Lecture III: Normal Form Games, Rationality

and Iterated Deletion of Dominated Strategies

Markus M. Mobius
February 7, 2003
1 Denition of Normal Form Game
Game theory can be regarded as a multi-agent decision problem. Its useful
to dene rst exactly what we mean by a game.
Every normal form (strategic form) game has the following ingredients.
1. There is a list of players D = {1, 2, .., I}. We mostly consider games
with just two players. As an example consider two people who want to
meet in New York.
2. Each player i can choose actions from a strategy set S
i
. To continue
our example, each of the players has the option to go the Empire State
building or meet at the old oak tree in Central Park (where ever that
is ...). So the strategy sets of both players are S
1
= S
2
= {E, C}.
3. The outcome of the game is dened by the strategy prole which
consists of all strategies chosen by the individual players. For example,
in our game there are four possible outcomes - both players meet at the
Empire state building (E, E), they miscoordinate, (E, C) and (C, E), or
they meet in Central Park (C, C). Mathematically, the set of strategy
proles (or outcomes of the game) is dened as
S = S
1
S
2
In our case, S has order 4. If player 1 can take 5 possible actions, and
player 2 can take 10 possible actions, the set of proles has order 50.
1
Figure 1: General 2 by 2 game
0,0 1,1
1,1 0,0 E
C
E C
4. Players have preferences over the outcomes of the play. You should
realize that players cannot have preferences over the actions. In a game
my payo depends on your action. In our New York game players just
want to be able to meet at the same spot. They dont care if they
meet at the Empire State building or at Central Park. If they choose
E and the other player does so, too, ne! If they choose E but the
other player chooses C, then they are unhappy. So what matters to
players are outcomes, not actions (of course their actions inuence the
outcome - but for each action there might be many possible outcomes -
in our example there are two possible outcomes per action). Recall, that
we can represent preferences over outcomes through a utility function.
Mathematically, preferences over outcomes are dened as:
u
i
: S R
In our example, u
i
= 1 if both agents choose the same action, and 0
otherwise.
All this information can be conveniently expressed in a game matrix as
shown in gure 1:
A more formal denition of a game is given below:
Denition 1 A normal (strategic) form game G consists of
A nite set of agents D = {1, 2, .., I}.
Strategy sets S
1
, S
2
, .., S
I
Payo functions u
i
: S
1
S
2
..S
I
R (i = 1, 2, .., n)
2
Well write S = S
1
S
2
.. S
I
and we call s S a strategy prole
(s = (s
1
, s
2
, .., s
I
)). We denote the strategy choices of all players except
player i with s
i
for (s
1
, s
2
, .., s
i1
, s
i+1
, ..s
I
).
2 Some Important Games
We already discussed coordination games. These are interesting games, be-
cause players have an incentive to work together rather than against each
other. The rst games analyzed by game theorists were just the opposite -
zero sum games, where the sum of agents utilities in each outcome sums up
to zero (or a constant).
2.1 Zero-Sum Games
Zero-sum games are true games of conict. Any gain on my side comes at
the expense of my opponents. Think of dividing up a pie. The size of the
pie doesnt change - its all about redistribution of the pieces between the
players (tax policy is a good example).
The simplest zero sum game is matching pennies. This is a two player
game where player 1 get a Dollar from player 2 if both choose the same
action, and otherwise loses a Dollar:
1,1 1,1
1,1 1,1 H
T
H T
2.2 Battle of the Sexes
This game is interesting because it is a coordination game with some elements
of conict. The idea is that a couple want to spend the evening together. The
wife wants to go to the Opera, while the husband wants to go to a football
game. Each get at least some utility from going together to at least one of
3
the venues, but each wants to go their favorite one (the husband is player 1
- the column player).
0,0 1,2
2,1 0,0 F
O
F O
2.3 Chicken or Hawk versus Dove
This game is an anti-coordination game. The story is that two teenagers
drive home on a narrow road with their bikes, and in opposite directions.
None of them wants to go out of the way - whoever chickens out loses his
pride, while the tough guy wins. But if both stay tough, then they break
their bones. If both go out of the way, none of the them is too happy or
unhappy.
0,10 5,5
1,1 10,0 t
c
t c
2.4 Prisoners Dilemma
This game might be the most famous of all. Its the mother of all cooperation
games. The story is that two prisoners are interrogated. If both cooperate
with the prosecution they get of with 1 year in prison. If both give each
other away (defect) they get 3 years in prison each. If one cooperates and
the other guy defects, then the cooperating guy is thrown into prison for 10
years, and the defecting guy walks free.
4
4,1 0,0
3,3 1,4 C
D
C D
Note, that the best outcome in terms of welfare is if both cooperate. The
outcome (D, D) is worst in welfare terms, and is also Pareto dominated by
(C, C) because both players can do better. So clearly, (D, D) seems to be a
terrible outcome overall.
Some examples of Prisoners dilemmas are the following:
Arms races. Two countries engage in an expensive arms race (corre-
sponds to outcome D,D). They both would like to spend their money
on (say) healthcare, but if one spends the money on healthcare and
the other country engages in arms build-up, the weak country will get
invaded.
Missile defence. The missile defence initiative proposed by the ad-
ministration is interpreted by some observers as a Prisoners dilemma.
Country 1 (the US) can either not build a missile defence system (strat-
egy C) or build one (strategy D). Country 2 (Russia) can either not
build any more missiles (strategy C) or build lots more (strategy D).
If the US does not build a missile system, and Russia does not build
more missiles then both countries are fairly well o. If Russia builds
more missiles and the US has no defence then the US feels very unsafe.
If the US builds a missile shield, and Russia does not missiles then the
US is happy but Russia feels unsafe. If the US builds missile defence
and Russia builds more missiles then they are equally unsafe as in the
(C,C) case, but they are much less well o because they both have to
increase their defence budget.
Driving a big SUV is a Prisoners Dilemma. I want my car to be as
safe as possible and buy an SUV. However, my neighbors who has a
Volkswagen Beetle suddenly is much worse o. If she also buys an SUV
5
2.5 Cournot Competition
This game has an innite strategy space. Two rms choose output levels q
i
and have cost function c
i
(q
i
). The products are undierentiated and market
demand determines a price p (q
1
+q
2
). Note, that this specication assumes
that the products of both rms are perfect substitutes, i.e. they are homoge-
nous products.
D = {1, 2}
S
1
= S
2
= R
+
u
1
(q
1
, q
2
) = q
1
p (q
1
+q
2
) c
1
(q
1
)
u
2
(q
1
, q
2
) = q
2
p (q
1
+q
2
) c
2
(q
2
)
2.6 Bertrand Competition
Bertrand competition is in some ways the opposite of Cournot competition.
Firms compete in a homogenous product market but they set prices. Con-
sumers buy from the lowest cost rm.
Remark 1 It is interesting to compare Bertrand and Cournot competition
with perfect competition analyzed in standard micro theory. Under perfect
competition rms are price takers i.e. they cannot inuence the market. In
this case there is not strategic interaction between rms - each rm solves a
simple prot maximization problem (decision problem). This is of course not
quite true since the auctioneer does determine prices such that demand and
supply equalize.
3 Experiment: Prisoners Dilemma
Students are asked which strategy they would play in the Prisoners dilemma.
The class was roughly divided in half - we calculated the expected payo from
both strategies if people in the class would be randomly matched against each
other. We found that strategy D was better - this is unsurprising as we will
see later since strategy C is strictly dominated by strategy D.
6
4 Experiment: Iterated Deletion Game
Students play the game given below and are asked to choose a strategy for
player 1.
D
C
B
A
A B C D
5,2 2,6 1,4 0,4
0,0 3,2 2,1 1,1
7,0 2,2 1,5 5,1
9,5 1,3 0,2 4,8
In class no student chooses strategy A which is weakly dominated by C. 2
students chose B, 9 students chose C because it looked safe and 16 students
chose D because of the high payos in that row. It turns out that only (B,B)
survives iterated deletion (see below).
5 Iterated Deletion of Dominated Strategies
How do agents play games? We can learn a lot by exploiting the assumption
that players are rational and that each player knows that other players are
rational. Sometimes this reasoning allows us to solve a game.
5.1 Rational Behavior
Assume that agent i has belief
i
about the play of her opponents. A belief
is a probability distribution over the strategy set S
i
. Since we are only
considering pure strategies right now this probability distribution is a point
distribution (i.e. puts all its weight on a single s
i
.
7
Denition 2 Player i is rational with beliefs
i
if
s
i
arg max
s

i
E

i
(s
i
)
u
i
(s

i
, s
i
) ,
or alternatively
s
i
maximizes

s
i
u
i
(s

i
, s
i
)
i
(s
i
) .
Note, that player i faces a simple decision problem as soon as she has
formed her belief
i
.
Denition 3 Strategy s
i
is stricttly dominated for player i if there is some
s

i
S
i
such that
u
i
(s

i
, s
i
) > u
i
(s
i
, s
i
)
for all s
i
S
i
.
Note that the inequality is strict for all s
i
. A strategy is weakly domi-
nated if the inequality is weak for all s
i
and strict for at least one s
i
.
Proposition 1 If player i is rational he will not play a strictly dominated
strategy.
Proof: If strategy s
i
is strictly dominated by strategy s

i
we can deduce that
for any belief of player i we have E

i
(s
i
)
u
i
(s

i
, s
i
) > E

i
(s
i
)
u
i
(s
i
, s
i
).
5.2 Iterated Dominance
The hardest task in solving a game is to determine players beliefs. A lot of
games can be simplied by rationality and the knowledge that my opponent
is rational. To see that look at the Prisoners Dilemma.
Cooperating is a dominated strategy. A rational player would therefore
never cooperate. This solves the game since every player will defect. Notice
that I dont have to know anything about the other player. This prediction
is interesting because it is the worst outcome in terms of joint surplus and
it would be Pareto improving if both players would cooperate. This result
highlights the value of commitment in the Prisoners dilemma - commitment
consists of credibly playing strategy C. For example, in the missile defence
example the ABM treaty (prohibits missile defence) and the START II agree-
ment (prohibits building of new missiles) eectively restrict both countrys
strategy sets to strategy C.
Now look at the next game.
8
D
U
L M R
2,2 1,1 4,0
1,2 4,1 3,5
1. If the column player is rational he shouldnt play M
2. Row player should realize this if he know that the other player is ra-
tional. Thus he wont play D.
3. Column player should realize that R knows that C is rational. If he
knows that R is rational he knows that R wont play D. Hence he wont
play R. This leaves (U,L) as only outcome for rational players.
Its worth while to discuss the level of knowledge required by players.
R has to know that C is rational. C has to know that R knows that C is
rational. This latter knowledge is a higher order form of knowledge. Its
not enough to know that my opponent is rational - I also have to be sure
that my opponent knows that I am rational. There are even higher order
types of knowledge. I might know that my opponent is rational and that he
knows that I am. But maybe he doesnt know that I know that he knows.
The higher the order of knowledge the more often the process of elimi-
nation can be repeated. For example, the game in section 4 of our second
experiment can be solved by the iterated deletion of dominated strategies.
An important concept in game theory is common knowledge (see next
lecture). We will assume throughout the course that rationality is common
knowledge between both players. Therefore, the iteration process can be
repeated arbitrarily often. However, the experiment showed that this as-
sumption might be too strong.
5.3 Formal Denition Of Iterated Dominance
Step I: Dene S
0
i
= S
i
9
Step II: Dene
S
1
i
=
_
s
i
S
0
i

i
S
0
i
u
i
(s

i
, s
i
) > u
i
(s
i
, s
i
) s
i
S
0
i
_
Step k+1: Dene
S
k+1
i
=
_
s
i
S
k
i

i
S
k
i
u
i
(s

i
, s
i
) > u
i
(s
i
, s
i
) s
i
S
k
i
_
S
k+1
i
is the set still not strictly dominated when you know your oppo-
nent uses some strategy in S
k
i
.
Note restrictions S
0
i
, S
1
i
, ..
Players know that opponents are rational, know that opponents know
that they are rational ..., e.g. rationality is common knowledge.
Step : Let S

i
=

k=1
S
k
i
.
Note, that the process must stop after nitely many steps if the strategy
set is nite because the sets can only get smaller after each iteration.
Denition 4 G is solvable by pure strategy iterated strict dominance if S

contains a single strategy prole.

Most games are not dominance solvable (coordination game, zero sum
game).
We have not specied the order in which strategies are eliminated. You
will show in the problem set that the speed and order of elimination does
not matter.
The same is not true for the elimination of weakly dominated strategies
as the next example shows.
B
M
T
L R
1,1 0,0
1,1 2,1
0,0 2,1
10
We can rst eliminate T and then L in which case we know that (2,1) is
played for sure. However, if we eliminate B rst and then R we know that
(1,1) is being played for sure. So weak elimination does not deliver consistent
results and is therefore generally a less attractive option than the deletion of
strictly dominated strategies.
6 Example: Cournot Competition
Remark 2 For the mathematically inclined: With nite strategy sets the set
S

is clearly non-empty because after each stage there must be some domi-
nant strategy left. For innite strategy sets this is not as obvious. However,
one can show that for compact strategy sets each nested subset S
k
i
is closed
and non-empty. Therefore the intersection of all nested subsets cannot be
empty.
Cournot competition with two rms can be solved by iterated deletion
in some cases. Specically, we look at a linear demand function p =
(q
i
+q
j
) and constant marginal cost c such that the total cost of producing
q
i
units is cq
i
. It will be usefull to calculate the best-response function
BR(q
j
) of each rm i to the quantity choice q
j
of the other rm. By taking
the rst-order condition of the prot function you can easily show that the
best-response function for both rms (there is symmetry!) is
BR
i
(q
j
) =
_
c
2

q
j
2
if q
j

c

0 otherwise
The best-response function is decreasing in my belief of the other rms
action. Note, that for q
j
>
c

rm i makes negative prots even if it

chooses the prot maximizing output. It therefore is better o to stay out of
the market and choose q
i
= 0.
Initially, rms can set any quantity, i.e. S
0
1
= S
0
2
=
+
. However, the
best-responses of each rm to any belief has to lie in the interval
_
q, q

with
q = 0 and q =
c
2
. All other strategies make negative prots, are therefore
dominated by some strategy inside this interval, and eliminated.
In the second stage only the strategies S
2
1
= S
2
2
= [BR
1
(q) , BR
1
_
q
_
]
survive, and in the third stage S
3
1
= S
3
2
= [BR
2
_
BR
1
_
q
__
, BR
2
(BR
1
(q))]
(note, that the BR function is decreasing!).
11
Therefore in the 2k + 1th stage only strategies in S
2k+1
1
= S
2k+1
2
=
[BR
2
_
..BR
1
_
q
__
, BR
2
(..BR
1
(q))] survive.
Its easy to show graphically that this interval shrinks in each iteration
and that the two limits converge to the intersection q

1
= q

2
of both best
response functions where q

2
= BR
2
(q

1
). Therefore, the Cournot game is
solvable through the iterated deletion of strictly dominated strategies.
Remark 3 It can be shown that the same game with three rms is NOT
dominance solvable. You have to show that on the problem set!
12