P3)

B1

B2

(A1, C1)

(A1, C2)

1 7 5
Consequently, v({1, 3})= val
= 2 .
3
1

for v({1, 2}) we have:

P3

(P1, P2)

C1

C2

(A1, B1)

(A1, B2)

(A2, B1)

(A2, B2)

and the reduced game

P3

(P1, P2)

C1

C2

(A1, B1)

(A1, B2)

a game with saddle point with v({1, 2})=3;

for v({2, 3}) we have:

P1

(P2, P3)

A1

A2

(B1, C1)

(B1, C2)

(B2, C1)

(B2, C2)

and the reduced game

P1

(P2, P3)

A1

A2

(B1, C2)

(B2, C1)

2 2
having the value v({2, 3})= val
=2.
5 1
Finally,
v()=0, v({1})=
v({1, 2, 3})=9.

1
5
3
, v({2})=0, v({3})= , v({1, 3})= , v({1, 2})=3, v({2, 3})=2,
2
4
2

2.1.2. Shapley Value of arbitrary sum many person cooperative games

The Shapley Value

Here is considered another approach to n-person games in characteristic function form:

the concept of value. In this approach, one tries to assign to each game in coalitional
form a unique vector of payoffs, called the value. The ith entry of the value vector may be
considered as a measure of the value or power of the ith player in the game.
Alternatively, the value vector may be thought of as an arbitration outcome of the game
decided upon by some fair and impartial arbiter. The central value concept in game
theory is the one proposed by Shapley in 1953.

The Shapley Axioms

A value function, , is a function that assigns to each possible characteristic function of
an n-person game, v, an n-tuple, (v) = (1(v), 2(v), . . . , n(v)) of real numbers. Here
i(v) represents the worth or value of player i in the game with characteristic function v.
The axioms of fairness are placed on the function, .
The Shapley Axioms for (v):
1. Efficiency.

i v = v(N).

iN

2. Symmetry. If i and j are such that v(S{i}) = v(S{j}) for every coalition S not
containing i and j, then i(v) = j(v).
3. Dummy Axiom. If i is such that v(S) = v(S {i}) for every coalition S not containing i,
then i(v) = 0.
4. Additivity. If u and v are characteristic functions, then (u + v) = (u) +(v).
Axiom 1 is group rationality, that the total value of the players is the value of the grand
coalition. The second axiom says that if the characteristic function is symmetric in
players i and j, then the values assigned to i and j should be equal. The third axiom says
that if player i is a dummy in the sense that he neither helps nor harms any coalition he
may join, then his value should be zero. The strongest axiom is number 4. It reflects the

feeling that the arbitrated value of two games played at the same time should be the sum
of the arbitrated values of the games if they are played at different times. It should be
noted that if u and v are characteristic functions, then so is u + v.

Theorem 1.
There exists a unique function satisfying the Shapley axioms.

Computation of the Shapley Value

Method 1
The computation of Shapley Value is based on the following relations:
cT =v(T)-

cS , for every TN,

S T
S T

i v

cS
, for i=1, 2, , n.
SN S

iS

Example

Consider the following characteristic function of a game in coalitional form:

v() = 0, v({1}) = 1, v({2}) = 0, v({3}) = 1, v({1, 2}) = 4, v({1, 3}) = 3, v({2, 3}) = 5,
v({1, 2, 3}) = 8.
Then, c{1} = v({1}) = 1, c{2} = 0 and c{3} = 1
c{1, 2} =v({1, 2}) c{1} c{2} = 4 1 0 = 3,
c{1, 3} = 3 1 1 = 1, and c{2, 3} = 5 0 1 = 4.
Finally,
cN = v(N) c{1,2} c{1,3} c{2,3} c{1} c{2} c{3} =
=8 3 1 4 1 0 1=2.
From this we find
1(v) = 1

3 1 2 7
,
2 2 3 3

4 3 2 17
2(v) =
2 2 3 6

3(v) = 1

1 4 2 17
.
2 2 3 6

14 17 17
Therefore, the Shapley value is = , , .
6 6 6

Method 2

There is an alternate way of arriving at the Shapley value that gives additional
insight into its properties. Suppose we form the grand coalition by entering the players
into this coalition one at a time. As each player enters the coalition, he receives the
amount by which his entry increases the value of the coalition he enters. The amount a
player receives by this scheme depends on the order in which the players are entered. The
Shapley value is just the average payoff to the players if the players are entered in
completely random order.
Theorem 2.
The Shapley value is given by = (1, 2, . . . , n), where for i = 1, . . . , n,
i(v) =

SN
iS

1! n - S !
n!

v S v S {i } (*)

The summation in this formula is the summation over all coalitions S that contain
i. The quantity, v(S) v(S {i}), is the amount by which the value of coalition S {i}
increases when player i joins it. Thus to find i(v), merely list all coalitions containing i,
compute the value of player is contribution to that coalition, multiply this by
(|S| 1)!(n |S|)!/n!, and take the sum.
The interpretation of this formula is as follows. Suppose we choose a random
order of the players with all n! orders (permutations) of the players equally likely. Then
we enter the players according to this order. If, when player i is enters, he forms coalition
S (that is, if he finds S {i} there already), he receives the amount [v(S) v(S {i})].

The probability that when i enters he will find coalition S {i} there already is

S 1! n - S ! . The denominator is the total number of permutations of the n players.

n!

The numerator is number of these permutations in which the |S| 1 members of S {i}
come first ((|S| 1)! ways), then player i, and then the remaining n |S| players ( in
(n |S|)! ways). So this formula shows that i(v) is just the average amount player i
contributes to the grand coalition if the players sequentially form this coalition in a
random order.

Example
As an illustration of the use of this formula, let us compute 1(v) again for the previous
example. The probability that player 1 enters first (S={1}) is

0!2! 1
, and then his payoff
3! 3

is v({1}) = 1. The probability that 1 enters second and finds 2 there (S={1, 2}) is

1!1! 1
,
3! 6

and his payoff is v({1, 2}) v({2}) = 4 0 = 4. The probability that 1 enters second and
1
, and the expected payoff is v({1, 3}) v({3}) = 3 1 = 2.
6

finds 3 there (S={1, 3}) is

The

probability

that

enters

last

is

2!0! 1
,
3! 3

and

then

his

payoff

is

v({1, 2, 3}) v({2, 3}) = 8 5 = 3. Player 1s average payoff is therefore

1
1
1
1
14
1(v) = 1 4 2 3
3
6
6
3
6

as found earlier.
The following table shows the computations for all three players simultaneously. The 6
different orders of the players are listed along with the payoffs to the players. In the first
row, the players enter in the order 1, 2, 3. Player 1 receives v({1}) = 1 upon entry; then
Player 2 receives v({1, 2}) v({1}) = 4 1 = 3; finally Player 3 receives v(N) v({1, 2})
= 8 4 = 4. Each of the six rows is equally likely, probability
is the average of the six numbers in each column.

1
each. The Shapley value
6

Player
Order of entry

Total

1 2 3

1 3 2

2 1 3

2 3 1

3 1 2

3 2 1

14
6

17
6

17
6

Average

Simple Games. The Shapley-Shubik Power Index.

The Shapley value has an important application in modeling the power of
members of voting games. This application was developed by Shapley and Shubik in
1954 and the measure is now known as the Shapley-Shubik Power Index. Players are
members of legislature or members of the board of directors of a corporation, etc. In such
games, a proposed bill or decision is either passed or rejected. Those subsets of the
players that can pass bills without outside help are called winning coalitions while those
that cannot are called losing coalitions. In all such games, we may take the value of a
winning coalition to be 1 and the value of a losing coalition to be 0. Such games are
called simple games.

Definition
A game (N, v) is simple if for every coalition S N, either v(S) = 0 or v(S) = 1.

In a simple game, a coalition S is said to be a winning coalition if v(S) = 1 and a losing

coalition if v(S) = 0. So in a simple game every coalition is either winning or losing. It
follows from superadditivity of v that in simple games every subset of a losing coalition
is losing, and every superset of a winning coalition is winning.

Typical examples of simple games are

(1) the majority rule game where v(S) = 1 if |S| > n/2, and v(S) = 0 otherwise;
(2) the unanimity game where v(S) = 1 if S = N and v(S) = 0 otherwise; and
(3) the dictator game where v(S) = 1 if 1 S and v(S) = 0 otherwise.
For simple games, formula (*) for the Shapley value simplifies because the difference
[v(S) v(S {i})] is always zero or one. It is zero if v(S) and v(S {i}) are both zero or
both one, and it is one otherwise. Therefore we may remove [v(S) v(S {i})] from
formula (*) provided we sum only over those coalitions S that are winning with i and
losing without i.
Formula (*) for the Shapley value (the Shapley-Shubik Index) becomes
i(v) =

1! n - S !
n!

S winning
S ilosing

(*)

There is a large class of simple games called weighted voting games. These games
are defined by a characteristic function of the form

1, if wi q

iS
v(S) =
0, if wi q
iS

for some non-negative numbers wi (the power of player i), called the weights, and a
positive number q, called the quota. If q =

1
wi , this is called a weighted majority
2 iN

game.

Examples
1) Consider the game with players 1, 2, 3, and 4, having 10, 20, 30, and 40 shares of
stock respectively, in a corporation. Decisions require approval by a majority (more than
50%) of the shares. This is a weighted majority game with weights w1 = 10, w2 = 20, w3 =
30, w4 = 40 and quota q = 50.
Let us find the Shapley value of this game. The winning coalitions are {2, 4}, {3, 4},
{1, 2, 3}, and all supersets (sets containing one of these).

For i = 1, v(S)v(S {1}) = 0 unless S = {1, 2, 3}. So

1(v) =

2!1! 1

4! 12

For i = 2, v(S) v(S {2}) = 0 unless S = {2, 4}, {1, 2, 3}, or {1, 2, 4}, so that
2(v) =

1!2!
2!1! 1
2

4!
4! 4

Similarly, for i = 3, v(S) v(S {3}) = 0 unless S = {3, 4}, {1, 2, 3}, or {1, 3, 4}, so
3(v) =

1!2!
2!1! 1
2
.
4!
4! 4

For i = 4, v(S)v(S {4}) = 0 unless S = {2, 4}, {3, 4}, {1, 2, 4}, {1, 3, 4} or {2, 3, 4}. So
4(v) = 2

1!2!
2!1! 5
23

4!
4! 12

1 3 3 5
The Shapley value is = , , , . Note that the value is the same for the players
12 12 12 12

2 and 3 although the player 3 has more shares.

2) Consider a group (Group I) of n1=100 individuals. Each of them is exposed to a
possible loss of 1, with a probability q1=0.1. Assume these persons decide to form a risk
retention group, a small insurance company to cover them against that risk. The premium
charged will be such that the ruin probability of the group is less than 0.001. By
considering that the risks are independent, and using the normal approximation of the
binomial distribution, the group must have the total funds equal to:

P1 n1q1 3 n1q1(1 q1) =10+9=19

Hence each person in Group I will pay, in addition to the net premium of 0.1, a safety
loading of 0.09.
Another group - Group II - consists of n2=100 persons exposed to a loss of 1 with a
probability q2=0.2. If they form their own retention group under the same conditions, the
total premium will be

P2 n 2q 2 3 n 2q 2 (1 q 2 ) =20+12=32.
Therefore, each person in Group II will pay the net premium of 0.2 and a safety loading
of 0.12.

Suppose now that the two groups decide to join and form one single company. In order to
ensure that the ruin probability shall be less than 0.001, the new company must have
funds amounting to:

P12 n1q1 n 2q 2 3 n1q1(1 q1) n 2q 2 (1 q 2 ) =10+20+15=45.

One can see from this example that it is to the advantage of the two groups to form one
single company. Total payment of premium will then be P12=45, while it will be
P1+P2=51 if each group forms its own company. The open question is how this advantage
shall be divided between the two groups.
The conventional method would be to divide the safety loading proportionally between
the two groups, that is, to let them pay total premium of 15 and 30, respectively.

This is a 2- person cooperative game with the characteristic function, v:

v()=0, v({1})=19, v({2})=32, v({1, 2})=45. The Shapley value is computed using the
table:

Player
Order of entry

Total

1 2

19

36

45

2 1

13

32

45

Average

16

29

45

So, the Shapley value is = (16, 29).

Consider another group - Group III with n3=120 persons and probability for a unit loss
q3=0.3. Then

P3 n 3q3 3 n 3q3 (1 q3 ) 36+15=51

P23 n 2q 2 n 3q3 3 n 2q 2 (1 q 2 ) n 3q3 (1 q3 ) 20+36+19.2=75.2
P13 n1q1 n 3q3 3 n1q1(1 q1) n 3q 3 (1 q 3 ) 10+36+17.5=63.5

P123 n1q1 n 2q 2 n 3q3 3 n1q1(1 q1) n 2q 2 (1 q 2 ) n 3q3 (1 q3 )

10+20+36+21=87
This is a 3- person cooperative game with the characteristic function, v:
v()=0, v({1})=19, v({2})=32, v({3})=51, v({1, 2})=45, v({1, 3})=63.5,
v({2, 3})=75.2, v({1, 2, 3})=87.The Shapley value is computed using the table:

Player
Order of entry

Total

1 2 3

19

26

42

87

1 3 2

19

23.5

44.5

87

2 1 3

13

32

42

87

2 3 1

11.8

32

43.2

87

3 1 2

12.5

23.5

51

87

3 2 1

11.8

24.2

51

87

14.5

26.9

45.6

87

Average

1
87.1
14.5,
19 19 13 11.8 12.5 11.8
6
6

1(v) =
2(v) =

1
161.2
26.9
26 23.5 32 32 23.5 24.2
6
6

3(v) =

1
273.7
45.6
42 44.5 42 43.2 51 51
6
6

so = (14.5, 26.9, 45.6).

The traditional pro rata approach leads to (13.2, 26.4, 47.4) which does not take into
account the savings each member brings to the grand coalition, or its threat possibilities.