Game Theory

--
Lecture 3

Patrick Loiseau
EURECOM
Fall 2016

1
 Lecture 2 recap
• Defined Pareto optimality
– Coordination games
• Studied games with continuous action space
– Always have a Nash equilibrium with some conditions
– Cournot duopoly example

à Can we always find a Nash equilibrium for all

games?
à How?

2
 Outline
1. Mixed strategies
– Best response and Nash equilibrium
2. Mixed strategies Nash equilibrium computation
3. Interpretations of mixed strategies

3
 Outline
1. Mixed strategies
– Best response and Nash equilibrium
2. Mixed strategies Nash equilibrium computation
3. Interpretations of mixed strategies

4
 Example: installing checkpoints
• Two road, Police choose on which to check,
Terrorists choose on which to pass
Terrorist
R1 R2 • Can you find a Nash
equilibrium?
R1 1 , -1 -1, 1

Police à Players must

R2 -1, 1 1, -1 randomize

5
 Matching pennies
• Similar examples:
Player 2 – Checkpoint placement
heads tails
– Intrusion detection
– Penalty kick
heads 1 , -1 -1, 1
– Tennis game
Player 1

tails -1, 1 1, -1 • Need to be unpredictable

6
 Pure strategies/Mixed strategies
• Game ( N, ( Ai )i∈N , (ui )i∈N )
• Ai: set of actions of player i (what we called Si
before)
• Action = pure strategy
• Mixed strategy: distribution over pure strategies
si ∈ Si = Δ(Ai )
– Include pure strategy as special case
– Support: supp si = {ai ∈ Ai : si (ai ) > 0}
• Strategy profile: s = (s1,, sn ) ∈ S = S1 ×× Sn 7
 Matching pennies: payoffs
• What is Player 1’s payoff if Player 2
plays s2 = (1/4, 3/4) and he plays:
Player 2
– Heads?
heads tails

heads 1 , -1 -1, 1
– Tails?
Player 1

tails -1, 1 1, -1 – s1 = (½, ½)?

8
 Payoffs in mixed strategies: general
formula
• Game ( N, ( Ai )i∈N , (ui )i∈N ) , let A = × Ai
i∈N
• If players follow a mixed-strategy profile s, the
expected payoff of player i is:
ui (s) = ∑ ui (a)Pr(a | s) where Pr(a | s) = ∏ si (ai )
a∈A i∈N

• a: pure strategy (or action) profile

• Pr(a|s): probability of seeing a given the
mixed strategy profile s
9
 Matching pennies: payoffs check
• What are the payoffs of Player 1
and Player 2 if s = ((½, ½), (¼, ¾))?
Player 2
heads tails

heads 1 , -1 -1, 1

Player 1

tails -1, 1 1, -1
• Does that look like it could be a
Nash equilibrium?

10
 Best response
• The definition for mixed strategies is
unchanged!

Definition: Best Response

Player i’s strategy ŝi is a BR to strategy s-i of other
players if:
ui(ŝi , s-i) ≥ ui(s’i , s-i) for all s’i in Si

• BRi(s-i): set of best responses of i to s-i

11
 Matching pennies: best response
• What is the best response of
Player 1 to s2 = (¼, ¾)?
Player 2
heads tails
• For all s1, u1(s1, s2) lie between
u1(heads, s2) and u1(tails, s2)
heads 1 , -1 -1, 1 (the weighted average lies
between the pure strategies
Player 1 exp. Payoffs)
tails -1, 1 1, -1

à Best response is tails!

12
 Important property
• If a mixed strategy is a best response then
each of the pure strategies in the mix must be
best responses
è They must yield the same expected payoff
Proposition:
For any (mixed) strategy s-i, if si ∈ BRi (s−i ), then
ai ∈ BRi (s−i ) for all ai such that si (ai ) > 0 .

In particular, ui(ai, s-i) is the same for all ai such that

si (ai ) > 0 13
 Wordy proof
• Suppose it were not true. Then there must be at least one
pure strategy ai that is assigned positive probability by my
best-response mix and that yields a lower expected payoff
against si
• If there is more than one, focus on the one that yields the
lowest expected payoff. Suppose I drop that (low-yield) pure
strategy from my mix, assigning the weight I used to give it to
one of the other (higher-yield) strategies in the mix
• This must raise my expected payoff
• But then the original mixed strategy cannot have been a best
response: it does not do as well as the new mixed strategy
• This is a contradiction

14
 Matching pennies again
• What is the best response
of Player 1 to s2 = (¼, ¾)?
Player 2
heads tails

heads 1 , -1 -1, 1
• What is the best response
of Player 1 to s2 = (½, ½)?
Player 1

tails -1, 1 1, -1

15
 Nash equilibrium definition
Definition: Nash Equilibrium
A strategy profile (s1*, s2*,…, sN*) is a Nash
Equilibrium (NE) if, for each i, her choice si* is a
best response to the other players’ choices s-i*

• Same definition as for pure strategies!

– But here the strategies si* are mixed strategies

16
 Matching pennies again
• Nash equilibrium:
Player 2 ((½, ½), (½, ½))
heads tails

heads 1 , -1 -1, 1

Player 1

tails -1, 1 1, -1

17
 Nash equilibrium existence theorem

Theorem: Nash (1951)

Every finite game has a Nash equilibrium.

• In mixed strategy!
– Not true in pure strategy
• Finite game: finite set of player and finite
action set for all players
– Both are necessary!
• Proof: reduction to Kakutani’s fixed-point thm 18
 Outline
1. Mixed strategies
– Best response and Nash equilibrium
2. Mixed strategies Nash equilibrium computation
3. Interpretations of mixed strategies

19
 Computation of mixed strategy NE
• Hard if the support is not known
• If you can guess the support, it becomes very
easy, using the property shown earlier:

Proposition:
For any (mixed) strategy s-i, if si ∈ BRi (s−i ), then
ai ∈ BRi (s−i ) for all ai such that si (ai ) > 0 .

In particular, ui(ai, s-i) is the same for all ai such that

si (ai ) > 0 (i.e., ai in the support of si) 20
 Example: battle of the sexes
Player 2
Opera Soccer
Opera 2,1 0,0
Player 1
Soccer 0,0 1,2

• We have seen that (O, O) and (S, S) are NE

• Is there any other NE (in mixed strategies)?

– Let’s try to find a NE with support {O, S} for each
player
21
 Example: battle of the sexes (2)
Player 2
Opera Soccer
Opera 2,1 0,0
Player 1
Soccer 0,0 1,2

• Let s2 = (p, 1-p)

• If s1 is a BR with support {O, S}, then Player 1
must be indifferent between O and S

à p = 1/3 22
 Example: battle of the sexes (3)
Player 2
Opera Soccer
Opera 2,1 0,0
Player 1
Soccer 0,0 1,2

• Similarly, let s1 = (q, 1-q)

• If s2 is a BR with support {O, S}, then Player 2
must be indifferent between O and S

à q = 2/3 23
 Example: battle of the sexes (4)
Player 2
Opera Soccer
Opera 2,1 0,0
Player 1
Soccer 0,0 1,2

• Conclusion: ((2/3, 1/3), (1/3, 2/3)) is a NE

24
 Example: prisoner’s dilemma
Prisoner 2
• We know that (D, D) is NE D C

• Can we find a NE with

support {C, D} with each? D -5, -5 0, -6

Prisoner 1

C -6, 0 -2, -2

• A NE in strictly dominant
strategies is unique! 25
 General methods to compute Nash
equilibrium
• If you know the support, write the equations
translating indifference between strategies in
the support (works for any number of
actions!)
• Otherwise:
– The Lemke-Howson Algorithm (1964)
– Support enumeration method (Porter et al. 2004)
• Smart heuristic search through all sets of support
• Exponential time worst case complexity

26
Complexity of finding Nash equilibrium
• Is it NP-complete?
– No, we know there is a solution
– But many derived problems are (e.g., does there
exists a strictly Pareto optimal Nash equilibrium?)
• PPAD (“Polynomial Parity Arguments on
Directed graphs”) [Papadimitriou 1994]
• Theorem: Computing a Nash equilibrium is
PPAD-complete [Chen, Deng 2006]

27
Complexity of finding Nash equilibrium
(2)

NP-hard

NP-complete

NP
PPAD

28
 Outline
1. Mixed strategies
– Best response and Nash equilibrium
2. Mixed strategies Nash equilibrium computation
3. Interpretations of mixed strategies

29
 Mixed strategies interpretations
• Players randomize
• Belief of others’ actions (that make you
indifferent)
• Empirical frequency of play in repeated
interactions
• Fraction of a population
– Let’s see an example of this one

30
 The Income Tax Game (1)
Tax payer
Honest Cheat

Auditor
A 2,0 4,-10 p

N 4,0 0,4 (1-p)

q 1-q

• Assume simultaneous move game

• Is there a pure strategy NE?
• Find mixed strategy NE

31
 The Income Tax Game: NE
computation
• Mixed strategies NE:

E[U1 ( A, (q,1 - q ))] = 2q + 4(1 - q) ü 2

ý2q = 4(1 - q) Þ q =
E[U1 (N , (q,1 - q ))] = 4q + 0(1 - q)þ 3
E[U 2 (H , ( p,1 - p ))] = 0 ü 2
ý4 = 14 p Þ p =
E[U 2 (C , ( p,1 - p ))] = -10 p + 4(1 - p)þ 7

Look at tax To find

payers auditors
payoffs mixing
32
The Income Tax Game: mixed strategy
interpretation
• From the auditor’s point of view, he/she is going
to audit a single tax payer 2/7 of the time
èThis is actually a randomization (which is applied
by law)

• From the tax payer perspective, he/she is going to

be honest 2/3 of the time
è This in reality implies that 2/3rd of population is
going to pay taxes honestly, i.e., this is a fraction
of a large population paying taxes

33
 The Income Tax Game (6)
• What could ever be done if one policy maker
(e.g. the government) would like to increase
the proportion of honest tax payers?

• One idea could be for example to “prevent”

fraud by increasing the number of years a tax
payer would spend in jail if found guilty

34
The Income Tax Game: Trying to make
people pay
Tax payer
Honest Cheat

Auditor
A 2,0 4,-20 p

N 4,0 0,4 (1-p)

q 1-q

• How to make people pay their taxes?

• One idea: increase penalty for cheating

• What is the new equilibrium? 35

 The Income Tax Game: new NE
E[U1 ( A, (q,1 - q ))] = 2q + 4(1 - q) ü 2
ý2q = 4(1 - q) Þ q =
E[U1 (N , (q,1 - q ))] = 4q + 0(1 - q)þ 3
E[U 2 (H , ( p,1 - p ))] = 0 ü 1 2
ý24 p = 4 Þ p = <
E[U 2 (C , ( p,1 - p ))] = -20 p + 4(1 - p)þ 6 7
• The proportion of honest tax payers didn’t change!
– What determines the equilibrium mix for the column
player is the row player’s payoffs
• The probability of audit decreased
– Still good, audits are expensive
• To make people pay tax: change auditor’s payoff
– Make audits cheaper, more profitable 36
 Important remark
• Row player’s NE mix determined by column
player’s payoff and vice versa
• Neutralize the opponent (make him
indifferent)
• In some sense the opposite of optimization
(my choice is independent of my own payoff)

37
 The penalty kick game
• 2 players: kicker and goalkeeper
• 2 actions each
– Kicker: kick left, kick right
– Goalkeeper: jump left, jump right
• Payoff: probability to score for the kicker,
probability to stop it for the goalkeeper
• Scoring probabilities: Goal keeper
L R

L 58.30 94.97
Kicker
R
92.91 69.92 38
 The penalty kick game: results
• Ignacio Palacios-Huerta. Professionals Play
Minimax. Review of Economics Studies (2003).
• Result:
Goal L Goal R Kicker L Kicker R
NE prediction 41.99 58.01 38.54 61.46
Observed freq. 42.31 57.69 39.98 60.02

• For a given kicker, his strategy is also serially

independent

39
 Summary
• Mixed strategies: distribution over actions
– A Nash equilibrium in mixed strategies always
exists for finite games
– Computation is easy if the support is known
• All pure strategies in the support of a best response are
also best responses
• Makes other player indifferent in his support
– Computation is hard if the support is not known
– Several interpretations depending on the game at
stake

40