EE6417: Incentive-Centered Design

Lecture 07

Bharadwaj Satchidanandan

Department of Electrical Engineering

Indian Institute of Technology Madras

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 1 / 12

A Slice of History (The 70s)

Most works until the 70s dealt with refinements of NE.

Aumann’s 1974 paper was one of the first attempts to
expand the definition of equilibrium. He introduced
Correlated Equilibrium, which generalizes Nash
equilibrium.
▶ Every NE is a CE, but not vice-versa.
Unlike NE, CE is computationally tractable, and emerges
as the limiting behavior of simple learning rules (more on
this in Part II of the course).
CE is also a rather natural equilibrium notion. In Nobel
laureate Roger Myerson’s words, “If there is intelligent life
on other planets, in a majority of them they would have
Robert J. Aumann discovered correlated equilibrium before Nash equilibrium.”
Nobel Laureate, 2005
CE also reconciles the “Bayesian worldview” with the
“Game theorist’s worldview” of how people behave.

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 2 / 12

Correlated Equilibrium
The Philosophy
The backdrop against which Nash equilibrium was developed assumes that
players sample from their mixed strategies independently.

Hence, Nash equilibrium is a plausible behavioral model in settings where

there are no means for players to correlate their strategies.

However, in many real-world scenarios, players have observations of some

common randomness, which is sometimes generically referred to as
weather.
Hence, each player in general could base her decision on the weather (in
addition to her anticipation of the others’ strategies), which in turn would
correlate the players’ strategies.

The equilibrium that arises in the presence of correlating signals is called

correlated equilibrium.

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 4 / 12

The Setup

Consider an n−player game with Si and ui denoting Player i’s strategy set
and utility function respectively.

There is a coordinator that provides each Player i a strategy

recommendation Sbi ∈ Si .

The strategy recommendation S b = [Sb1 , . . . , Sbn ] is chosen at random by the

coordinator according to the joint distribution p : S1 × . . . , Sn → [0, 1].
Hence, p(b
s1 , . . . , sbn ) is the probability that the coordinator
recommends the strategy sb1 to Player 1, the strategy sb2 to Player 2,
and so on.

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 5 / 12

The Definition

Definition (Correlated Equilibrium)

A joint distribution of strategies p is a correlated equilibrium (CE) if for all
i ∈ {1, . . . , n} and all sbi , si′ ∈ Si ,

si , Sb−i )|Sbi = sbi ≥ E ui (si′ , Sb−i )|Sbi = sbi ,

   
E ui (b (1)

where (Sb1 , . . . , Sbn ) ∼ p.

In words, if everybody else is known to follow the coordinator’s recommendation,

then it is best for the player in question to also follow the coordinator’s
recommendation regardless of what the recommendation is.

Note that since a single entity, namely the coordinator, chooses the strategy
recommendation for all players, correlated recommendations can be attained. I.e.,
one can attain joint distributions of players’ strategies that are not equal to the
product of their marginals. Hence the term “correlated equilibrium.”

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 6 / 12

An Example

The simplest example of a correlated equilibrium is what happens in a traffic intersection.

Consider a traffic intersection with two directions: Up and Across.
Traffic can flow only in one of the directions at any time or there will occur a collision.
Vehicles that arrive at the intersection face a decision problem: Should I go through the
intersection or should I wait for the other vehicle to pass?
This is a strategic decision-making problem and can be modeled as a game with the
following payoff matrix.

Up Vehicle
Go Wait
Go (−1, −1) (1, 0)
Across Vehicle
Wait (0, 1) (0, 0)

Let us give this game a name: The Traffic Intersection game.

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 7 / 12

An Example
The Traffic Intersection game has two pure strategy Nash equilibria:
(Go, Wait) and (Wait, Go).
   
0.5 0.5
It also has a unique mixed strategy Nash equilibrium: , .
0.5 0.5

Corresponding to each Nash equilibrium is a joint distribution of the

players’ strategies:

Go Wait Go Wait Go Wait

Go 0 1 Go 0 0 Go 0.25 0.25
Wait 0 0 Wait 1 0 Wait 0.25 0.25
Nash equilibrium 1 Nash equilibrium 2 Nash equilibrium 3

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 8 / 12

An Example

Suppose now that a traffic signal gets installed at the intersection. Let us assume
that the signal flips a coin to select which direction to signal green.
The players can observe the signal. However, they can choose whether to obey it
or not.
Let us analyze the game from one of the players’, say Player 1’s, perspective.
If the signal recommends Player 1 to wait, then he knows that it is recommending
the Player 2 to go and vice-versa.
Therefore, if Player 2 obeys the signal, then it is best for Player 1 to also obey it.
And if Player 1 obeys the signal, then it is best for Player 2 to also obey it.
Obeying the signal results in a correlated equilibrium.

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 9 / 12

An Example

The joint distribution of the players’ strategies if they obey the signal is

Go Wait
Go 0 0.5
Wait 0.5 0
Correlated equilibrium

Note that the above joint distribution is not attained by any Nash
equilibrium.

Hence, the set of behaviors that can be implemented as an equilibrium

with a coordinating device is strictly larger than the set of behaviors that
can be implemented as an equilibrium without a coordinating device.

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 10 / 12

A Second Example
Consider a 3-player game with the payoff matrix

Left Right Left Right Left Right

Up 0, 0, 3 0, 0, 0 Up 2, 2, 2 0, 0, 0 Up 0, 0, 0 0, 0, 0
Down 1, 0, 0 0, 0, 0 Down 0, 0, 0 2, 2, 2 Down 0, 1, 0 0, 0, 3

PNE are (D, L, A), (U, R, A), (D, L, C ), (U, R, C ) with corresponding payoffs
(1, 0, 0), (0, 0, 0), (0, 1, 0), (0, 0, 0).

Suppose Players 1 and 2, with the help of a correlating device, play (L, U) with
probability 0.5 and (R, D) with probability 0.5. Also, Player 3 is not told the
recommendation of the correlating device. Then, Player 3’s optimal strategy is to play
Matrix 2. Moreover, this is a correlated equilibrium.

The CE payoffs are (2, 2, 2) which is better than the Nash equilibrium payoffs for
everybody.

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 11 / 12

Information Asymmetry improves Welfare
An interesting aspect of the above example is that Player 3 should not be told the
outcome of the correlating event. Only then will (2, 2, 2) be an equilibrium payoff.

If Player 3 is told the outcome, then he has an incentive to deviate to either

Matrix 1 or Matrix 3 as the case may be. And if Players 1 and 2 know that Player
3 is told the outcome of the correlating event, then they cannot trust him to play
Matrix 2. That in turn destroys the equilibrium property of payoff (2, 2, 2), and
the society degenerates to a suboptimal equilibrium.

Hence, Player 3 will not even want to know the outcome of the correlating event,
for his own good.

Open question: Let’s wear the hat of a policy-maker. What are the implications
of this observation for media regulation? Should everybody be told everything in
a society? Is free flow of information welfare-optimal? Can issues surrounding
media regulation be formulated in a Game-theoretic framework and optimal
policies be devised?

Bharadwaj Satchidanandan EE6417: Incentive-Centered Design Lecture 07 12 / 12