2nd year Master University Amar Télidji Laghouat

Decision Support
Systems
Decision Under Uncertinity
Game Theory
Younes Guellouma
Chapter 4: Decision Making under
Uncertainty
Game Theory

Younes Guellouma

1
Introduction
Introduction

Game theory is a mathematical paradigm that means “a study of

mathematical models of conflicts and cooperation that may exist
between a group of rational decision-makers”. Game theory is used
primarily in economics, political science, psychology, as well as
logic, and especially in computer science. This theory experienced
rapid development starting in the 1950s. Many schools adopted
new techniques in this area. This field of application has been the
recipient of numerous Nobel Prizes in economics.

2
Introduction (2)

Game theory, in its form, is similar to decision making in an

uncertain environment, except that the states of nature are
replaced by decision makers in direct confrontation. Each decision
of a decision-maker is influenced by the decisions of other
opponents. This is referred to as “strategies”. Therefore, game
theory is primarily used to model situations where social actors
make separate individual decisions, but these decisions have a
combined impact on the actors.

3
Introduction (3)

Here, we give a generic description of a game.

1. There are at least two players.

2. The game begins with a choice made by a player from several
alternatives.
3. This first choice defines a certain situation that determines
the next choice (who and what).
4. The choices made by one player may or may not be known to
the other players.
5. If a game is described by these successive choices, there is a
rule for the end of the game.
6. The final situation of a game determines the reward for each
participant.
4
Cooperative and Competitive
Games
Cooperative Games

A game is said to be cooperative when players can freely

communicate with one another and make agreements (e.g., in the
form of a contract). They then form a coalition and seek the
collective interest, followed by a distribution of the gains among all
players.
In a non-cooperative game, players (who do not communicate or
cannot communicate with each other) act based on the principle of
economic rationality: each one seeks to make the best decisions for
themselves (that is, each selfishly seeks to maximize their
individual gains). This latter type of game involves probabilities.

5
Competitive Games

A game is called a ”zero-sum” game when the sum of the players’

gains is constant (or can be made constant through the careful
choice of a utility function). In other words, what one player gains
is necessarily lost by another (e.g., chess).
Board games are often zero-sum games, but real-world situations
are often better described by non-cooperative, non-zero-sum
games, as some outcomes can be beneficial for everyone or harmful
for everyone (e.g., political life, business situations).

6
Formalism
Formalism

Definition
A normal-form game (or strategic-form game) is defined by:

• A finite set of players P = {1, . . . , n}.

• A set of actions Si , 1 ≤ i ≤ n (pure strategies) for each player i.
• The set of all possible strategies

n
Y
S= Si = S1 × . . . × Sn
i=1

.
• The utility function ui : S → R for player i, 1 ≤ i ≤ n.
Interpretation: For player i and the tuple (s1 , . . . , si , . . . , sn ), called the strategy
profile, ui represents the payoff for player i if every player k, k ̸= i, chooses the
strategy sk .
• The overall utility U : S → Rn . For each profile − → s = (s1 , . . . , sn ) belonging to
S, U(− →
s ) = (u (−→
1s ), . . . , u (−
n
→
s )).

7
Formalism (2)

Example
In the game of chess, two players compete, where each player tries to make a move
based on what their opponent plays. Here, we have:

• The set of players is P = {1, 2} where |P| = 2. 1 controls the light pieces, and
2 controls the dark ones.
• For player 1, the set of possible actions S1 is calculated at each step. For
example, if pawn P5 is located on square e4, the only accessible position for it
is f 5. Therefore, the strategy e4f 5 belongs to S1 .
• A strategy profile includes one action from 1 and another from 2. For example,
the profile −
→
s = (e4f 5, f 8a3) indicates that the light player moves the 5th pawn
to f 5 capturing the dark player’s pawn in that move; and 2 moves the first
bishop to a3.
• The utility function can be defined based on the capture of opponent pieces.
For example, if capturing a pawn is worth 1 point, then the overall utility of the
profile −
→
s is: U(−→s ) = (1, −1).

8
9
Representation of Utility Function
Representation of Utility Function

There are two ways to represent a game: the matrix form, which is
generally used for simultaneous games, and the extensive form,
which is useful for representing sequential games.

10
Normal Form

The Payoff Matrix is a double-entry table that lists on each side

the possible strategies of the respective players. In the cell at the
intersection of two strategies, the pair of payoffs for the two
players is recorded.

e o
e 1,-1 -1,1
o -1,1 1,-1

11
Extensive Form

Extensive Form is a tree representation which is suitable for

sequential games. Each level represents the choices of a player. A
leaf contains the utility of the profile present in the branch between
the root and the leaf.

12
Example

Example
“Dell versus Compaq” Game: Both companies want to choose an operating system
for their machines. The constraints of the game are as follows :

• Each company must choose its operating system (Linux or Windows).

• Dell prefers Windows, and Compaq prefers Linux.
• Both manufacturers benefit from having the same OS.

The formalism of the game is :

• P = {Dell, Compaq},
• SDell = SCompaq = {windows, linux}
• S = {(windows, windows), (windows, linux), (linux, windows), (linux, linux)}
• U(windows, windows) = (600, 200), U(linux, linux) =
(200, 600), U(windows, linux) = U(linux, windows) = (100, 100).

13
Normal Form of the game

Dell / Compaq Windows Linux

Windows (600, 200) (100, 100)
Linux (100, 100) (200, 600)

14
Extensive Form of the game

15
Game Solution
Dominant Strategy

We now begin the study of predicting the outcomes of normal-form

games. Given a game, and assuming the players are rational, what can
we predict they will play, or what can we predict they will not play? If we
were to play such a game, what choice would we make? Throughout the
following, we will use the following notations:

n
Y
S= Si
i=1

We define

Y
S−i = Sj
j̸=i

Let s = (s1 , . . . , si−1 , si+1 , . . . , sn ) ∈ S−i . The profile s · t with t ∈ Si is

s · t = (s1 , . . . , si−1 , t, si+1 , . . . , sn )
16
Prisoner’s dilemma

Prisoner’s dilemma
Consider the following example (the Prisoner’s Dilemma):
Two suspects, 1 and 2, are arrested by the police. However, the
officers do not have enough evidence to convict them, so they
interrogate them separately, offering them the same deal: “If you
testify (t) and the other one stays silent (s), you will be set free,
and the other will receive 10 years. If you both testify, you will
both receive 5 years. If neither of you testify, you will both
receive 6 months”.

17
Example (2)

1/2 Stays silent Testifies

Stays silent (−1/2, −1/2) (−10, 0)
Testifies (0, −10) (−5, −5)

18
Dominant Strategy

Definition
A strategy s ∈ Si for a player i is said to be strictly dominant if,
for every other strategy t ∈ Si , we have:

∀s−i ∈ S−i , ui (s−i · t) < ui (s−i · s)

19
Example

Example
Taking the same example as before. For player 1, we notice that:

u1 (t, t) = −5 > u1 (s, t) = −10 (1)

u1 (t, s) = 0 > u1 (s, s) = −1/2 (2)

We notice that the strategy ’testify’ is better than ’stay silent’ for
1, regardless of what 2 chooses.

20
Uniqueness theorem

Theorem
If a strategy s ∈ Si is strictly dominant for a player i, then it is
the only one with this property.

21
Dominated Strategy

Definition
A strategy s ∈ Si for a player i is said to be strictly dominated if
there exists a strategy t ∈ Si such that

∀s−i ∈ S−i , ui (s−i · t) > ui (s−i · s)

22
Example (3)

Let us choose the strategy t (stay silent) for player 1. The other
strategy that player 1 can choose is d (testify). We can observe
that
u1 (s, t) = −10 < u1 (t, t)

We can say that the strategy “stay silent” is a dominated strategy

for player 1.”

23
Nash Equilibrium

This concept describes a stable outcome when individuals or

institutions make rational choices based on how they believe others
will act.
Definition
Let s = (s1 , . . . , sn ) be a strategic profile. The tuple s−i ∈ S−i
for 1 ≤ i ≤ n is equal to (s1 , . . . , si−1 , si+1 , . . . , sn ). The profile s
is said to be a Nash equilibrium if:

∀i ∈ {1, . . . , n}, ∀ti ∈ Si , ui (s−i · si ) ≥ ui (s−i · ti ) (3)

24
Example

Example
Consider the prisoner dilemma.

1. Let us try the profile (s, s). We notice that for 1:

u1 (s, s) = −1/2 < u1 (t, s) = 0, and for 2 we have
u2 (s, s) = −1/2 < u2 (s, t) = 0. Hence, (s, s) is not a Nash
equilibrium.
2. For the profile (s, t), we have for 2:
u2 (s, t) = 0 > u2 (s, s) = −1/2, but for 1:
u1 (s, t) = −10 < u1 (t, t) = −5. Here, although the condition
holds for 2, it does not hold for 1.
3. Only the profile (t, t) is a Nash equilibrium.

25
Theorem
If there exists a dominant strategy si for every player i, then the
strategic profile (s1 , . . . , sn ) is a Nash equilibrium.

26
Pareto Efficiency

Let s, s ′ ∈ S be two strategic profiles. We say that s

Pareto-dominates s ′ if and only if:

∀i ∈ {1, . . . , n}, ui (s) ≥ ui (s ′ ) ∧ ∃j ∈ {1, . . . , n}, uj (s) > uj (s ′ )

A strategic profile is a Pareto optimum if no other profile

Pareto-dominates it.

27
Example

We notice that the Nash equilibrium from the previous example

((t, t)) is Pareto-dominated by the profile (s, s).

28
Example (2)

J1 / J2 A B C D
E (5, 4) (8, 1) (2, 2) (5, 1)
F (3, 5) (3, 6) (1, 7) (5, 5)
G (1, 3) (2, 8) (6, 3) (4, 5)

29
Example (2)

J1 / J2 A B C D
E (5, 4) (8, 1) (2, 2) (5, 1)
F (3, 5) (3, 6) (1, 7) (5, 5)
G (1, 3) (2, 8) (6, 3) (4, 5)

The Pareto optima are : (E , B), (F , B), (F , D), (G , B) et (G , C )

29
Pareto Front

30