Game Theory

Seller interdependence
• If D Mart offers deep discounts on soft drinks, will Big Bazar
follow suit?
• Indigo Airlines “loyalty” miles do not expire—how did Air
India, Akasha, et al react?
• Airtel “Connected everywhere” ad campaign prompted an
effective retaliatory ad strategy by Vodafone, Jio.
• Ultratech decision to add production capacity is conditioned
upon the investment plans of rival Cement producers.
Game Theory and
Seller Interdependence

Selecting a course of
action in a situation in
which rival players are
selecting strategies that
suit their interests is the
basic problem of game
theory.
Players and
their actions
A situation of competitive rivalry must involve
two or more players whose choice of actions
affect each other.
• A “player” can be a firm, an interest group or coalition, a military
leader, government official.
• Games generally consider only one kind of action—e.g., for Indigo
it will be number of daily departures, fares, in-flight services,
schedules, advertising, choice of hubs, ordering planes, expanding
terminals, mergers and acquisitions, and human resource
decisions.
• For IPL, it maybe pitch condition, form of rival’s players, Weather,
Psychological conditions of the team (ex of T-20 world cup by
captain Rohit Sharma, Shreya’s AI tool)
 Outcomes and Payoffs

The firm’s action, together with the actions of its rivals,

determine its payoff

• In the standard “business” game, the payoff can be in the

form of profit, market share, ratings points,
• In war games, the payoff might be measured in enemy
killed or territory seized.
• In political games, payoffs may be measured in votes or
campaign contributions.
 Underlying “rules”
The rules of the game define the range of
possible outcomes and payoffs

• For example, collusion to fix prices or a merger among direct

rivals in a concentrated market structure may be against the
rules.
• Another set a rules specifies whether players move
sequentially or simultaneously, who moves first, and what
does each player know about the other players’ preference
and prior to actions?
Prisoner’s dilemma

Kalu and Bhola have been charged with

bank robbery. But lacking a confession,
the Police can only get a “disruption of peace” charge
to stick. So the police play one suspect off against
the other.
Let’s make a deal

OK, Kalu. Confess,

on Bhola, and you get a
reduced sentence of one
year in prison.
What will Bhola
do?
 The payoff matrix

Bhola

Stay Confess
Mum
Stay 2 years, 8 years,
Mum 2 years 1 year
Kalu
Confess 1 year, 5 years,
8 years 5 years
 Game Theory

 Probability Describes what an agent should

Theory believe based on evidence.

+
 Utility Theory Describes what an agent wants.

 GameTheory Describes what an agent should

do.
Utility Theory
Utility Theory based on:
rationality
maximization of utility
It is a quantification of a person's
preferences with respect to
certain objects.
How would game theory view
this?
Indiana Jones Scenario
Rationality
Assumptions:
humans are rational beings
humans always seek the best alternative in
a set of possible choices
Why assume rationality?
narrow down the range of possibilities
predictability
What is Game Theory?

Game theory is a study of how to

mathematically determine the
best strategy for given conditions
in order to optimize the outcome
 Game Theory

Finding acceptable, if not optimal,

strategies in conflict situations.
Abstraction of real complex
situation
Game theory is highly
mathematical
Game theory assumes all human
interactions can be understood
and navigated by presumptions.
Why is game theory
important?
Allintelligent beings make
decisions all the time.
AI needs to perform these tasks
as a result.
Helps us to analyze situations
more rationally and formulate an
acceptable alternative with
respect to circumstance.
 Zero-Sum Games

The sum of the payoffs remains

constant during the course of the
game.
Two sides in conflict
Being well informed always helps a
player
Non-zero Sum Game
The sum of payoffs is not
constant during the course of
game play.
Players may co-operate or
compete
Being well informed may harm a
player.
 Environments

 Decision under
certainty,
e.g., Product mix
problems
 , Decision under risk
e.g., Decision tree
 Decision under
uncertainty,
e.g., game theory
Decision Under Risk
Two concepts Here:
 Actual
Money: .5X1,00,000=50,000,
where .5 is probability of
occurrence of gaining 1,00,000 or
zero (both equal) if one invests
20,000 in a venture
Utility: This proposal will have no
utility for some one having only
20,000 in bank as compare to
Contd.
Mathematically let Z be a
random variable with expected
value E (z) and variance ², the
sample ave Z¯ will have variance
²/n, as n → ∞, ² becomes 0 and
Z¯ approaches E (z)
This is useful when a single stage
decision is taken. To handle
multiple stage decision process,
decision tree is used
 Decision Tree
Example :
High Demand (0.75)
1,000,000/year
2
i ld 300,000/year
u
B rge t 00 Low Demand (0.25)
0
la lan 00,
p 5,0 High demand
1 (0.75) 900,000/year
nd Low demand
1,

Expa 5
00

(0.25)
Bu nt

200,000/year
0,

,000
pla

ar
i ld

250,000/ye
00

4 4, 20 0
0
sm

d
High deman 6 (0.75) 250,000/year
all

3 Don’t e
xpand
( 0.75) Low demand
Lo 200,000/year
(0 wd (0.25)
.2 em
5) an
d 200,000/year

2 years ( stage 1 ) 8 years ( stage 2)

1. Full size plant with high (low ) demand will yield
1,000,000(300,000)per year
2. Small plant and low demand will yield
200,000/year.
3. Small plant with high demand will yield 250,000
for each of the 10 years.
4. Expanded small plant with high ( low) demand
will yield 900,000(200,000)/ year.
5. Small plant with no expansion and high demand
in the first for years followed by low demand will
yield 200,000 in each of the remaining 8 years.
 Evaluation

At node 4 :
(900,000 x 0.75 + 200,000 X 0.25) X 8 –
4,200,000 = 1,600,000.
With no expansion
(250,000 x 0.75 + 200,000 x 0.25 ) x 8 =
1,900,000.
 Replace all branches after 4 by 1,900,000 /-.

At node 1 :
(1,000,000x 0.75 + 300,000 x 0.25) x 10 –
5,000,000 = 3,250,000.
And
 Method to handle Decision
under uncertainty

The Laplace criterion

(more optimistic condition)
The Mini Max criterion
(minimize the maximum loss)
The savage criterion
(less conservative –have a regret matrix)
The Hurwitz criterion
(attitude ranging from most optimistic to most
pessimistic)
 Laplace Criterion
This criterion is based on what is
known as the principle of insufficient
reason.
Since their probability of occurrence of
events Q1, Q2, …. Qn are unknown, we
don’t have enough reason to conclude
that these probability will be different.
For this reason they are considered to
be equally likely to occur.
 The Laplace criterion

 Ex.: The following matrix given cost that will be

incurred if supply deviates from ideal supply for a
given customer footfall.
Customer foot fall
(which may refer to 200, 250, 300, 350 foot
fall)
Q1 Q2 Q3 Q4
a1 5 10 18 25

Supply Level a2 8 7 8 23

a3 21 18 12 21

a4 30 22 19 15
 Itassumes that Q1, …Q4 are equally likely to
occur with associated probability of 1/4.
 The expected costs of different action is
calculated as
E(a1) = ¼ [ 5 + 10 + 18 + 25 ] = 14.5
E(a2) = ¼ [ 8 + 7 + 8 + 23 ] = 11.5
E(a3) = ¼ [ 21 + 18 + 12 + 21 ] = 18
E(a4) = ¼ [ 30 + 22 + 19 + 15 ] =21.5
 The best level of inventory according to
Laplace criterion is ‘a2’.
Minimax Criterion
 This is most comparative since it is based on
making the best out of the worst possible
conditions.
 For the same example :
◦ Maximum com for row ‘1’ ( supply / inventory q 1 ) =
25
◦ Maximum com for row ‘2’ ( supply / inventory q 2 ) =
23
◦ Maximum com for row ‘3’ ( supply / inventory q 3 ) =
21
◦ Maximum com for row ‘4’ ( supply / inventory q 4 ) =
30
 This criterion given inventory level as q3.
 Savage Minimax regret criterion
 Consider the following loss matrix

Q1 Q2
a1 1100 90
0
a2 1000 1000
0 0
 The above minimax criteria will yield ‘a2’, which
logically is incorrect as with a1, with Q2 state you
loose only 90, which is very low.
 Thetoo conservative approach of minimax is
addressed in this method by constructing another
matrix, i.e. regret Matrix by subtracting smallest
element in a column to all other elements in the
same column. In theQ1 above
Q2 ex. this will lead to
a1 1000 0 1000

a2 0 9910 9910

 The minimax now given a1’ as answer since regret

matrix represent the loss (whether profit/
cost/opportunity cost ).
 Always minimax criterion is used.
Hurwicz Criterion
The criterion represents range of
attitude from the most optimum
to the most pessimistic.
It strikes a balance between
extreme conditions by weighing
the above by respective weights.
 Min Max Weighted
values
for ∞= .5
a1 5 25 15

a2 7 23 15

a3 12 21 16.5

a4 15 30 22.5
 Game theory
Based on Mini Max
Each player has finite or infinite number
of choices called strategy
The outcome or payoff of a game is
summarized as function of strategies for
each player
 Assumptions
Finite no. of players
Choices are made simultaneously, i.e.,
opponents choice is not known
List of possible course of action is
available to each player
Every combination of course of action is
associated with an outcome
Zero sum game
• Pay off matrix
• ( +ve for player A – Maxi Min player
-ve for player B - Mini Max player )
Ex: Two players tossing coins
B
H T

A H 1 -1
T -1 1
Bidding problem

Two items: 1 II
Rs. 75 Rs. 125

Two bidders: A 100/-

War chest is usually
known to some degree of
B 130/-
certainty in real life……
Example

‘A’ has two ammunition stores, one of

which is twice as valuable as the
other. ‘B’ is an attacker who can
destroy an under undefended store
but he can only attack one of them
‘A’ knows that ‘B’ is about to attack
but does not know which. ‘A’ can
successful defend only one store at a
time.
 Attack smaller
Attack larger

Defend smaller store 0

-2

Defend larger store -1

0
Solution of 2x2 game
SaddlePoint
Row Dominance
Column Dominance
Graphical solution of (2xN) and (Mx2)
games
 Graphical solutions are only applicable to games in
which at least one of the players has two strategies
only. Consider the following (2xN) game.
B
y1 y2 … yn
x1 a11 a12 … a1n
A
x2=1-x1 a21 a22 … a2n

It is assumed that the game does not have a saddle point.

Since A has two strategies, it follows that x2 = 1 – x1; x1 ≥ 0,
x2 ≥ 0, His expected payoffs corresponding to the pure
strategies of B are given by:

B’s pure strategy A’s expected payoff

1 (a11 - a21)x1 + a21
2 (a12 - a22)x1 + a22
.
.
.
n (a1n – a2n)x1 + a2n
Example: Consider the following (2x4) game.
B
1 2 3 4
A 1 2 2 3 -1
2 4 3 2 6
This games does not have a saddle point. Thus
A’s expected payoffs corresponding to B’s pure
strategies are given as follows:
B’s pure strategies A’s
expected payoffs
1 -2x1 + 4
2 -x1 + 3
3 x1 + 2
4 -7x1 + 6
 +6
4
+5

+4 1 maximin

Average +3 2
payoff 3
+2 v* = 5/2

+1

X1=0
X1=1
-1 X1 * = 1/2
-2
Example: Consider the following (4x2) game:
B
1 2
1 2 4
A 2 2 3
3 3 2
4 -2 6
This game does not have a saddle point. Let y1 and y2(=1-
y1) be B’s mixed strategies. Thus
A’s pure strategy B’s expected payoff
1 -2y1 + 4
2 -y1 + 3
3 Y1 + 2
4 -8y1 + 6
 The point x1* is determined by solving:
-x1* + 3 = 2x1* + 2

+6
Upper
4 envelope
+5
minimax
+4 1
Average
2
payoff +3
3 v* = 8/3
+2

+1

Y1 = 0

Y1* = 2/3
-1

-2
 LP Approach to Game Theory
1 2 - - n
1 a11 a12 - - a1n
2 a21 a22 - - a2n
` ` ` `
` ` ` `
m am1 am2 - - amn
 = total number of alternatives for player ‘A’
= total number of alternatives for player ‘B’
= payoff matrix
= value of game
= probability of selection of alternative by player ‘A’;

= probability of selection of alternativeby player ‘B’;

 Development of LP from A’s perspective

Player ‘B’ Alternative Strategy Expected gain function

of ‘A’

`
`
Descriptive Model for ‘A’:
,. . . } ]
subject to

The descriptive model is converted into LP model for ‘A’:

Let

The LP may be:

subject to
;
After Dividing both sides of constraints by
The constraints can be transformed into:

. . . . (eq. 1)
Let; for

[,,...]
and

] (From eq.1)
The model is now:
+...+
subject to

The Value of ‘v’ can be obtained by

and for
Example: Consider the following (3x3) game:
B
1 2 3 Row minimum
1 3 -1 -3 -3
A 2 -3 3 -1 -3
3 -4 -3 3 -4

Column 3 3 3
maximum
Since the maximin value is -3, it is possible
that the value of the game may be negative or
zero. Thus a constant K, which is at least equal
to the negative of the maximin value, is added
to all the elements of the matrix, that is K ≥ 3.
Let K = 5. The preceding matrix becomes
 B
1 2 3
1 8 4 2
A 2 2 8 4
3 1 2 8
B’s linear programming problem is thus given as:
maximize w = y1 + y2 + y3

Subjected to: 8y1 + 4y2 + 2y3 ≤ 1

2y1 + 8y2 + 4y3 ≤ 1
y1 + 2y2 + 8y3 ≤ 1
y 1, y 2 , y 3 ≥ 0
The final optimal tableau for this problemis given by:

Basic y1 y2 y3 s1 s2 s3
solution
w 0 0 0 5/49 11/196 1/14 45/196
y1 1 0 0 1/7 -1/14 0 1/14
y2 0 1 0 -3/98 31/196 -1/14 11/196

y3 0 0 1 -1/98 -3/98 1/7 5/49

Thus, for the original problem,

v* = - K = 196/45 – 5 = -29/45
y1* = = = 14/45
y2* = = = 11/45
y3* = = = 20/45
Thanks…