Economic Theory and Analysis

Bryony Reich

Topic 6:
Bayesian Games
Examples
To avoid distracting others, please turn off or silence your
phone and put it in your bag. Laptops/tablets should be in
airplane mode and not distract others in the class.
Example: To Fight or Not to Fight
Player 2
Don’t fight Fight

Don’t fight 0,-2 -10,-1

Player 1
Fight -1,-10 -5,-5

• With probability 𝜇, players 1 and 2 play the simultaneous move game above. With probability 1 − 𝜇 the
game below. This is common knowledge. Which game they are playing is private information to player 2.
• Describe in words the preferences of Player 1 and the two types of Player 2.
Player 2
Don’t Fight Fight

Don’t Fight 0,-2 -10,-7

Player 1
Fight -1,-10 -5,-11
Example: To Fight or Not to Fight
• Do we have enough information to define this problem as a Bayesian game?

Type T (prob 𝜇) Type B (prob 1 − 𝜇)

Defining a Bayesian Game

Definition.
A Bayesian Game 𝐺 = (𝑁, {𝐴𝑖 }𝑛𝑖=1 , {𝛩𝑖 }𝑛𝑖=1 , {𝑢𝑖 }𝑛𝑖=1 , 𝑝) consists of

1. A set of players 𝑁 = {1,2, … , 𝑛};

2. A set 𝐴𝑖 of actions for each player 𝑖 ∈ 𝑁, where 𝑎𝑖 ∈ 𝐴𝑖 denotes an action;
3. A set 𝛩𝑖 of types for each player 𝑖 ∈ 𝑁, where 𝜃𝑖 ∈ 𝛩𝑖 denotes a type;
4. A payoff function for each player 𝑢𝑖 (𝑎1 , . . , 𝑎𝑛 , 𝜃𝑖 );
5. A joint probability distribution 𝑝 𝜃1 , … , 𝜃𝑛 over types.

• The probability distribution over types, 𝑝 𝜃1 , … , 𝜃𝑛 , is common knowledge among the players – this is taken as
standard and typically not specified as part of the definition.
Example: To Fight or Not to Fight
• Do we have enough information to define this problem as a Bayesian game?

1. Players: player 1 and player 2.

2. A set of actions for each player: 𝐴𝑖 = {𝐹𝑖𝑔ℎ𝑡, 𝐷𝑜𝑛′ 𝑡 𝐹𝑖𝑔ℎ𝑡} for each 𝑖 ∈ 1,2.

3. Player 2’s set of types is 𝛩2 = {𝑇, 𝐵}. Player 1 has one type 𝛩1 = {1}.

Type T (prob 𝜇) Type B (prob 1 − 𝜇)

Example: To Fight or Not to Fight
• Do we have enough information to define this problem as a Bayesian game?

4. Payoff function for player 1:

𝑢1 𝐷, 𝐷 = 0, 𝑢1 𝐷, 𝐹 = −10, 𝑢1 𝐹, 𝐷 = −1, 𝑢1 𝐹, 𝐹 = −5

Type-contingent payoff function for player 2:

𝑢2 (𝐷, 𝐷, ) 𝑇 = −2, 𝑢2 𝐷, 𝐹 , 𝑇 = −1, 𝑢2 𝐹, 𝐷 , 𝑇 = −10, 𝑢2 𝐹, 𝐹 , 𝑇 = −5
𝑢2 (𝐷, 𝐷, ) 𝐵 = −2, 𝑢2 𝐷, 𝐹 , 𝐵 = −7, 𝑢2 𝐹, 𝐷 , 𝐵 = −10, 𝑢2 𝐹, 𝐹 , 𝐵 = −11

5. Joint probability distribution over types 𝑝 1, 𝑇 = 𝜇, 𝑝 1, 𝐵 = 1 − 𝜇 which is common knowledge.

Type T (prob 𝜇) Type B (prob 1 − 𝜇)

Example: To Fight or Not to Fight
𝟏
• Find all the Bayes-Nash equilibria if 𝝁 > .
𝟔

Type T (prob 𝜇) Type B (prob 1 − 𝜇)

Example: To Fight or Not to Fight
𝟏
• Find all the Bayes-Nash equilibria if 𝝁 > .
𝟔

• In any BNE:
• Type T player 2 must play Fight (strictly dominant strategy)
• Type B player 2 must play Don’t Fight (strictly dominant strategy)
• So player 2’s strategy is (F,DF).

• Expected payoff to player 1 of DF if player 2 plays (F, DF): 𝜇 ⋅ −10 + 1 − 𝜇 ⋅ 0 = −10𝜇

• Expected payoff to player 1 of F if player 2 plays (F,DF): 𝜇 ⋅ −5 + 1 − 𝜇 ⋅ −1 = −1 − 4𝜇

• The unique BNE is (F,(F,DF)) if the probability of type T is high enough. What is the intuition for this result?

Type T (prob 𝜇) Type B (prob 1 − 𝜇)

Example: Armies and Conflict
• There are two armies, 𝑖 and 𝑗.
• Each simultaneously chooses to attack or not attack, A or NA.
• Each army is either strong or weak with equal probability.
• The realisation of one army’s strength is independent from the other’s strength.
• Whether an army is strong or weak is private information to that army.

Payoffs:
• If army 𝑖 plays NA, then 𝑖’s payoff is 0.

• If army 𝑖 plays A
• and army 𝑗 plays NA then 𝑖’s payoff is M

• and army 𝑗 plays A then 𝑖’s payoff is

• 𝑀 − 𝑠 if 𝑖 is strong and 𝑗 is weak
• −𝑠 if 𝑖 is strong and 𝑗 is strong
• −𝑤 if 𝑖 is weak
𝑀
• Where 0 < < 𝑠 < 𝑀 < 𝑤.
2
Example: Armies and Conflict
• Describe this as a Bayesian game.

1. Players: army 𝑖 and army 𝑗.

2. Action set for each player: Ai = Aj = {𝐴, 𝑁𝐴}

3. A set of types for each player: Θ𝑖 = Θ𝑗 = {𝑠𝑡𝑟𝑜𝑛𝑔, 𝑤𝑒𝑎𝑘}.

4. A type-contingent payoff function for each player:

𝑢𝑖 𝑁𝐴,⋅ , ⋅,⋅ = 0, 𝑢𝑖 𝐴, 𝑁𝐴 , ⋅,⋅ = 𝑀, 𝑢𝑖 𝐴, 𝐴 , 𝑠𝑡𝑟𝑜𝑛𝑔, 𝑤𝑒𝑎𝑘 = 𝑀 − 𝑠,
𝑢𝑖 𝐴, 𝐴 , 𝑠𝑡𝑟𝑜𝑛𝑔, 𝑠𝑡𝑟𝑜𝑛𝑔 = −𝑠, 𝑢𝑖 𝐴, 𝐴 , 𝑤𝑒𝑎𝑘,⋅ = −𝑤.

1 1
5. Joint probability distribution over types: 𝑝 𝑤𝑒𝑎𝑘, 𝑤𝑒𝑎𝑘 = , 𝑝 𝑤𝑒𝑎𝑘, 𝑠𝑡𝑟𝑜𝑛𝑔 = ,
4 4
1 1
𝑝 𝑠𝑡𝑟𝑜𝑛𝑔, 𝑠𝑡𝑟𝑜𝑛𝑔 = 4, 𝑝 𝑠𝑡𝑟𝑜𝑛𝑔, 𝑤𝑒𝑎𝑘 = 4.
Example: Armies and Conflict
• Find the Bayesian Nash equilibria of this game.

• Player 𝑖 and player 𝑗 are symmetric in this game.

• Player 𝑖 has 4 possible strategies (where the first element refers to the action if strong):
𝑁𝐴, 𝑁𝐴 , 𝑁𝐴, 𝐴 , 𝐴, 𝑁𝐴 , (𝐴, 𝐴).

• Let us find player 𝑗’s best response to each of these strategies by the opponent.
Example: Armies and Conflict
• Find the Bayesian Nash equilibria of this game.

• Player 𝑖 plays (𝑁𝐴, 𝑁𝐴).

• If player 𝑗 plays NA she gets 0, if she plays A she gets M>0.
• Her best response is A whether she is weak or strong, (𝐴, 𝐴).

• Player 𝑖 plays (𝑁𝐴, 𝐴).

𝑀 𝑀 𝑠 M 𝑤
• If player 𝑗 plays NA she gets 0, if she plays A she gets + − > 0 if she is strong and − <0
2 2 2 2 2
if weak.
• Her best response is A if strong and NA if weak, (𝐴, 𝑁𝐴).
Example: Armies and Conflict
• Find the Bayesian Nash equilibria of this game.

• Player 𝑖 plays (𝐴, 𝑁𝐴).

𝑠 𝑀 𝑤 𝑀
• If player 𝑗 plays NA she gets 0, if she plays A she gets − + > 0 if strong and − + < 0 if
2 2 2 2
weak.
• Her best response is A when strong and NA when weak, (𝐴, 𝑁𝐴).

• Player 𝑖 plays (𝐴, 𝐴).

𝑀
• If player 𝑗 plays NA she gets 0, if she plays A she gets − 𝑠 < 0 if strong and −𝑤 < 0 if weak.
2
• Her best response is NA when strong and NA when weak, (𝑁𝐴, 𝑁𝐴).
Example: Armies and Conflict
• Find the Bayesian Nash equilibria of this game.

• Summarise the best responses.

• Player 𝑖 plays (𝑁𝐴, 𝑁𝐴) → best response (𝐴, 𝐴).

• Player 𝑖 plays (𝑁𝐴, 𝐴) → best response (𝐴, 𝑁𝐴).

• Player 𝑖 plays (𝐴, 𝑁𝐴) → best response (𝐴, 𝑁𝐴)

• Player 𝑖 plays 𝐴, 𝐴 → best response (𝑁𝐴, 𝑁𝐴).

• Therefore there are 3 pure strategy BNE: 𝐴, 𝑁𝐴 , 𝐴, 𝑁𝐴

𝑁𝐴, 𝑁𝐴 , 𝐴, 𝐴
𝐴, 𝐴 , 𝑁𝐴, 𝑁𝐴
Example: Continuous Type Space

• Two countries, A and B, are in conflict over a resource.

• Each country chooses fight, F, or retreat, R, simultaneously.
• The cost of choosing to fight is 𝑐𝑖 for country 𝑖 ∈ {𝐴, 𝐵}, where 1 > 𝑐𝑖 > 0 (even if the other
country retreats).

• If a country retreats it gets payoff 0.

• If both fight then neither wins the resource but each pay their cost 𝑐𝑖 .
• If only one country fights they win the resource which is valued at 1.
Example: Continuous Type Space

• Suppose costs 𝑐𝐴 and 𝑐𝐵 are common knowledge.

• Describe this game as a simultaneous move game of complete information.

• Players: country A and country B.

• Action set for each player: 𝐴𝐴 = 𝐴𝐵 = {𝐹, 𝑅}.
• Payoffs for each player given by

Country B
Fight Retreat

Fight −𝑐𝐴 , −𝑐𝐵 1 − 𝑐𝐴 ,0

Country A
Retreat 0,1 − 𝑐𝐵 0,0
Example: Continuous Type Space

• Suppose costs 𝑐𝐴 and 𝑐𝐵 are common knowledge.

• Describe this game as a simultaneous move game of complete information.

Country B
Fight Retreat

Fight −𝑐𝐴 , −𝑐𝐵 1 − 𝑐𝐴 ,0

Country A
Retreat 0,1 − 𝑐𝐵 0,0
Example: Continuous Type Space

• Suppose costs 𝑐𝐴 and 𝑐𝐵 are common knowledge.

• What are the pure strategy Nash equilibria?

Country B
Fight Retreat

Fight −𝑐𝐴 , −𝑐𝐵 1 − 𝑐𝐴 ,0

Country A
Retreat 0,1 − 𝑐𝐵 0,0
Example: Continuous Type Space

• Suppose costs 𝑐𝐴 and 𝑐𝐵 are common knowledge.

• What are the pure strategy Nash equilibria?

• Players: country A and country B.

• Action set for each player: 𝐴𝐴 = 𝐴𝐵 = {𝐹, 𝑅}.
• Payoffs for each player given by

Country B
Fight Retreat

Fight −𝑐𝐴 , −𝑐𝐵 1 − 𝑐𝐴 ,0

Country A
Retreat 0,1 − 𝑐𝐵 0,0
Example: Continuous Type Space

• Now suppose the costs of fighting 𝑐𝐴 , 𝑐𝐵 are private information so that a country only knows its
own costs. The costs 𝑐𝐴 and 𝑐𝐵 are independent draws from a uniform distribution on 0,1 and
this is common knowledge.

• Describe this game as a Bayesian game.

Example: Continuous Type Space

• Now suppose the costs of fighting 𝑐𝐴 , 𝑐𝐵 are private information so that a country only knows its
own costs. The costs 𝑐𝐴 and 𝑐𝐵 are independent draws from a uniform distribution on 0,1 and
this is common knowledge.

• Describe this game as a Bayesian game.

1. Players: country A and country B.

2. Action set for each player: 𝐴𝐴 = 𝐴𝐵 = {𝐹, 𝑅}.

3. Type set for each player: Θ𝐴 = Θ𝐵 = [0,1]

4. Payoffs for each player given by 𝑢𝑖 (𝑎𝐴 , 𝑎𝐵 , 𝑐𝑖 ), which is dependent on the player’s type 𝑐𝑖 . (The payoffs
are the same as in the matrix in the previous slide.)

5. The joint probability distribution over types is given by independent uniform distributions for each
type.
Example: Continuous Type Space

• Suppose country B’s strategy is to fight if and only if 𝒄𝑩 < 𝜶 for some 𝟎 < 𝜶 < 𝟏. What is
country A’s best response?
Example: Continuous Type Space

• Suppose country B’s strategy is to fight if and only if 𝒄𝑩 < 𝜶 for some 𝟎 < 𝜶 < 𝟏. What is
country A’s best response?

• The probability that country B fights is the probability that 𝑐𝐵 < 𝛼 which is 𝛼 (since 𝑐𝐵 is
uniformly distributed on [0,1].

• Country A’s expected payoff from playing F is 𝛼 −𝑐𝐴 + 1 − 𝛼 1 − 𝑐𝐴

• Country A’s expected payoff from playing R is 0.

• Country A will play F if 𝛼 −𝑐𝐴 + 1 − 𝛼 1 − 𝑐𝐴 ≥ 0 so if 𝑐𝐴 ≤ 1 − 𝛼, and will play R otherwise.

Example: Continuous Type Space

• What is country B’s best response to country A’s strategy (play F if 𝒄𝑨 ≤ 𝟏 − 𝜶 and play R
otherwise)?
Example: Continuous Type Space

• What is country B’s best response to country A’s strategy (play F if 𝒄𝑨 ≤ 𝟏 − 𝜶 and play R
otherwise)?

• The probability that country A fights is the probability that 𝑐𝐴 < 1 − 𝛼 which is 1 − 𝛼 (since 𝑐𝐴 is
uniformly distributed on the interval [0,1]).

• Country B’s expected payoff from playing F is 1 − 𝛼 −𝑐𝐵 + 𝛼(1 − 𝑐𝐵 ).

• Country B’s expected payoff from playing R is 0.

• Country B plays F if 1 − 𝛼 −𝑐𝐵 + 𝛼 1 − 𝑐𝐵 ≤ 0 that is if 𝑐𝐵 ≤ 𝛼.

• We have a BNE: Country A fights if and only if 𝑐𝐴 ≤ 1 − 𝛼 and country B fights if and only if 𝑐𝐵 ≤
𝛼 and these are a best response to each other.

• This result does not specify the value of 𝛼 and this may be a BNE for multiple values of 𝛼.
Summary

• Example: Conflict model with private information.

• Example: Armies and Conflict.

• Example: Continuous type space.