0% found this document useful (0 votes)
3 views

01 - introduction

The document introduces Game Theory and Information Economics, emphasizing the importance of interactive decision-making in various contexts such as auctions and strategic interactions. It outlines the course structure, covering static and dynamic games, incomplete information, and applications like signaling and moral hazard. Additionally, it explains the concept of mixed strategies and provides examples of classic games like Matching Pennies and the Prisoners' Dilemma.

Uploaded by

qygkwjj6rt
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
3 views

01 - introduction

The document introduces Game Theory and Information Economics, emphasizing the importance of interactive decision-making in various contexts such as auctions and strategic interactions. It outlines the course structure, covering static and dynamic games, incomplete information, and applications like signaling and moral hazard. Additionally, it explains the concept of mixed strategies and provides examples of classic games like Matching Pennies and the Prisoners' Dilemma.

Uploaded by

qygkwjj6rt
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 30

G5212 (Game Theory & Info Econ): Introduction

Guillaume Haeringer

Spring 2025

1/30
Why Game Theory?

Basic microeconomics (consumer, firm, competitive eq.) does not


consider interactive decision making:

the outcome for each person depends on their actions and the
actions of the other person

I People bidding for the same item on eBay

I People working together on a joint project

I Generals deciding where to position their armies

I Oligopolistic firms setting prices for similar products

2/30
Why Game Theory?

For interactive settings, optimization (like for consumer or the


firm) will not get us very far:
I Best bid of bidder 1 depends on bid of bidder 2
I Best bid of bidder 2 depends on bid of bidder 1

→ Need some way of solving both problems together

This (basically) is what game theory studies

Applied to: mating displays of birds, banking crises, spectrum


auctions, kidney exchanges, insurance, school choice, political
platforms, sport, war, and more

3/30
Why mathematical models?

Empiricism (descriptive models):


I May want to make predictions about the world
I Math can do heavy lifting, as for physicists

Advice (normative models):


I Modeling others’ behavior helps optimize yours
I Math can be a parsimonious language

Storytelling (rhetorical models):


I Model can highlight hypothetical cause and effect
I Math forces a story to be internally consistent

For all purposes: Forced to make implicit assumptions explicit

4/30
The Plan (For The Course)

Part 1: Game Theory (first half of course)

I Focus on tools
I Static games of complete information
I Dynamic games of complete information
I Games of incomplete information
I Solution concepts and refinements

I With a few applications along the way


I Bargaining
I Auctions
I Voting

5/30
The Plan (For The Course)
Part 2: Information Economics
Focus on applications in the face of asymmetric information
I Example 1: Signaling
I Companies want to recruit people of high ability
I But they cannot observe ability directly
I Can education be used by high ability candidates to signal that
they have high ability?
I Example 2: Moral Hazard
I A boss wants to encourage their worker to work hard
I But they cannot observe effort directly, only outcomes (which
have a random component)
I How should they design their incentive scheme?
I Example 3: Optimal selling mechanisms
I Seller makes a product, doesn’t know how much it’s worth to
prospective buyers
I How to design auction rules?
6/30
The Plan (for today)

I A gentle introduction!
I Talk through some ‘classic’ games
I Formal definition of a game
I Mixed strategies

7/30
Example: Matching Pennies

Bob
H T
H 1, −1 −1, 1
Alice
T −1, 1 1, −1

I Two players (Alice and Bob) each reveal a penny showing


heads or tails
I If the pennies match then Bob pays Alice a dollar
I If not, Alice pays Bob a dollar
I This is the matrix form of this game

8/30
Example: Prisoners’ Dilemma

Bob
Defect Cooperate
Defect −6, −6 0, −9
Alice
Cooperate −9, 0 −1, −1

I Classic story of two prisoners who must decide whether or not


to confess
I But many other (more economically interesting) applications:
e.g. pollution, traffic, pricing (collusion)

9/30
Example: partnership

I Effort E produces an output of 6 at a cost of 4, with output


shared equally;
I shirking S produces an output of 0 at a cost of 0.

Bob
S E

6
S 0 − 0, 0 − 0 2 − 0, 62 − 4
Alice

6
E 2 − 4, 62 − 0 2× 6
2 − 4, 2 × 6
2 −4

That’s also a Prisoner’s dilemma

10/30
Example: partnership

After some heavy and tedious calculations we get a simpler version


of the game:

Bob
S E
Alice S 0, 0 3, −1
E −1, 3 2, 2

11/30
Prisoner’s dilemma

The “general” form of the prisoner’s dilemma is given by


Bob
Defect Cooperate
Defect P, P T,S
Alice
Cooperate S, T R, R

where

T(emptation) > R(eward) > P(unishment) > S(ucker)

12/30
Defining a Game
A simultaneous-move game (a.k.a strategic form game, normal
form game) consists of 3 elements:

1. The players

2. The actions that each player can take

3. The payoffs associated with each set of actions

We are initially making some ‘hidden’ assumptions:


I Players move at the same time
I All payoffs are known to all players

Later in the course we will relax these assumptions, and so a


description of the game will also include
I The sequence of play
I Who knows what
13/30
Defining a Game
Definition
An n-player normal form (or strategic form) game G is an n-tuple
{(S1 , u1 ) , ..., (Sn , un )} , where for each i

(1) Si is a nonempty set, called i’s strategy space, and

(2) ui : nk=1 Sk → R is called i’s payoff function.


Q

Notation
I S := nk=1 Sk
Q

I s := (s1 , ..., sn ) ∈ S

I S−i := k6=i Sk
Q

I (si0 , s−i ) := (s1 , ..., si−1 , si0 , si+1 , ..., sn )

I u : S → Rn
14/30
What are we assuming about individuals?

Preferences are transitive


I The cornerstone of rationality

Preferences are complete


I Sometimes without loss (but not always!)

Preferences are quantifiable


I Essentially without loss

Players only care about....


I ... money? NOT assuming that!
I ... themselves? NOT assuming that!

15/30
Payoffs in games

A player’s payoff (function) in a game comes from her preferences


over outcomes.

Those preferences may depend on:


I The player’s ‘direct’ outcome (the object of money she gets)
I Other players’ outcomes (which may include subjective
elements).

So, a player
I only looks at her own payoff to evaluate an outcome
I BUT may want to look at the others’ payoffs to ‘guess’ what
they’ll do.

16/30
Example: Second-Price Sealed-Bid Auction

I A seller has one indivisible object.

I There are n bidders with respective valuations for the object

0 ≤ v1 ≤ · · · ≤ vn

(which are common knowledge).


I The bidders simultaneously submit bids.

I The highest bidder wins the object and pays the second
highest bid.
In the case of a tie all winning bidders are equally likely to
have their bid accepted.

17/30
Example: Second-Price Sealed-Bid Auction
Formally, the 2nd price auction is described as follows:

I Players: 1, ...., n

I Strategy set for each player i = 1, . . . , n:

Si = [0, ∞)

I Payoffs:
Given a profile of bids, s, the set of highest bidders is

W (s) ≡ {k : ∀j, sk ≥ sj }

Then we have

 vi − maxj6=i sj if si > maxj6=i sj
1
ui (si , s−i ) = (vi − si ) if si = maxj6=i sj
 |W (s)|
0 if si < maxj6=i sj .
18/30
Example: Cournot Duopoly

I There are two firms, call them 1 and 2, producing perfectly


substitutable products.
I The market demand is P(Q) = max {a − Q, 0}, Q = q1 + q2 .

I The cost of producing qi is given by C (qi ) = cqi , 0 < c < a.

I The two firms choose quantities simultaneously.

19/30
Example: Cournot Duopoly

Formally, we have
I Players: 1,2
I Strategies Si = [0, ∞).
I Payoffs
ui (q1 , q2 ) = [P (q1 + q2 ) − c] qi .

20/30
Example: Voting

I There are three players i = 1, 2, 3.

I There are two candidates a and b which they can vote for.

I The voting rule is the majority rule.

I Voters’ preferences are as follows (with preferred candidate


listed higher)
1 2 3
a b b
b a a

I A player receives a payoff of 1 if his favorite candidate wins; 0


else.

21/30
Example: Voting

I Players: 1,2,3

I Strategies Si = {a, b}

I Payoffs (for Player 1):

u1 (a, a, a) = 1 u1 (b, a, a) = 1
u1 (a, a, b) = 1 u1 (b, a, b) = 0
,
u1 (a, b, a) = 1 u1 (b, b, a) = 0
u1 (a, b, b) = 0 u1 (b, b, b) = 0

22/30
Mixed Strategies

Consider again the matching pennies game


Here is another action Bob could take: rather than put the coin
down H or T, he could flip it, and play whichever way the coin falls

This is a new strategy: it is not H or T, but a 50% chance of


H and a 50% chance of T

More generally, we might like to extend the player’s strategy space


to allow them to randomize between pure strategies

These are ‘Mixed Strategies’

They will be useful going forward. . .

23/30
Mixed Strategies

Definition
Suppose {(S1 , u1 ) , ..., (Sn , un )} is an n-player normal-form game.
A mixed strategy for player i is a probability distribution over
elements of Si , denoted by σi ∈ ∆ (Si ) .
Strategies in Si are called pure strategies.

Remarks
I In most cases, we assume Si is finite. Then σi : Si → [0, 1]
P
s.t. si ∈Si σi (si ) = 1.
I When Si is not countable, it’s more work/math to define
mixed strategies (mixed strategy sets become infinite
dimensional. . . )
→ We will not worry about this ;-)

24/30
Mixed Strategies
Qn
Extend ui to j=1 ∆ (Sj ) by taking expected values. If Si is finite:
X X
ui (σ1 , . . . , σn ) := ··· ui (s1 , ..., sn ) σ1 (s1 ) σ2 (s2 ) · · · σn (sn ) .
s1 ∈S1 sn ∈Sn

So, we’re simply using expected utility (i.e., we’re vNM utilities,
but does not mean we’re assuming risk neutrality).

Notation:
X Y
ui (si , σ−i ) := ui (si , s−i ) σj (sj )
j6=i
s−i ∈S−i
X
ui (σi , σ−i ) := ui (si , σ−i ) σi (si )
si ∈Si

25/30
Mixed Strategies

Games are usually described with pure strategy sets:

{(S1 , u1 ) , ..., (Sn , un )}

And then we say we’ll allow players to use mixed strategies:

{(∆(S1 ), u1 ) , ..., (∆(Sn ), un )}

26/30
Mixed Strategies

Do mixed strategies mean that all distributions over strategy


profiles can arise?

No, because we don’t allow for correlation:


Yn Y n 
∆ (Sj ) 6= ∆ Sj = ∆ (S)
j=1 j=1
| {z } | {z }
No correlation Correlation

We have the same for strategy profiles:


Y
(σ1 , ..., σi−1 , σi+1 , ..., σn ) = σ−i ∈ ∆ (Sj ) 6= ∆ (S−i )
j6=i

27/30
Correlated v. non-correlated strategies
Bob
L R
U α β
Alice
D γ δ

SA = {U, D} ⇒ ∆(SA ) is a probability distribution over {U, D}


SB = {L, R} ⇒ ∆(SB ) is a probability distribution over {L, R}.

So, we have

Yn
∆ (Si ) = ∆(SA ) × ∆(SB )
i=1
= a probability Alice plays U (and D)
AND a probability Bob plays L (and R)

⇒ Alice and Bob randomize independently (non-correlated). 28/30


Correlated v. non-correlated strategies

If instead we look at
Yn 
∆ Si = ∆(SA × SB )
i=1
= a probability Alice and Bob play (U, L),
and (U, R), and (D, L), and (D, R))

⇒ Alice and Bob randomize together (correlated).

29/30
Summary

Key things from today

I Understand what a game is

I Understand how to translate a story into a game

I Understand what mixed strategies are

I Understand how to translate a game in pure strategies into a


game with mixed strategies

30/30

You might also like