Dynamic Games with perfect information. Commitment.

Backward induction. Stackelberg’s model

Extensive form
• This form is pictured by way of a game tree that starts from a unique node
called the root
• Out of the root come several branches and at the end of each branch is a
decision node.
• In turn, branches emanate from each of these decision nodes and end in yet
another set of nodes.
• A decision node is a point in the game where one player—and only one player
—has to make a decision
• Each branch of the tree emanating from that node corresponds to one of
his choices.
• If a node has no branches emerging from it then it is called a terminal node
Extensive form game
Information set
Information set (Prisoners Dilemma)
A formal treatment
In order for a tree to represent a game, the nodes and the branches need
to satisfy three consistency requirements:
• Single Starting Point: there must be one, and only one, starting point.
• No Cycles: It is important that we not hit an impasse while playing the
game; it must not be possible for the branches of a tree to double back
and create a cycle.
• One Way to Proceed: there must not be two or more branches leading
to a node
.
No single starting point
One way to proceed not fulfilled
Predecessor node
• The predecessors of a node, say α, are those nodes from which you
can go (through a sequence of branches) to α.
To guarantee the three consistency requirements, the
following restrictions are imposed on predecessor nodes:
• A node cannot be a predecessor of itself
• A predecessor’s predecessor is also a predecessor: if a node β is a
predecessor of α, and a node γ is, in turn, a predecessor of β, then γ is
a predecessor of α as well.
• Predecessors can be ranked: if β and γ are both predecessors of α, it
must be the case that either β is a predecessor of γ or vice versa
• There must be a common predecessor: Consider any two nodes, α
and β, neither of which precedes the other. Then there must be a node
γ that is a predecessor of both α and β
There cannot be two or more roots
• Restriction 4, by itself, implies that there cannot be two or more roots
in a tree.
• If there are two roots, then there are two nodes neither of which
precedes the other.
• But then there must be yet a third node that precedes them both, and
that is a logical contradiction.
There cannot be a cycle
• Restrictions 1 and 2 together imply that there cannot be a cycle
• Suppose it were the case that β is a predecessor of α, γ is a
predecessor of β, and so on until we reach a node λ for which α is a
predecessor. (This, after all, is what we mean by a cycle.)
• But then, by restriction 2, α is a predecessor of α.
• And that result violates the first restriction.
One way to proceed
• Finally, restriction 3 implies that there cannot be two or more branches
leading to α.
• If there were, then there would be two associated nodes, say, β and γ ,
that are predecessors of α.
• However, it must then be the case that either β is a predecessor of γ
or vice versa.
• Put differently, the road from, say, γ , has to come through β
Strategies, Mixed Strategies, and Chance Nodes
Strategies
• A player’s strategy is a complete, conditional plan of action.
• It is conditional in that it tells a player which branch to follow out of a
decision node if the game arrives at that node.
• It is complete in that it tells him what to choose at every relevant
decision node
Example: Extensive form game example
• Suppose that two players are on their way to see a concert that is in
great demand
• The demand is so great that there is exactly one ticket left
• Whoever arrives first will get that ticket (game)
• The first player (player 1) leaves home a little earlier than player 2;
• In that sense he makes his choice at the root of the game tree and
subsequently the other player makes her transportation
choice.
Example: Extensive form
Conditional Strategies
• Player 1 makes only one choice—at the root
• Hence he has three strategies to pick from—take bus (b), take cab (c), or take
subway (s)
• Player 2 is faced with three possible conditionalities: what to do if player 1 takes
the bus, what if he took a cab, and, what if player 1 hopped the subway.
• Hence, each strategy of player 2 has three components, one component for every
conditionality.
• A representative strategy is; take a cab if player 1 takes the bus, bus if player 1
takes a cab, and subway if player 1 hops a subway as well.
• Since there are three possible ways to choose in every conditionality, player 2 has
3 × 3 × 3, that is, 9 such strategies.
Mixed Strategies in extensive form
game
• A mixed strategy is defined in exactly the same way as in the strategic
form
• It is simply a probability distribution over the pure strategies
• So in the sequential theater game a mixed strategy for player 1 is given
by two numbers p and q
• These are respectively, the probabilities with which b and c are chosen
(and 1 − p − q is the probability with which s is picked)
Chance Nodes
• We can also build uncertainty that is inherent to the game
• Opposed to uncertainty that the players introduce via mixed
strategies into the extensive form
• For instance, the amount of time it takes on the subway might depend
on whether or not there is a rush-hour delay in the subway system
• One way to model that possibility is to allow for a third kind of
node, called a chance node
• This is a node whose branches represent several random
possibilities
Example: Chance Node
• Suppose that there are two possible subway outcomes—delay or no
delay.
• This uncertainty needs to be incorporated into the extensive form.
• Exactly how it will be incorporated will depend on when this
uncertainty is resolved
• Do the players know whether or not there is a delay before they make
their choices, and so on.
• For simplicity: suppose that when our theater-goers make their
transportation choice they do know whether there is a delay or not
Chance Node
Perfect Information Games
• One in which there is no information set (with multiple nodes).
• If an information set has three nodes, then a player cannot tell which
of the three immediately preceding nodes is the one that was actually
played, although she knows that one of them must have been played
• If, on the other hand, an information set has a single node then there
is no such ambiguity
• Any time a player has to move she knows exactly the entire history of
choices that were made by all previous players
Example (Entry 1)
• Consider A firm—say, Coke—is debating whether or not to enter a new market—
say, the Former Soviet Union (FSU)—where the market is dominated by its rival,
Pepsi.
• Coke’s decision is guided by the potential profitability of this new market
• Depends principally on how Pepsi is going to react to Coke coming into its market.
• If Pepsi mounts acts “tough” (big advertising campaign, spends a lot of money
upgrading facilities, ties up retailers with exclusive contracts)—then Coke will lose
money
• If Pepsi were not to mount such a tough counterattack—Coke would make money
• E (for enter) and O (for stay out) stand for Coke’s alternatives, T (for tough) and A
(for accommodate) refer to Pepsi’s two choices.
Perfect Information (Entry 1)
Example (Entry 2)
• Suppose that after Pepsi’s decision, Coke has a further decision to
make
• It has to decide whether or not it will itself mount an aggressive
advertising campaign and spend a lot of money on facilities, and the
like
• In other words, suppose that after observing Pepsi’s response,
Coke will itself have to act “tough” or “accommodate”
Example (Entry 2)
Example (Entry 3)
• Suppose that, should Coke enter the market, both Coke and Pepsi will
make a decision about how much to invest in this market
• That is, whether to act tough or accommodate.
• However, unlike Example 2, suppose these decisions are taken
simultaneously (not a game of perfect information)
Example (Entry 3)
Backward Induction
• What is a reasonable prediction about play in examples 1 and 2?
• It will turn out that this is really a question about sequential rationality
• It will involve rationality because a player will pick the best action
available to him at a decision node, given what he thinks is going to be the
future play of the game
• It will involve sequentially because a player will infer what this future is
going to be knowing that, in the future, players will reason in the same
way
• In particular, the decision maker at a subsequent node will pick the best
available action given what he, in turn, believes about the remaining
future of the game
Strategic form representation
Coke\Pepsi Tough Accomodate

Enter -2,-1 1,2

Out 0,5 0,5

Nash Equilibrium
• Two Nash equilibria of this game: (enter, accommodate) and (out, tough)
• The Nash equilibrium (out, tough) is, however, unreasonable: Pepsi undertakes
to fight Coke if Coke were to enter the market
• But if Coke were to enter the market, Pepsi would be better off accommodating.
• Indeed Pepsi’s strategy, tough, is a best response—to out—only because that
strategy is actually never used, since, anticipating a tough response, Coke chooses
to stay out of the market
• However, Coke might not find a tough stand by Pepsi credible precisely for this
reason
• By this line of logic, the only reasonable equilibrium behavior is
for Pepsi to accommodate; hence, (enter, accommodate) is the only reasonable
Nash equilibrium
Strategic form representation:
Example 2
Coke\Pepsi Tough Accomodate

ETT -2,-1 0,-3

ETA -2,-1 1,2

EAT -3,1 0,-3

EAA -3,1 1,2

OTT 0,5 0,5

OTA 0,5 0,5

OAT 0,5 0,5

OAA 0,5 0,5

Strategic form
• Note that every strategy of Coke’s must have three components
• First component tells Coke whether or not to enter the market
• Second tells it whether or not to act “tough” if Pepsi acts “tough,”
• Third specifies behavior if Pepsi accommodates
• For example, EAT means (1) enter,
(2) against a tough Pepsi, accommodate, (3) against an accommodating
Pepsi, act tough.
• Pepsi has exactly two strategies—either to act tough or to accommodate
There are essentially three pure-strategy Nash
equilibria of the strategic form
• Nash equilibria in which Pepsi plays T and Coke plays any one of the
(four) strategies in which it stays out—OTT , OTA, OAT , or OAA
• (ETA, A)—with outcome that Coke enters and both firms
accommodate
• (EAA, A)—with the same outcome as in the second equilibrium.
Logic of sequential rationality
• What should Pepsi’s action be?
• Depends on whether Coke will subsequently act tough or
accommodate
• For example, it is more profitable for Pepsi to accommodate if it
thinks Coke will accommodate as well
• But it is better for Pepsi to fight if it thinks Coke will act tough.
• In order to determine which of these options will be chosen by Coke—
and therefore what Pepsi should do—we can apply the logic of
sequential rationality twice
• Suppose that Pepsi accommodates
• At this point it is more profitable for Coke to accommodate than to fight
• Hence, the only credible choice for Coke is to accommodate.
• On the other hand, if Pepsi acts tough, Coke will find it more profitable
to fight
• Knowing Coke’s responses, Pepsi now has to compare the profits from
(T , T ) against (A, A), that is, the two profit levels of −1 and 2.
• Pepsi will therefore accommodate.
Sequentially rational strategy
• One can, finally, back this logic out to Coke’s initial decision
• Coke can either enter—and then it expects Pepsi to accommodate and
expects to do the same thing itself.
• Or it can stay out
• The profit level in the first case is 1, while it is 0 for the second option
• Coke enters.
• The only sequentially rational strategy for Coke is ETA, while for
Pepsi it is to play A
• The only one of the three types of Nash equilibria in the strategic
form that is sequentially rational is the second one
The Power of Commitment
• Having fewer choices is typically worse than having more choices
• You might think that this should be true for games as well.
• However extensive form games, and in particular games of perfect information,
provide examples where less (choices) can mean more (equilibrium payoffs).
• This statement may sound paradoxical at first, but the intuition is actually
straightforward.
• If a player has more options later, she may behave very differently in the future
than she would if she had fewer options.
• In turn, this behavior will affect current play by her as well as by the other
players.
• This change can, be beneficial or harmful to the player with enlarged options.
Example 1: Only Tough Pepsi
• After Coke enters the market
• Pepsi has no choice but to play tough
(i.e., let us reduce Pepsi’s options by eliminating the choice accommodate)
• Since Pepsi has only one option, it will necessarily exercise that option
and fight Coke’s entry.
• So Coke suffers a loss if it enters this market (Coke will prefer to stay out)
• By having fewer options—or, by committing to one option in the future—
Pepsi is able to increase its (backward induction) equilibrium payoffs.
Example 1: Only Tough Pepsi
Example 2: Only Tough Coke
• Suppose after Coke enters the market
• Coke has no choice but to play tough (i.e., suppose Coke has one
option less; it cannot accommodate)
• This fact implies that while making its choice Pepsi knows that it will
surely face a tough opponent in the future
• Pepsi’s best response is therefore also to play tough
• Consequently, at the entry stage Coke is better off deciding to stay out,
since coming in will entail a loss of 2
• Here fewer options for Coke benefits Pepsi because it renders Pepsi’s
threat to be tough credible
Example 2: Only Tough Coke
Backward Induction: A General
Result
• The solution concept employed in the two examples can be
generalized
• The generalization goes by the name backward induction
• The logic:
1. Suppose the game is at a final decision node
2. Any decision by the player who chooses at that node terminates the
game
3. The only reasonable prediction for play is that this player will pick
that action which maximizes her payoffs
Logic of backward induction:
Example
• For instance, in example 2,
• Coke as the final decision maker gets a higher profit from playing
tough if Pepsi is being tough
• Hence, a rational Coke must pick tough.
• In the other final decision node, that which follows Pepsi
accommodating, Coke must pick accommodate because it gets a
higher profit by doing so
Logic of backward induction:
Example
• Consider now a penultimate decision node
• —for instance, Pepsi’s decision node in example 2.
• At any such node, the decision maker knows the exact consequence of each
of his choices because he knows the subsequent decision that will be taken
at the final decision node.
• For example, Pepsi’s tough stance will be reciprocated; so will an accom-
modating stance.
• Hence a decision maker at the penultimate decision node can compute
the exact payoff to each of his choices, and he must make the best
available choice
Backward Induction: A General
Result
• By similar logic, at decision nodes three steps removed from a
terminal node, the decision maker knows the exact consequence of her
choices.
• This is the case because she knows the choice that will be made by any
player in a penultimate decision node as well as the choice that will be
made at the consequent final decision node.
• Hence such a three-step removed decision maker has a best choice.
And so on.
Backward Induction: A General
Result
• We fold the game tree back one step at a time till we reach the beginning.
• The fact that we start at the end of the tree is the backward part of the
terminology
• The fact that we work one step at a time in doing the tree folding is
why
it is called an induction procedure.
• Note that this procedure works as long as there is a last node to start from
• The above arguments yield the result called Kuhn’s theorem.
• A special case of this result was proved in 1913 by E. Zermelo, who
showed that the game of chess must always have a winning strategy
Kuhn’s (and Zermelo’s) Theorem
• Every game of perfect information with a finite number of nodes has a
solution to backward induction.
• Indeed, if for every player it is the case that no two payoffs are the
same, then there is a unique solution to backward induction.
Connection with IEDS in the
strategic form
• Backward induction in the extensive form of a game turns out to be
exactly the same as solving the game by iterated elimination of
dominated strategies (IEDS) in the strategic form.
Coke and Pepsi (Entry 1)
Coke\Pepsi Tough Accomodate

Enter -2,-1 1,2

Out 0,5 0,5

Backward induction similar to IEDS
• The backward induction outcome was for Coke to enter and for Pepsi
to accommodate.
• Note that tough is a dominated strategy for Pepsi.
• Hence, the IEDS outcome is indeed (enter, accommodate).
The first, third, and fourth strategies of Coke
are dominated by ETA
Coke\Pepsi T A

ETT -2,-1 0,-3

ETA -2,-1 1,2

EAT -3,1 0,-3

EAA -3,1 1,2

OTT 0,5 0,5

OTA 0,5 0,5

OAT 0,5 0,5

OAA 0,5 0,5

Reduced Matrix
Coke\Pepsi T A
ETA -2,-1 1,2
OTT 0,5 0,5
OTA 0,5 0,5
OAT 0,5 0,5
OAA 0,5 0,5
Solution
• For Pepsi, T is now dominated by A.
• The IEDS outcome is in fact (ETA, A) exactly as we saw by way of
backward induction in the extensive form.
Poison pill and other takeover
deterrents
Legal Poison pill 1
Solution
• A company N has to decide whether or not to fight company C
• On one hand, if they do decide to fight (and make a share offer), C
can either play tough by refusing to negotiate, or they may
accommodate
• Suppose that with a poison pill provision C is committed to fight
• Additionally, now suppose that there is an initial choice that C has to
make, and that is to decide whether or not to arm themselves with the
poison pill
Solution
• Without the poison pill, C will accommodate and hence N will enter
the former’s profits are therefore 2.
• With the poison pill, N will choose to stay away from the takeover,
and hence C will make profits of 3.
• Clearly, C prefers to adopt the poison pill, and this commitment nets
them an extra dollar (billion dollars) of profits.
Legal Poison pill 2
Legal Poison pill 2
• But didn’t Norfolk Southern actually make a takeover offer?
•
Well, perhaps then the payoffs—in the extensive form—are slightly different.
• As before, without the poison pill, C ’s payoff is 2.
• Now, though, despite the poison pill, N finds it profitable to enter (since they make
profits of 0.5).
• However, C makes a profit of 2.5, which is still better than the 2 they would make
by not adopting the poison pill.
• So the backward induction outcome is that C prefers to adopt the poison
pill provisions,
• N prefers to make a share offer (mount a takeover of C), and C fights that offer
Exercise 1: strategic investment to deter entry
Exercise 2: Agency
Exercise 3: A public decision process
Summary
• The extensive form representation specifies who moves at different
points in a game, what their choices are at such moves, and what the
eventual consequences are for all players.
• Restrictions are placed on the precedence between decision nodes to
ensure that there is a well-defined play for every choice of strategies.
• A game of perfect information is one in which every information set
has a single element.
Summary
• Backward induction is a general solution concept for games of perfect
information.
• Every finite game of perfect information has a backward induction solution.
• Backward induction in the extensive form of a game of perfect information is
identical to Iterated Elimination of Dominated Strategies in the associated
strategic form.
• In a perfect information game, being able to commit to having fewer choices in
the future may be beneficial for a player.
• Poison pills are commitment devices that restrict management’s options in the
event of an unfriendly takeover. Consequently, they may increase the payoffs of
a target firm
Application
• A technological breakthrough can be profitably utilized by more than
one company, and hence each company has an incentive to let somebody
else make a costly breakthrough. Patents are a way of solving this
incentive problem.
• In many industries, firms battle furiously to win patents on new
technologies; high-definition television (HDTV) is an example of an
ongoing patent race.
• A two-firm patent race in which the competitors take turns making R&D
investments can be modeled as a game of perfect information.
Furthermore, this game can be solved using backward induction on
location space.
Application
• The backward induction solution has the feature that two firms that
are similar distances from project completion invest heavily to get an
R&D advantage. A firm that falls sufficiently behind is better off
dropping out of the patent race.
• Dropping out is less likely the higher the value of the patent and the
lower the costs of doing R&D. These characteristics describe the
current HDTV race.