2 Game Theory

2.
- Game Theory Overview
Game Theory studies situations in which

the outcome of the decisions of an indi-
vidual depends on the decisions taken by
other individuals as well as on his own
- Strategic Interdependence -
2.- Game Theory Overview
Plan
2.1 Introduction
2.2 Static Games of Complete Information: Nash Equilibrium
2.3 Dynamic Games of Complete Information: Backwards Induction and Subgame Perfection
2.4 Static Games of Incomplete Information:Bayesian Equilibrium
2.5 Dynamic Games of Incomplete Information: Perfect Bayesian Equilibrium and Sequential
Equilibrium
2.1 Introduction
Any (non cooperative) game has the following elements
1. The Players Who is to take decisions
2 The Rules Who moves first, who moves next, what can be done and what can not, ...
3 The Strategies What actions are available to the players
4 The Outcomes What is the consequence of each combination of actions by the players
5 The Payoffs What each player obtains in each possible outcome

2.- Game Theory Overview 2.1 Introduction
Example: Rock - Scissors - Paper
1. The Players Player 1 and Player 2
2 The Rules Simultaneously, both players show either rocks, scissors or paper
3 The Strategies Each player can choose among showing rocks, scissors or paper
4 The Outcomes Rock beats scissors, which beats paper, which beats rock
5 The Payoffs The winning player receives 1 from the loser (thus getting -1)
2.- Game Theory Overview 2.1 Introduction
Game Scenarios
Complete Information Incomplete Information

Static Players have all relevant information Not all players have all the information
and move simultaneously and move simultaneously
Dynamic Players have all relevant information Not all players have all the information
and move in turns and move in turns
Examples
Complete Information Incomplete Information

Static Rock, Scissors, Paper Sealed-bid Auction
Dynamic Chess English Auction
2.2 Static Games of Complete Information
Players take decisions without knowing the actions taken by other players
Basic Assumptions
1. Players are rational and self-interested (Rationality)
2. Players have full information about all the elements of the game (Complete
Information)
3. Players know that players are rational and fully informed, and they know
that others know that players are rational and fully informed, and they
know that others know that they know that players are rational and fully
informed, and ... (Common Knowledge)
2.- Game Theory Overview 2.2 Static Games of Complete Information
Common Knowledge examples
1. The Byzantine Generals and the Coordinated Attack Problem
2. The Three Hats Problem
3. The Princes Bride
4. La vida es sueño (Pedro Calderón de la Barca - 1636)

Notation
• I = {1, 2, . . . , I} is the Set of Players
• S i = {si1, si2, . . .} is the Set of Pure Strategies of player i ∈ I
• S = S 1 × S 2 × · · · × S I is the Set of pure strategy profiles
• (s∗i , s−i) = (s1, s2, . . . , si−1, s∗i , si+1, . . . , sI )
• ui : S → R is the (von Neumann - Morgenstern) Payoff function of player i ∈ I
A Game is fully characterized
Γ = (I, {S i}i∈I , {ui}i∈I )
This representation of a game is know a normal form game or strategic form game,.
Example 1: The Prisoners’ Dilemma
• 2 players I = {1, 2}
• 2 (identical) strategies for each player S i = {C, D} (Cooperate, Defect) i ∈ I
• Strategy profiles in S = {C, D} × {C, D} = {(C, C), (C, D), (D, C), (D, D)}
• Symmetric payoffs given by

– u1(C, C) = 3 u²(C, C) = 3
– u¹(C, D) = 0 u²(C, D) = 5
– u¹(D, C) = 5 u²(D, C) = 0
– u¹(D, D) = 1 u²(D, D) = 1
This is usually represented by a bi-matrix table

Example 1: The Prisoners’ Dilemma (continued)

1\2 C D
C 3, 3 0, 5
D 5, 0 1, 1
Independently of the other player strategy, each player will strictly prefer to play Def ect (D)
Any game satisfying the relationships below is “a prisoners’ dilemma”
1\2 s21 s22
s11 R, R S, T
s12 T, S P, P
with T > R > P > S
T → Temptation payoff R → Reward payoff
P → Punishment payoff S → Sucker payoff
This is a game in which there are gains from cooperation, but a binding agreement is not possible
(neither stable !). Situations such as price-fixing/collusion and free-riding can be understood
using a prisoner’s dilemma model.
It is a compelling example that individual self-interest might not lead to social
efficiency
Example 2: The Battle of the Sexes
I ={1, 2}, S i = {F, O} (Football, Opera)

1\2 F O
F 2, 1 0, 0
O 0, 0 1, 2
The players want to coordinate, but have conflicting interests. Any coordination is a possible
result.
Example 3: Meeting in Barcelona
I ={1, 2}, S i = {P, R} (Plaça Catalunya, Rambles)
1\2 P R
P 2, 2 0, 0
R 0, 0 1, 1
The players want to coordinate, and have common interests. Any coordination is a possible
result, but one of them is “more possible” than the other (Schelling’s Focal Point)
Example 4: Rock-Scissors-Paper
I ={1, 2}, S i = {R, S, P } (Rock, Scisors, Paper)
1\2 R S P
R 0, 0 1, −1 −1, 1
S −1, 1 0, 0 1, −1
P 1, −1 −1, 1 0, 0
In this case, what is possible outcome of the game ? (Zero-Sum Game)

Main Question
• What behavior should we expect form

a rational, fully informed, and self-
interested player ?
• If all players behave in such way, what

will be the outcome of the game ?
- EQUILIBRIUM -
• People (players) don’t do silly things (elimination)
• People’s behavior will be stable if nobody has incentives to change

The Equilibrium
Dominant and Dominated Strategies
The simplest way to consider rational behavior by the players is that of Dominant and Dominated
Strategies.
Consider the static game
Γ = (I, {S i}i∈I , {ui}i∈I )
Definition. A strategy si ∈ Si is a Strictly Dominant Strategy for player i ∈ I if, for any
other s′ ∈ Si (s′i 6= si),
ui(si, s−i) > ui(s′i, s−i)
for all s−i ∈ S−i
Example The strategy D (Defect) is a Strictly Dominant Strategy for both players in the
Prisoners’ Dilemma
Definition. A strategy si ∈ Si is a Strictly Dominated Strategy for player i ∈ I if there exists

another strategy s′ ∈ Si (s′i 6= si) such that, for all s−i ∈ S−i
ui(s′i, s−i) > ui(si, s−i)
In such case, we say that the strategy s′i strictly dominates si for player i
Example The strategy C (Cooperate) is a Strictly Dominated Strategy for both players in the
Prisoners’ Dilemma. (The strategy D dominates C)
Note A strategy si ∈ Si is a Strictly Dominant Strategy for player i ∈ I if it strictly dominates

every other strategy of player i
Definition. A strategy si ∈ Si is a Weakly Dominant Strategy for player i ∈ I if, for any
other s′ ∈ Si (s′i 6= si),
ui(si, s−i) ≥ ui(s′i, s−i)
for all s−i ∈ S−i
Definition. A strategy si ∈ Si is a Weakly Dominated Strategy for player i ∈ I if there exists

another strategy s′ ∈ Si (s′i 6= si) such that, for all s−i ∈ S−i
ui(s′i, s−i) ≥ ui(si, s−i)
In such case, we say that the strategy s′i weakly dominates si for player i
Example
1\2 L R
T 1, 3 0, 2
B 0, 2 0, 3
T weakly dominates B. That is, B is weakly dominated by T . Or, T is a weakly dominant
strategy
Comments
• A strictly dominant strategy, if it exists, must be unique
• Weakly dominant strategies, if any, might not be unique
• Dominated strategies (both strictly and weakly) might not be unique

Iterated elimination of (strictly) dominated strategies
Strictly dominated strategies should never be used by a rational player. The assumption of
common knowledge allows players to effectively “delete” such strategies. Sometimes, this leads
to some reduction of the possible outcomes of the game
Example
1\2 L C R
T 3, 0 0, −5 0, −4
M 1, −1 3, 3 −2, 4
B 2, 4 4, 1 −1, 8
R strictly dominates C. Hence,
1\2 L R
T 3, 0 0, −4
M 1, −1 −2, 4
B 2, 4 −1, 8
B strictly dominates M . Hence

1\2 L R
T 3, 0 0, −4
B 2, 4 −1, 8
T strictly dominates B. Hence
1\2 L R
T 3, 0 0, −4
Finally, L strictly dominates R. Thus, we finally have
1\2 L
T 3, 0
The final outcomes of the game are reduced to (only) (T, L) which produces payoffs (3, 0)
Comments
• Common knowledge is needed for this to make sense
• In many cases it is not of any help (Battle of the Sexes)
• The order of elimination is not important for the final result
• Strategies that are dominant at the origin will remain dominant through all the process
Iterated elimination of (weakly) dominated strategies
The “deletion” of weakly dominated strategies does not have as much sense as in the case of
strictly dominated strategies. The reason is that, in this case, we can not say that “a rational
player will never use such strategy” as before. Furthermore, the process might produce a strange
result
Example
1\2 L R
T 5, 1 4, 0
M 6, 0 3, 1
B 6, 4 4, 4
Strategies T and M are weakly dominated

a) If T is removed,
1\2 L R
M 6, 0 3, 1
B 6, 4 4, 4
Now L is weakly dominated,
1\2 R
M 3, 1
B 4, 4
and then M can be eliminated to finally obtain
1\2 R
B 4, 4
b) If M is removed,
1\2 L R
T 5, 1 4, 0
B 6, 4 4, 4
Now R is weakly dominated
1\2 L
T 5, 1
B 6, 4
And now T can be deleted to finally get
1\2 L
B 6, 4
Rationalizable Strategies
Consider the following game

1\2 L R
T 4, 0 0, 2
M 1, 2 2, 1
B 0, 1 3, 2
Notice that,
• No player has a Dominant Strategy
• No player has a Dominated Strategy
• Yet, strategy M for the first player is never going to be used (It is never a good option by a
rational player)
We could extend the concept of Dominated Strategy to strategies that are not going to be used
by a rational player.
Definition. A strategy si ∈ Si is a best response for player i ∈ I to the strategies s−i of his
opponents if
ui(si, s−i) ≥ ui(s′i, s−i)
for any s′i ∈ Si
Definition. A strategy si ∈ Si is never a best response if there is no s−i for which si is a

best response
We can, therefore, replace the process of iterated elimination of dominated strategies by a process
of iterated elimination of strategies that are never a best response. Notice that:
• A strictly dominated strategy is never a best response
• A strategy that is never a best response might not be a strictly dominated strategy
Thus, more strategies are eliminated in this process o elimination of strategies that are never a
best response
Definition. The strategies in Si that survive the process of iterated elimination of strategies
that are never a best response are called player’s i rationalizable strategies
Alternatively, we can say that player’s i rationalizable strategies are those strategies in Si that
are a best response to some s−i by player’s i opponents.
In the previous example, if we eliminate M for player 1 we get
1\2 L R
T 4, 0 0, 2
B 0, 1 3, 2
Now L is strictly dominated for player 2
1\2 R
T 0, 2
B 3, 2
Now we can eliminate T
1\2 R
B 3, 2
Thus, rationalizable strategies are B for player 1 and R for player 2
Nash Equilibrium
It is clear that in many cases the process of iterated elimination of strategies cannot be started
and, thus, no prediction can be made about the outcome of the game
The Battle of the Sexes

1\2 F O
F 2, 1 0, 0
O 0, 0 1, 2
Yet, it is clear that rational outcomes are (F, F ) and (O, O), and that none of them is “more
rational” that the other.
The two outcomes might derive from a process of self-fulfilling conjectures made by rational
players under common knowledge
• If my partner chooses F, F is my best option, and if F is my best option, my partner will

chose F (Same goes for O) [Stable]
• If my partner choose F,O is NOT my best option ... [Unstable]

Definition. Given a game Γ = (I, {S i}i∈I , {ui}i∈I ), a strategy profile s∗ ∈ S is a Nash

Equilibrium if for every i ∈ I,
ui(s∗i , s∗−i) ≥ ui(si, s∗−i)
for any other si ∈ Si
Thus, given that all are playing as in s∗, everybody’s best option is to play as in s∗.
Definition. Given a game Γ = (I, {S i}i∈I , {ui}i∈I ), the best response correspondence for
player i, bi : S−i → Si, is given by
bi(s−i) = {si ∈ Si | u1(si, s−i) ≥ ui(s′i, s−i) for all s′i ∈ Si}
Consequently,
Definition. Given a game Γ = (I, {S i}i∈I , {ui}i∈I ), a strategy profile s∗ ∈ S is a Nash
s∗i ∈ bi(s∗−i)
Examples
Battle of the Sexes

1\2 F O
F 2, 1 0, 0
O 0, 0 1, 2
Two Nash Equilibria: (F, F ) and (O, O)
Meeting in Barcelona
1\2 P R
P 2, 2 0, 0
R 0, 0 1, 1
Two Nash Equilibria: (P, P ) and (R, R)
Comments
• A Nash equilibrium might not exist (for the moment ...)
• A Nash equilibrium might not be unique
• A Nash equilibrium needs not to be efficient
• In a Nash equilibrium no player has incentives to deviate
• Nash equilibrium are minimal requirements for strategic stability and self-enforcement
• Strictly dominant strategies are always present in a Nash equilibrium.
• No strictly dominated strategy can be part of a Nash equilibrium.
• Weakly dominated strategies may be part of a Nash equilibrium (!)

Mixed Strategies
How to play ?
1\2 R S P
R 0, 0 1, −1 −1, 1
S −1, 1 0, 0 1, −1
P 1, −1 −1, 1 0, 0
If my opponent plays ...
1
p(R) = p(S) = p(P ) =
3
I get ...
1 1 1
u(R) = 0 + (−1) + 1 = 0
3 3 3
1 1 1
u(S) = (−1) + 0 + 1 = 0
3 3 3
1 1 1
u(P ) = 1 + (−1) + 0 = 0
3 3 3
Thus, any strategy is a best response to p(R) = p(S) = p(P ) = 13 .

1
But ... Is p(R) = p(S) = p(P ) = 3 my opponent’s best response to any of my strategies ?
1
Only playing p(R) = p(S) = p(P ) = 3 is best response to playing p(R) = p(S) = p(P ) =
1
3.
1
That is, only p(R) = p(S) = p(P ) = 3 is a Nash Equilibrium.
It’s an equilibrium in mixed strategies
Notation
• A mixed strategy for player i ∈ I is a probability distribution over S i = {si1, si2, . . .}

X
i
σ = (σii, σ2i , . . .), σji ≥ 0, σji = 1
• The set of all mixed strategies of player i is the Simplex on S i, denoted ∆(S i)
• ∆(S) = ∆(S 1) × ∆(S 2) × · · · × ∆(S I ) is the Set of strategy profiles
• (σ i∗, σ −i ) = (σ 1, σ 2, . . . , σ i−1, σ i∗, , . . . , σ I )
• ui : S → R is the (von Neumann - Morgenstern) Payoff function of player i ∈ I
ui(σ i) = σ1i ui(si1) + σ2i ui(si2) + · · ·

Example: The Battle of the Sexes
1\2 F O
F 2, 1 0, 0
O 0, 0 1, 2
We know that σ 1 = σ 2 = (1, 0) and σ 1 = σ 2 = (0, 1) are two Nash equilibria.
For another strategy profile (σ 1 , σ 2) to be an equilibrium, it must be the case that
u1(F, σ 2) = u1(O, σ 2)
u2(σ 1, F ) = u2(σ 1 , O)
That is,
2σ12 + 0(1 − σ12) = 0σ12 + 1(1 − σ12 )

1σ11 + 0(1 − σ11) = 0σ11 + 2(1 − σ11 )
Solving we get that

1 2 1 2 1 2
σ = ( , ) and σ = ( , )
3 3 3 3
is also a Nash equilibrium.
Interestingly...
1 2
3 3
1\2 F O
2 2 4
3 F 9 9
1 1 2
3 O 9 9
there will be more miss-coordinations that coordinations

Nash Equilibrium
Definition. Given a game Γ = (I, {S i}i∈I , {ui}i∈I ), a strategy profile σ ∗ ∈ ∆(S) is a
Nash Equilibrium if for every i ∈ I,
ui(σ i∗, σ −i∗) ≥ ui(σ i , σ −i∗)
for any other σ i ∈ S i

Thus, given that all are playing as in σ ∗, everybody’s best option is to play as in σ ∗ .
Definition. Given a game Γ = (I, {S i}i∈I , {ui}i∈I ), the best response correspondence for
player i, bi : ∆(S −i ) → ∆(S i), is given by
bi(σ −i ) = {σ i ∈ ∆(S i) | ui(σ i , σ −i ) ≥ ui(σ ′i, σ −i) for all σ ′i ∈ ∆(S i)}
Consequently,
Definition. Given a game Γ = (I, {S i}i∈I , {ui}i∈I ), a strategy profile σ ∗ ∈ S is a Nash
σ i∗ ∈ bi(σ −i∗)
Existence of Nash Equilibrium
Theorem 1. Any finite game has a mixed strategy Nash equilibrium

' $
Theorem 2. (Kakutani’s fix point theorem)

Let f : A ⇉ A be a correspondence, with x ∈ A ⇒ f (x) ⊂ A, satisfying the following
conditions:
(i) A is a compact, convex, and non-empty subset of a finite dimensional Euclidean space
(ii) f (x) is non-empty: ∀x ∈ A, f (x) is well defined
(iii) f (x) is convex: ∀x ∈ A, f (x) is a convex-valued correspondence
(iv) f (x) is an upper-hemicontinous correspondence: If {xn, y n} → {x, y} with y n ∈ f (xn),

then y ∈ f (x) (the graph of f is closed), and the image of any compact set is bounded
Then, ∃x∗ ∈ A such that x∗ ∈ f (x∗)
& %
Existence of Nash Equilibrium
Proof. Let b : ∆(S) ⇉ ∆(S) be a correspondence defined by:
b(σ) = (b1(σ −1 ), b2(σ −2 ), . . . , bI (σ −I )), ∀σ ∈ ∆(S)
The proof proceeds by showing that b(σ) satisfies the conditions of Kakutani’s theorem to
conclude that ∃σ ∗ ∈ ∆(S) such that σ ∗ ∈ b(σ ∗ )
(i) ∆(S) is compact, convex, and non-empty

This is true by definition of ∆(S)
(ii) b(σ) is non-empty

By definition, bi(σ −i ) = {σ i ∈ ∆(S i) | ui(σ i, σ −i ) ≥ ui(σ ′i, σ −i ) for all σ ′i ∈ ∆(S i)},
that is, bi(σ −i ) is the solution to the maximization of a continuous function ui on a non-
empty and compact set. Then, by the Weirstrass theorem, bi(σ −i ) is non-empty
(iii) b(σ) is convex

We need to prove that bi(σ .i) is convex ∀i ∈ I. Let σ ′i, σ ′′i ∈ bi(σ −i ). Then,
ui(σ ′i, σ −i ) ≥ ui(µi, σ −i ) ∀µi ∈ ∆(S i)

ui(σ ′′i, σ −i ) ≥ ui(µi, σ −i ) ∀µi ∈ ∆(S i)
Then, ∀λ ∈ [0, 1],
λui(σ ′i, σ −i ) + (1 − λ)ui(σ ′′i, σ −i ) ≥ ui(µi, σ −i) ∀µi ∈ ∆(S i)
Now, since ui(·, σ −i ) is linear in σ i,
ui(λσ ′i + (1 − λ)σ ′′i, σ −i ) ≥ ui(µi, σ −i ) ∀µi ∈ ∆(S i)
Therefore, λσ ′i + (1 − λ)σ ′′i ∈ bi(σ −i ) and thus b(σ) is convex

(iv) b(σ) is an upper-hemicontinous correspondence

Since b(σ) : ∆(S) ⇉ ∆(S), and ∆(S) is a compact set, it is clear that the image of
any compact set will be bounded. Hence, to show that b(σ) is upper-hemicontinuous we
only need to prove that the graph is closed. Assume to the contrary that there exists a
convergent sequence of pairs of mixed strategy profiles (σ n, σ̂ n) → (σ, σ̂) such that ∀n ∈
{1, 2, . . .} σ̂ n ∈ b(σ n) but σ̂ 6∈ b(σ). Along the sequence, because σ̂ n ∈ b(σ n), for each
player i ∈ I we have that
ni i n−i
σ̂ ∈ b (σ )
At the limit, because we are assuming that σ̂ is not a best response to σ, at least one
player must have a better strategy than σ̂ i against σ −i . That is, because σ̂ 6∈ b(σ), ∃i ∈
I, ∃σ̄ i ∈ ∆(S i) such that
ui(σ̄ i, σ −i ) > ui(σ̂ i, σ −i)
n−i
n
Now, since σ → σ we have that σ → σ −i . Thus, since ui is continuous in σ (it’s
n−i
linear), we have that for n large enough u (σ̄ , σ ) will be arbitrarily close to ui(σ̄ i, σ −i ).
i i
n n i ni n−i
Analogously, since (σ , σ̂ ) → (σ, σ̂), we have that for n large enough u (σ̂ , σ ) will
be arbitrarily close to ui(σ̂ i, σ −i ). Alltogether we have that, for n sufficiently large,
i i n−i i ni n−i
u (σ̄ , σ ) > u (σ̂ , σ )
i −i
that contradicts σ̂ n ∈ bi(σ n ). Therefore, b(σ) must be upper-hemicontinuous.
Thus, we have verified that the mixed strategy profiles set ∆(S) together with the best re-
sponse correspondence b satisfy the hypothesis of the Kakutani’s fixed-point theorem. There-
fore, ∃σ ∗ ∈ ∆(S) such that
σ ∗ ∈ b(σ ∗)
that is, σ ∗ is a Nash equilibrium
2
Correlated Equilibrium
Recall the mixed strategy equilibrium in the Battle of the Sexes game.
1\2 F O
F 2, 1 0, 0
O 0, 0 1, 2
∗1 2 1 ∗2 1 2
σ = ( , ) and σ = ( , )
3 3 3 3
A possible interpretation of this equilibrium is that is corresponds to a steady state of a stochastic

process according to which each player plays a best response to the beliefs about the future
behavior of other players based on some“signal”received from nature (for instance, the empirically
observed frequency of the play of each player in the past) In equilibrium these frequencies must
remain constant over time and are stable in the sense that any action taken with positive
probability by a player is optimal given the steady state beliefs.
An assumption that underlies this steady state interpretation is that no player detects any cor-
relation among the other players’ actions or between the other players’ actions and his own
behavior.
full correlation
One of the main assumptions of non-cooperative game theory is that players take actions in-
dependently, without communicating, without any sort of coordination. On the other extreme
would be the case when players can talk to coordinate their actions. Aumann (1974) suggested
that between “no coordination” (no communication) and “full coordination” (talking) there might
exist a range of correlations between their actions. This is the idea of the correlated equilibrium.
Back to the Battle of the Sexes example, imagine that there are only two possible states of
nature, x and y (for instance, sunny and rainy ) and that the realized state can be fully observed
by the two players. Suppose further that Player 1 announces that he is going to go to the
Football if x occurs and to the Opera if y occurs. Then the best response for Player 2 is to
mimic what Player 1 is doing. The combination of these behaviors is an equilibrium in the sense
that, given that the other player is doing, no one wants to change his or her behavior. The
equilibrium is achieved by means of some “coordination device” (a weather report, for example)
that implements a random variable whose realizations are observable to the two players. This
is the case of “perfect correlation” (or perfect coordination), when all the players have access to
exactly the same information.
less than full correlation
Consider, on the contrary, the following case. There are three possible states of nature: x, y,
and z (sunny, cloudy, and rainy), but players have access to the information in different ways
(different weather reports). For instance, imagine that Player 1 can tell whether the state of
nature is x or an element of [y, z] (sunny or not), and that Player 2 can distinguish only between
[x, y] and z (rainy or not). Then, the strategy for Player 1 has two elements: what to do if he
learns that the state of nature is x and what to do if the state of nature is in [y, z]. Similarly
for Player 2: what to do when the state of nature is in [x, y] and what to do if it is x.
A player’s strategy is optimal, given the strategy of the other(s) player(s), if for any realization
of his information he can do no better by choosing an action different from that dictated by his
strategy
Imagine, for example, that the probabilities of the states x, y, and z are px, py , and
pz respectively, and that Player 2 announces that she is going to the Football if it is not go-
ing to rain (that is, [x, y] occurs) and to the Opera if it is going to rain (x occurs). Then,
if Player 1 is informed that it is not going to be sunny (that is, [y, z] occurs)1, he will have
to choose the action that is optimal against his partner going to the Football with probability
1
What will be Player 1’s optimal choice if he is informed that x occurs ?
p2(F |[y, z]) and to the Opera with probability p2(O|[y, z]), where:
2 py
p (F |[y, z]) = p([x, y]|[y, z]) =
py + pz
pz
p2(O|[y, z]) = p(z|[y, z]) =
py + pz
This is one example in which full coordination is not possible (players observe different signals)
yet there is some correlation between their actions.
Formally, we can give the following definition of a correlated equilibrium
Definition of Correlated Equilibrium

Definition. Correlated Equilibrium
A correlated equilibrium of a a game Γ = (I, {S i}i∈I , {ui}i∈I ) consists of
(i) A finite probability space (Ω, p) (Ω is the set of states of nature and p is a probability
measure on Ω)
(ii) For each player i ∈ I a partition P i of Ω (P i represents Player i’s information struc-
ture)
(iii) For each player i ∈ I a function σ̄ i : Ω → S i such that σ̄ i (ω) = σ̄ i(ω ′) whenever
ω, ω ′ ∈ Pki for some Pki ∈ P i (σ̄ i is the strategy of Player i)
such that for every player i ∈ I and for any other σ i : Ω → S i such that σ i(ω) = σ i (ω ′)
whenever ω, ω ′ ∈ Pki for some Pki ∈ P i (that is, for any strategy of Player i) we have
X X
−i
i i
p(ω)u (σ̄ (ω), σ̄ (ω)) ≥ p(ω)ui(σ i(ω), σ̄ −i (ω)) (1)
ω∈Ω ω∈Ω
✔ Notice that the probability space and the information partitions are part of the equilibrium
Proposition. Any Nash equiulibrium “is” a Correlated equilibrium

For every Nash equilibrium σ ∗ of the game Γ = (I, {S i}i∈I , {ui}i∈I ) there is a correlated
equilibrium (Ω, p), {P i}i∈I , σ̄ in which for each player i ∈ I the distribution on S i induced
by σ̄ i is σ ∗i
Proof. Let Ω = S = S 1 × S 2 × · · · × S I and define p by p(s1, s2, . . . , sI ) = σ ∗1 (s1) ×

σ ∗2 (s2)×· · ·×σ ∗I (sI ). For each player i ∈ I and for all strategy si ∈ S i define P i(si) = {s ∈
S | si = si} and let P i = {P i(si1), P i(si2), . . . , P i(si|S i|)}. Finally, define σ̄ i (s1, s2, . . . , sI ) =
si .
Then, h(Ω, p), {Pi}i∈I , σ̄i is a correlated equilibrium since for each player i ∈ I (1) is satisfied
for every other strategy σ i:
• The left hand side is the payoff of player i in the mixed strategy Nash equilibrium σ ∗
• The right hand side is the payoff of player i when using the mixed strategy that chooses
the pure strategy σ i(s1, s2, . . . , sI ) with probability σ ∗i(si) and everybody else is playing
according to the mixed strategy σ ∗j
Finally, the distribution on S i induced by σ̄ i is σ ∗i 2

Remember, again, the mixed strategy Nash equilibrium in the Battle of the Sexes
1 2
3 3
1\2 F O
2
3 F 2, 1 0, 0
1
3 O 0, 0 1, 2
A correlated equilibrium can be constructed as follows


' $
• Ω = {(F, F ), (F, O), (O, F ), (O, O)}
2
• p(F, F ) = 3 × 13 = 2
9
2
p(F, O) = 3 × 32 = 4
9
1
p(O, F ) = 3 × 13 = 1
9
1
p(O, O) = 3 × 32 = 2
9
• P 1 = {{(F, F ), (F, O)}, {(O, F ), (O, O)}}

P 2 = {{(F, F ), (O, F )}, {(F, O), (O, O)}}
σ̄ 1(F, F ) = F σ̄ 2(F, F ) = F
σ̄ 1(F, O) = F σ̄ 2 (F, O) = O
•
σ̄ 1(O, F ) = O σ̄ 2 (O, F ) = F
σ̄ 1(O, O) = O σ̄ 2(O, O) = O
& %
In this case, the equilibrium condition

X X
−i
i i
p(ω)u (σ̄ (ω), σ̄ (ω)) ≥ p(ω)ui(σ i (ω), σ̄ −i(ω))
ω∈Ω ω∈Ω
for player 1 boils down to
2 4 1 2 2 4 1 2
2+ 0+ 0+ 1≥ 0+ 1+ 2+ 0
9 9 9 9 9 9 9 9

It is also easy to check that there is another correlated equilibrium in the Battle of the Sexes
game that yields a payoff of ( 23 , 32 ).
Indeed, consider
' $
• Ω = {x, y}
1
• p(x) = 2
1
p(y) = 2
• P 1 = {{x}, {y}}
P 2 = {{x}, {y}}
σ̄ 1(x) = F σ̄ 2(x) = F
•
σ̄ 1 (y) = O σ̄ 2 (y) = O
& %
✔ Notice that p(x) and p(y) are not important for the equilibrium strategies, only for the final
payoff
Proposition. NOT every Correlated equilibrium is a Nash Equilibrium

Any convex combination of correlated equilibrium payoffs profiles of a game Γ =
(I, {S i}i∈I , {ui}i∈I ) is a correlated equilibrium payoff profile of Γ
Proof. Let (u1, u2, . . . , uK ) a collection of K profiles of payoffs corresponding to K correlated

equilibria
For each value of k ∈ [1, K], let (Ωk , pk ), {Pik }i∈I , σ̄ k be the correlated equilibrium that
generates the payoff profile uk .
P k k
The following defines a correlated equilibrium whose payoff is k λ u :
• Ω = Ω1 ∪ . . . ∪ ΩK
k k k
λk = 1 and k is such that ω ∈ Ωk
P
• For each ω ∈ Ω, p(ω) = λ p (ω), where λ ≥ 0,
(assume that all Ωk are disjoint)
• For each i ∈ I, P i = ∪k P ik and define σ i by σ i(ω) = σ ik (ω), where k is such that ω ∈ Ωk
2
It is like if first we randomly select the correlated equilibrium to play, and second we implement
it
In the case of the Battle of the Sexes we have
Example: The Game of Chicken (Wikipedia)

Consider the game of chicken
1\2 D C
D 0, 0 7, 2
C 2, 7 6, 6
In this game two individuals are challenging each other to a contest where each can either dare
or chicken out. If one is going to Dare, it is better for the other to chicken out. But if one is
going to chicken out it is better for the other to Dare. This leads to an interesting situation
where each wants to dare, but only if the other might chicken out.
In this game, there are three Nash equilibria. The two pure strategy Nash equilibria are (D, C)
and (C, D). There is also a mixed strategy equilibrium where each player Dares with probability
1
3.
Now consider a third party (or some natural event) that draws one of three cards labeled: (C, C),
(D, C), and (C, D). After drawing the card the third party informs the players of the strategy
assigned to them on the card (but not the strategy assigned to their opponent). Suppose a player
is assigned D, he would not want to deviate supposing the other player played their assigned
strategy since he will get 7 (the highest payoff possible). Suppose a player is assigned C. Then
the other player will play C with probability 21 and D with probability 12 . The expected utility
of Daring is 0 21 + 7 12 = 3.5 and the expected utility of chickening out is 2 21 + 6 12 = 4. So, the
player would prefer to Chicken out.
Since neither player has an incentive to deviate, this is a correlated equilibrium. Interestingly,
the expected payoff for this equilibrium is 7 13 + 2 31 + 6 31 = 5 which is higher than the expected
payoff of the mixed strategy Nash equilibrium.
Trembling-Hand Equilibrium
• Rationality does not rule out weakly dominated strategies
• In fact, NE can include weakly dominated strategies
• Example: (B, R) in
1\2 L R
T 1, 1 0, −3
B 3, 0 0, 0
• But should we expect players to play weakly dominated strategies?
• Players should be completely sure of the choice of the others
• But, what if there is some risk that another player makes a ”mistake”?
Trembling-Hand Perfection
Trembling hand perfect equilibrium is a refinement of Nash Equilibrium due to Reinhard Selten
(1975). A trembling hand perfect equilibrium is an equilibrium that takes the possibility of off-
the-equilibrium play into account by assuming that the players, through a ”slip of the hand” or
tremble, may choose unintended strategies, albeit with negligible probability.
We first define what is a “Perturbed Game”
Definition. Perturbed Game
For any game Γ = (I, {S i}i∈I , {ui}i∈I ), we can create a Perturbed Game Γε =
(I, {∆ε(S i)}i∈I , {ui}i∈I ) by endowing each player i ∈ I with numbers εi i
(s ) ∈ (0, 1) for
any si ∈ S i in such a way that si∈S i εi(si) < 1 and
P
X
i i i i i i i i i
∆ε(S ) = {σ ∈ ∆(S ) | , σ (s ) ≥ ε (s ) for all s ∈ S , and σ i(si) = 1}
si ∈S i
• Interpretation:
Each strategy si is played with some minimal probability: this is the unavoidable probability
of a mistake
Trembling-Hand Perfect Equilibrium
Definition. Trembling-Hand Perfect Equilibrium (Selten 1975)

A Nash equilibrium σ ∗ of the game Γ = (I, {S i}i∈I , {ui}i∈I ) is a Trembling-Hand Perfect
∞
Equilibrium if there is a sequence of perturbed games {Γεk }k=1 that
kconverges to Γ for wich
∞
there is some associated sequence of mixed strategy Nash equilibria σ k=1 that converges to
σ∗
Proposition. Computation of Trembling-Hand Perfect Equilibrium

A Nash equilibrium σ ∗ of the game Γ = (I, {S i}i∈I , {ui}i∈I ) is a Trembling-Hand
k ∞ Perfect
Equilibrium if and only if there is a sequence of totally mixed strategies σ k=1 converging
∗ i∗
to σ
k ∞ such that, for any player i ∈ I, σ is a best response to every element of the sequence
σ−i k=1
• NOTICE:
This rules out (B, R) in the previous example
Corollary.
σ i in a trembling-hand perfect equilibrium cannot be weakly dominated. No weakly dominated
pure strategy can be played with positive probability
Proposition. Existence of Trembling-Hand Perfect Equilibrium

Every finite game Γ = (I, {S i}i∈I , {ui}i∈I ) has a Trembling-Hand Perfect Equilibrium
Corollary.
Every finite game Γ = (I, {S i}i∈I , {ui}i∈I ) has at least a Nash equilibrium in which no player
plays a weakly dominated strategy with possitive probability
Proper Equilibrium
Conside the game

1\2 L R
T 1, 1 0, 0
B 0, 0 0, 0
It has 2 Nash Equilibria: (T, L) and (B, R) but only one Trembling-hand perfect equilibrium
(T, L)
Let us add two weakly dominated strategies
1\2 L R R+
T 1, 1 0, 0 −9, −9
B 0, 0 0, 0 −7, −7
B+ −9, −9 −7, −7 −7, −7
Now it has 3 Nash Equilibria: (T, L), (B, R), and (B +, R+) and 2 Trembling-hand perfect
equilibrium (T, L) and (B, R)!!!
(B, R) is indeed a Trembling-Hand perfect equilibrium. Consider the totally mixed strategy
(ε, 1 − 2ε, ε) for both players. Deviating to T (or L) would yield (ε − 9ε), while B has a payoff
of −7ε. Thus, playing B ((0, 1, 0)) is a best response to (ε, 1 − 2ε, ε) for any ε and, clearly,
(ε, 1 − 2ε, ε) →ǫ→0 (0, 1, 0)
• Idea of proper equilibrium: more likely to tremble to better strategies:

– second-best actions assigned at most e times the probability of third-best actions,
– fourth-best actions assigned at most e times the probability of third-best actions,
– etc.
• (B, R) is not a proper equilibrium: if player 2 puts weight ε on L and ε2 on R then deviating
to T for player 1 will produce a payoff of (ε − 9ε2), while B has a payoff of −7ε. In this
case, (ε − 9ε2) > −7ε for ε small enough
Definition. ε-proper strategies (Myerson 1978)

A totally mixed strategy profile σ of the game Γ = (I, {S i}i∈I , {ui}i∈I ) is said to be an
ε-proper if there is a parameter ε > 0 such that for any player i ∈ I and for any pair of pure
strategies si, s′i ∈ S i for which
ui(si, σ −i) < ui(s′i, σ −i )
we have that
σ i (si) ≤ εσ i(s′i)
Definition. Proper Equilibrium (Myerson 1978)

A Nash equilibrium σ ∗ of the game Γ = (I, {S i}i∈I , {ui}i∈I ) is said to be a Proper Equilibrium
if it is the limit point of a sequence of ε-proper totally mixed strategy profiles
2.3 Dynamic Games of Complete Information
Players take decisions one after another. Thus, some players may know the actions taken by
other players before choosing its own
Basic Assumptions
2. Players have full information about all the elements of the game (Complete
Information)
2.- Game Theory Overview 2.3 Dynamic Games of Complete Information
Extensive Form Game Representation
Dynamic games, because of its sequential structure, are better represented using trees. Formally,
it is called Extensive Form Representation and consists of:
(i) A finite set of nodes N , a finite set of possible actions A, and a finite set of players
I = {1, ..., I}.
(ii) A function p : N → N ∪ ∅ specifying a single immediate predecessor of each node x;

p(x) 6= ∅ for all x ∈ N but one, the initial node x0.
The immediate successor nodes of x are then s(x) = p−1(x), and the set of all predecessors
and all successors of node x can be found by iterating p(x) and s(x). To have a tree
structure, we require that these sets be disjoint.
The set of terminal nodes is T = {x ∈ N |s(x) = ∅}. All other nodes inN \T are decision
nodes.
(iii) A function α : N \{x0}→A giving the action that leads to any noninitial node x from its
immediate predecessor p(x) and satisfying the property that if x′, x′′ ∈ s(x) and x′ 6= x′′
then α(x′) 6= α(x′′).
The set of choices/actions available at decision node x is c(x) = {a ∈ A | a =
α(x′) for some x′ ∈ s(x)}.
(iv) A collection of information sets H, and a function H : N → H assigning each decision

node x to an information set H(x) ∈ H.
Thus, the information sets in H form a partition of N .
We require that all decision nodes assigned to a single information set have the same
choices/actions available; formally, c(x) = c(x′) if H(x) = H(x′ ).
We can therefore write the choices/actions available at information set H as C(H) = {a ∈
A | a ∈ c(x) for all x ∈ H} .
(v) A function ι : H → I assigning each information set in H to a player (or to nature:

formally player 0) who moves at the decision nodes in the set.
We can denote the collection of player i’s information sets by Hi = {H ∈ H|i = ι(H)}.
(vi) A function ρ : H0 × A → [0, 1] assigning probabilities to actions at information sets where

nature moves and satisfying ρ(H, a) = 0 if a ∈
P / C(H), ρ(H, a) ≥ 0 if a ∈ C(H), and
x∈C(H) ρ(H, a) = 1 for all H ∈ H0 .
(vii) A collection of payoff functions u = {u1(·), ..., uI (·)} assigning utilities to the players for
each terminal node that can be reached,ui : T → R (vNM utility functions!).
Thus, formally, ΓE = (N, A, I, p(·), α(·), H, H(·), ι(·), ρ(·), u) completely describes an exten-
sive form game.
As long as the set N is finite, the extensive form game is finite.
Example: The Truth Game
There are two players, called 1 and 2, and a game-master. The game-master has a coin that is
bent in such a way that, flipped randomly, the coin will come up ”heads” 80% of the time. (The
bias of the coin is known to both players.)
The game-master flips this coin, and the outcome of the coin flip is shown to player 1. Player
1 then makes an announcement to player 2 about the result of the coin flip; player 1 is allowed
to say either ”heads” or ”tails” (and nothing else). Player 2, having heard what player 1 says but
not having seen the result of the coin flip, must guess what the result of the coin flip was, either
”heads” or ”tails”. That ends the game.
Payoffs are made as follows. For player 2 things are quite simple; player 2 gets $1 if his guess
matches the actual result of the coin flip, and he gets $0 otherwise. For player 1 things are more
complex. She gets $2 if player 2’s guess is that the coin came up ”heads”, and $0 if player 2
guesses ”tails”, regardless of how the coin came up. In addition to this, player 1 gets $1 (more)
if what she (player 1) says to player 2 matches the result of the coin flip, while she gets $0 more
if her message to player 2 is different from the result of the coin flip.
Figure 1: The Truth Game. Extensive Form Representation

Information
Notice, in the Truth Game, that Player 1 has full information on the move chosen by nature
(the coin flip), whereas Player 2 doesn’t. In other words, when its Player 1 turn to move, he has
FULL information on the history of the game until that moment, while this is not true for Player
2.
This is shown in the tree by means of the dotted lines that joints the decision nodes of Player
2. Such “dotted lines” correspond to the information sets “owned” by Player 2
Definition. Perfect/Imperfect Information

An extensive form game ΓE is one of perfect information if each information set contains a
single decision node. Otherwise it is a game of imperfect information.
Thus, the Truth Game is a game of Imperfect Information as the information sets of Player 2
contain more than one decision node.
The difference between Perfect/Imperfect Information should not be mistaken by the difference
between Complete/Incomplete Information
Definition. Perfect/Imperfect Recall

A game is of Perfect Recall if players never forget information acquired throughout the game,
including the strategies he chose.
Figure 2: A game without Perfect Recall
We will consider games with Perfect Recall only

Strategies
Loosely speaking, a strategy for a player is a contingent plan, that is, a specification of what
action to take at each of his/her information sets. Thus, a strategy in a dynamic game is a
“plan” that says what to do at every point in the game the player has to choose an action.
This is (can be) very complicated (think Chess)
Definition. Pure and Mixed Strategies
A (pure) strategy for player i is a function si : Hi → A such that si(Hi) ∈ C(Hi ). Similarly
as for normal form games a mixed strategy σi is a probability distribution over agent i‘s pure
strategies.
' $
Example. The Truth Game

In The Truth Game players have the following (pure) strategies:
S1 = {HH, HT, T H, T T }
S2 = {HH, HT, T H, T T }
where, for each player, each pair of actions indicates the action to choose at each of their
information sets.
& %
Note that any mixed strategy σi in an extensive form game will induce a probability distribution on
the actions player i can choose in any of his information sets Hi. So, rather than randomizing
over the potentially large set of pure strategies, player i could randomize separately over the
possible actions in C(Hi) at each of his information sets Hi.
Definition. Behavioral Strategies

A behavioral strategy for player i specifies for every information set Hi ∈ Hi a probability
distribution
P bi(·, Hi) over C(Hi), i.e., for each Hi ∈ Hi and a ∈ C(Hi ), bi(a, Hi) ≥ 0, and
a∈C(Hi) bi (a, Hi ) = 1.
Formally, they are equivalent
Theorem. In an extensive form game with perfect recall, mixed and behavioral strategies are
equivalent (with respect to the probability distribution on the terminal nodes they induce).
' $

Notice, for instance, that the mixed strategy ( 21 , 12 , 0, 0) for Player 2 is equivalent to the behav-
ioral strategy
b2(H, H21) = 1, b2(T, H21) = 0

1 1
b2(H, H22) = , b2(T, H22) =
2 2
where H21is the information set owned by Player 2 following the choice of Heads by Player
1 (Left on Figure 1) and H22is the information set owned by Player 2 following the choice of
T ails by Player 1 (Right on Figure 1)
& %
Normal Form Game Representation
As well as Static Games, Dynamic Games can be represented in Normal (Strategic) Form.
The only drawback is that with this representation we loose all the information regarding the
sequentiality and informational structure of the game.
' $
Claim
• Any Dynamic Game can be represented in Normal (Strategic) Form in a unique way (up to
re-labeling of the strategies) .
• Any Static Game can be represented in Extensive Form, but the representation might not be
unique.
& %

Recall that in The Truth Game players have the following (pure) strategies:
S1 = {HH, HT, T H, T T }
S2 = {HH, HT, T H, T T }
where, for each player, each pair of actions indicates the action to choose at each of their
information sets.
The table below represents the same game in Normal Form
HH HT TH TT
14 4 14 4 4 1 4 1
HH 5,5 5,5 5, 5 5, 5
15 4 13 5 7 5 1
HT 5,5 5,5 5, 0 5, 5
10 4 2 8 1
TH 5,5 5, 0 5, 5 0, 51
11 4 1 1 11 4 1 1
TT 5,5 5, 5 5,5 5, 5
Table 1: Normal Form Representation of the Truth Game

• The fact that any dynamic game can be represented in Normal Form guarantees that any
dynamic game has a Nash equilibrium (at least in mixed strategies)
• The problem, though, is that typically dynamic games have “too many” Nash equilibria, some
of which do not make sense.
It’s easy to check that The Truth Game has infinitely many Nash Equilibria:
P layer 1 : HH
P layer 2 : anything
' $
Example. The Market Entry Game

A firm wants to enter a new market, and its main concern is about the reaction of an incumbent
company that currently is making a profit $100,000. If the reaction is aggressive, the challenger
will suffer a loss of $10,000, and the incumbent’s profits will be reduced to only $20,000 (because
of the costs of the fight). On the other hand, if the incumbent chooses to accommodate to
the new market scenario, then the two companies will share the $100,000 profit ($50,000 each)
Obviously, the challenger can always choose to stay out if that seems to be the best choice (with
a profit of $0)
The tree in Figure 3 corresponds to the Extensive Form Representation of this game
& %
Figure 3: The Market Entry Game

In this case each player only has one information set and therefore strategies and actions coincide
in this game. Thus, the Normal Form Representation is given in the table below
Aggressive Accommodate
Enter -10,20 50,50
Don’t 0,100 0,100
Table 2: Normal Form Representation of the Market Entry Game
Clearly, the game has 2 equilibria in pure strategies (as well as another one in mixed strategies)
EQ1 : (Enter, Accommodate)

EQ2 : (Don′t, Aggressive)
• EQ1 is sequentially rational in the sense that if the players actually were to follow the
strategies prescribed by this equilibrium, it would be rational for the two of them to do so in
every step of the game
• EQ2 is NOT sequentially rational as if the players actually were to follow the strategies
prescribed by this equilibrium, then it wouldn’t be rational for Player 2 to play Aggressive
once Player 1 has chosen Enter (Player 2 would get a strictly lower payoff that playing
Accommodate) Being “Aggressive” in step 2 is a threat to Player 1, but it’s an incredible
threat since it would also harm Player 2 and therefore a “rational” Player 2 would never carry
it out.
Definition. Sequential Rationality

A profile of strategies satisfies sequential rationality if at any information set players following
those strategies specify optimal actions at that point in the game.
We will consider different equilibrium refinements to eliminate those Nash equilibria that are not
sequentially rational:
• Backward Induction: When the game is of Perfect Information
• Subgame Perfection: When the game is of Imperfect Information

Equilibrium by Backward Induction
Definition. Equilibrium by Backward Induction

The (pure) strategy profile s is a backward induction strategy profile for the finite extensive form
game ΓE of perfect information if it is derived as follows.
1. Call a node x penultimate in ΓE if all nodes immediately following it are terminal nodes.
2. For every penultimate node x, let sι(x)(x) be an action leading to a terminal node that
maximizes player ι(x)’s payoff on c(x).
3. Let ux denote the resulting payoff vector.
4. Remove the nodes and actions strictly following every penultimate node x in ΓE and assign
the payoff ux to x, which then becomes a terminal node.
5. Repeat the above described process until an action has been assigned to each decision node
Example
Consider the game in the tree below. At the first “penultimate” node, Player 1 will choose R′
Example
The game reduces to the tree below. At the second penultimate node, Player 1 will chose L′′
Example
The game reduces to the tree below. Now, at the new penultimate node, Player 2 will chose r
Example
The game reduces to the tree below. At the second new penultimate node, Player 2 will chose
l′
Example
The game reduces to the tree below.

Example
Finally, Player 1 will chose R
Thus, the Equlibrium by Backward Induction corresponds to:
{RR′L′′, rl′}
Theorem. Kuhn/Zermelo
Every finite game of perfect information has a pure strategy Nash equilibrium that can be derived
through backward induction.
Notice that only one of the two Nash Equilibria of the Market Entry Game (Figure 3) is obtained
by Backward Induction, EQ1 : (Enter, Accommodate)
Subgame Perfect Nash Equilibrium
In games of Imperfect Information, like the Truth Game, Backward Induction is not possible.The
concept of Subgame is needed to extend the principle of Backward Induction to this class of
games
Definition. Subgame
A subgame of an extensive form game ΓE is a subset of the game having the following properties:
(i) It begins with an information set containing a single decision node, contains all the decision
nodes that are successors of this node, and contains only these nodes.
(ii) If decision node x is in the subgame, then every x ∈ H(x) also is; e.g., there are no “broken”
information sets.
Definition. Subgame Perfect Nash Equilibrium

A strategy profile σ for ΓE is a subgame perfect Nash equilibrium (SPNE) if it induces a Nash
equilibrium in every subgame of ΓE .
Example
Consider the game represented in the tree below

Example
There are two subgames in this game:\Gamm
1. The one that starts at the second information set of Player 1, which is a “proper” subgame
2. The game itself
Thus, to find the Subgame Perfect Nash Equilibrium we must
1. Find all the Nash Equilibria of the static game induced by the “proper” subgame
2. Proceed by Backward Induction to find all the Nash Equilibria of the game that results from
the substitution of that “proper” subgame by each of its Nash Equilibria
Example
The Normal Form Representation of the proper subgame is given in the table 3 below
l r
L 1,3 0,0
R 0,0 3,1
Table 3: Normal From Representation of the proper subgame
This (sub)game has two Nash Equilibria in pure strategies: {(L, l), (R, r)} and one in mixed
strategies: (σ 1, σ 2) = (( 41 , 34 ), ( 43 , 14 ))
Thus, moving by Backward Induction we find three possible games (one for each of the three
Nash equilibria of the subgame)
Hence, we have three Subgame Perfect Nash Equilibria of this game
EQ1 : {Out L, l}
EQ1 : {In R, r}
1 3 3 1
EQ3 : {Out ( , ), ( , )}
4 4 4 4

Question:
How many Subgame Perfect Nash Equilibria has The Truth Game ?

Repeated Games
• A special class of Dynamic Games is that of Repeated Games, when players play the same
(static) stage game in every period
• When the actions of the players are observed after each period, players can condition their
play on the past history of the game, which can lead to equilibrium outcomes that do not
arise when the game is played only once
• In the Prisoners’ Dilemma, for example, one might think that cooperation is possible based
on some strategy of punishment and reward (stick and carrot)
• Repeated games can be a good representation of long-term relationships

Notation
• Let Γ = (I, {Ai}i∈I , {ui}i∈I ) be the stage game. For convenience, strategies in the stage-
game will be called actions (A = A1 × · · · × AI ) in the repeated-game
• Let ht ∈ (A)t be a history of the game up to period t. For completeness, h0 = ∅ represents

the history when the game starts at t = 0 (null history). Thus, Ht = (A)t ∩ ∅ represents
the space of all possible period-t histories
i
i ∞
• A pure strategy s for player i is a sequence of maps st t=0 (one for each period t) that
map possible histories at period t (ht ∈ Ht) to actions ai ∈ Ai
sit : Ht → Ai
• Consequently,
i ∞ a mixed (behavioral) strategy in the repeated game is a sequence of maps
βt t=0 that map possible histories at period t (ht ∈ Ht) to mixed actions σ i ∈ ∆(Ai)
βti : Ht → ∆(Ai)
• Notice that a player’s strategy can not depend on the past values of his opponents mixed
actions σ−i, it can depend only on the past values of a−i.
• Notice that each period begins a proper subgame. Moreover, since the stage game is static,
these are the only subgames of the game.
• Games can be repeated a finite (fixed) number of times T or an infinite (undecided or un-
known) number of times
Infinitely Repeated Games
Correspond to the case when players think that the game can always extend one more period
with positive probability
The payoff that the players (expect to) obtain at the end of the game is “some” weighted average
of the per period payoffs
∞
X
U i = Eσ (1 − δ) δ tui(σt(ht))
t=0
where
• δ ∈ (0, 1) is the discount factor that represents the fact that payoffs lose value as time passes
(because they actually do or because players are impatient)
• (1 − δ) is the normalization factor that serves to measure the stage-game and the repeated-
game payoffs in the same units
Continuation Payoffs
Since each period begins a proper subgame, for any strategy profile β and history ht we can
compute the players’ expected payoff from period t on. This is the continuation payoff, which
will be re-normalized so that it is measured in period-t units
∞
X
Uti = Eσ (1 − δ) δ τ −tui(σt(ht))
τ =t
' $
Alternative Payoff specification

When players are “infinitely patient” (δ → 1), we can use the limit of the average criterion
T
1 X
U i = lim inf Eσ ( ) ui(σt(ht))
T →∞ T t=0
& %
Example: The Repeated Prisoners’ Dilemma

Basic Results (i)
The main results from the theory of infinitely repeated games are
Claim
∗
i∗ ∞ i∗ ∞
(i) If σ is a Nash equilibrium of the stage game, then the strategies βt t=0 = σ t=0
for every player i ∈ I consitute a subgame-perfect Nash equilibrium of the repeated game
∗ k
(ii) If σj j=1
are k different Nash equilibria of the stage game, then for any assginment map
i∗ ∞
ρ : {0, 1, 2, . . .} → {1, 2, . . . , k} (from time periods to indices) the strategies βt t=0 =
n o∞
i∗
σρ(t) for every player i ∈ I consitute a subgame-perfect Nash equilibrium of the
t=0
repeated game
To see that this Claim is true, note that with these strategies the future play of player i’s
opponents is independent of how he plays today. Thus, it is optimal for player i to maximize his
−i∗
current period’s payoff, that is, to play a static best response to σρ(t)
Thus, the repeated play of a game does not reduce the set of equilibrium payoffs
Basic Results (ii)
As a matter of fact, the only reason not to play static best responses is the concern about the
future.
Hence, if the discount factor is small enough (the future does not matter much), only strategies
that specify a static equilibrium at every history to which the equilibrium gives positive probability
are subgame perfect Nash equilibria of the repeated game. On the other hand, for a large enough
discount factor almost any payoff can be achieved by means of some subgame perfect Nash
equilibrium (using strategies that punish out-of-equilibrium behaviors for long enouch periods)
These observations are formalized in the “Folk Theorem(s) for Infinitely Repeated Games”
Basic Results (iii)
' $
Definition. Reservation Utility (minmax payoff) and Individualy Rational Payoffs

• Define player i’s reservation utility (or minmax payoff) to be

v i = min max ui(σ i, σ −i )
σ −i σi
That is, the lowest payoff player i’s opponents can hold him to by any choice of σ −i , provided
that player i correctly foreses σ −i and plays a best response to it (Worst Case Scenario).
• A payoff v i is individualy rational for player i if v i ≥ v i
• If m−i,i denotes the minmax profile agains player i, let mi ∈ Ai be a strategy for player
i such that ui(mi, m−i,i) = v i
& %

1\2 C D
C 3, 3 0, 5
D 5, 0 1, 1
In this case we have:
v 1 = v 2 = 1 and mi = m−i = D (i = 1, 2)
'
Basic Results (iv) $
Definition. Feasible payoff

Under some technical assumptions (see Fudenberg-Tirole, pp. 151-152 for details), the set of
feasible payoffs vectors of the repeated game is
V = convex hull v ∈ ℜI | ∃a ∈ A with ui(a) = v i ∀i ∈ I

& %
Any payoff combination inside the figure is feasible. Only payoff vectors above (v 1, v 2) ≥ (1, 1)
are individually rational
Basic Results (v)
Theorem. The Folk Theorem

For every possible payoff vector v with v i ≥ v i for all player i ∈ I, there exists a δ < 1 such
that for all δ ∈ (δ, 1) there is a Nash equilibrium of the repeated game with payoffs v
Proof. Assume that there is a pure action profile a ∈ A such that ui(a) = v i ∀i and consider
the following strategies for each player i
“Play ai in period t = 0 and continue playing ai as long as:
(i) the realized action in the previous period was a
(ii) the realized action in the previous period differed from a in two or more components
Otherwise, if player j was the only one not to follow profile a, then play mi,j (the minmax
strategy against j) for the rest of the game”
Can player i gain by deviating from this strategy ?. If he does:
• He will get at most maxb∈Ai ui(b, a−i) in the period he deviates

• He will get at most v i in all periods after his first deviation since his opponents will minmax
him afterwards
Thus, if player i deviates in period t he obtains at most
(1 − δ t)v i + δ t(1 − δ) max ui(b, a−i ) + δ t+1v i

b∈Ai
which will be less than v i if
(1 − δ t)v i + δ t(1 − δ) max ui(b, a−i ) + δ t+1v i ≤ v i

b∈Ai
−δ tv i + δ t(1 − δ) max ui(b, a−i ) + δ t+1v i ≤ 0

b∈Ai
−v i + (1 − δ) max ui(b, a−i) + δv i ≤ 0

b∈Ai
(max ui(b, a−i) − v i ) ≤ δ(max ui(b, a−i) − v i)

b∈Ai b∈Ai
(maxb∈Ai ui(b, a−i) − v i)

≤ δ
Call
i (maxb∈Ai ui(b, a−i) − v i )
δ =
Notice that since v i > v i we have that δ i < 1.

Let δ = maxi δ i. Then, it is clear that for δ > δ, no player wants to deviate from the strategy 2
2.4 Static Games of Incomplete Information
Players take decisions without knowing the actions taken by other players
Basic Assumptions
2. Some players don’t have full information about all the elements of the
game (Incomplete Information)
2.- Game Theory Overview 2.4 Static Games of Incomplete Information
Notation
• I = {1, 2, . . . , I} is the Set of Players
• S i = {si1, si2, . . .} is the Set of Pure Strategies of player i ∈ I
• S = S 1 × S 2 × · · · × S I is the Set of pure strategy profiles
• T i is the set of possible “types” for player i ∈ I
• T = T 1 × T 2 × · · · × T I is the set of types profiles (usually assume that T is finite)
• ui : S × T → R is the (von Neumann - Morgenstern) Payoff function of player i ∈ I that

depends on both, the strategy profile and the types profile
• p is the common prior distribution (known to all the players) over the set of possible types
profiles T such that ∀t ∈ T, p(t) > 0
• Each player i ∈ I knows his type ti, but there is uncertainty about all other agents types
t−i. Let agent i be of type t̄i. Now, based on the common prior p and t̄i, we can capture
agent i’s belief about t−i by the conditional probability distribution that is induced by p and
t̄i :
p(t̄i, t−i ) p(t̄i, t−i )
p(t−i|t̄i) = = P
p(t̄i) t−i ∈T−i p(t̄i, t−i )
A Bayesian Game is fully characterized
ΓBN = (I, {S i}i∈I , {T i}i∈I , {ui}i∈I , p)
This representation of a game is known as the “normal form bayesian game” or “strategic form
bayesian game”.
Equilibria in Games of Incomplete Information
• What solution concept can we use for incomplete information games?
• We simply can use the solution concepts that we developed so far.
• Trick: Interpret each type as a separate player. This implies a larger set of players Ĩ and an
enlarged game Γ̃BN . Nature, represented by p chooses which types are actually realized.
The strategies for the large game Γ̃BN are obtained from the original game.
Notice
• Any solution for Γ̃BN implies a solution for any game that is actually realized.
• A Nash equilibrium in Γ̃BN induces a Nash equilibrium in the game that actually is to be
played after “nature’s move”.
The Associate Strategic Form Game
• J = {j | j = (i, ti) where i ∈ I and ti ∈ T i} is the extended set of players
• Rj = S i is the set of strategies for player j = (i, ti) ∈ J
• R = R1 × R2 × · · · × RJ is the Set of pure strategy profiles (where J = |J | = |I| × |T |)
• ũj : R → R is the (von Neumann - Morgenstern) Payoff function of player j ∈ J that is

given by
X
j
ũ (r) = p(t−i|ti)u(r|agents in type profile (ti,t−i ); (ti , t−i ))
t−i∈T−i
The game Γ̃BN = (J , {Rj }j∈J , {ũj }j∈J ) is the strategic form game associated with the
game of incomplete information ΓBN
Example
MWG, p. 254, Example 8.E.1.
With probability µ prisoner 2 is of type I (“normal”) and with probability 1 − µ prisoner 2 is of
type II. Payoffs are:
1\(2,I) C D 1\(2,II) C D
C 0,-2 -10,-1 C 0,-2 -10,-7
D -1,-10 -5,-5 D -1,-10 -5,-11
In this case,
• J = {1, (2, I), (2, II)}
• R1 = S 1 = {C, D}, R(2,I) = S 2 = {C, D}, R(2,II) = S 2 = {C, D}
• ũ1(r1, r(2,I), r(2,II)) = µu1(r1, r(2,I); 1) + (1 − µ)u1(r1, r(2,II); 1)
• ũ(2,I)(r1, r(2,I),r(2,II)) = u2(r1, r(2,I)); I)
• ũ(2,II)(r1, r(2,I),r(2,II)) = u2(r1, r(2,II)); II)

Bayesian Nash Equilibrium
Definition. Bayesian Nash Equilibrium

A Bayesian-Nash equilibrium of a game of incomplete information ΓBN is a Nash equilibrium of
the associated strategic form game Γ̃BN .
Note that in a Bayesian Nash equilibrium, each player must be playing a best response to the
conditional distribution of his opponent strategies for each type that he might end up having.
Theorem. Existence of Bayesian Nash equilibria

Every finite strategic game of incomplete information has at least one Bayesian Nash equilibrium.
Proof. Trivial 2
Example: Mixed strategies revisited
Reconsider the Battle of the Sexes Game, including the following incomplete information (ti’s
are independently drawn from the uniform distribution on [0, ε]):
F O
F 2 + t1, 1 0, 0
O 0, 0 1, 2 + t2
• Find a pure strategy Bayesian Nash equilibrium
• Show that if the incomplete information disappears (ε → 0) the players behavior in the pure
BNE approaches their behavior in the mixed-strategy NE in the original complete information
game.
Harsanyi (1973) showed that this result is quite general: a mixed strategy Nash equilibrium in
a game of complete information can (almost always) be interpreted as a pure strategy Bayesian
Nash equilibrium in a closely related game with a small amount of incomplete information.
Notice that in both cases one player is uncertain about the other player’s choice, but there may
be different reasons for the uncertainty: it can arise because of randomization or because of a
little incomplete information.
2.5 Dynamic Games of Incomplete Information
Players take decisions one after another. Thus, some players may know the actions taken by
other players before choosing its own
Basic Assumptions
2. Some players don’t have full information about all the elements of the
game (Incomplete Information)
2.- Game Theory Overview 2.5 Dynamic Games of Incomplete Information
Similarly as before, we can apply the solution concept developed for normal form games to
extensive form games as well. However, from our experience with complete information dynamic
games, we know that some information is lost when considering an extensive form game in its
normal form.
To rule out “unreasonable” equilibria we considered refinements using backwards induction and
subgame perfection. Therefore, when extending the normal form solution concept of Bayesian
Nash equilibria to the extensive form, we at the same time try to extend the idea of subgame
perfection.
In order to capture sequential rationality in dynamic games with incomplete information, we will
use the concept of a Perfect Bayesian equilibrium. To develop this notion, we start with a weaker
version that also applies to dynamic games with complete information.
Weak Perfect Bayesian Equilibrium
Consider the game below
Figure 4:
Notice that there is a unique subgame: the game itself. Thus, any Nash equilibrium of the game
will be a Subgame perfect Nash equilibrium as well.
Using the normal form representation

l m r
L 2,2 2,2 2,2
M 4,0 -1,1 0.4
R 0,4 -1,1 4.0
Table 4: The strategic form representation of the game in Figure4
We find that (L, m) is the unique Nash equilibrium in pure strategies and, thus, the unique
Subgame perfect Nash equilibrium (in pure strategies)
Suppose now that Player 2 “believes” that the game is at the right node of her information set
with probability µ(and at the left node with probability 1 − µ)2
For what values of µ is m a “sequentially rational” strategy for Player ? That is, for what values
of µ is the expected payoff of m larger than the expected payoff of l and r?
2
Also, M and R can be interpreted as the same strategy for 2 different types of Player 1. Them, µ is the common prior
distribution over the set of types
Eµ(u2(l)) = 4(1 − µ)
Eµ(u2(m)) = 1
Eµ(u2(r)) = 4µ
Now,
2 2 3
Eµ(u (m)) > Eµ(u (l)) ⇒ 1 > 4(1 − µ) ⇒ µ >
4
2 2 1
Eµ(u (m)) > Eµ(u (r)) ⇒ 1 > 4µ ⇒ µ <
4
Thus, m is never sequentially rational ! It cannot be “rationalized” by any “beliefs” Player 2 might
have regarding her position in his information set.

We will require that a player’s strategy is optimal for some belief that he or she might have
about his opponents strategies. This criterium leads to Weak perfect Bayesian equilibria.

Definition. System of Beliefs

A system of beliefs µ in an extensive form game ΓE is a function µ : N → P[0, 1] that specifies
a probability for each decision node such that for all information sets H, x∈H µ(x) = 1.
Such a system of beliefs µ embodies the beliefs of all players at each of their information sets
about the history of play up to this point.
By ui(H, µ, σ) we denote player i’s expected utility starting at information set H if his beliefs
regarding the conditional probabilities of being at the various nodes of H are given by µ, he
plays strategy σi, and all other players play according to σ−i .
Definition. Sequential Rationality

A strategy profile σ in ΓE is sequentially rational at information set H for a system of beliefs µ
if for the player ι(H) who moves at H and for any of his strategies σ̃ι(H) we have
uι(H)(H, µ, (σι(H), σ−ι(H))) ≥ uι(H)(H, µ, (σ̃ι(H), σ−ι(H) ))

The definition of a Weak perfect Bayesian equilibrium requires that
• strategies are sequentially rational given beliefs µ
• whenever possible, beliefs are consistent with the given strategies: in equilibrium, players
should have correct beliefs about their opponents’ strategy choices.
Definition. Weak perfect Bayesian equilibrium

A profile of strategies and a system of beliefs (σ, µ) is a Weak perfect Bayesian equilibrium
(weak PBE) for ΓE if it has the following properties:
(i) the strategy profile is sequentially rational given belief system µ and
(ii) the system of beliefs µ is derived through Bayes’ rule whenever possible. That is, for any
information set H that can be reached with positive probability under σ (i.e., p(H|σ) > 0)
we have that for all x ∈ H:
p(x|σ)
µ(x) =
p(H|σ)
Theorem. Existence of Weak perfect bayesian equilibria

Every finite strategic game of incomplete information has at least one Weak perfect Bayesian
equilibrium.
The following result sheds some light on the conditions that a Nash equilibrium must verify to
be a Weak perfect bayesian equilibrium. It also shows that Weak perfect bayesian equilbria are
refinements of Nash equilibria
Proposition. Relation between Weak perfect bayesian equilibrium and Nash equi-
librium
A strategy profile is a Nash equilibrium of ΓE if and only if there exists a system of beliefs µ
such that
(i) the strategy profile is sequentially rational given belief system µ
at all information sets H such that p(H|σ) > 0 and
(ii) the system of beliefs µ is derived through Bayes’ rule whenever possible.
Example
Consider the game in Figure 4 whose strategic form representation is at Table 4

The game has a total of 6 Nash equilbria
1 1 1
σL σM σR σl2 2
σm σr2
1 1 0 0 0 1 0
3 2
2 1 0 0 5 5 0
1 1
3 1 0 0 2 0 2
2 3
4 1 0 0 0 5 5
1 1 1 1 1
5 3 3 3 2 0 2
1 1 1 1
6 0 2 2 2 0 2
Which one/ones is/are Weak perfect bayesian equilibria ?

Let’s start with the strategies of Player 2 that are sequentially rational at her information set. For
convenience, let’s call x(M ) the left node on the information set of Player 2 and x(R) the right
node. Let the beliefs µ of Player 2 be such that p(x(M )|H) = µ (and p(x(R)|H) = 1 − µ).
Remember that m is never sequentially rational and thus we must consider only l and r with
possitive probability. The corresponding expected payoffs (according to µ) are
Eµ(l) = 4(1 − µ)
Eµ(r) = 4µ
clearly,
1
Eµ(l) > Eµ(r) ⇔ µ <
2
1
Eµ(l) < Eµ(r) ⇔ µ <
2
1
Eµ(l) = Eµ(r) ⇔ µ =
2
From here we can consider 3 cases
a) µ < 12 In this case, σ 2 = (1, 0, 0). Then, given σ 2 , the best reply for Player 1 is σ 1 =
(0, 1, 0). But his can not be an equilibrium because in such case we would have that
p(x(M )|σ) 1
= 6= µ
p(H|σ) 1
b) µ > 21 In this case, σ 2 = (0, 0, 1). Then, given σ 2, the best reply for Player 1 is σ 1 =
(0, 0, 1). But his can not be an equilibrium because in such case we would have that
p(x(M )|σ) 0
= 6= µ
p(H|σ) 1
c) µ = 12 In this case Player 2 is indiferent between l and r. Thus, any σ 2 such that σm² = 0
is sequentially rational for Player 2. Now, for the belief µ = 12 to be consistent with the
1 1
equilibrium path it must be the case that σM = σR , and this happens at all Nash equilibria.
Thus, the Weak perfect bayesian equilibria are all the Nash equilibria such that σm² = 0,
that is:
1 1 1
σL σM σR σl2 2
σm σr2
1 1
3 1 0 0 2 0 2
1 1 1 1 1
5 3 3 3 2 0 2
1 1 1 1
6 0 2 2 2 0 2
Notice that all other Nash equilibria are sequentially rational along the equilibrium path, but not
off the equilibrium path.
Strengthening the Weak perfect bayesian equilibrium concept

One problem of weak PBE is that no restrictions are placed on beliefs “off the equilibrium path”
Thus, extra consistency restrictions on beliefs are often added, which result in Perfect Bayesian
equilibria (Fudenberg and Tirole 1991). Other refinements were obtained by Kreps and Wilson
(1982, Sequential equilibria) and Selten (1975, Trembling hand perfection).
Definition. Sequential Equilibrium
A profile of strategies and a system of beliefs (σ, µ) is a Sequential Equilibrium for ΓE if it has
the following properties:
(i) the strategy profile is sequentially rational given belief system µ and
(ii) there exists a sequence of completely mixed strategies {σ k }∞

k=1 → σ such that µ =
limk→∞ µk , where
k p(x|σ k )
µ (x) =
p(H|σ k )
Proposition.
Every sequential equilibrium is both Weak perfect bayesian equilibrium and Subgame perfect
Nash equilibrium
Example: Signaling Games
These sender-receiver game features two players. Player 1, the sender, has private information
regarding his own type and sends a signal to Player 2, the receiver, who observes Player 1’s
action (but not his type) before choosing her own
' $
Examples
• The sender is a worker who knows his productivity and must choose a level of education. The
receiver is a firm, which observes the workers education, but not the productivity level, and
then decides on the wage to offer.
• The sender is a driver who knows his probability of having an accident and must choose an
insurance policy. The receiver is an insurance company, which observes the policy proposed
by the sender, but no the probability of an accident, and then decides whether to accept it
or not
& %
A signaling game consists of
• Player 1, who is the sender, has private information on his type θ ∈ Θ (the use of θ instead
of t1 is common in signaling games) and chooses an action a1 ∈ A1
• Player 2, whose type is common knowledge, is the receiver who observes a1 and chooses
a2 ∈ A2
• Mixed strategies and their spaces are denoted by αi ∈ Ai . The payoff associated with mixed
strategy profile (α1, α2) and type θ is ui(α1, α2, θ)
• Before the game begins, it is common knowledge that Player 2 holds prior beliefs p over
Player 1’s type θ
• Player 1’s strategies:σ1 : Θ → A1; i.e., a strategy prescribes a probability distribution σ1(·|θ)
over actions in A1 for each type θ ∈ Θ.
• Player 2’s strategies: :σ2 : A1 → A2 ; i.e., a strategy prescribes a probability distribution

σ2(·|a1) over actions in A2 for each action a1.
• Payoffs given σ = (σ1, σ2) and type θ are

Player 1
XX
1
U (σ|θ) = σ1(a1|θ)σ2(a2|a1)u1(a1, a2, θ)
a1 a2
Player 2
 
X XX
2
U (σ|θ) = p(θ)  σ1(a1|θ)σ2(a2|a1)u2(a1, a2, θ)
θ a1 a2
Perfect Bayesian Equilibrium
Definition. Perfect Bayesian Equilibrium of a signaling game

A profile of strategies σ ∗and a system of (posterior) beliefs µ(·|a1) is a Perfect Bayesian Equi-
librium of a signaling game if it has the following properties:
(i) for all θ,
σ1∗(·|θ) ∈ arg max u1(α1, σ2∗, θ)
α1
(ii) for all a1, X

σ2∗(·|a1) ∈ arg max µ(θ|a1)u2(a1, α2, θ)
α2
θ
′
)σ1∗ (a1|θ ′) > 0, then
P
(iii) If θ ′∈Θ p(θ
p(θ)σ1∗(a1|θ)
µ(θ|a1) = P ′ ∗ ′
θ ′ ∈Θ p(θ )σ1 (a1 |θ )
In practical terms
Definition. Perfect Bayesian Equilibrium of a signaling game

A profile of pure strategies σ ∗and a system of (posterior) beliefs µ(·|a1) is a Perfect Bayesian
Equilibrium of a signaling game if it has the following properties:
(i) for all θ,
σ1(θ) ∈ arg max u1(a1, σ2(a1), θ)
a1
(ii) for all a1, X

σ2(a1) ∈ arg max µ(θ|a1)u2(a1, a2, θ)
a2
θ
′
P
(iii) If θ ′∈{Θ|σ1 (θ ′)=a1} p(θ ) > 0, then
p(θ)
µ(θ|a1) = P
θ ′ ∈{Θ|σ1(θ ′)=a1 } p(θ ′)

2 Game Theory

Uploaded by

Copyright:

Available Formats

2 Game Theory

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2 Game Theory

Uploaded by

Copyright:

Available Formats

2.

- Game Theory Overview

Game Theory studies situations in which

2.2 Static Games of Complete Information: Nash Equilibrium

2.4 Static Games of Incomplete Information:Bayesian Equilibrium

Any (non cooperative) game has the following elements

1. The Players Who is to take decisions

3 The Strategies What actions are available to the players

5 The Payoffs What each player obtains in each possible outcome

Example: Rock - Scissors - Paper

1. The Players Player 1 and Player 2

Complete Information Incomplete Information

Complete Information Incomplete Information

2.2 Static Games of Complete Information

1. Players are rational and self-interested (Rationality)

Common Knowledge examples

1. The Byzantine Generals and the Coordinated Attack Problem

2. The Three Hats Problem

3. The Princes Bride

4. La vida es sueño (Pedro Calderón de la Barca - 1636)

• I = {1, 2, . . . , I} is the Set of Players

• S i = {si1, si2, . . .} is the Set of Pure Strategies of player i ∈ I

• S = S 1 × S 2 × · · · × S I is the Set of pure strategy profiles

• (s∗i , s−i) = (s1, s2, . . . , si−1, s∗i , si+1, . . . , sI )

• ui : S → R is the (von Neumann - Morgenstern) Payoff function of player i ∈ I

A Game is fully characterized

Γ = (I, {S i}i∈I , {ui}i∈I )

Example 1: The Prisoners’ Dilemma

• 2 (identical) strategies for each player S i = {C, D} (Cooperate, Defect) i ∈ I

• Symmetric payoffs given by

This is usually represented by a bi-matrix table

Example 1: The Prisoners’ Dilemma (continued)

Example 2: The Battle of the Sexes

I ={1, 2}, S i = {F, O} (Football, Opera)

I ={1, 2}, S i = {R, S, P } (Rock, Scisors, Paper)

In this case, what is possible outcome of the game ? (Zero-Sum Game)

• What behavior should we expect form

• If all players behave in such way, what

• People (players) don’t do silly things (elimination)

• People’s behavior will be stable if nobody has incentives to change

Dominant and Dominated Strategies

Definition. A strategy si ∈ Si is a Strictly Dominated Strategy for player i ∈ I if there exists

ui(s′i, s−i) > ui(si, s−i)

Note A strategy si ∈ Si is a Strictly Dominant Strategy for player i ∈ I if it strictly dominates

Definition. A strategy si ∈ Si is a Weakly Dominated Strategy for player i ∈ I if there exists

ui(s′i, s−i) ≥ ui(si, s−i)

• A strictly dominant strategy, if it exists, must be unique

• Weakly dominant strategies, if any, might not be unique

• Dominated strategies (both strictly and weakly) might not be unique

Iterated elimination of (strictly) dominated strategies

B strictly dominates M . Hence

• Common knowledge is needed for this to make sense

• In many cases it is not of any help (Battle of the Sexes)

• The order of elimination is not important for the final result

Iterated elimination of (weakly) dominated strategies

Strategies T and M are weakly dominated

Now L is weakly dominated,

and then M can be eliminated to finally obtain

Now R is weakly dominated