Game
Theory
Lecture
1
Pietro Michiardi
alpha
B -
my
pair
beta
alpha
me
beta
alpha
B -
beta
B +
me
beta
B +
my grades
pairs grades
beta
This
is
an
outcome
matrix:
alpha
B - , B -
beta
C , A
A , C
me
1st
grade
Row
player
B + , B +
2nd
grade
Column
player
beta
alpha
0 , 0
3, -1
beta
-1, 3
1,1
#
of
uGles,
or
uGlity:
(A,C)
3
(B-,
B-)
0
Hence
the
preference
order
is:
A
>
B+
>
B-
>
C
me
Deni4on:
We
say
that
my
strategy
alpha
strictly
dominates
my
strategy
beta,
if
my
payo
from
alpha
is
strictly
greater
than
from
beta,
regardless
of
what
others
do.
Lesson
1
Do
not
play
strictly
dominated
strategies
Lesson
2
RaEonal
choice
(i.e.,
not
choosing
a
dominated
strategy)
can
lead
to
outcomes
that
suck!
beta
alpha
0 , 0
-1, -3
beta
-3,-1
1,1
#
of
uGles,
or
uGlity:
(A,C)
3
4
=
-1
my
A
my
guilt
(C,
A)
-1
2
=
-3
my
C
my
indignaGon
This
is
a
coordina4on
problem
me
Lesson
3:
Payos
ma6ers!
<<
you
cant
get
what
you
want
unless
you
know
what
you
want
>>
my
pair
(AltruisGc)
alpha
beta
0 , 0
3, -3
-1,-1
1,1
alpha
Me
(Selsh)
beta
my
pair
(Selsh)
alpha
beta
alpha
0 , 0
-1, -1
Me
By
thinking
of
what
my
opponent
will
do
(AltruisGc)
I
can
decide
what
to
do.
beta
-3,3
1,1
Lesson
4:
Put
yourself
in
others
shoes
and
try
to
gure
out
what
they
will
do
ObservaGons
In
realisGc
selngs:
It
is
omen
hard
to
determine
what
are
the
payos
of
your
opponent
It
is
easier
to
gure
out
my
own
payos
2. Price compeGGon
3. Common resource
Carbon
emissions
Fishing
Nota4on
Players
i, j,
You all
Strategies
si:
a
parGcular
strategy
of
player
i
13
Payos
{1, 2, , 100}
s:
a
parGcular
play
of
the
game
strategy
prole
(vector,
or
list)
$5
-
error
ui(s)
=
0
AssumpGons
We
assume
all
the
ingredients
of
the
game
to
be
known
Everybody
knows
the
possible
strategies
everyone
else
could
choose
Everybody
knows
everyone
elses
payos
Example
1
2
C
L
1
T
B
5,
-1
6,
4
11,
3
0,
2
0,0
2,
0
Players
1, 2
Strategy sets
S1={T,B}
S2={L,C,R}
Payos
U1(T,C) = 11
U2(T,C) = 3
Example
conGnued
How
is
the
game
going
to
be
played?
Does
player
1
have
a
dominated
strategy?
Does
player
2
have
a
dominated
strategy?
For
a
strategy
to
be
dominated,
we
need
another
strategy
for
the
same
player
that
does
always
beWer
(in
terms
of
payos)
E
H
1,
1
0,
2
1,
1
2,
0
To
sum
up
(1)
Weve
explored
a
bit
the
idea
of
deleGng
dominated
strategies
Look
at
a
game
Figure
out
which
strategies
are
dominated
Delete
them
Look
at
the
game
again
Look
at
which
strategies
are
dominated
now
and
so
on
To
sum
up
(2)
Itera4ve
dele4on
of
dominated
strategies
seems
a
powerful
idea,
but
its
also
dangerous
if
you
take
it
literally
In
some
games,
iteraGve
deleGon
converges
to
a
single
choice,
in
others
it
may
not
(well
see
shortly
an
example)
HINT:
try
to
idenGfy
all
dominated
strategies
at
once
before
you
delete,
this
may
save
you
some
rounds
10
RIGHT
WING
u1(1,1) = 50 %
<
u1(2,1) = 90%
Vs. 2
u1(1,2) = 10 %
<
u1(2,2) = 50%
Vs. 3
u1(1,3) = 15 %
<
u1(2,3) = 20%
Vs. 4
u1(1,4) = 20 %
<
u1(2,4) = 25%
U1(2,1) = 90 %
>
U1(3,1) = 85%
u1(2,2) = 50 %
<
u1(3,2) = 80%
Vs. 3
u1(2,3) = 20 %
<
u1(3,3) = 50%
Vs. 4
u1(2,4) = 25 %
<
u1(3,4) = 30%
Vs. 5
U1(2,5) = 30 %
<
u1(3,5) = 35%
WHY?
This
is
going
to
be
(one
of)
your
next
homework
CriGcs
(1)
We
have
been
using
a
model
of
a
real-world
situaGon,
and
tried
to
predict
the
outcome
using
game
theory
What
is
missing?
Is
there
anything
wrong
with
the
model?
CriGcs
(2)
Voters
are
not
evenly
distributed
Many
voters
do
not
vote
There
may
be
more
than
2
candidates
See
these
in
your
homework
There
may
be
higher
dimensions
to
the
problem
CriGcs
(3)
So
if
were
missing
so
many
things,
our
model
is
useless,
and
in
general
modeling
(as
an
abstracGon
eort)
is
useless!!
No:
rst,
analyze
a
problem
with
simplifying
assumpGons,
then
relax
them
and
see
what
happens
E.g.:
would
a
dierent
voters
distribuGon
change
the
result?
l
U
1
M
D
5,1
1,3
4,2
0,2
4,1
2,3
l
U
1
M
D
5,1
1,3
4,2
0,2
4,1
2,3
l
U
1
M
D
5,1
1,3
4,2
0,2
4,1
2,3
Best
Response
Obviously,
the
50%
l
50%
r
is
just
a
belief
I
could
believe
my
opponent
would
lean
to
lem,
e.g.
with
a
75%
l
25%
r
probabiliGes
Can
we
use
a
representaGon
to
sum
up
all
these
possibiliGes
and
come
up
with
a
predicGon?
Expected
Payo
5
u1(U,l)
u1(M,r)
u1(D,l)
u1(D,r)
u1(U,r)
u1(M,l)
0
x
BR
is
U
0.5
BR
is
D
1
BR
is
M
p(r)
belief
Pietro Michiardi
Recap
We
introduced
the
idea
of
best
response
(BR)
do
the
best
you
can
do,
given
your
belief
about
what
the
other
players
will
do
We
saw
a
simple
game
in
which
we
applied
the
BR
idea
and
worked
with
plots
4,
-4
6,
-6
9,
-9
goalie
9,
-9
6,
-6
4,
-4
Expected
Payo
10
E[u1(L, p(r) )]
E[u1(R, p(r) )]
8
6
4
E[u1(L, p(r) )]
0.5
p(r)
belief
Lesson
1:
Do
not
choose
a
strategy
that
is
never
a
BR
to
any(*)
belief
Expected
Payo
10
8
6
4
2
0.5
p(r)
belief
s2
5
BR1(s2)
4
3
BR2(s1)
2
1
s1
s2
5
BR1(s2)
4
3
2
1
s1
s2
5
4
3
BR2(s1)
2
1
s1
s2
5
BR1(s2)
4
3
BR2(s1)
2
1
s1
s2
BR1(s2)
2
7/4
BR2(s1)
6/4
5/4
5/4
6/4
7/4
s1
Introducing
NE
So
in
the
partnership
game
weve
seen
what
a
NE
is
Recall
the
numbers
game:
what
was
the
NE
there?
Did
you
play
a
NE?
Although
NE
is
a
central
idea
in
game
theory,
be
aware
that
it
is
not
always
going
to
be
played
By
repeaLng
the
numbers
game,
however,
weve
seen
that
we
were
converging
to
the
NE
NE:
observaLons
It
is
not
always
the
case
that
players
play
a
NE!
E.g.:
in
the
numbers
game,
we
saw
that
playing
NE
is
not
guaranteed
s2
4
3
2
1
s1
Finding
NE
point(s)
Next
we
will
play
some
very
simple
games
involving
few
players
and
few
strategies
Get
familiar
with
nding
NE
on
normal
form
games
We
will
have
a
glimpse
on
algorithmic
ways
of
nding
NE
and
their
complexity
0,4
4,0
3,5
Player
2
c
r
4,0
0,4
3,5
5,3
5,3
6,6
0,4
4,0
3,5
Player
2
c
r
4,0
0,4
3,5
5,3
5,3
6,6
BR1(l)
=
M
BR2(U)
=
l
BR1(c)
=
U
BR2(M)
=
c
BR1(r)
=
D
BR2(D)
=
r
What
is
the
NE?
Why?
Player
2
c
r
0,2
2,3
11,1
3,2
0,3
1,0
4,3
0,0
8,0
alpha
beta
beta
0,0
3,-1
-1,3
1,1
Player 1
l
r
1,1
0,0
0,0
0,0
Lets
check:
If
everyone
invests,
none
would
have
regrets,
and
indeed
the
BR
would
be
to
invest
If
nobody
invests,
then
the
BR
would
be
to
not
invest
Recap
(1)
Weve
seen
the
investment
game,
which
is
one
instance
of
a
coordina?on
game
Lesson
1:
communica?on
can
help
CoordinaLon
games
are
very
dierent
from
Prisoners
Dilemma
games
Do
you
remember
why?
Recap
(2)
In
the
Prisoners
Dilemma
game
communicaLon
cannot
help
one
strategy
is
dominated
nobody
can
oblige
you
to
play
a
dominated
strategy
Recap
(3)
Player
2
Player 1
U
D
1,1
0,0
0,0
1,1
Recap
(4)
We
introduced
the
noLon
of
strategic
complements
Investment
game:
the
more
people
invest
the
more
likely
you
are
to
invest
Partnership
game:
the
more
the
other
person
does,
the
more
likely
for
me
to
do
more
s2
BR1(s2)
4
3
BR2(s1)
2
1
s1
SW
Bourne UlLmatum
Snow White
Player 1
SW
BU
GS
SW
Player 1
Player 1
BU
GS
BU
GS
2,1
0,0
0,0
1,2
Player 1
BU
GS
BU
GS
2,1
0,0
0,0
1,2
Player 1
BU
GS
BU
GS
2,1
0,0
0,0
1,2
Player 2
Player 1
U
D
1,1
0,0
0,0
1,1
CoordinaLon
Games
Pure
coordinaLon
games:
there
is
no
conict
whether
one
NE
is
befer
than
the
other
E.g.:
in
the
investment
game,
we
all
agreed
that
the
NE
with
everyone
invesLng
was
a
befer
NE
Slope:
-b
a
q1 + q2
u1(q1,q2) = a * q1 b * q21 b * q1 q2 c * q1
q2
q1
What
if
Firm
2
didnt
produce
at
all?
q2
=
0
What
is
the
best
response
for
Firm
1
q2
q1
MONOPOLY:
When
marginal
revenue
=
Marginal
cost
Demand
curve
Slope:
-b
Marginal
revenue
Slope:
-2b
Marginal
cost
q1
MONOPOLY
q2
Lets
nd
this
point
with
some
math
0
How
much
quanLty
should
Firm
2
produce
for
Firm
1
not
to
produce
at
all?
q2
such
as
q1
would
be
the
best
response
for
Firm
1
q1
q2
q1
PERFECT
COMPETITION:
When
demand
=
Marginal
cost
Demand
curve
Slope:
-b
Marginal
revenue
Slope:
-2b
Marginal
cost
q1
MONOPOLY
PERFECT
COMPETITION
PERFECT
COMPETITION:
When
demand
=
Marginal
cost
Demand
curve
Slope:
-b
Marginal cost
q1
If
Firm
1
would
produce
more,
the
selling
price
would
not
cover
her
costs
BR1
Monopoly
BR2
0
Perfect
compeLLon
q1
q2
BR1
Monopoly
NE
BR2
Perfect
compeLLon
q1
Cournot
Duopoly:
what
about
the
market?
Lets
now
take
the
perspecLve
of
the
market
and
not
of
a
single
player
What
about
the
total
industry
prots?
Are
they
maximized?
Where
on
the
plot
we
drew
before,
are
industry
prots
maximized?
q2
BR1
Industry
prots
are
maximized
BR2
0
q1
q2
BR1
Both
rms
produce
half
of
the
monopoly
quanLty
BR2
0
q1
Perfect
CompeLLon
Cournot
QuanLty
Cournot
QuanLty
Monopoly
Monopoly
QUANTITIES
PRICES
Recap
(1)
We
started
to
study
imperfect
compe++on
We
used
the
Cournot
Model
What
happens
between
monopoly
and
perfect
compeBBon?
Recap
(2)
This
is
the
general
approach
you
use
when
modeling
a
problem
with
Game
Theory
You
formally
solve
the
problem,
with
handy
mathemaBcal
tools
You
graphically
solve
the
problem
to
gain
insights
and
intuiBon
to
the
problem
You
translate
your
ndings
in
the
real
world
and
gure
out
if
they
make
sense
Recap
(3)
In
Cournot
Equilibrium
we
have:
Saddle
points
sit
in
between
monopoly
and
perfect
compeBBon
QuanBty
produced
is
less
than
perfect
compeBBon
but
more
than
monopoly
Industry
prot
is
less
than
monopoly
but
larger
than
perfect
compeBBon
Bertrand
CompeBBon
What
if
we
take
a
somehow
more
realisBc
view
and
let
Firms
decide
on
prices
instead
of
quanBBes?
Companies
compete
on
prices
and
let
quanBBes
seTle
as
a
consequence
of
imperfect
compeBBon
Lets
dene
the
game
Revenues
Costs
Firm
2
y
GeneralizaBon:
The
linear
city
model
can
be
used
to
think
about
a
dimension
of
the
product
(think
of
beer,
Bud
light
vs.
Guinness)
Right Wing
Payos:
Prize
if
win
=
B
Cost
for
running
=
C,
with
B
=
2
C
DisuBlity
(cost)
if
winning
candidate
is
far
from
your
poliBcal
posiBon
=
|x-y|
R
Who
is
going
to
win
the
elecBon?
Is
this
a
Nash
Equilibrium?
Y
runs
payo
=
-
$1000
-
$240
Y
stays
put
payo
=
-$240
A
player
can
deviate
by
staying
put
R
Who
is
going
to
win
the
elecBon?
Is
this
a
Nash
Equilibrium?
Y
runs
payo
=
-
$1000
-
$60
Y
stays
put
payo
=
-$60
A
player
can
deviate
by
staying
put
Is
this
an
equilibrium?
What
happens
if
someone
else
runs?
Who
is
the
winner
in
case
orange
player
runs?
DeviaBon
3:
X
drops
X
runs,
then
expected
payo
is
50%
B-C
,
50%
-C
-
|x-y|
X
drops,
then
payo
is
C
-
|x-y|
for
sure!!
1/6
2/6
1/2
4/6
5/6
1/2
70
140
#
of
your
type
in
your
town
Tall
player
Short
player
West
Town
East
Town
Tall
player
Short
player
West
Town
East
Town
First
itera6on
For
tall
players
Theres
a
minority
of
east
town
giants
to
begin
with
For
short
players
Theres
a
minority
of
west
town
dwarfs
to
begin
with
short
players
will
want
to
switch
towns
(WE)
tall
players
will
want
to
switch
towns
(EW)
Tall
player
Short
player
West
Town
East
Town
Second
itera6on
We
keep
the
same
trend
short
players
will
want
to
switch
towns
(WE)
tall
players
will
want
to
switch
towns
(EW)
Tall
player
Short
player
West
Town
East
Town
Third
itera6on
What
happened?
People
got
segregated
Warning
Note
how
a
Bny
modeling
detail
ended
up
being
very
important
We
added
randomizaBon
to
add
things
up
correctly
It
turned
out
to
yield
an
equilibrium
in
the
game
Mixed
Strategies
The
pure
strategies
are
embedded
in
our
mixed
strategies
QuesBon:
What
are
the
payos
from
playing
mixed
strategies?
In
parBcular,
what
is
the
expected
payo?
A
B
Player 2
2,1
0,0
0,0
1,2
1/5
4/5
Example (2)
Example (3)
Example
(4)
Lets
focus
on
player
1s
expected
payo
3/5
Obviously
we
have:
The
weighted
average
must
lie
between
the
two
pure
strategies
expected
payos
ObservaBon
The
expected
payo
from
mixed
strategies
must
lie
between
the
pure
strategies
expected
payos
This
simple
observaBon
turns
out
to
be
the
key
to
compute
mixed
strategies
NE
If
a
mixed
strategy
is
a
best
response
then
each
of
the
pure
strategies
in
the
mix
must
be
best
responses
They
must
yield
the
same
expected
payo
ObservaBon
Our
informal
lesson
in
red
before
implies
that
VENUS
RIGHT
LEFT
SERENA
right
leq
Viewpoint
L
R
Serena
50,50
80,20
90,10
20,80
q
p
(1-p)
(1-q)
L
R
Serena
30,70
80,20
90,10
20,80
q
p
(1-p)
(1-q)
VENUS
RIGHT
LEFT
SERENA
right
leN
Viewpoint
L
R
Serena
50,50
80,20
90,10
20,80
q
p
(1-p)
(1-q)
Discussion
(1)
Can
anybody
suggest
some
other
places
where
we
see
randomiza:on
or
at
least
mixed
strategy
equilibria
in
spor:ng
events?
Pick
your
own!!
Discussion
(2)
Since
were
in
a
mixed
strategy
equilibrium,
it
must
be
the
case
that
the
payos
are
equal
Indeed,
if
it
was
not
the
case,
then
you
shouldnt
be
randomizing!!
Discussion
(3)
ANer
the
security
problems
at
U.S.
and
worldwide
airports
due
to
high
risks
of
ajacks,
the
need
for
devices
capable
of
inspec:ng
luggage
has
raised
considerably
The
problem
is
that
there
are
not
enough
of
such
machines
Wrong
statements
have
been
promoted
by
local
governments:
Discussion
(4)
The
problem
with
that
line
of
reasoning
was
that
the
concept
of
mixed
strategy
was
not
adopted
What
if
you
wouldnt
no:fy
where
you
would
actually
put
the
check
devices,
which
boils
down
to
randomizing?
The
hard
thing
to
do
in
prac:ce
is
how
to
mimic
randomiza:on!!
INTERPRETATIONS
TO
MIXED
STRATEGIES
A
B
Player 2
2,1
0,0
0,0
1,2
1-q
p
(1-p)
Any volunteer?
Player 2 perspec:ve, nd NE p:
p 1-p
Player 2
1-q
Player 2
Mixed
Strategies:
Interpreta:on
#2
Rather
than
thinking
of
players
actually
randomizing
over
their
strategies,
we
can
think
of
them
holding
beliefs
of
what
the
other
players
would
play
What
weve
done
so
far
is
to
nd
those
beliefs
that
make
players
indierent
over
what
they
play
since
theyre
going
to
obtain
the
same
payos
Mixed
Strategies:
Interpreta:on
#3
We
could
actually
think
in
terms
of
frac<on
of
a
popula<on
when
we
discussed
mixed
strategies
Lets
mo:vate
this
line
of
thinking
through
an
example/game
Auditor
A
N
2,0
4,-10
4,0
0,4
q
p
(1-p)
1-q
Lets
focus
on
a
simultaneous
move
game
(despite
in
this
case
its
not
realis:c)
The
auditor
can
decide
to
audit
or
not
a
tax
payer
The
tax
payer
can
decide
to
be
honest
or
to
cheat
in
declaring
income
tax
Take
a
look
at
the
payos
Auditor
A
N
2,0
4,-10
4,0
0,4
q
p
(1-p)
1-q
Auditor
A
N
2,0
4,-20
4,0
0,4
q
p
(1-p)
1-q
To
sum
up
Lesson
1:
mixed
strategies
can
have
dierent
interpreta:ons
(frac:on
of
popula:on)
Lesson
2:
we
can
verify
a
mixed
strategy
NE
is
eec:vely
one
simply
by
checking
pure-
strategy
devia:ons
Lesson
3:
Row
players
payos
impact
Column
players
mixing
probability
and
vice-versa
Evolu:on
(1)
Concept
related
to
a
specic
branch
of
Biology
Relates
to
the
evolu:on
of
the
spices
in
nature
Powerful
modeling
tool
that
has
received
a
lot
of
ajen:on
lately
by
the
computer
science
community
Why
look
at
evolu:on
in
the
context
of
Game
Theory?
Evolu:on
(2)
Game
Theory
had
a
tremendous
inuence
on
evolu:onary
Biology
Study
animal
behavior
and
use
GT
to
understand
popula:on
dynamics
Idea:
Relate
strategies
to
phenotypes
of
genes
Relate
payos
to
gene:c
tness
Strategies
that
do
well
grow,
those
that
obtain
lower
payos
die
out
Important note:
Evolu:on
(3)
Examples:
Group
of
lions
deciding
whether
to
ajack
in
group
an
antelope
Ants
deciding
to
respond
to
an
ajack
of
a
spider
Mobile
ad
hoc
networks
TCP
varia:ons
P2P
applica:ons
Evolu:on
(4)
Evolu:onary
biology
had
a
great
inuence
on
Game
Theory
Similar
ideas
as
before,
relate
strategies
and
payos
to
genes
and
tness
Example:
Simplifying
assump:ons
When
studying
evolu:on
through
the
lenses
of
GT,
we
need
to
make
some
assump:ons
to
make
our
life
easy
We
will
relax
these
assump:ons
later
on
2. Asexual
reproduc:on
We
assume
no
gene
redistribu:on
Example
(1)
Player
2
Cooperate
Defect
Player
1
C
D
2,2
3,0
0,3
1,1
1-
Example
(2)
Player
strategy
hardwired
C
Spa<al
Game
All
players
are
coopera:ve
and
get
a
payo
of
2
What
happens
with
a
muta:on?
Example
(3)
Player
strategy
hardwired
C
Player
strategy
hardwired
D
Focus
your
ajen:on
on
this
random
tournament:
Coopera:ng
player
will
obtain
a
payo
of
0
Defec:ng
player
will
obtain
a
payo
of
3
Survival
of
the
jest:
D
wins
over
C
Example
(4)
Player
strategy
hardwired
C
Player
strategy
hardwired
D
Example
(5)
Player
strategy
hardwired
C
Player
strategy
hardwired
D
Example
(6)
Player
strategy
hardwired
C
Player
strategy
hardwired
D
A
small
ini:al
muta:on
is
rapidly
expanding
instead
of
dying
out
Lets
now
try
to
be
a
lijle
bit
more
formal
Example
(7)
Player
2
Cooperate
Defect
Player
1
C
D
2,2
3,0
0,3
1,1
1-
1-
Is
coopera:on
ES?
C
vs.
[(1-)C
+
D]
(1-)2
+
0
=
2(1-)
D
vs.
[(1-)C
+
D]
(1-)3
+
1
=
3(1-)+
3(1-)+
>
2(1-)
Example
(8)
Player
2
Cooperate
Defect
Player
1
C
D
2,2
3,0
0,3
1,1
1-
1-
Is
defec:on
ES?
D
vs.
[(1-)D
+
C]
(1-)1
+
3
=
(1-)+3
C
vs.
[(1-)D
+
C]
(1-)0
+
2
=
2
(1-)+3
>
2
Observa:ons
Lesson
1:
Nature
can
suck
Another
e
xample
(
1)
a
b
a
b
c
Another
e
xample
(
2)
a
b
a
b
c
Another
e
xample
(
3)
a
b
a
b
c
Is
(c,c)
a
NE?
No,
because
b
is
a
protable
devia:on
Observa:ons
Lesson
3:
If
s
is
not
Nash
(that
is
(s,s)
is
not
a
NE),
then
s
is
not
evoluHonary
stable
(ES)
a
b
1,1
0,0
Is
b
ES?
b
0
a
(1-)
0
+
1
=
>
0
0,0
0,0
1-
a
b
1,1
0,0
0,0
0,0
1-
Deni:on
2
In
a
symmetric
2
player
game,
the
pure
strategy
is
ES
(in
pure
strategies)
if:
A)
and
B)
Theorem
(1)
Deni:on
1
Deni:on
2
Lets
see
Def.
2
Def.
1
Sketch
of
proof:
Fix
a
a
nd
suppose
(,)
is
NE,
that
is
Theorem
(2)
Case
1:
u( s, s ) > u( s, s) s
the
mutant
dies
out
because
she
meets
oNen
Case
2:
a
b
1,1
1,1
1,1
0,0
1-
Example
con:nued
Player
2
a
b
Player
1
a
b
1,1
1,1
1,1
0,0
1-
2,2
0,0
0,0
1,1
2,2
0,0
0,0
1,1
2,2
0,0
0,0
1,1
2,2
0,0
0,0
1,1
0,0
1,2
2,1
0,0
Biology interpreta:on:
0,0
1,2
2,1
0,0
0,0
1,2
2,1
0,0
0,0
1,2
2,1
0,0
0,0
1,2
2,1
0,0
Deni:on
2bis
In
a
symmetric
2
player
game,
the
mixed
strategy
p
is
ES
(in
mixed
strategies)
if:
A)
and
B)
0,0
1,2
2,1
0,0
0,0
1,2
2,1
0,0
Recap
on
Chickens
It
turns
out
that
in
many
cases
that
arise
in
nature,
the
only
equilibrium
is
a
mixed
equilibrium
But
what
does
it
mean
to
have
a
mix
in
nature?
It
could
mean
that
the
gene
itself
is
randomizing,
which
is
plausible
It
could
be
that
there
are
actually
two
types
surviving
in
the
popula:on,
and
this
is
connected
to
our
alterna:ve
interpreta:on
of
mixed
strategies
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
H is ESS if v c
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
What
if
c
>
v?
We
know
H
is
not
ESS
and
D
is
not
ESS
What
about
a
mixed
strategy?
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
(v-c)/2,
(v-c)/2
0,
v
v,0
v/2,
v/2
One
last
e
xample
(
1)
S
B
S
B
T
Assume
1<v<2
What
is
this
game?
One
last
e
xample
(
2)
S
B
S
B
T
One
last
e
xample
(
3)
S
B
S
B
T
Lets
assume
p
=
S
There
is
no
ESS!!
(0,0)
$0
1
$1
2
$3
2
$1
(1, 1.5)
- $1
(-1, 1)
$3
(3, 2)
- $3
(-3, 3)
Backward
InducUon
Start
with
the
last
player
and
chose
the
strategies
yielding
higher
payo
This
simplies
the
tree
ConUnue
with
the
before-last
player
and
do
the
same
thing
Repeat
unUl
you
get
to
the
root
This
is
a
fundamental
concept
in
game
theory
(0,0)
$0
1
$1
2
$3
2
$1
(1, 1.5)
- $1
(-1, 1)
$3
(3, 2)
- $3
(-3, 3)
(0,0)
$0
1
$1 2
(1, 1.5)
$3
2
(-3, 3)
(0,0)
$0
1
$1
2
$3
2
$1
(1, 1.5)
- $1
(-1, 1)
$3
(3, 2)
- $3
(-3, 3)
(0,0)
$0
1
$1
2
$3
2
$1
(1, 1.5)
- $1
(-1, 1)
$3
(3, 2)
- $3
(-3, 3)
Moral
Hazard
One
player
(the
borrower)
has
incenUves
to
do
things
that
are
not
in
the
interests
of
the
other
player
(the
lender)
By
giving
a
too
big
loan,
the
incenUves
for
the
borrower
will
be
such
that
they
will
not
be
aligned
with
the
incenUves
on
the
lender
NoUce
that
moral
hazard
has
also
disadvantages
for
the
borrower
Introduce
laws
Similarly
to
what
we
discussed
for
the
PD
Today
we
have
such
laws:
bankruptcy
laws
But,
there
are
limits
to
the
degree
to
which
borrowers
can
be
punished
The
lender
can
say:
I
cant
repay,
Im
bankrupt
And
he/shes
more
or
less
allowed
to
have
a
fresh
start
Limits/restricUons
on
money
Another
way
could
be
to
asking
the
borrowers
a
concrete
plan
(business
plan)
on
how
he/she
will
spend
the
money
This
boils
down
to
changing
the
order
of
play!
But,
whats
the
problem
here?
Lack
of
exibility,
which
is
the
moUvaUon
to
be
an
entrepreneur
in
the
rst
place!
Problem
of
Uming:
it
is
someUmes
hard
to
predict
up-front
all
the
expenses
of
a
project
(0,0)
$0
1
$1
2
$3
2
$1
(1, 1.5)
- $1
(-1, 1)
$3
(1.9, 3.1)
- $3
(-3, 3)
Examples
of
incenUves
IncenUves
in
contracts
for
CEOs
Bad
interpretaUon,
they
screw
up
the
world
Mild
interpretaUon,
they
align
CEOs
acUons
towards
the
interests
of
the
shareholders
Beyond
incenUves
Can
we
do
any
other
things
rather
than
providing
incenUves?
Ever
heard
of
collateral?
Example:
subtract
house
from
run
away
payos
Lowers
the
payos
to
borrower
at
some
tree
points,
yet
makes
the
borrower
befer
o!
Collateral
example
The
borrower
could
re-design
the
payos
of
the
game
in
case
the
project
is
successful
2
(0,0)
$0
1
$1 2
$1
-
$1
$3
2
(1,
1.5)
(-1,
1
-
HOUSE)
$3
(3,2)
- $3
(-3, 3 - HOUSE)
Collaterals
They
do
hurt
a
player
enough
to
change
his/
her
behavior
Lowering
the
payos
at
certain
points
of
the
game,
does
not
mean
that
a
player
will
be
worse
o!!
Collaterals
are
part
of
a
larger
branch
called
commitment
strategies
Next,
an
example
of
commitment
strategies
N
invade
ght
ght
run
run
N
ght
run
(0,0)
(1,2)
(2,1)
(1,2)
N
invade
ght
ght
run
run
N
ght
run
(0,0)
(1,2)
(2,1)
(1,2)
N
invade
ght
run
(1,2)
N
(2,1)
N
invade
ght
ght
run
run
N
ght
run
(0,0)
(1,2)
(2,1)
(1,2)
ght
run
run
N ght
Not
burn
boats
run
Burn
boats
S
ght
ght
run
N ght
N ght
(0,0)
(1,2)
(2,1)
(1,2)
(0,0)
(2,1)
ght
run
run
N ght
(1,2)
(2,1)
Not
burn
boats
Burn
boats
S
ght
run
N ght
N ght
(0,0)
(2,1)
(1,2)
Not
burn
boats
Burn
boats
S
(2,1)
ght
run
run
N ght
Not
burn
boats
run
Burn
boats
S
ght
ght
run
N ght
N ght
(0,0)
(1,2)
(2,1)
(1,2)
(0,0)
(2,1)
Lesson
learned
SomeUmes,
gegng
rid
of
choices
can
make
me
befer
o!
Commitment:
Slope:
-b
a
q1 + q2
u1(q1,q2) = a * q1 b * q21 b * q1 q2 c * q1
q2
BR1
Monopoly
NE
BR2
Perfect
compeUUon
q1
q2
q2
BR2
q2
0
q1
q1
q1
q2
q2
BR2
q2
0
q1
q1
q1
ObservaUons
(1)
Is
what
weve
looked
at
really
a
sequenUal
game?
Despite
we
said
rm
1
was
going
to
move
rst,
theres
no
reason
to
assume
shes
really
going
to
do
so!
What
do
we
miss?
ObservaUons
(2)
We
need
a
commitment
In
this
example,
sunk
cost
could
help
in
believing
rm
1
will
actually
play
rst
Assume
rm
1
was
going
to
invest
a
lot
of
money
in
building
a
plant
to
support
a
large
producUon:
this
would
be
a
credible
commitment!
ObservaUons
(3)
Lets
make
an
example:
assume
the
two
rms
are
NBC
and
Murdoch
trying
to
gain
market
shares
for
newspapers
producUon
in
a
city
Suppose
theres
a
board
meeUng
where
the
strategy
of
the
rms
are
decided
What
could
Murdoch
do
to
deviate
from
Cournot?
ObservaUons
(4)
An
example
would
be
to
be
somehow
dishonest
and
hire
a
spy
to
gain
more
informaUon
on
NBCs
strategy!
To
make
the
scenario
even
more
intriguing,
lets
assume
NBC
knows
that
theres
a
spy
in
the
board
room
What
should
NBC
do?
Examples
of
games
NIM,
that
we
played
earlier
Tic-tac-toe:
Two
players
Perfect
informaUon
Finite
Three
outcomes
1
W1
L1
L1
L1
L1
L1
W1
L1
T
L1
1
T
L1
Example
of
a
more
complex
game
1
1
1
2
1
What
is
the
maximum
length
of
the
game?
2
1
1
1
2
1
W1
L1
Example
(1)
Strategies
1
2
1
U
D
l
r
(1,0)
(2,4)
(3,1)
(0,2)
Player
2:
[l],
[r]
Player
1:
[U,u],
[U,d]
[D,
u],
[D,d]
Example
(2)
Note:
1
2
1
U
D
l
r
(1,0)
(2,4)
(3,1)
(0,2)
In
this
game
it
appears
that
player
2
may
never
have
the
possibility
to
play
her
strategies
This
is
also
true
for
player
1!
Example
(3)
Backward
InducUon
1
2
1
U
D
l
r
(2,4)
(3,1)
(0,2)
(1,0)
Summarize
game
r
higher
payo
Summarize
game
D
higher
payo
BI :: {[D,d],r}
Example
(4)
l
1
2
1
U
D
l
r
(1,0)
u
d
(0,2)
(2,4)
(3,1)
U u
2,4
0,2
U d
3,1
0,2
D u
1,0
1,0
D d
1,0
1,0
Example
(4)
l
1
2
1
U
D
l
r
u
d
(2,4)
(3,1)
(0,2)
(1,0)
U u
2,4
0,2
U d
3,1
0,2
D u
1,0
1,0
D d
1,0
1,0
Backward Induc9on
Nash Equilibrium
{[D, d],r}
{[D,
d],r}
{[D,
u],r}
NF
(0,3)
(-1,0)
(1,1)
NF
OUT
NF
IN
-1,0
1,1
OUT
0,3
0,3
(-1,0)
(1,1)
(0,3)
Backward Induc9on
Nash Equilibrium
(IN, NF)
(IN,
NF)
(OUT,
F)
NF
(0,3)
(-1,0)
(1,1)
NF
OUT
NF
IN
-1,0
1,1
OUT
0,3
0,3
(-1,0)
(1,1)
(0,3)
Backward Induc;on
Nash Equilibrium
(IN, NF)
(IN,
NF)
(OUT,
F)
REPUTATION
ReputaKon
(1)
To
make
our
intuiKon
work,
lets
try
to
introduce
a
new
idea
Assume
theres
a
small
chance
(say
1%),
that
the
monopolist
rm
is
crazy
This
implies
that
the
payos
are
not
exactly
the
same
as
in
the
single
market
game
we
looked
at
Its
like
the
monopolist,
someKmes,
actually
prefers
to
ght
just
for
the
fun
of
it
ReputaKon
(2)
Now,
lets
look
at
the
1st
entrant:
she
knows
theres
a
1%
possibility
that
the
monopolist
will
go
bonkers
and
ght
If
there
was
only
one
market,
we
would
be
done:
the
entrant
should
step
in
When
there
are
10
markets,
things
are
dierent.
Lets
see
why
ReputaKon
(3)
Assume
that
the
1st
entrant
thinks
that
with
0.9
probability
the
monopolist
will
not
ght
and
she
enters
What
if
the
monopolist
goes
crazy
and
ght?
What
happens
to
the
other
9
entrants?
Subsequent
entrants
would
modify
their
beliefs
and
assume
a
higher
probability
for
the
monopolist
to
go
bonkers
ReputaKon
(4)
Entrants
would
start
believe
that
the
monopolist
is
indeed
crazy
and
step
out
of
the
markets!
The
small
possibility
that
the
monopolist
would
be
crazy
allowed
to
build
a
reputaBon
that
keeps
entrants
out
of
the
markets
ReputaKon
(5)
Now,
this
argument
can
be
strengthened
Assume
that,
in
fact,
the
monopolist
is
not
crazy
By
acKng
as
if
she
was
crazy
early
on
in
the
game,
she
was
able
to
scare
the
entrants
In
game
theory,
we
always
assume
players
know
how
to
put
themselves
in
others
shoes
Hence,
the
entrants
should
know
that
theres
a
possibility
that
the
monopolist
is
not
crazy
but
shes
acKng
out
ReputaKon
(6)
Entrants
stay
out
not
only
because
they
think
the
monopolist
is
crazy
They
stay
out
also
because
they
think
that
even
if
the
monopolist
was
not
crazy,
she
would
ght
in
any
case!
We
just
saw
that
irrespecKvely
of
the
monopolist
being
crazy
or
not,
she
will
act
like
crazy.
Hence,
a
raKonal
entrant
would
know
that
and
would
not
update
her
beliefs
entrants
learn
nothing
by
observaKon
Ques;on:
Can
this
be
an
equilibrium?
ReputaKon
(7)
To
answer
the
quesKon,
lets
use
backward
inducKon
and
go
to
the
10th
market
The
10th
entrant,
learns
nothing
about
the
monopolist
being
crazy
or
not
As
such,
she
would
sKll
believe
the
monopolist
could
be
sane
with
.99
probability
The
10th
entrant
would
step
in,
and
this
would
unravel
back
to
the
1st
market
The
answer
to
the
quesKon
involves
mixed
strategies:
a
monopolist
should
(with
some
probability)
act
like
crazy,
build
a
reputaKon
and
keep
entrants
out
of
the
markets
DUELS
Pi (d) 1 P j (d 1) Pi (d) + P j (d 1) 1
If
P
2
(2)
+
P
1
(1)
1
then
player
1
should
shoot
BARGAINING GAMES
If
player
2
accepts:
Player
1
gets
S,
player
2
gets
1-S
If
player
2
declines:
Player
1
and
player
2
get
nothing
If
player
2
accepts:
Player
1
gets
S1,
player
2
gets
S2=1-S1
If
player
2
declines:
We
ip
roles
and
play
the
game
again
This
is
the
second
stage
of
the
game
Example:
If
I
give
you
$1
today,
thats
what
you
get
If
I
give
you
$1
in
1
month,
we
assume
its
worth
less,
say
< 1
<1
1-period
1-period
2-period
<1
Lets
be
careful:
In
the
second
round
of
the
two-period
game,
player
2
makes
the
oer
about
the
whole
pie
We
know
that
this
is
going
to
be
an
ulKmatum
game,
so
player
2
will
keep
the
whole
pie
and
player
1
will
accept
(by
BI)
However,
seen
from
the
rst
round,
the
pie
in
the
second
round
that
player
2
could
get,
is
worth
less
than
$1
1-period
1
1
2-period
3-period
1 (1 )
Receivers split
0
<1
(1 )
NOTE: in the table, we report the split player 1 should oer in the rst
Oerer
Receiver
1
0
1-period
1
2-period
3-period
1 (1 )
4-period
?
<1
(1 )
NOTE:
give
people
just
enough
today
so
theyll
accept
the
oer,
and
just
enough
today
get
tomorrow
discounted
by
delta
is
whatever
they
You
dont
need
to
go
back
all
the
way
up
to
period
1
1-period
2-period
3-period
4-period
Oerer
1
Receiver
0
1 +2
1 +2 3
2 +3
S1(10) = 1 + 2 3 + 4 + ... 9
= 1 + + + ...
S1(10) = (1 + 2 3 + 4 + ... 9 )
(10)
1
(1+ )S
=1
10
(10)
1
1
=
1+
10
()
1
S2()
1
1
=
=
1+
1+
()
= 1 S1 =
=
1+
1+
S2()
1
1
1
=
1+
2
1
1
=
1+
2
IMPERFECT INFORMATION
IntroducKon
We
have
seen
simultaneous
move
games,
in
which
players
cant
observe
strategies
and
have
to
reason
based
on
the
idea
of
best
response
We
have
seen
sequenBal
move
games,
in
which
observaBon
is
allowed,
and
players
reason
using
backward
inducKon
Now,
lets
study
a
class
of
games
in
which
these
two
approaches
are
blended
Example
(1)
2
U
1
M 2
(4,0)
(0,4)
(0,4)
u
d
D
2
(4,0)
(1,2)
(0,0)
Example
(2)
2
U
1
M
2
D
2
(4,0)
d
u
(0,4)
(0,4)
(4,0)
(1,2)
(0,0)
If
player
1
chooses
D,
player
2
can
observe
it
If
player
1
chooses
U
or
M,
player
2
doesnt
know
which
choice
was
made
Example
(3)
InformaKon
set
U
1
(4,0)
(0,4)
(0,4)
u
d
D
2
(4,0)
(1,2)
(0,0)
Example
(4)
InformaKon
set
U
1
M
D
2
(4,0)
d
u
(0,4)
(0,4)
(4,0)
(1,2)
(0,0)
Example
(1)
U
1
(2,2)
(-1,3)
2
D
(3,-1)
(0,0)
Example
(2)
U
1
(2,2)
(-1,3)
2
D
(3,-1)
(0,0)
Example
(3)
Player
2
l
U
1
(2,2)
(-1,3)
Player
1
D
2
D
(3,-1)
(0,0)
2,2
3,-1
-1,3
0,0
Example
(4)
Player
2
l
U
1
(2,2)
(-1,3)
2
D
U
Player
1
D
2,2
3,-1
-1,3
0,0
(3,-1)
(0,0)
Example
(1)
l
U
1
(a1,a2)
m
(b1,b2)
(c1,c2)
2
D
l
r
(d1,d2)
(e1,e2)
(f1,f2)
As
in
the
previous
example,
player
2
does
not
know
if
player
1
chooses
up
or
down
Player
2
has
just
three
choices
Our
goal
now
is
to
transform
the
game
into
a
matrix
Example
(2)
l
U
1
(a1,a2)
m
(b1,b2)
(c1,c2)
U
D
a1,a2
d1,d2
b1,b2
e1,e2
c1,c2
f1,f2
2
D
l
r
(d1,d2)
(e1,e2)
(f1,f2)
Example
(3)
U
D
a1,a2
d1,d2
b1,b2
e1,e2
c1,c2
f1,f2
l
m
1
(a1,a2)
(d1,d2)
(b1,b2)
U
D
U
(e1,e2)
(c1,c2)
(f1,f2)
ObservaKons
What
mamers
is
not
Bme,
but
informaBon
By
the
end
of
today
we
will
set-up
the
machinery
to
analyze
such
games
and
predict
what
it
is
going
to
happen
Example
(1)
l
U
1
(4,2)
u
(0,0)
1 d
(1,4)
r
2
(0,0)
(2,4)
Example
(2)
l
l
(4,2)
u
Ud
r
1
d
(1,4)
Du
2
D
(0,0)
Uu
Dd
l
(0,0)
(2,4)
4,2
4,2
0,0
0,0
0,0
1,4
2,4
2,4
Example
(3)
Nash
Equilibria:
l
Uu
Ud
Du
Dd
4,2
4,2
0,0
0,0
0,0
1,4
2,4
2,4
(Uu,l)
(Du,r)
(Dd,r)
Example
(4)
Backward Induction
l
U
1
(0,0)
1 d
(1,4)
r
2
(4,2)
sub-game
l
(0,0)
(2,4)
Example
(5)
Nash
Equilibria:
Were
not
saying
these
are
not
NE,
its
just
that
they
are
inconsistent
with
what
we
could
predict
with
BI
We
need
a
new
noKon
of
soluKon,
that
is
able
to
treat
games
that
have
both
sequenKal
moves
and
simultaneous
moves
Example
(1)
(1,0,0)
A
B`
u
2
(0,1,1)
(0,0,2)
3
d
l
(0,0,-1)
(2,1,0)
Example
(2)
Player
1:
A
l
U
2
1,0,0
1,0,0
Player
1:
B
3
1,0,0
1,0,0
l
U
2
D
0,1,1
0,0,-1
0,0,2
2,1,0
Example
(3)
(1,0,0)
A
B`
u
2
(0,1,1)
(0,0,2)
3
Player
3
d
l
r
(0,0,-1)
(2,1,0)
U
Player
2
D
Sub-game
1,1
0,-1
0,2
1,0
Example
(4)
Player
3
U
Player
2
D
1,1
0,-1
0,2
1,0
Deni2on:
sub-games
A
sub-game
is
a
part
of
the
game
that
looks
like
a
game
within
the
tree.
It
saKses
the
three
following
properKes:
1. It
starts
from
a
single
node
2. It
comprises
all
successors
to
that
node
3. It
does
not
break
up
any
informaKon
set
Examples
of
sub-games
2
3
2