Network Science

Lecture 4
Game Theory for Networks (I)

Prof. Chun-Hung Liu

Dept. of Electrical and Computer Engineering
National Chiao Tung University
Spring 2016

2016/3/27 Lecture 4: Game Theory for Networks (I) 1

 Motivation
• In the context of social networks, or even communication networks,
agents make a variety of choices.
• For example:
• What kind of information to share with others you are connected to.
• How to evaluate information obtained from friends, neighbors, coworkers
and media.
• Whether to trust and form friendships.
• Which of the sellers in your neighborhood to use.
• Which websites to visit.
• How to map your drive in the morning (or equivalently how to route your
network traffic).

• In all of these cases, interactions with other agents you are connected
to affect your payoff, well-being, utility.
• How to make decisions in such situations?
• Game theory or Multi-agent decision theory

2016/3/27 Lecture 4: Game Theory for Networks (I) 2

 Rational Decision Making
• Powerful working hypothesis in economics: individuals act rationally in
the sense of choosing the option that gives them higher “payoff”.
• Payoff here need not be monetary payoff. Social and psychological factors
influence payoffs and decisions.
• Nevertheless, the rational decision-making paradigm is useful because it
provides us with a (testable) theory of economic and social decisions.

• We often need only ordinal information; i.e., two options a and b, and we
imagine a (real-valued) utility function u(·) that represents the ranking of
different options, and we simply check whether

u(a) u(b), and u(a)  u(b),

• But in game theory we often need cardinal information because

decisions are made under natural or strategic uncertainty. The theory
of decision-making under uncertainty was originally developed by John
von Neumann and Oskar Morgenstern.

2016/3/27 Lecture 4: Game Theory for Networks (I) 3

 Decision-Making under Uncertainty (I)
• von Neumann and Morgenstern posited a number of “reasonable” axioms
that rational decision-making under uncertainty should satisfy. From
these, they derived the expected utility theory.
• Under uncertainty, every choice induces a lottery, that is, a probability
distribution over different outcomes.
• e.g., one choice would be whether to accept a gamble which pays $10
with probability 1/2 and makes you lose $10 with probability 1/2.

• von Neumann and Morgenstern’s expected utility theory shows that

(under their axioms) there exists a utility function (also referred to as
Bernoulli utility function) u(c) , which gives the utility of consequence
(outcome) c.
• Then imagine that choice a induces a probability distribution Fa (c) over
consequences.
• Then the utility of this choice is given by the expected utility according to
the probability distribution Fa (c) : Z
U (a) = u(c)dFa (c)
2016/3/27 Lecture 4: Game Theory for Networks (I) 4
 Decision-Making under Uncertainty (II)
• In other words, this is the expectation of the utility u(c), evaluated
according to the probability distribution Fa (c).
• More simply, if Fa (c) is a continuous distribution with density fa (c),
then Z
U (a) = u(c)fa (c)dc,

• Given expected utility theory and our postulate of “rationality,” single

person decision problems are (at least conceptually) simple.

• If there are two actions, a and b, inducing probability distributions Fa (c)

and Fb (c) , then the individual should choose a over b only if
Z Z
U (a) = u(c)dFa (c) U (b) = u(c)dFb (c).

2016/3/27 Lecture 4: Game Theory for Networks (I) 5

 From Single Person to Multi-person Decision Problems
• But in a social situation (or more specifically, in a “social network”
situation), the utility of an agent or probability distribution over
outcomes depends on actions of others.
• A simple game of “partnership” represented as a matrix game:

• Here the first number is the payoff to player (partner) 1 and the second
number is the payoff to player 2. More formally, the cell indexed by row x
and column y contains a pair, (a, b) where
a = u1 (x, y) and b = u2 (x, y)
• These numbers could be just monetary payoffs, or it could be inclusive of
“social preferences” (the fact that you may be altruistic towards your
partner or angry at him or her).
• Should you play “work hard” or “shirk”?
2016/3/27 Lecture 4: Game Theory for Networks (I) 6
 Strategic Form Games (I)
• Let us start with games in which all of the participants act simultaneously
and without knowledge of other players’ actions. Such games are referred
to as strategic form games - or as normal form games or matrix games.
• For each game, we have to define
• The set of players.
• The strategies.
• The payoffs.
• More generally, we also have to define the game form, which captures the
order of play (e.g., in chess) and information sets (e.g., in asymmetric
information or incomplete information situations). But in strategic form
games, play is simultaneous, so no need for this additional information.
• Definition: A strategic form game is a triplet hI, (Si )i2I , (ui )i2I i where
• I is a finite set of players, I = {1, 2, . . . , I} .
• Si is a non-empty set of available actions for player i. Q
• ui : S ! R is the payoff (utility) function of player i where S = i Si .

2016/3/27 Lecture 4: Game Theory for Networks (I) 7

 Strategic Form Games (II)
• For strategic form games, we will use the terms action and (pure) strategy
interchangeably.
• We denote by si 2 Si an action for player i, and by s i = [sj ]j6=i a vector
of actions for all players except i. We refer to the tuple (si , s i ) 2 S as an
Q
action (strategy) profile, or outcome. We also denote by S i = j6=i Sj the
set of actions (strategies) of all players except i.
• Our convention throughout will be that each player i is interested in
action profiles that “maximize” his utility function ui (·) .
• Example (Finite Strategy Sets): A two-player game where the strategy set Si
of each player is finite can be represented in matrix form. We adopt the
convention that the rows (columns) of the matrix represent the action set of
player 1 (player 2).The cell indexed by row x and column y contains a pair,
(a, b), where a is the payoff to player 1 and b is the payoff to player 2, i.e.,
a = u1 (x, y) and b = u2 (x, y). This class of games is sometimes referred to
as bimatrix games. For example, consider the following game of “Matching
Pennies.

2016/3/27 Lecture 4: Game Theory for Networks (I) 8

 Examples of Finite and Infinite Games

• This game represents “pure conflict,” in the sense that one player’s utility
is the negative of the utility of the other player, i.e., the sum of the utilities
for both players at each outcome is “zero.”
• This class of games is referred to as zero-sum games (or constant-sum
games) and has been extensively studied in the game theory literature.
• Example (Infinite Strategy Sets): The strategy sets of players can also
have infinitely many elements. Consider the following game of Cournot
Competition, which models two firms producing the same homogeneous
good and seeking to maximize their profits. The formal game
G = hI, (Si ), (ui )i consists of:

2016/3/27 Lecture 4: Game Theory for Networks (I) 9

 Examples of Finite and Infinite Games
• A set of two players, I = 1, 2.
• A strategy set Si = [0, 1) for each player i, where si 2 Si .
represents the amount of good that the player produces.
• A utility function ui for each player i given by its total revenue
minus its total cost, i.e.,
ui (s1 , s2 ) = si p(s1 + s2 ) ci si
where p(q) represents the price of the good (as a function of the total
amount of good q), and ci is the unit cost for firm i.
• For simplicity, we consider the case where both firms have unit
cost, c1 = c2 = 1, and the price function is piecewise linear and is given
by p(q) = max{0, 2 q} .
• We analyze this game by considering the best-response correspondences
for each of the firms. For firm i, the best-response correspondence Bi (s i )
is a mapping from the set S i into set Si such that

2016/3/27 Lecture 4: Game Theory for Networks (I) 10

 Examples of Finite and Infinite Games
Bi (s i ) = {si 2 Si |ui (si , s i ) ui (s0i , s i ), 8s0i 2 Si }.
Note that for this example, the best response correspondences are unique-
valued (hence can be referred to as best-response functions), since for
all s i , there is a unique action that maximizes the utility function for firm
i. In particular, we have:
Bi (s i ) = arg max{si p(si + s i ) si }.
si 0

• When s i > 1 , the action that maximizes the utility function of firm i,
i.e., its best response, is 0. More generally, it can be seen that for
any s i 2 S i , the best response of firm i is given by

Bi (s i ) = arg max(si (2 si s i) si )
si 0
(
1 s
2
i
, if s i  1,
=
0, otherwise

2016/3/27 Lecture 4: Game Theory for Networks (I) 11

 Examples of Finite and Infinite Games

Figure: Best response functions for

the Cournot Competition game.

• The above figure illustrates the best response functions as a function of s1

and s2 . Intuitively, we expect the outcome of this game to be at the point
where both of these functions intersect. In the following, we will argue
that this intersection point is a reasonable outcome in this game.

2016/3/27 Lecture 4: Game Theory for Networks (I) 12

 Dominant and Dominated Strategies (I)
• In some games, it may be possible to predict the outcome assuming that
all players are rational and fully knowledgeable about the structure of the
game and each other’s rationality.
• This is the case, for instance, for the well-studied Prisoner’s Dilemma
game. The underlying story of this game is as follows: Two people are
arrested for a crime, placed in separate rooms, and questioned by
authorities trying to extract a confession.
• If they both remain silent (i.e., cooperate with each other), then the
authorities will not be able to prove charges against them and they will
both serve a short prison term, say 2 years, for minor offenses.
• If only one of them confesses (i.e., does not cooperate), his term will be
reduced to 1 year and he will be used as a witness against the other person,
who will get a sentence of 5 years.
• If they both confess, they both get a smaller sentence of 4 years. This
game can be represented in a matrix form as follows:

2016/3/27 Lecture 4: Game Theory for Networks (I) 13

 Dominant and Dominated Strategies (II)
(Confess)

(Confess)

• In this game, regardless of the other players decision, playing Don’t

Cooperate yields a higher payoff for each player. Hence, the strategy
Don’t Cooperate is strictly dominant, i.e., no matter what strategy the
other player chooses, this strategy always yields a strictly better outcome.
• We can therefore infer that both players will choose Don’t Cooperate
and spend the next four years in jail while if they both chose the strategy
Cooperate, they could have ended up in jail only for two years!
• Prisoner’s Dilemma is a paradigmatic example of self-interested rational
behavior not leading to jointly (socially) optimal outcomes.

2016/3/27 Lecture 4: Game Theory for Networks (I) 14

 Dominant and Dominated Strategies (III)
• A compelling notion of equilibrium in games would be the dominant
strategy equilibrium, where each player plays a dominant strategy as
formalized in the next definition.
• Definition (Dominant Strategy): A strategy si 2 Si is a dominant strategy
for player i if
ui (si , s i ) ui (s0i , s i ) for all s0i 2 Si and for s i 2S i

It is strictly dominant if this relation holds with a strict inequality

• Definition (Dominant Strategy Equilibrium): A strategy profile s⇤ is a

(strictly) dominant strategy equilibrium if for each player i, s⇤i is a
(strictly) dominant strategy.
• In the Prisoner’s Dilemma game, (Don’t Cooperate, Don’t Cooperate) is
a strictly dominant strategy equilibrium. Though compelling, dominant
strategy equilibria do not always exist, as illustrated by the Matching
Pennies game.
2016/3/27 Lecture 4: Game Theory for Networks (I) 15
 Dominant and Dominated Strategies (IIII)
• Consider a slightly modified Prisoner’s Dilemma game in which players
also have the strategy Suicide leading to the following payoff structure:

• This payoff matrix models a scenario in which if one player chooses the
strategy Suicide, then, due to lack of witnesses, the other player gets off
free if he remains silent (cooperates). In this game, there is no dominant
strategy equilibrium because of the additional strategy Suicide.
• Notice, however, that the strategy Suicide is the worst possible option
for a player, no matter what the other player does. In this sense, Suicide
is strictly dominated by the other two strategies. More generally, we say
that a strategy is strictly dominated for a player if there exists some
other strategy that yields a strictly higher payoff regardless of the
strategies of the other players.
2016/3/27 Lecture 4: Game Theory for Networks (I) 16
 Dominant and Dominated Strategies (V)
• Definition (Strictly Dominated Strategy): A strategy si 2 Si is strictly
dominated for player i if there exists some s0i 2 Si such that

ui (s0i , s i ) > ui (si , s i ) for all s i 2S i

• Definition (Weakly Dominated Strategy): A strategy si 2 Si is weakly

dominated for player i if there exists some s0i 2 Si such that

ui (s0i , s i ) ui (si , s i ) for all s i 2S i

and
ui (s0i , s i ) > ui (si , s i ) for some s i 2S i

• It is plausible to assume that no player chooses a strictly dominated

strategy. Moreover, common knowledge of payoffs and rationality leads
players to do iterated elimination of strictly dominated strategies, as
illustrated next.

2016/3/27 Lecture 4: Game Theory for Networks (I) 17

 Iterated Elimination of Strictly Dominated Strategies
• In the Prisoner’s Dilemma with Suicide game, the strategy Suicide is a
strictly dominated strategy for both players. Therefore, no rational player
would choose Suicide.
• Moreover, if player 1 is certain that player 2 is rational, then he can
eliminate her opponent’s Suicide strategy. We can use a similar reasoning
for player 2. After one round of elimination of strictly dominated
strategies, we are back to the Prisoner’s Dilemma game, which has a
dominant strategy equilibrium.
• Thus, iterated elimination of strictly dominated strategies leads to a
unique outcome, (Don’t Cooperate, Don’t Cooperate) in this modified
game. We say that a game is dominance solvable if iterated elimination of
strictly dominated strategies yields a unique outcome.
• Consider next another game. Player 2
Player 1

2016/3/27 Lecture 4: Game Theory for Networks (I) 18

 Iterated Elimination of Strictly Dominated Strategies
• In this game, there are no strategies that are strictly dominated for
player 1 (the row player). On the other hand, the strategy Middle is
strictly dominated by the strategy Right for player 2 (the column player).
Thus, we conclude that it is not rational for player 2 to play Middle and
we can therefore remove this column from the game, resulting in the
following reduced game.

Game after one removal of strictly dominated strategies.

• Now, note that both strategies Middle and Down are strictly dominated
by the strategy Up for player 1, which means that both of these rows can
be removed, resulting in the following game.

2016/3/27 Lecture 4: Game Theory for Networks (I) 19

 Iterated Elimination of Strictly Dominated Strategies

Game after three iterated removals of strictly dominated strategies.

• We are left with a game where player 1 does not have any choice in his
strategies, while player 2 can choose between LEFT and RIGHT. Since
LEFT will maximize the utility of player 2, we conclude that the only
rational strategy profile in the game is (Up, LEFT).
• More formally, we define the procedure of iterated elimination of
strictly dominated strategies (or iterated strict dominance) as follows:

2016/3/27 Lecture 4: Game Theory for Networks (I) 20

 Iterated Elimination of Strictly Dominated Strategies

• Theorem: Suppose that either (i) each Si is finite, or (ii) each ui (si , s i )
is continuous and each Si is compact. Then Si1 is nonempty for each i.

• We next apply iterated elimination of strictly dominated strategies to the

Cournot Competition game. Recall that we use the notation Sik to denote
the set of strategies of player i that survive iterated elimination of strictly
dominated strategies at step k. In the first step, we note that both firms
must choose a quantity between [0,∞), i.e.,
S11 = [0, 1], and S21 = [0, 1].

2016/3/27 Lecture 4: Game Theory for Networks (I) 21

Iterated Elimination of Strictly Dominated Strategies
• Since the range of the best response function of player 1 is [0, 1/2], any
strategy outside this range is never a best response and therefore is
strictly dominated. The same reasoning holds for player 2. Thus, at the
second step, we have
S22 = [0, 1/2]. and S12 = [0, 1/2].

• Given that player 2 only chooses actions in the interval [0, 1/2], then
player 1 can restrict the domain of his best response function to only
these values. Using his best response function, this implies that the
strategies outside the interval [1/4, 1/2] are strictly dominated for player
1. Applying the same reasoning for player 2, we obtain

S13 = [1/4, 1/2] and S23 = [1/4, 1/2]

• It can be shown that in the limit, the endpoints of the intervals converge
to the point where the two best response functions intersect.

2016/3/27 Lecture 4: Game Theory for Networks (I) 22

Iterated Elimination of Strictly Dominated Strategies

Figure: Elimination of strictly dominated strategies for the Cournot

competition game.

2016/3/27 Lecture 4: Game Theory for Networks (I) 23

 Nash Equilibrium
• The Cournot Competition game is another example of a dominance
solvable game. Most games, however, are not solvable by iterated strict
dominance; therefore, we need a stronger equilibrium notion to predict
the outcome.
• We next introduce the fundamental solution concept for strategic form
games, the notion of a Nash equilibrium. A Nash equilibrium captures a
steady state of the play in a strategic form game such that each player
acts optimally and forms correct conjectures about the behavior of the
other players.
• Definition (Nash Equilibrium): A (pure strategy) Nash equilibrium of a
strategic form game hI, (Si ), (ui )i2I i is a strategy profile s⇤ 2 S such
that for all i 2 I , we have
ui (s⇤i , s⇤ i ) ui (si , s⇤ i ) for all si 2 Si
• Hence, a Nash equilibrium is a strategy profile s⇤such that no player i
can profit by unilaterally deviating from his strategy s⇤i , assuming every
other player j follows his strategy s⇤j .
2016/3/27 Lecture 4: Game Theory for Networks (I) 24
 Nash Equilibrium
• Example (Battle of the Sexes): Consider a two player game with the
following payoff structure:

• This game, referred to as the Battle of the Sexes game, represents a

scenario in which the two players wish to coordinate their actions, but
have different preferences over their actions. This game has two pure
Nash equilibria, i.e., the strategy profiles (BALLET, BALLET) and
(SOCCER, SOCCER).

2016/3/27 Lecture 4: Game Theory for Networks (I) 25

 Nash Equilibrium
• Example (Second Price Auction - with Complete Information):
We consider a second price auction: There is a single indivisible object to
be assigned to one of n players. Player i’s valuation of the object is
denoted by vi . We assume without loss of generality that
v1 v2 · · · vn > 0 and that each player knows all the
valuations v1 , . . . , vn , i.e., it is a complete information game. The rules
of this auction mechanism are described as follows:
• The players simultaneously submit bids, b1 , . . . , bn
• The object is given to the player with the highest bid (or to a
random player among the ones bidding the highest value).
• The winner pays the second highest bid.
• This mechanism induces a game among the players in which the strategy
of each player is given by her bid, and her utility for a bid profile is given
by her valuation of the object minus the price she pays, i.e., if player i is
the winner, her utility is vi bj where j is the player with the second
highest bid; otherwise, her utility is zero.
2016/3/27 Lecture 4: Game Theory for Networks (I) 26
 Nash Equilibrium
• We first show that the strategy profile (b1 , . . . , bn ) = (v1 , . . . , vn ) is a Nash
equilibrium. First note that if indeed everyone plays according to this
strategy profile, then player 1 receives the object and pays a price v2 .
Hence, her payoff will be v1 v2 > 0 , and all other payoffs will be 0.
Now, player 1 has no incentive to deviate since her utility cannot increase.
Similarly, for all other players i 6= 1 , in order for player i to change her
payoff, she needs to bid more than v1, in which case her payoff will
be vi v1 < 0 .Therefore, no player has an incentive to unilaterally
deviate, showing that this strategy profile is a Nash equilibrium.
• We next show that the strategy profile ( v1 , 0, 0,…, 0) is also a Nash
equilibrium. As before, player 1 will receive the object, and will have a
payoff of v1 0 = v1 . Using the same argument as before, we conclude
that none of the players have an incentive to deviate, and this strategy
profile is a Nash equilibrium.

2016/3/27 Lecture 4: Game Theory for Networks (I) 27

 Nash Equilibrium
• So far, we have shown that the game induced by the second price
auction has multiple Nash equilibria. We finally show that for each
player i, the strategy of bidding her valuation, i.e., bi = vi, in fact,
weakly dominates all other strategies.
• Given a bid profile, let B ⇤ denote the maximum of all bids excluding
player i’s bid, i.e.,
B ⇤ = max bj .
j6=i

• Assume that player i’s valuation is given by v ⇤. The following figure

illustrates the utility of player i as a function of B ⇤, when she bids her
valuation, bi = v ⇤ , less than her valuation, bi < v ⇤, and more than her
valuation bi > v ⇤ . In the second graph, which represents the case when
she bids bi < v ⇤, notice that whenever bi  B ⇤  v ⇤ , player i receives
zero utility since she loses the auction to whoever bid B ⇤. If she would
have bid her valuation, she would have positive utility in this region (as
depicted in the first graph). These figures show that for each player,
bidding her own valuation weakly dominates all her other strategies.

2016/3/27 Lecture 4: Game Theory for Networks (I) 28

 Nash Equilibrium

ui (bi ) ui (bi ) ui (bi )

⇤ ⇤
v B ⇤ bi B ⇤
v v⇤ bi B ⇤
⇤ ⇤ ⇤
bi = v bi < v bi > v

Figure: The utility of player i as a function of B ⇤, the maximum of all

bids except i’s bid.

2016/3/27 Lecture 4: Game Theory for Networks (I) 29

 Nash Equilibrium
• Given that Nash equilibrium is a widely used solution concept in
strategic form games, a natural question is why one should expect the
Nash equilibrium outcome in a strategic form game.
• One justification is that since it represents a steady state situation,
rational players should somehow reason their way to Nash equilibrium
strategies; that is, Nash equilibrium might arise through introspection.
This justification requires that players are rational and know the
payoff functions of all players, that they know their opponents are
rational and know the payoff functions, and so on.
• A second justification is that Nash equilibria are self-enforcing. That is,
if players agree on a strategy profile before independently choosing
their actions, then no player has an incentive to deviate if the agreed
strategy profile is a Nash equilibrium.

2016/3/27 Lecture 4: Game Theory for Networks (I) 30

Mixed Strategy and Mixed Strategy Nash Equilibrium
• Recall the game of Matching Pennies:

• It is easy to see that this game does not have a pure Nash equilibrium,
i.e., for every pure strategy in this game, one of the parties has an
incentive to deviate. However, if we allow the players to randomize over
their choice of actions, we can determine a steady state of the play.
• Assume that player 1 picks Heads with probability p and Tails with
probability 1 p , and that player 2 picks both Head and Tail with
probability 1/2. Then, with probability
1 1 1
p + (1 p) =
2 2 2
2016/3/27 Lecture 4: Game Theory for Networks (I) 31
Mixed Strategy and Mixed Strategy Nash Equilibrium
player 1 will receive a payoff 1. Similarly, she will receive a payoff -1 with
the same probability. This implies that if player 2 plays the strategy (1/2,
1/2), then no matter what strategy player 1 chooses, she will get the same
payoff. Due to the symmetry of the game, we conclude that the
randomized strategy (( 1/2 , 1/2 ), (1/2,1/2)) is a stable play of the game.
• We next formalize the notion of “randomized" or mixed strategies. Let ⌃i
denote the set of probability measures over the pure strategy (action)
set Si . We use i 2 ⌃i to denote the mixed strategy of player i. When Si
is a finite set, a mixed strategy is a finite dimensional probability vector,
i.e., a vector whose elements denote the probability with which a
particular action will be played.
• If Si has two elements, the set of mixed strategies i is the one-
dimensional probability simplex, i.e., ⌃i = {(x1 , x2 )|xi 0, x1 + x2 = 1}
Q
• We use 2 ⌃ = i2I ⌃i to denote a mixed strategy profile. Note that
this implicitly assumes that players randomize independently. We
Q
similarly denote i 2⌃ i = j6=i ⌃j .

2016/3/27 Lecture 4: Game Theory for Networks (I) 32

Mixed Strategy and Mixed Strategy Nash Equilibrium
• Following vonNeumann-Morgenstern expected utility theory, we
extend the payoff functions ui from S to ⌃ by
Z
ui (s)d (s),
S

i.e., the payoff of a mixed strategy σ is given by the expected value of

pure strategy payoffs under the distribution σ.
• Definition (Mixed Strategy Nash Equilibrium) A mixed strategy profile ⇤

is a mixed strategy Nash equilibrium (or mixed Nash equilibrium) if for

each player i,
ui ( i⇤ , ⇤ i ) ui ( i , ⇤ i ) 8 i 2 ⌃i

• Proposition 1: A mixed strategy profile ⇤ is a (mixed strategy) Nash

equilibrium if and only if for each player i,
⇤ ⇤ ⇤
ui ( i, i) ui (si , i) 8si 2 Si

2016/3/27 Lecture 4: Game Theory for Networks (I) 33

Mixed Strategy and Mixed Strategy Nash Equilibrium
• Proposition 2: Let G = hI, (Si )i2I , (ui )i2I i be a finite strategic form
game. Then, ⇤ 2 ⌃ is a Nash equilibrium if and only if for each
player i 2 I , every pure strategy in the support of i⇤ is a best response
to ⇤ i .
• Example: Let us return to the Battle of the Sexes game.

• Recall that this game has 2 pure Nash equilibria. Using the characterization
result in Proposition 2, we show that it has a unique mixed strategy Nash
equilibrium (which is not a pure strategy Nash equilibrium). First, by using
Proposition 2 (and inspecting the payoffs), it can be seen that there are no Nash
equilibria where only one of the players randomizes over its actions

2016/3/27 Lecture 4: Game Theory for Networks (I) 34

Mixed Strategy and Mixed Strategy Nash Equilibrium
• Now, assume instead that player 1 chooses the action Ballet with
probability p 2 (0, 1) and Soccer with probability 1-p, and that player
2 chooses Ballet with probability q 2 (0, 1) and Soccer with probability
1-q. Using Proposition 2 on player 1’s payoffs, we have the following
relation
2q + 0 ⇥ (1 q) = 0 ⇥ q + 1 ⇥ (1 q)

Similarly, we have
1 ⇥ p + 0 ⇥ (1 p) = 0 ⇥ p + 2 ⇥ (1 p)
We conclude that the only possible mixed strategy Nash equilibrium is
given by q = 1/3 and p = 2/3 .

2016/3/27 Lecture 4: Game Theory for Networks (I) 35

 Appendix:
Existence of A Nash Equilibrium

2016/3/27 Lecture 4: Game Theory for Networks (I) 36

 Existence of A Nash Equilibrium
• Example (Pricing-Congestion Game): We consider a price competition
model. Consider a parallel link network with I links. Assume that d
units of flow are to be routed through this network. We assume that
this flow is the aggregate flow of many infinitesimal users.

l1 (x), p1

l2 (x), p2
Figure: A price competition model over congested networks.

• Let li (xi ) denote the latency function of link i, which represents the
delay or congestion cost as a function of the total flow xi on link i.

2016/3/27 Lecture 4: Game Theory for Networks (I) 37

 Existence of A Nash Equilibrium
• Assume that the links are owned by independent service providers,
which set a price pi per unit of flow on link i (see Figure in above). The
effective cost of using link i is pi + li (xi ). Users have a reservation
utility equal to R, i.e., if pi + li (xi ) > R , then no traffic will be routed
on link i.
• We consider a special case of this model with two links and latency
functions l1 (x1 ) = 0 and l2 (x2 ) = 3x2 2 .
• For simplicity, we assume that R = 1and d = 1. Given the prices
(p1 , p2 ) , we assume that the flow is allocated according to a Wardrop
equilibrium, i.e., the flows are routed along minimum effective cost paths
and the effective cost cannot exceed the reservation utility.
• Formally, a flow vector x = [xi ]i=1,...,I is a Wardrop equilibrium if
PI
xi  1 and pi + li (xi ) = min{pj + lj (xj )},
i=1 j
pi + li (xi )  1,
PI
with i=1 xi = 1 and minj {pj + lj (xj )} < 1
2016/3/27 Lecture 4: Game Theory for Networks (I) 38
 Existence of A Nash Equilibrium
• We use the preceding characterization to determine the flow
allocation on each link given prices 0  p1 , p2  1 :
(
2
3 (p1 p2 ), p1 p2
x2 (p1 , p2 ) =
0, otherwise

x1 (p1 , p2 ) = 1 x2 (p1 , p2 )
• The payoffs for the providers are then given by:

u1 (p1 , p2 ) = p1 ⇥ x1 (p1 , p2 )
u2 (p1 , p2 ) = p2 ⇥ x2 (p1 , p2 )

• We find the pure strategy Nash equilibria of this game by

characterizing the best response correspondences, Bi (p i ) for each
player.

2016/3/27 Lecture 4: Game Theory for Networks (I) 39

 Existence of A Nash Equilibrium
• In particular, for a given p2 , B1 (p2 ) is the optimal solution set of the
following optimization problem
max p1 x 1
0p1,0x1

3
s.t. p1 = p2 + (1 x1 ).
2

Solving the preceding optimization problem, we find that

⇢
3 p2
B1 (p2 ) = min 1, + .
4 2
Similarly, B2 (p1 ) = p1
2

2016/3/27 Lecture 4: Game Theory for Networks (I) 40

 Existence of A Nash Equilibrium

Figure: Best response correspondences.

2016/3/27 Lecture 4: Game Theory for Networks (I) 41

 Existence of A Nash Equilibrium
• The above figure illustrates the best response correspondences as a
function of p1 and p2 . These correspondences intersect at the unique
point (p1 , p2 ) = (1, 1/2) , which is the unique pure strategy Nash
equilibrium.
• We next consider a similar example with latency functions given by
(
0, if 0  x  1/2
l1 (x) = 0, l2 (x) = x 1/2
✏ , x 1/2,

for some sufficiently small ✏ > 0. For this example, we show that no
pure Nash equilibrium strategy exists for small ✏ . The following list
considers all candidate Nash equilibria (p1 , p2 ) and profitable
unilateral deviations for ✏ sufficiently small, thus establishing the
nonexistence of a pure strategy Nash equilibrium:

• p1 = p2 = 0 : A small increase in the price of provider 1 will generate

positive profits, thus provider 1 has an incentive to deviate.

2016/3/27 Lecture 4: Game Theory for Networks (I) 42

 Existence of A Nash Equilibrium
• p1 = p2 > 0 : Let x be the corresponding flow allocation. If x1 = 1,
then provider 2 has an incentive to decrease its price. If x1 < 1, then
provider 1 has an incentive to decrease its price.
• 0  p1 < p2 : Player 1 has an incentive to increase its price since its
flow allocation remains the same.
• 0  p2 < p1 : For ✏ sufficiently small, the profit function of player 2,
given p1 , is strictly increasing as a function of p2, showing that
provider 2 has an incentive to increase its price.
• Theorem (Weierstrass): Let A be a nonempty compact subset of a finite
dimensional Euclidean space and let f : A ! R be a continuous
function. Then there exists an optimal solution to the optimization
problem
minimize f (x)
subject to x 2 A.
• Theorem: Any finite strategic game has a mixed strategy Nash
equilibrium.
2016/3/27 Lecture 4: Game Theory for Networks (I) 43