Beruflich Dokumente
Kultur Dokumente
Introduction
We will first look at situations where players or agents -more than one-
choose actions simultaneously ( a term to be made precise soon). Importantly,
the utility or payoff to any player is not determined just by that players
choice of action - it depends on other players choice of action. So, there
is strategic interaction, and each player has to form conjectures about what
other players will choose to play. Complete Information implies that every
player knows the entire structure of payoffs, not just her own, but also that
of others- otherwise conjectures about other players actions are harder to
make.1
A convenient representation of strategic possibilities is the normal or
strategic form.
A game in strategic form is (N, {Ai , }iN , u)
(i) N = {1, 2, . . . , n} is the finite set of players.
(ii) For each i N , Ai is the pure strategy set or action set of player i.
Each strategy is a description of the action that player wants to take.
1
The set up is the same. But, the highest bidder pays the bid announced
by the second-highest bidder.
2
Example 4 (Prisoners Dilemma)
C D
C (1, 1) (1, 2)
D (2, 1) (0, 0)
Example 5:
L R
U (2, 1) (1, 0)
M (0, 0) (0, 2)
D (1, 1) (2, 0)
H T
H (1, 1) (1, 1)
T (1, 1) (1, 1)
i can guarantee himself only a payoff of 1.
But, what if he keeps opponent guessing?
Such games provide a rationale for randomization or using mixed strate-
gies.
3
Mixed Strategies
A mixed strategy is a probability distribution over the set of pure strate-
gies.
For simplicity, let us assume that each Ai is finite.
Let (Ai ) i denote the set of probability distributions over Ai , and
iN i .
i (ai ) is the probability with which ai is played.
Player is randomization is statistically independent of those of other
players.
The payoffs to a profile of mixed strategies are the expected values of the
corresponding pure-strategy profiles.
X n
Y
ui () = j (aj ) ui (a) (1)
sS j=1
Solution Concepts
How should rational players play a game? Rationality is equivalent to
saying that players choose actions in order to maximise their own (expected)
payoffs given the prevailing information structure. So, as outside observers
can we predict the outcome(s) that are likely to arise in a given game? This
requires us to specify the likely choice of actions of rational players in a game
and leads to the notion of a solution concept. Of course, an appropriate
solution concept will depend on assumptions made about the information
structure.
Consider the Prisoners Dilemma game - it is in the interest of each player
to play D. But, (C,C) gives strictly higher payoffs to both players. Why cant
the 2 players cooperate and decide to play (C,C)? This is because players
cannot sign binding contracts. We are in a non-cooperative world in which
any contract has to be self-enforcing - no one should be to deviate unilaterally
from the equilibrium play and become better off.2
2
This is also why splitting up the monopoly profit equally between two Cournot
duopolists cannot be a Cournot equilibrium.
4
Domination
ui (ai , ai ) ui (a0i , ai )
The strategy ai is weakly dominated if there exists i0 such that (2) holds
with weak inequality, and the inequality is strict for at least one ai .
A pure strategy can be strictly dominated by a mixed strategy even it
is not dominated by any pure strategy in the support of the mixed strategy.
(Check Example 5).
Claim: A rational player can never play a strictly dominant strategy
in any self-enforcing equilibrium. (Check that this assertion makes sense.)
Correspondingly, a rational player should play a strictly dominant strat-
egy if it exists.
Question : Can there be more than one strictly dominant strategy?
But, most interesting games of economic interest do not have strictly
dominant strategies. However, there are games which do possess weakly
dominant strategies. Here are a couple of examples.
5
Sealed Bid Second -price auction
vi bj < 0
ui (vi , bi ) = vi bj > 0
6
Provision of Discrete Public Good
Agent i values a discrete public good at vi . This is her net willingness
to pay for the public good. Note that vi can be negative since cost shares
may be incorporated into the description of the project.3
Efficiency : supply iff iN vi 0.
P
3
That is, the government asks the following question. If we build the bridge and
ask you to pay a sum of xi if the bridge is built, how much (in money terms) additional
utility do you get?
7
Iterated elimination of strictly dominated strategies
It is reasonable to assert players will use dominant strategies if they
exist - only need to know strategy sets, their own payoffs.
Also, it is reasonable that players will not use dominated strategies.
Suppose more information is available - the entire payoff matrix is
known, and there is common knowledge of rationality
If ai is strictly dominated, then j should know that i will not play ai .
This might result in some of js strategies which were not strictly dom-
inated earlier being dominated once ai is eliminated.
This opens up the possibility of iteratively eliminating dominated strate-
gies.
8
Example : The Cournot Duopoly
Consider 2 duopolists producing a homogeneous good, with constant
marginal cost of production
ci (qi ) = 10qi
q1 22.5
So, the second round of elimination implies that qi [22.5, 45]. This process
can be continued. Using the fact that the 2 firms are symmetric, the process
converges to an interval [qmin , qmax ] satisfying
90 qmax
qmin =
2
90 qmin
qmax =
2
The only solution to these is
q = 30 (4)
9
There could be a problem with iterated elimination of weakly dominated
strategies - order of deletion may matter, as shown in the enxt example.
Example 7:
L R
U (1, 1) (0, 0)
M (1, 1) (2, 1)
D (0, 0) (2, 1)
10
The Notion of Best Response
The fundamental difference between a one-person decision problem and
game theory is the strategic interdependence between players. In a decision
problem, knowledge of the decision problem allows the agent to calculate
optimal decisions. This is not gnerally true in game theory. Optimal decisions
depend on the structure of the game as well as the choice of action of other
players unless players have dominant strategies.
This is well-illustrated in the game below. (Can you explain why?)
Example 8: (Battle of the Sexes)
O F
O (2, 1) (0, 0)
F (0, 0) (1, 2)
Here, R wants C to play O and then he will also play O. (On the other
hand, C wants both players to play to F.) But, R will play O only if he
believes that C will play O. In other words, optimal actions must depend
upon beliefs about others choice of actions.
Definition 4 A belief of player i about other players actions is a probability
distribution i over Ai .
That is, i (ai ) is the probability with which i believes that others will play
ai .
Notice that this definition allows for the possibility that player 1 may
have beliefs which dont pin down the others strategies. For example, R may
believe that C will play both F and O with probability half.
Given a particular belief about other players actions, player i should then
play a best response. That is, if is belief is represented by i , then i should
play the optimal action given this belief. So, if R believes that C is playing
F and O with equal probability, then he should play O. (Why?)
Definition 5 An action ai is a best response to belief i if no other action
a0i gives him a strictly higher expected payoff given this belief. That is, ai is
a best response to i if
X X
ui (ai , i (ai )ai ) ui (ai , i (ai )ai ) for all ai Ai
ai ai
Sometimes, more than one action may be a best response to some belief.
This gives rise to the best response correspondence BRi (i ).
11
Rationalizability
Rationalizability is a weak solution concept- it is one that derives
restrictions on players choice of actions from assumptions that rationality
and payoffs are common knowledge.
It asks : what are all the strategies that a rational player can possibly
play?
Player cant reasonably play a strategy which is not a best response to
some beliefs about his opponents strategies. Moreover, rationality means
that opponents will also use only strategies which are best responses to some
beliefs.
An action ai is never a best response if ai / BR(i ) for any belief mui .
Clearly, i should not use ai if it is never a best response. Neither should
player i expect his opponents to use actions that are in their never best
response sets. Eliminating these actions gives rise to smaller reduced sets
of actions. Clearly, one can then iteratively delete never best actions in the
same way as in iterative elimination of strictly dominated strategies. This
will give rise to the set of rationalizable strategies.
Definition 6 An action aj is rationalizable in the game (N, (Ai )iN , (ui )iN )
if for each j N , there is a set Zj Aj such that
(i) aj Zj ,
(ii) every action aj Zj is a best response to a belief j of player j whose
support is a subset of Zj .
12
: What can be the set of rationalizable strategies
: What are the set of strategies which survive iterated elimination of
strictly dominated strategies?
13
Relationship between Rationalizability and Iterated Elimination
of Strictly Dominated Strategies
14
Nash Equilibrium
All the solution concepts discussed so far have no predictive power in the
BOS game. The Nash equilibrium solution concept ties down actions and
beliefs in a more precise manner. The concept is built around the notion
that
(i) Actions must be a best response to beliefs
(ii) Beliefs must be correct in equilibrium.
Formally,
15
Example 8: (Battle of the Sexes)
B F
F (0, 0) (2, 1)
B (1, 2) (0, 0)
There are two pure strategy Nash equilibria - (F,F) and (B,B).
16
Existence
Use of fixed point theorems to prove existence of equilibrium.
Define the set-valued function B : 7 by B() = iN Bi (i ).
A Nash equilibrium is an action profile such that B( ).
That is, must be a fixed point of the mapping B.
Example: Let f : [0, 1] [0, 1] be a continuous function. Then, there
must be a number x [0, 1] such that f (x) = x.
Kakutani Fixed Point Theorem: Let X be a compact, convex, subset
of some Euclidean space <M , and let f : X X be a set-valued function
such that
(ii) the graph of f is closed. That is, for all sequences {xn }, {yn } such that
yn f (xn ) for all n, xn x0 , yn y0 , we have y0 f (x0 ).
Theorem 3 The game (N, {Ai }iN , {ui }iN ) has a (pure strategy) Nash equi-
librium if for all i N ,
Ai is a non-empty compact convex subset of a Euclidean space.
Each ui is continuous and quasi-concave on Ai .
Hence, ai Bi (ai ).
Now, take a sequence (aki , aki )k converging to (ai , ai ) with aki Bi (aki )
for each k.
This implies that ui (aki , aki ) ui (a0i , aki ) for all a0i Ai .
17
By continuity of ui , this implies that ui (ai , ai ) ui (a0i , ai ).
Hence, ai Bi (ai ).
So, the graph of B is closed, and B satisfies all the conditions for a fixed
point to exist by Kakutanis theorem.
This fixed point is a Nash equilibrium.
Theorem 4 Every strategic game in which each player has a finite number
of pure actions has a mixed strategy Nash equilibrium.
So, B is convex-valued.
Closedness also follows from continuity of ui . Hence, a fixed point must
exist.
18
Here are some obvious facts about Nash equilibria. Suppose is a
Nash equilibrium. Then,
(ii) All pure strategies in the support of i must give individual i the same
(expected) payoff.
19
Can we identify games which will have pure strategy Nash equilibria?
Potential Games (Monderer and Shapley, Games and Economic Be-
havior, 1996)
where F (.) is the inverse demand function and c is the constant marginal
cost.
Check that this is a potential.
20
Congestion game: defined by a set of resources E.
The strategy set of player i is Ai 2E .
For each resource e E, a cost function ce : N R gives the cost of
resource e as a function of xe , the number of players using the resource.
The cost function of each player is given by
X
ci (a) = ce (xe )
eai
Proof.
For any strategy vector, define
xe
XX
P (a) = ce (j)
eE j=1
/ ai , e a0i }
E + = {e E|e
E = {e E|e ai , e
/ a0i }
For every resource that is both in ai and a0i or neither in ai or a0i , there is
no effect of the deviation on P . So,
P (a) P (a0i , ai ) =
X X
ce (xe ) ce (xe + 1)
eE eE +
= c(ai , ai ) c(a0i , ai )
21
Supermodular Games
These are games where playerss strategies can be ordered, and where
each players marginal utility of increasing his strategy rises with increases
in his rivals strategies : strategic complements.
Implication best response of a player is a nondecreasing function of
other players strategies.
Here is a definition of supermodular games (not the most general one
possible).
Suppose each Ai is a compact interval of <, and ui is twice continuously
differentiable.
2u
The game is supermodular iff a i ai
j
0.
Examples
1. Bertrand game with heterogeneous goods:
Let demand functions be:
X
Di (pi , pi ) = ai + bi pi + dij pj
j6=i
2 ui
>0
pi pj
.
2. Cournot duopoly game:
Suppose inverse demand function Pi (qi , qj ) and firm is marginal revenue
Pi + qi P
qi
i
are decreasing in qj . The payoffs are
22
Tragedy of the Commons :
Nash equilibria are often Pareto suboptimal.
There are n farmers in a village.
Value per cow = v(C), when C is the total number of cows grazing.
cj ) + ci v 0 ( cj ) d = 0
X X
v(
j j
Cv(C) dC
v(C ) + C v 0 (C ) d = 0
n
c > C - overgrazing.
P
So, i
j=1
23
Correlated Equilibrium
Consider the BOS game described below.
B F
F (6, 6) (2, 7)
B (7, 2) (0, 0)
There are 3 Nash equilibria, with payoffs being (7,2), (2,7) and (14/3,
14/3).
Pre-play communication allows players to improve payoffs.
They can correlate their strategies by agreeing to play the joint actions
(B,F), (F,F) and (B,B), each with probability 1/3.
Notice that the expected payoff is then (5,5), which is not in the convex
hull of the Nash equilibrium payoffs.
Using mediator to enforce correlated strategies.
Once 1 hears the recommendation F, what can she infer about what
player 2 has received?
Check that she has no incentive to deviate from the recommendation.
A correlated strategy is a probability distribution over A.
Given any , let S() denote the support of , That is,
S() = {a A|(a) > 0}
Definition 9 A correlated strategy of the game (N, (Ai )iN , (ui )iN ) is a
correlated equilibrium if for all a S(), for all players i, and for all a0i Ai ,
ui (a0i , bi )(ai , bi )
X X
ui (ai , bi )(ai , bi )
bi Ai bi Ai
Two observations:
1. A Nash equilibrium must be a correlated equilibrium. (Why?)
2. The set of correlated equilibria is a convex set.
Theorem 7 Every action used with positive probability by some player in a
correlated equilibrium of a finite strategic game is rationalizable.
24