Professional Documents
Culture Documents
P2 1. Rationalizable Behavior - Class Slides
P2 1. Rationalizable Behavior - Class Slides
Rationalizable Behavior
Summary
Benjamin Bernard
Big Picture
Big Picture Mixing and Conjectures Formal Notation Examples
Game:
• We model a strategic interaction with a game G = I, (Ai ), (ui ) .
• Each player i ∈ I is rational (aims to maximize ui ) and aware of G.
Solution concept:
• We solve a game with a solution concept.
• Different solution concepts differ in the assumptions required about
the players’ knowledge and beliefs.
Examples:
• No knowledge: strict dominant-strategy equilibrium.
• Common knowledge of rationality: rationalizability.
L C R
T 4, 3 5, 1 6, 2
B 2, 1 8, 4 3, 6
IESDS:
• Very intuitive: eliminate clearly suboptimal pure strategies iteratively.
• Find set Σ∞ ∞
i of undominated mixed strategies in ∆(Si ).
• Unfortunately, in general the remaining strategies in Σ∞
i need not be
∞
best responses to a conjecture on Σ−i .
Rationalizability:
• In each step we wonder: is it justifiable to play a certain strategy? Is
there a valid conjecture, to which this strategy is a best response?
• Each σi ∈ Rki is a best responses to conjecture on Rk−1
−i .
2
Big Picture Mixing and Conjectures Formal Notation Examples
L C R
T 4, 3 5, 1 6, 2
B 2, 1 8, 4 3, 6
Rationalizability:
• In each step we wonder: is it justifiable to play a certain strategy?
• L is justified by conjecture δT , R is justified by conjecture δB .
• T is justified by conjecture δL , B is justified by conjecture δC .
• Problem: we also have to do it for mixed strategies x T + (1 − x )B.
Mixed Actions
L R L R
T 25% 25% 50% T 0% 50% ?%
50% 50% ?% ?%
Action profiles:
• The set of pure action profiles is A = A1 × · · · × An = ×ni=1 Ai .
• A pure action profile is a vector a = (a1 , . . . , an ), where ai ∈ Ai .
• The set ∆(Ai ) of all distributions on Ai is the set of i’s mixed actions.
×ni=1 ∆(Ai ).
• A mixed action profile is a vector α = (α1 , . . . , αn ) ∈
The set ∆(A) of all distributions on A is larger than ×i=1 ∆(Ai ).
n
•
4
Big Picture Mixing and Conjectures Formal Notation Examples
Mixed Strategies
L R L R
T 25% 25% 50% T 0% 50% ?%
50% 50% ?% ?%
Strategy profiles:
• The set of pure strategy profiles is S = S1 × · · · × Sn = ×nj=1 Si .
• A pure strategy profile is a vector s = (s1 , . . . , sn ), where si ∈ Si .
• The set ∆(Si ) of all distributions on Si is the set of i’s mixed strategies.
×ni=1 ∆(Si ).
• A mixed strategy profile is a vector σ = (σ1 , . . . , σn ) ∈
The set ∆(S) of all distributions on S is larger than ×i=1 ∆(Si ).
n
•
5
Big Picture Mixing and Conjectures Formal Notation Examples
Conjectures
L R L R
T 25% 25% 50% T 10% 15% 25%
Conjectures:
• Beliefs about opponent j’s strategy are an element of ∆(∆(Sj )).
• Example: Player i believes Player 1 plays σ1L and σ1R each with 50%.
• A conjecture πi = (πi1 , . . . , πii−1 , πii+1 , . . . , π n ) is a vector of beliefs πij
what each opponent j ̸= i will do: an element of j̸=i ∆(∆(Sj )). ×
×
• The set ∆( j̸=i ∆(Sj )) of all distributions over opponent’s mixed
strategy profiles is larger than the set of conjectures.
6
Big Picture Mixing and Conjectures Formal Notation Examples
Mixing:
• Suppose we want to play 21 R + 12 P.
• Let mi (for mixing) be the outcome of a coin flip, then we can choose
(
R if mi = H,
σi (mi ) =
P if mi = T .
• Importantly, for strategic purposes only the distribution σi of the chosen
pure strategies matters, not how these were realized.
• Mostly, we don’t write the dependence of σi on i’s mixing device.
7
Big Picture Mixing and Conjectures Formal Notation Examples
Definition 1.5
3. The set of player i’s mixed strategies Σ∞ i that survive IESDS is given
by all σi ∈ ∆(Si∞ ) that are not strictly dominated in S ∞ .
Interpretation:
1. In each step, eliminate strictly dominated pure strategies.
2. Repeat until the process converges.
3. Find undominated mixed strategies.
9
Big Picture Mixing and Conjectures Formal Notation Examples
Rationalizability
Definition 1.8
×
( )
k k−1 there exists πi ∈
j̸ =i
∆(Rk−1
j ),
Ri := σi ∈ Ri .
for which σi is a best response
Interpretation:
1. In each step, we find all mixed strategies that are a best response to
some plausible conjecture.
2. Repeat until the process converges.
10
Big Picture Mixing and Conjectures Formal Notation Examples
A B C D
T 0, 3 4, 0 1, 2 3, 1
M 3, 0 2, 0 1, 2 0, 1
B 4, 0 0, 3 1, 2 2, 1
L R
T 5, 2 1, 3
M 3, 4 2, 0
B 2, 1 4, 2
Problem:
• Which strategies are rationalizable?
• Since it is a 2-player game, we apply IESDS.
12
Big Picture Mixing and Conjectures Formal Notation Examples
Cournot Duopoly
$80
$60
$40
$20
09/21/2019 09/21/2021
Cournot duopoly:
• Each firm i = 1, 2 can produce quantity qi ∈ [0, 200] at cost c(qi ) = 10qi .
• Suppose the market price is p(q) = 100 − q1 − q2 so that i’s payoff is
Best responses:
• Suppose Firm 2 produces quantity q2 . How should Firm 1 respond?
• Extreme value theorem: the maximum of u1 (q1 , q2 ) is attained at:
• a point where u ′1 (q) = 0,
• a boundary point,
• a point where u1 is not differentiable.
14
Big Picture Mixing and Conjectures Formal Notation Examples
u1 q
b1
q1
Best responses:
• If u1 is maximized at an interior point qb1 , we must have
∂u1 (q) 90 − q2
0= = 90 − 2b q 1 − q2 , ⇒ qb1 = .
∂q1
q1 =b
q1 2
• We verify this is a local maximum with the second-order condition
∂ 2 u1 (q)
= −2 < 0.
∂q 21
• u1 is maximized at q1 = 0 if qb1 ≤ 0.
15
Big Picture Mixing and Conjectures Formal Notation Examples
q1
best responses
90−q2
q1 = 2
q2
Best responses:
• In summary, the Firm 1’s best response is
(90 − q2 )+
q1∗ = max{b
q 1 , 0} = .
2
• Any strategy q1 ∈ [0, 45] is a best response to a Dirac conjecture δq2 .
• By Theorem 1.11, we conclude R1i = [0, 45].
16