7 Bayesian Games

So far, we have only analysed games where each player knows the correct and full information
about their opponents. But there are situations where players only have incomplete information,
and these roughly go under the umbrella of Bayesian games.

Example 7.1 (Alice and Bob face a showdown). Sheri↵ Bob faces an armed suspect called
Alice and they have to decide simultaneously whether to shoot or not. The suspect is either a
criminal with probability p or innocent with probability 1 p, and payo↵s are as follows:

Game against criminal Game against innocent person

shoot dont shoot shoot dont shoot
shoot 0, 0 2, 2 shoot 3, 1 1, 2
dont shoot 2, 1 1, 1 dont shoot 2, 1 0, 0

Payo↵s reflect the following assumptions.

• The Sheri↵ would rather shoot if the suspect shoots, but not otherwise.
• A criminal would rather shoot even if the sheri↵ does not.
• An innocent suspect would rather not shoot even if the sheri↵ shoots.
A natural question then is to determine those p for which Sheri↵ Bob should shoot.
What happens if Bob decides to shoot? There are two outcomes to consider: criminal Alice
will shoot, while innocent Alice will not shoot back. The first case gives Bob a payo↵ of 0 with
probability p and the second case yields 1 with probability 1 p, giving Bob a total expected
payo↵ of 1 + p.
If Bob decides not to shoot, then criminal Alice will shoot and innocent Alice will not shoot
as before; Bob’s expected utility in this case is 2p. So he shoots when (1 p) 2p, i.e.,
when p 1/3. (Note that the response from Alice remains the same because shooting strictly
dominates not shooting for criminal Alice, while not shooting dominates shooting for innocent
Alice.)

Remark 7.2. The preceding method applies whenever a player has to decide which action to
take when the type of opponent is not definite: simply take the expected payo↵s for each action
assuming best response from each possible type of opponent, and then choose one when the
expected payo↵ is optimal. This will often involve determining range of probabilities when a
particular action gives better payo↵.

Definition 7.3. A two-player Bayesian game, with players A and B, consists of the following
data.
• A set of actions for each player: S for player A and T for player B.
• A finite set of types for each player: ⇥A for A and ⇥B for B.

45
 • A set of possible states of the world ⌦. Each state is a game in strategic form played by
specific types of A and B with action sets S and T (i.e., the actions are fixed regardless
of type). Formally, we are given:
– functions ⌧A : ⌦ ! ⇥A and ⌧B : ⌦ ! ⇥B ;
– a payo↵ function for each player whose domain is S ⇥ T ⇥ ⌦.
• Probabilities P(!|✓) for all ✓ 2 ⇥A [ ⇥B and ! 2 ⌦. In other words, we are given the
probability with which a given type of player plays a strategic form game.

Example 7.1, continued. Let us consider how the above definition relates to our example. The
players are now Alice (player A) and Bob (player B), and both have action sets

S = T = {shoot, don’t shoot}.

The possible types are given by ⇥A = {innocent, criminal} and ⇥B = {Sheri↵}.

There are two possible states of the world: one in which Sheri↵ Bob is playing against an
innocent Alice and another one in which Bob plays against a criminal Alice. We write !1 for
the game/state in which Alice is innocent and !2 represents the world state in which Alice is
a criminal. These data are captured by introducing ⌦ = {!1 , !2 } and functions ⌧A : ⌦ ! ⇥A
and ⌧B : ⌦ ! ⇥B given by

⌧B (!1 ) = ⌧B (!2 ) = Sheri↵, ⌧A (!1 ) = innocent, ⌧A (!2 ) = criminal.

Payo↵ functions uA and uB are then read o↵ from the table: for example,

uA (shoot, shoot, !1 ) = 3, uB (shoot, shoot, !1 ) = 1, etc.

Finally, the associated probabilities are as follows:

P(!1 |Sheri↵) = 1 p, P(!1 |innocent) = 1, P(!1 |criminal) = 0,

P(!2 |Sheri↵) = p, P(!2 |innocent) = 0, P(!2 |criminal) = 1.

Bayes-Nash equilibrium. To discuss equilibrium in Bayesian games, we need to define strate-

gies and construct uitlities. As in the setting of sequential and repeated games, strategies have
to specify actions for every possible scenario. A little more formally, we have the following
definition.

Definition 7.4. A strategy for a player in a Bayesian game specifies an action for each possible
type the player can be. Or more formally, a strategy is a function from the set of its types to
the set of its actions.

Notation. If the players are Alice and Bob with action sets S and T , types ⇥A and ⇥B , then
S ⇥A and T ⇥B represent the set of strategies for Alice and Bob, respectively. (In general, the

46
set of functions from a set X to a set Y is denoted by Y X .) A strategy profile is a choice of a
strategy for each player, i.e., an element in S ⇥A ⇥ T ⇥B .
Given a strategy profile (s( ), t( )) 2 S ⇥A ⇥ T ⇥B , Alice’s expected payo↵ when she has type
✓A 2 ⇥A is
X
P(!|✓A )uA (s(✓A ), t(⌧B (!)), !)
!2⌦

and Bob’s expected payo↵ when he has type ✓B 2 ⇥B is

X
P(!|✓B )uB (s(⌧A (!)), t(✓B ), !).
!2⌦

Note that payo↵s are attached to each possible type a player can take.

Definition 7.5. The strategy profile (s( ), t( )) 2 S ⇥A ⇥ T ⇥B is a Bayes-Nash equilibrium if

s = s(✓A ) maximises
X
P(!|✓A )uA (s, t(⌧B (!)), !)
!2⌦

for all ✓A 2 ⇥A and t = t(✓B ) maximises

X
P(!|✓B )uB (s(⌧A (!)), t, !)
!2⌦

for all ✓B 2 ⇥B .

This definition says that a strategy profile is a Bayes-Nash equilibrium if each type separately
would not change its action as specified in the corresponding strategy. So for practical uses we
can look for Bayes-Nash equilibria by replacing a given Bayesian game with a new game whose
set of players are all the types in the original game.

Example 7.6. Alice has a dispute with Bob, who is either strong (S) or weak (W ). Alice
believes that Bob is strong with probability p. Each person can either fight or yield; we denote
these by F for fight and Y for yield. Strong Bob plays game G1 and weak Bob plays game G2
given below:
G1 F Y G2 F Y
F 1, 1 1, 0 F 1, 1 1, 0
Y 0, 1 0, 0 Y 0, 1 0, 0

Here the sets of actions for both players is {F, Y }. Alice has just one possible state ⇥A = {A},
but ⇥B = {S, W } (Bob can either be strong or weak). The states are given ⌦ = {G1 , G2 }
where G1 represents game in which Bob is strong and G2 represents game in which Bob is
weak. A little more formally, the functions ⌧A and ⌧B are given by
(⌧A ,⌧B ) (⌧A ,⌧B )
G1 ! (A, S); G2 ! (A, W ).

As for strategies, Alice has to decide between F or Y , but strategies for Bob need to specify
what action is taken by each type of Bob. Thus Alice has two strategies: (A ! F ) and (A !

47
Y ); and, Bob has four strategies: (S ! F, W ! F ), (S ! F, W ! Y ), (S ! Y, W ! F ),
and (S ! Y, W ! Y ). We may abbreviate these to F and Y for Alice, and [F, F ] (or even
F F ) for (S ! F, W ! F ), and so on for Bob.
We compute expected payo↵s of all strategy profiles:

[F, F ] [F, Y ] [Y, F ] [Y, Y ]

F 1 2p, (1, 1) 1 2p, (1, 0) 1, (0, 1) 1, (0, 0)
Y 0, (1, 1) 0, (1, 0) 0, (0, 1) 0, (0, 0)

In each payo↵ box, the first entry is Alice’s payo↵ and the two bracketed terms indicate Bob’s
payo↵ dependent on whether Bob is strong or weak.
If p < 1/2, then 1 2p > 0 and the first row dominates the second, and we obtain a Bayes-
Nash equilibrium at (F, [F, Y ]). If p > 1/2, then 1 2p < 0 and we obtain a Bayes-Nash
equilibrium at (Y, [F, F ]).

Example 7.7 (Where more information may hurt!). Alice and Bob play one of the games below:

G1 L M R G2 L M R
U 1, 2x 1, 0 1, 3x U 1, 2x 1, 3x 1, 0
D 2, 2 0, 0 0, 3 D 2, 2 0, 3 0, 0

where 0  x < 1/2. The players do not know which game is played and assign probability 1/2
to either game. Here each player has one type, ⌦ = {G1 , G2 }.
Strategies can be identified with actions and the expected payo↵s of this Bayesian game are
given by
L M R
U 1, 2x 1, 3x/2 1, 3x/2
D 2, 2 0, 3/2 0, 3/2
Column L dominates the others and we obtain a Bayes-Nash equilibrium (D, L) which results
in a payo↵ of 2 for Bob.
Suppose now that Bob knows which game he is playing. The states remain the same (there
are only two kinds of games in consideration), but there now two types of Bob we need to
account for. We can call these B1 and B2 , and set ⌧B (G1 ) = B1 and ⌧B (G2 ) = B2 . In writing
down strategies for Bob, we now need to specify an action for B1 (Bob’s action in game G1 )
and an action for B2 (which action Bob should play in game G2 ). There are 9 such strategies:
(L, L), (L, M ), . . . .
Strategy R is a dominant strategy for B1 , and M is a dominant strategy for B2 . We can
therefore eliminate strategy L from both games. Now Alice’s strategy U dominates, and we end
up in a Bayes-Nash equilibrium (U, (R, M )). If Bob is of type B1 , then his payo↵ is obtained
from the profile (U, R) in game G1 ; for type B2 , Bob’s payo↵ is obtained from the profile
(U, M ) in game G2 . In either case, Bob obtains a payo↵ of 3x. Now 3x < 2 (as x < 12 ), and
so Bob’s additional knowledge has put him in worse o↵ position!

48