Forward Induction, Signaling

and Reputation

14.126 Game Theory

Sergei Izmalkov
Muhamet Yildiz

Road Map
1. Forward Induction
2. Signaling games
1. Sequential Equilibria
2. Intuitive Criteria
3. Reputation
1. Chain-store paradox, finitely repeated games
2. Centipede game with incomplete information
3. Finitely repeated entry-deterrence game with incomplete
information.

1
 Forward Induction

The Battle of the Sexes with

outside options
1

B S
2,2
B 3,1 0,0
S 0,0 1,3

2
 Forward Induction
• One ought to interpret the actions as
outcomes of conscious choice even off the
path.
• Intuitive criterion
• Mistaken theories

Strong belief in rationality

At any history of the game, each agent is
assumed to be rational if possible. (That is, if
there are two strategies s and s’ of a player i
that are consistent with a history of play, and
if s is strictly dominated but s’ is not, at this
history no player j believes that i plays s.)

3
 Examples
1 1

2 2
5 2
0 1 0 1
1 1
2 2
3 2 3 2
1 3 1 3

Burning Money
1
0 D BB BS SB SS

B S B S 0B
B 3,1 0,0 B 2,1 -1,0 0S
S 0,0 1,3 S -1,0 0,3 DB

DS

O T H E R

4
 Table for the bidding game
Ui = 20(2+2minjbidj – bidi)

min 1 2 3
bid
1 60 - -

2 40 80 -

3 20 60 100

Nash equilibria of bidding game

• 3 equilibria: s1 = everybody plays 1; s2 = everybody
plays 2; s3 = everybody plays 3.
• Assume each player trembles with probability ε < 1/2,
and plays each unintended strategy w.p. ε/2, e.g., w.p.
ε/2, he thinks that such other equilibrium is to be played.
– s3 is an equilibrium iff

– s2 is an equilibrium iff

– s1 is an equilibrium iff

5
 1

0.8

0.6

0.4

0.2

0.047
0

-0.2
n
(1-ε/2 ) -1/2
-0.4

-0.6

-0.8 n n
(1-ε) +(1-ε/2 ) -1

-1
0 0.032 0.05 0.1 0.15 0.2

Bidding game with entry fee

Each player is first to decide min 1 2 3
whether to play the Bid
bidding game (E or X); if 1 60 - -
he plays, he is to pay a fee 2 40 80 -
p > 60.
3 20 60 100

For each m =1,2,3, ∃SPE: (m,m,m) is played in the bidding

game, and players play the game iff 20(2+m) ≥ p.
Forward induction: when 20(2+m) < p, (Em) is strictly dominated
by (Xk). After E, no player will assign positive probability to
min bid ≤ m. FI-Equilibria: (Em,Em,Em) where 20(2+m) ≥ p.
What if an auction before the bidding game?

6
 Signaling

Model
• Players: (S)ender, (R)eceiver
1. Nature selects t from T – the probability
distribution is π;
2. S observes t, and sends message m from a set M;
3. R observes m – but not t – and takes an action a;
4. S gets US(t,m,a) and R gets UR(t,m,a).
This is common knowledge.

7
 Beer – Quiche
0 1
1 du l 1
el e
beer quiche du

2 t {.1} don 3
’
0 don tw
’t
0

1 0
0 ts
0

el
du

du
el

beer {.9} quiche

3 ’ t don 2
1 don ’ t 1

Good equilibrium
0 1
1 du l 1
el e
beer quiche du
{.1} don
2 ’t 3
0 don tw
’t
0

1 0
0 ts
0
el
du

du
el

beer {.9} quiche

3 ’t don 2
1 don ’ t 1

8
 Bad equilibrium
0 1
1 du l 1
el e
beer quiche du

2 t {.1} don 3
’
0 don tw
’t
0

1 0
0 ts
0

el
du

du
el

beer {.9} quiche

3 ’ t don 2
1 don ’ t 1

Beer – Quiche – M
0 1
1 du l 1
el e
beer quiche du
1000002 $1M $1M 1000003
-999999 2 {.1} don 3 -1000000
’t
0 don tw
’t
0

1 0
0 ts
0
el
du

du
el

$1M beer {.9} quiche

1000003 $1M 1000002
-1000000 3 don 2 -999999
’t
1 don ’ t 1

9
 Cho -- Kreps
• T(m); M(t); A(m)
• Action set is finite;
• ρ(m;t) = the probability that t sends m;
• φ(a;m) = the probability that R chooses a,
receiving m;
• BR( µ , m) = arg max a∈A( m ) ∑U R (t, m, a) µ (t)
• For subset I of T, t∈T ( m)
BR( I , m) = U BR( µ , m)
{ µ :µ ( I )=1}
• MBR

Sequential equilibrium

• Beliefs:  π (t) ρ (m; t)

 if ∑π (t' ) ρ (m;t' ) > 0
µ (t | m) =  ∑ π (t' ) ρ (m;t' ) t '∈T ( m)

 t '∈T ( m )
 something otherwise

• ρ(m,t) > 0 => ΣaUS(t,m,a)φ(a;m) is

maximized at m;
• φ(.;m) is in MBR(µ(.|m),m)

10
 Testing an equilibrium
• U*(t) = expected utility of type t in equilibrium;
1. Pick a criterion, saying that particular out-of
equilibrium message (OEM) cannot be sent by
some type t. Also, say that a will not bet taken in
response to m if a is not in BR(T(m),m). Iterate.
[Ts(m)]
2. For each OEM m, consider all sequential
equilibrium responses of R to m in the original
game. Are all of of these are sequentially
rational, given Ts(m). If not, FAIL.

Dominance
• For any OEM m, eliminate t if ∃m’ s.t.

min a∈A( m ') U S (t , m' , a ) > max a∈A( m ) U S (t , m, a )

-1
a1
m’ m 0
0
0 {.9} 1
tw a2
1
a1 -1
m’ ts m
0 {.1} 1
0 a2 -1
0

11
 Equilibrium Dominance &
Intuitive Criterion
Equilibrium Domination: ∀OEM m, eliminate
(t,m) if

U * (t) > max a∈A( m) U S (t, m, a ).

Intuitive Criterion: ∀OEM m, define
T (m) = {t |U * (t) > max a∈BR(T ( m),m) U S (t, m, a )}.
~
If ∃(t’,m) s.t.
U * (t' ) < min a∈BR (T ( m )\T~ ( m ),m) U S (t', m, a ),
then the equilibrium fails the Intuitive Criterion.

Bad equilibrium
0 1
1 du l 1
el e
beer quiche du
{.1} don
2 ’t 3
0 don tw
’t
0

1 0
0 ts
0
el
du

du
el

beer {.9} quiche

3 ’t don 2
1 don ’ t 1

12
 Good equilibrium
0 1
1 du l 1
el e
beer quiche du

2 t {.1} don 3
’
0 don tw
’t
0

1 0
0 ts
0

el
du

du
el

beer {.9} quiche

3 ’ t don 2
1 don ’ t 1

Beer – Quiche – M
0 1
1 du l 1
el e
beer quiche du
1000002 $1M $1M 1000003
-999999 2 {.1} don 3 -1000000
’t
0 don tw
’t
0

1 0
0 ts
0
el
du

du
el

$1M beer {.9} quiche

1000003 $1M 1000002
-1000000 3 don 2 -999999
’t
1 don ’ t 1

13
 Reputation

Entry deterrence
1 Enter 2 Acc.
(1,0)

X Fight

(0,1) (-1,-1)

14
 Entry deterrence, repeated twice,
many times
1 Enter 2 Acc. 1 Enter 2 Acc.
2,0

X Fight X Fight
Acc. 2 Enter 1
1,1 (1,1) 0,-1
1 Enter 2 Acc.
Fight X 0,-1

X Fight
-1,0 0,2

-1,0 (-2,-2)
What would happen if repeated n times?

Twice-repeated PD
1
C D
2
C D C D

1 1 1 1
D C D C
C C D D
2 2 2 2

C D C D D C D D C D
C D C D C C

10 5 11 6 5 0 6 1 11 6 12 7 6 1 7 2
10 11 5 6 11 12 6 7 5 6 0 1 6 7 1 2

What would happen if T = {0,1,2,…,n}?

15
 Centipede Game
1 2 1 1 2 1 2 100
100
…

1 0 2 98 97 99 98
1 3 2 98 100 99 101

Centipede Game – with doubt

197 5 4 3 2 1 =n
{.999}
1 2 2 1 2 1 2 100
µ5 µ3 µ1 100
…

1 0 96 98 97 99 98
1 3 99 98 100 99 101

1 2 2 1 2 1 2 0
{.001} 100
…

-1 0 0 -1 0 -1 0
1 3 99 98 100 99 101

16
 Facts about the Centipede
• Every information set of 2 is reached with positive
probability.
• 2 always goes across with positive probability.
• If 2 strictly prefers to go across at n, then
– 1 must strictly prefer to go across at n+1,
– 2 must strictly prefer to go across at n+2,
– her posterior at n is her prior.
• For any n > 2, 1 goes across with positive
probability. If 1 goes across w/p 1 at n, then 2’s
posterior at n-1 is her prior.

Centipede Game – with doubt

197 5 4 3 2 1 =n
{.999}
1 2 2 1 2 1 2 100
100
…

1 0 96 98 97 99 98
1 3 99 98 100 99 101

1 2 µ5 2 1 µ3 2 1 µ1 2 0
{.001} 100
…

-1 0 0 -1 0 -1 0
1 3 99 98 100 99 101

17
 If 2’s payoff at any n is x and 2 is mixing,
then
x = µn(x+1) + (1- µn)[(x-1)pn +(1-pn)(x+1)]
= µn(x+1) + (1- µn)[ (x+1) -2pn]
= x+1 - 2pn(1- µn)
Ù(1- µn) pn= 1/2
µn µn
µn−1 = = = 2 µn
µn + (1 − µn )(1 − pn ) µn + (1 − µn ) − pn (1 − µn )
µn−1
µn =
2

0.5

0.45

0.4

0.35

0.3

0.25
Go Across
0.2

0.15

0.1

0.05
Mix
0
0 20 40 60 80 100

18
 Entry-deterrence with doubt
1 Enter 2 Acc.
{0.999}
(1,0)
Weak
X Fight

(0,2) (-1,-1)
µ
1 Enter 2
{0.001} Strong
X Fight

(0,2) (-1,-1)

Entry-deterrence with doubt

• Incumbent is as before.
• Each day there is a new entrant.
• Two types of entrants:
– W/p q < 1/2, tough, u(enter, fight) = 1;
– W/p 1-q, weak, u(enter, fight) = 1.

19
 Sequential Equilibrium
• Tough entrant (tE) • In the last period,
always enters; strong – wI accommodates;
incumbent (sI) always – wE enters iff
fights.
− 1µ0 + (1 − µ0 ) ≥ 0 ⇔ µ0 ≤ 1/ 2
• If any entrant (E) is
accommodated; it • if µ0 < 1/ 2, at n=1, wI
becomes common accommodates => µ0 = 1.
knowledge that • if µ0 > 1/ 2, at n=1, wI
incumbent (I) is weak fights => µ0 = 0.001.
…
• µ0 = 1/ 2.

• µ0 = µ1/(µ1 + φ1(1−µ1)) =1/2

Ùµ1 = φ1(1−µ1)
=>Prob(fight|µ1) = 2µ1.
=>at n=1, wE enters iff 2µ1 < ½ i.e., µ1 < ¼.
• Similarly, µ1 = ¼.
• Proceed as before.

20