Simultaneous Games- Mixed Strategies

Abdul Quadir
XLRI

24 December, 2019
Reading: Dixit and Skeath, 7
Mixed Strategies

I Suppose players are choosing their pure strategies

unsystematically.
I This means in a sense that they are choosing their pure
strategies randomly.
I Randomization over pure strategies is known as mixed
strategies.
I In other words, mixed strategies are probabilities associated
with each strategies.
I Note that since mixed strategies are probabilities, they must
lie in [0, 1].
I Therefore, there are like the continuous strategies.
I Pure strategies are special case of mixed strategies.
Expected Payoff

I We assume independent randomization among players.

I How one player randomizes over his set of strategies will not
depend on how others randomize.
I Consider the following game:
C D
A (50,50) (80,20)
B (90,10) (20,80)
3
I Suppose Player 1 plays strategy A with probability 4 and
strategy B with probability 14 .
1
I Player 2 plays C with probability 4 and D with probability 34 .
I What are the expected payoffs of players playing their
strategies?
Expected Payoffs
I Row player (Player 1) is playing A with with probability 34 and
strategy B with probability 14 .
I Given this strategy, the expected payoff of Player 2 is given:

3 1
Playing C : × 50 + × 10 = 40
4 4
3 1
Playing D : × 20 + × 80 = 35
4 4
 
I Against the mixed strategy 43 , 41 for player 1, the best for
Player 2 is to play C .
I Therefore, if Player 2 plays D for the above strategy, the
expected payoff of Player 1:
3 1
× 50 + × 90 = 60 if player 2 plays C
4 4
3 1
× 80 + × 20 = 65 if player 2 plays D
4 4
Indifference Principle

I Note that Player 2 can exploit player 1 by playing C for this

mixed strategy.
I Thus, this is not optimal mix for Player 1.
I The ideal mixed strategy will be where there is no exploitation
of player 1 by player 2 and vice versa.
I This will arise if Player 1 choose a mixed strategy that leave
player 2 indifferent in choosing strategy C and D.
I In other words, player 1 will choose the mixed strategy that
makes the expected payoff of player 2 same for C and D.
I This principle is known as indifference property.
I This is the key for mixed strategy equilibria in non-zero sum
game.
Computation of Mixed Strategy
I To compute exploitation proof mixed strategy, we assign
probability p to the strategy A of player 1.
I Thus, the expected payoff of Player 2 is:
Playing C : p × 50 + (1 − p) × 10
Playing D : p × 20 + (1 − p) × 80
I Player 2 will be indifferent in playing C and D if the expected
payoffs are same for either playing C or D.
I Therefore, we have
p × 50 + (1 − p) × 10 = p × 20 + (1 − p) × 80
I Solving this we get p = 0.7.
I Thus, the exploitation proof mixed strategy for player 1 is
(0.7, 0.3).
I Compute the expected payoff of the mixed strategy (0.7, 0.3)
and compare it with the expected payoffs of pure strategy and
the strategy (0.75,0.25).
Best Response

I Suppose player 2 plays C with probability q and D with

probability (1 − q).
I What is the best response of Player 1 p for the choices of q of
player 2?
I First compute the expected payoff of player 1 if Player 2 plays
C or D.

p × 50 + (1 − p) × 90 if Player 2 plays C
p × 80 + (1 − p) × 20 if Player 2 plays D

I What is the expected payoff of Player 1 against the mixed

strategy q of player 2?

[50p + 90(1 − p)]q + [80p + 20(1 − p)](1 − q)

Best Response

I Suppose player 2 plays C with probability q and D with

probability (1 − q).
I What is the best response of Player 1 p for the choices of q of
player 2?
I First compute the expected payoff of player 1 if Player 2 plays
C or D.

p × 50 + (1 − p) × 90 if Player 2 plays C
p × 80 + (1 − p) × 20 if Player 2 plays D

I What is the expected payoff of Player 1 against the mixed

strategy q of player 2?

[50p + 90(1 − p)]q + [80p + 20(1 − p)](1 − q)

Best Response

I Suppose player 2 plays C with probability q and D with

probability (1 − q).
I What is the best response of Player 1 p for the choices of q of
player 2?
I First compute the expected payoff of player 1 if Player 2 plays
C or D.

p × 50 + (1 − p) × 90 if Player 2 plays C
p × 80 + (1 − p) × 20 if Player 2 plays D

I What is the expected payoff of Player 1 against the mixed

strategy q of player 2?

[50p + 90(1 − p)]q + [80p + 20(1 − p)](1 − q)

Best Response

I Suppose player 2 plays C with probability q and D with

probability (1 − q).
I What is the best response of Player 1 p for the choices of q of
player 2?
I First compute the expected payoff of player 1 if Player 2 plays
C or D.

p × 50 + (1 − p) × 90 if Player 2 plays C
p × 80 + (1 − p) × 20 if Player 2 plays D

I What is the expected payoff of Player 1 against the mixed

strategy q of player 2?

[50p + 90(1 − p)]q + [80p + 20(1 − p)](1 − q)

Best Response

Rearranging the above equation

[50p + 90(1 − p)]q + [80p + 20(1 − p)](1 − q) = 50pq + 90q − 90pq+

80p + 20 − 20p − 80pq − 20q + 20pq

This equates to

70q − 100pq + 60p + 20 = [20 + 70q] + [60 − 100q]p

I Note that we are trying to find p to maximize the Player 1’s

expected payoff given q.
I See the coefficient of p, that is 60 − 100q.
Best Response

I If 60 − 100q > 0, player 1 can increase his expected payoff by

increasing p.
I Thus, if q < 0.6, p = 1.
I If 60 − 100q < 0, player 1 can increase his expected payoff by
decreasing p.
I Thus, if q > 0.6, p = 0.
I For q = 0.6, any p-mix is the best response.
I We can summarize it as:
If q < 0.6, the best response is p = 1.
If q = 0.6, any p-mix is a best response. If q > 0.6, best
response is p = 0.
I This is the best response of Player 1 for Player 2’s strategies.
Best Response Graph

0.6 1 q
Best Response

I Similarly we can compute the best response for player 2.

I The expected payoff of player 2 is:

[50q + (1 − q)20]p + [10q + 80(1 − q)](1 − p)

I Rearranging we get

(80 − 60p) + (100p − 70)q

I Note that if the coefficient of q is positive, then player 2 will

increase q.
I Thus,we have
If p < 0.7, best response is q = 0
If p = 0.7, any q-mix is a best response.
If p > 0.7, best response is q = 1.
Best Response

I Similarly we can compute the best response for player 2.

I The expected payoff of player 2 is:

[50q + (1 − q)20]p + [10q + 80(1 − q)](1 − p)

I Rearranging we get

(80 − 60p) + (100p − 70)q

I Note that if the coefficient of q is positive, then player 2 will

increase q.
I Thus,we have
If p < 0.7, best response is q = 0
If p = 0.7, any q-mix is a best response.
If p > 0.7, best response is q = 1.
Best Response Graph

0.7 1 p
Nash Equilibrium
q

0.6

0.7 1 p
 
Mixed strategy Nash equilibrium is (0.7, 0.3), (0.6, 0.3) .
Nash Equilibrium
q

0.6

0.7 1 p
 
Mixed strategy Nash equilibrium is (0.7, 0.3), (0.6, 0.3) .
Mixed Strategy
I We can also use the indifference property and calculate the
optimal mixed strategy q.
I Thus, the expected payoff of Player 1 is:

Playing A : q × 50 + (1 − q) × 80
Playing B : q × 90 + (1 − p) × 20

I Player 1 will be indifferent in playing A and B if the expected

payoffs are same for either playing A or B.
I Therefore, we have

q × 50 + (1 − q) × 80 = q × 90 + (1 − q) × 20

I Solving this we get q = 0.6.

I If probabilities are coming out to be negative or greater than
1, then there is no mixed strategy Nash equilibrium.
Mixed Strategy NE in Coordination Games

I Consider the coordination

X Y
X (1,1) (0,0)
Y (0,0) (3,3)
I Suppose Player 1 and Player 2 are playing X with probability
p and q, respectively.
I What is the mixed strategy Nash equilibria?
I Player 1:

X : 1 × q + (1 − q) × 0 = q
Y : 0 × q + 3 × (1 − q) = 3(1 − q)

I Thus, q = 3(1 − q) implies q ∗ = 43 .

I Similar calculation will yield p ∗ = 34 .
Mixed Strategy NE in Coordination Games

I Consider the coordination

X Y
X (1,1) (0,0)
Y (0,0) (3,3)
I Suppose Player 1 and Player 2 are playing X with probability
p and q, respectively.
I What is the mixed strategy Nash equilibria?
I Player 1:

X : 1 × q + (1 − q) × 0 = q
Y : 0 × q + 3 × (1 − q) = 3(1 − q)

I Thus, q = 3(1 − q) implies q ∗ = 34 .

I Similar calculation will yield p ∗ = 34 .
Best Response

I The expected payoff of Player 1 if player 2 plays X is p and if

he plays Y , then it is 3(1 − p)
I Thus, the expected payoff of Player 1 is

pq + 3(1 − p)(1 − q) = 3(1 − q) + (4q − 3)p

I This implies that If q < 34 , best response is p = 0

If q = 34 , any p-mix is a best response.
If q > 34 , best response is p = 1.
I Since Player 2 is symmetric, his best response is similar to
player 1.
If p < 34 , best response is q = 0
If p = 43 , any q-mix is a best response.
If p > 34 , best response is q = 1.
Best Response

I The expected payoff of Player 1 if player 2 plays X is p and if

he plays Y , then it is 3(1 − p)
I Thus, the expected payoff of Player 1 is

pq + 3(1 − p)(1 − q) = 3(1 − q) + (4q − 3)p

I This implies that If q < 34 , best response is p = 0

If q = 34 , any p-mix is a best response.
If q > 34 , best response is p = 1.
I Since Player 2 is symmetric, his best response is similar to
player 1.
If p < 34 , best response is q = 0
If p = 43 , any q-mix is a best response.
If p > 34 , best response is q = 1.
Best Response Graph for Player 1

0.75 1 q
Best Response Graph for Player 2

0.75

1 q
Nash Equilibria
p

0.75

0.75 1 q

Three Nash equilibria (0, 0), (1, 1), (0.75, 0.75).

Nash Equilibria
p

0.75

0.75 1 q

Three Nash equilibria (0, 0), (1, 1), (0.75, 0.75).

Exercise

1. Compute the Nash equilibria of matching pennies game and

draw the best response functions of both the players.
2. Compute the mixed strategy Nash equilibia of the following
matrix game:
X Y
X (2,1) (0,0)
Y (0,0) (1,2)
Also draw the best response functions of both the players on a
graph.
Dominated and Mixed Strategies

I A pure strategy that is not dominated by any pure strategy

may be dominated by a mixed strategy.
I Consider the following game:
L R
T (3,1) (0,4)
M (0,2) (3,1)
B (1,0) (1,2)
I No strategy dominates B.
I However, the mixed strategy 12 T and 21 M strictly dominates
the pure strategy B.
I A strictly dominated strategy is never used with positive
probability in a mixed strategy equilibrium.
Dominated and Mixed Strategies

I An easy way to figure out dominated strategy is to compare

expected payoffs.
I Let us consider Player 2’s mixed strategy as prob(L) = q.
I Then the expected payoffs of player 1 are:

U1 (T , q) = 3 × q + (1 − q) × 0 = 3q
U1 (M, q) = 0 × q + 3 × (1 − q) = 3(1 − q)
U1 (B, q) = 1 × q + 1 × (1 − q) = 1
Dominated and Mixed Strategies

U1 (·, q)

U1 (B, q)
3

1 U1 (B, q)

1 q
Dominated and Mixed Strategies

I Consider the following game:

L R
T (1,1) (1,0)
M (3,0) (0,3)
B (0,1) (4,0)
I No action dominated T .
I However, mixed strategy 12 M and 21 B is strictly dominates T .
Dominated and Mixed Strategies

I Consider the following game:

L R
T (1,1) (1,0)
M (3,0) (0,3)
B (0,1) (4,0)
I No action dominated T .
I However, mixed strategy 12 M and 21 B is strictly dominates T .
Real Life Example
I In a paper Ignacio Palacios-Huerta (2003) considered 5 years
worth of penalty kicks.
I Empirical scoring probabilities
L R
L (58,42) (95,5)
R (93,7) (70,30)
I R is the natural side of the kicker.
I Kicker must be indifferent
58q + 95(1 − q) = 93q + 70(1 − q) ⇒ q = 0.42
I Goal keeper must be indifferent
42p + 7(1 − p) = 5p + 30(1 − p) ⇒ p = 0.39
I Kicker’ theoretical prediction is 39% and data also provides
40%.
I Goallie’s theoretical prediction is 42% and data provides 42%.
Real Life Example
I In a paper Ignacio Palacios-Huerta (2003) considered 5 years
worth of penalty kicks.
I Empirical scoring probabilities
L R
L (58,42) (95,5)
R (93,7) (70,30)
I R is the natural side of the kicker.
I Kicker must be indifferent
58q + 95(1 − q) = 93q + 70(1 − q) ⇒ q = 0.42
I Goal keeper must be indifferent
42p + 7(1 − p) = 5p + 30(1 − p) ⇒ p = 0.39
I Kicker’ theoretical prediction is 39% and data also provides
40%.
I Goallie’s theoretical prediction is 42% and data provides 42%.