134 PrciiJllillaries

I

2 '\1
'2
()
4 :1 :1 8
"j 3 2
FIGURE 7.3 A game tree wilh a '\lmnIllY" player.
7 0 2
<1
:1
The precedence relation, together with the way in which players are assigned
to nodes, describes the way in which the game unfolds. For example, in Figure 7.3
player I is assigned to the root, so i (xo) = I. His action set at the root, AI (xo), includes
two choices that determine whether the game will terminate at node X4 with payoffs
(2,0, 4, 3), or whether player 2 will gel to play at node xI' Player 2 then can choose
whether player 4 will play at or x3, and at each of these nodes player 4 has two
choices that both end in termination of the game. Player 3 has no moves to make.
2
There is still one missing component: how do we describe the knowledge of each
player when it is his turn to move? In Figure 7.3 we sec that player 2 moves after
player I. Would player 2 know what player I did? In this example we would think
the answer should be obvious: if player 1 chose his other action, the game would end
and player 2 would not have the option of choosing an action. Moving down the game
tree to player 4, however, raises some questions. It seems implicit in the way we drew
the game tree in Figure 7.3 that player 4 knows what happened before he moves, that
is, if he is at node X2 or at node X3' Hence if player 4 knows where he is in the game
tree then he must know what player 2 did before him.
Perhaps the game that is played actually calls on player 4 to make his move without
knowing what player 2 did at Xl' How can we describe this situation in a game tree?
Clearly we need to find a way to represent the case in which player4 cannot distinguish
between being at x2 and being at x3' That is, we need to be able to make statements
like "I know that I am at either X2 or X3, but I don't know at which of the two I am."
We proceed to put structure on the information that a player has when it is his turn
to move. A player can have very fine information and know exactly where he is in
the game tree, or he may have coarser information and not know what has happened
before his move, therefore not knowing exactly where he is in the game tree. We
introduce the following definition:
2, Note that the moves from any nontenninal node result in a move to another node in the game. Thus
we can save on notation and give moves "names" that are consistent with the nodes in which they
will result. For our current example we could write A I (xo) {xj, x4}, A
2
(Xl) {x2, x3}, A
4
(x2) =
{xs, x61, and A
4
(x3) {x7, Xg}.
7.1 'fllc Exlcllsivl:hmn(lame T.E) 4>
()
()
I
o 2 o 2
FIGURE 7.4 'rhesimultaneous-moveBattleoftheSexesgame.
Definition 7.2 Every player i has a collection of information sets hi E Hi Ihal
partitionthenoder;of/hegameatwhichplayeri moveswiththefollowingproperties:
1. Ifhi is asingletonthatincludes only x then playeri who moves at x knowr;
thatheis atx.
2. If x I=- Xl andifbothx E hi andXl E hi thenplayeri whomovesatx doesnot
know whetherheis atx orXl.
3. Ifx I=- Xl andifbothx E hi andXl E hi then A;(x') Ai(x).
Theformaldefinitionbuildsonasimpleidea.Considerthesequential-moveBattle
ofthe Sexes gamein Figure7.2 and observe thatplayer2 moves atXl' We wantto
describewhetherornotheknowsthatheisatXl'Ifwewrite112 = [Xl}'thismeansthat
the information setat Xl is asingleton (itincludes only the nodeXl)' Henceplayer
2 has information that says "} am atxJ," which is captured by property (1) ofthe
definition. In this caseit will follow thatplayer2will haveanotherinformationset,
h; = {x2}'
If, in contrast, we want to represent a game in which player 2 does not know
whetherheis atXI orx2, then itmustbethecasethathis informationis "1 knowthat
IamateitherXl orX2, butIdon'tknowatwhichofthetwoIam."Thuswewill write
h2 {Xl> x2}, whichexactlymearisthatplayer2cannottellwhetherheis atXl orX2'
Thisis theessenceofproperty(2) ofthedefinition.
Finally,property(3)isalsoessentialtomaintainthelogicofinformation.If instead
X E hi andx' E hi butAi (x') Ai (x), thenbythemerefactthatplayeri hasdifferent
actions from which to choose at each ofthe nodes X and x', he should be ableto
distinguish betweenthesetwo nodes. Itwouldthereforebe illogical to assume that
hecannotdistinguishbetweenthem.
Wearelefttoconstructagraphicalrepresentationtoshowwhichnodesbelongin
thesameinformationset.InFigure7.4wepresentthetwocommonwaysofdepicting
this using theextensive-formrepresentationofthesimultaneous-moveBattleofthe
Sexesgamethatwehavealreadyseenandanalyzed.Player1choosesfromtheaction
setAl= {O, F}, andplayer2choosesfromA2 = {o, n, without observing thechoice
ofplayer 1. On theleftsideofFigure7.4weuseanellipsetodenotean information
set, andall thenodesthatareinthesameellipsebelongtothesameinformation set.
Inthisexampleplayer2cannotdistinguishbetweenXl andX2, sothath2 {Xl, X2}'
OntherightsideofFigure7.4isanothercommonwayofdepictinginformationsets,
accordingtowhichthedashedlineconnectingXl withX2 denotesthatbothareinthe
sameinformationset. Anotherexampleofasimultaneous-movegamedepictedas a
136 ... (:lWpltT 7 Preliminaries
R // P

2
o --J o -J -J o
o o -I o
FIGURE 7.5 Game tree of rock-paper-scissors.
game tree is the game of depicted in Figure In this game
player J chooses from the set A I {R, P, S} while player 2 chooses from the set
A2 /r, p, s}, without ohserving the choice made by player I.
7.1.2 Imperfect versus Perfect Information
We defined games of complete in Chapter 3 as the situation in which
each player i knows the action set and the pay01T function of each and every player
.i N, and this itself is common knowledge. This definition sufficed for the normal-
form representation. For extensive-form games, however, it is useful to distinguish
between two different types of complete-information games:
Definition 7.3 A game of complete information in which every information set is a
singleton and there are no moves of Nature is called a game of perfect information.
A game in which some information sets contain several nodes or in which there are
moves of Nature is called a game of imperfect information.
In a game of perfect information every player knows exactly where he is in the
game by knowing what occurred before he was called on to move. Examples would
be the trust game in Figure 7.1 and the sequential-move Battle of the Sexes game
in Figure 7.2. In a game of (complete but) imperfect information some players do
not know where they are because some information sets include more than one node.
This happens, for example, every time they move without knowing what some players
have chosen previously, implying that any game is a game of
imperfect information. Examples include the simultaneous-move Battle of the Sexes
game shown in Figure 7.4 and the rock-paper-scissors game depicted in Figure 7.5.
Games of imperfect information are also useful to capture the uncertainty a player
may have about acts of Nature. For example, imagine the following card game: There
is a large deck that includes an equal number of only kings and aces, from which
player 1 pulls out a card without looking at it. The probability of getting a king is 0.5,
and we can think of this as Nature's move. Hence player 1 moves after Nature and
does not know if Nature chose a king or an ace. After drawing the card, player 1 can
call (C) or fold (F). If he folds, he pays $1 to player 2. If he calls, he pays $2 to player
2 if the card is a king, while player 2 pays him $2 if the card is an ace.
We can accurately describe this game in one of the two ways depicted in Figure 7.6.
To see this, consider the game tree in the left panel. The order of appearance is loyal
to the story: N atme chose K or A. Player 1 docs not know what happened, but he
knows that he is at either node in his information set with probability i. Then player
1 makes his move and the game ends. The game on the right looks different but is
SI r:ilcgies and Nash Equilibriulll "131
N
c F
2
FIGURE 7.6 A card game.
strat.egically equivalent. Player I makes his move (C or F) without knowing which
card was drawn, and then Nature draws the card, K or A, with equal probability.
Therefore in both games we are depicting exactly (he same information and moves,
ami hence we are geUing the story right.)
Therc is an insight worth emphasizing here. A game will be one of impcrl'ecl
information when a player must make a move either without knowing the move 01'
another player or without knowing the realization or a choice of Nature. Uncertainly
over the choice of Nature, or exogcnous ullccrtainty, is at the heart of the si ngle
person decision problems that were described in Chapter 2. (And the card game
in Figure 7.6 is effectively a decision problem because player 2 has
no choices to make.) Uncertainty over the choice of another player, or endogenous
uncertainty, is the subject of games like the simultaneous-move
Battle of thc Sexes game in Figure 7.4. Notice, however, that both situations share
a common feature: occurrences that some player does not know are captured by
uncertainty over where he is in the game, be it from exogenous or endogenous
uncertainty. In either case a player must form beliefs about the unobserved actions,
of Nature or of other players, in order to analyze his situation.
Strategies and Nash Equilibrium
Now that we have the structure of the extensive-form game well defined, and have
developed game trees to represent this structure, we move to the next important step
of describing strategies. Recall that in Section 3.1 we argued that "a strategy is often
defined as a plan of action intended to accomplish a spec{jic goal." In the
form game it was very easy to define a strategy for a player: a pure strategy was
some element from his set of actions, Ai' and a mixed strategy was some probability
distribution over these actions. As we will now see, a strategy is more involved in
extensive-form games.
7.2.1 Pure Strategies
Consider the Battle of the Sexes game described again in Figure 7.7.
(We will now seldom include the names of nodes in our game trees because they have
... ...... ---
3. This is the reason that the definition of games of perfect or imperfect information explicitly requires
the reference to moves of Nature. It is sometimes possible to include them in a game tree in which
all the information sets are singletons.
'138 .. Clwpler7
() F
2 2
2 o o
o o 2
FIGURE 7.7 The Battleofthe Sexesgame.
norealconsequencefortheissllcswithwhichweareconcerned.)Player] hasasingle
information scI with one node, sofor him a purestrategy is as simple as "play 0"
or"play F."Forplayer2, however, thingsarea bitmoreinvolved. Player2 has two
information sets, each associated with a difrerentactionofplayer 1. Hence the two
simplcstalements"play0"and"playf" donot seem toexham;tall thepossibilities
forplayer2. In particular,player2canchoosethefollowingratherattractivestrategy:
"Ifplayer I plays 0 then 1will play0, whilcifplayer I plays F thenI willplay f."
This simpleexample demonstrates that when a player's move follows aftcrthe
realizationofpreviollsevents in thegame,and iftheplayercan distinguishbetween
thesepreviollsevents(Lhey resultin differentinformationsets),thenhecancondition
his behavior011 theevents thathappened. A strategyis thereforeno longera simple
statement ofwhataplayerwill do,asinthenormal-formsimultaneous-movegame.
Instead wehave
Pure Strategies in Games A pure strategy for player i is a
completeplan that describes whichpureaction playeri willchooseateach
ofhis informationsets.
If we consider the simultaneous-move Battle ofthe Sexes game in Figure 7.8,
the pure strategies for player I are Sj = {O, F}, and those for player 2 are S2
{o, fJ. Becauseeachplayerhasonlyoneinformationset,theextensive-formgameis
identicaltothesimplenormal-formgamewehavealreadyencountered.Incontrast,in
thesequential-moveBattleoftheSexesgameinFigure7.7,player2hastwodistinct
information setsin whichhecanchoose() or f,eachinformationsetresultingfrom
the previously made choice ofplayer 1. Thereforea "completeplan ofplay" must
accommodatea strategy that directs what player 2 will choose for each choice of
o F
2 o o
1 o o 2
FIGURE 7.8 Thesimultaneous-moveBattleoftheSexesgame.
7.2 Strategies "lid Nash EqllilihriulIl "139
player I. 'fhat is, player 2' s c1lOice or nel iOll "rom the set I0, I J can he Ilwde
on what player I does, Udlllillillg Ilie possibility oCtile fOlTn "Ifplaycr I
plays 0 thell J will play 0, wilile if player I plays F thell I will play f."
}"or this example, we can describe the set of pure strategies for p!;lyer 2 as follows:
,)'2 {00, oI, fo, .II},
where a pure strategy "ah" is shorthand for "I will playa if player I plays 0 and I
wi II pJ ay h if he plays F." r;'or player I the pure strategy set remaills S1'= {O, F}.
We now introduce some notation that builds on what we have already developcd
in order to define formally a pure strategy. Let Hi be the collection of all informalion
sets at which player i plays, and let hi E Hi be one of i 's information sets. Let Ai (hi)
be the actions that player i can take at hi, and let Ai be the set of al I actions of pI ayer
i, Ai = U
h
.
0
l/A
i
(h
i
) (i.e., the union of all the elements in all the sets Ai(h
j
. We can
I . I
now define a pure strategy as follows:
Definition 7.4 A pure strategy for player i is a mapping .}'j : Ai that assigns
an action siehl) E Ai(h;) for every information set hi E Hi' We denote by Si the scI
of al1 pure-strategy mappings Sf E Si'
A final point is in order, and the sequential-move Battle of the Sexes game ill
li'igure 7.7 is useful to illustrate it. Notice that even though player 2 has only two
actions from which to choose, by moving after observing what player J has chosen,
his strategy defines actions that are conditional on his information about where he is
in the game. In this example the two actions translate into four pure strategies because
he has two information sets.
This observation implies that a potentially small set of moves can translate into a
much larger set of strategies when sequential moves are possible, and when players
have knowledge of what preceded their play. In general assume that player i has k > ]
information sets, the first with m J actions from which to choose, the second with IIl2'
and so on until Ink' Letting lSi I denote the number of clements in Si' the total number
of pure strategies player i has is
lSi 1= m I x m2 x ... x mk
For example, a player with 3 information sets, 2 actions in the first, 3 in the second,
and 4 in the third will have a total of 24 pure strategies.
7.2.2 Mixed versus Behavioral Strategies
Now that we have defined pure strategies, the definition of mixed strategies follows
immediately, just like in the normal-form game:
Definition 7.5 A mixed strategy for player i is a probability distribution over his
pure strategies Si E Sf
How do we interpret a mixed strategy? In exactly the same way that we did for
the normal form: a player randomly chooses between all his pure strategies-in this
case all the complete plans of play-and once a particular plan is selected the player
follows it.
You may notice that this interpretation takes away some of the dynamic flavor
that we set out to capture with extensive-form games. More precisely, when a mixed
140 ~ ChaplDI' 7 Pre! illlillaric:-;
slr:licgy is used, the player selecls (I plan randolllly he/ore the gallic is played and (hen
rollows a jJ(/rticular pure strategv.
This description of mixed :-;Iralcgies wa:-; scn:-;iblc for nonnal-form games hecause
there it was a oncc-and-for-all choice to be macle. In a game trec, however, fhe
player may want to randomize at i-lomc nodes, independently of what he did al curlier
nodes. In other words, tbe player may wanl 10 "cross the bridge when he gels there."
This cannot be captured by mixed i-llrategies as previously deflned because oncc the
randomizal ion is over, the player is choosing a pure plan of action.
To illustrate this point, consider again the sequential-move Battle of the Sexes
game in Figure 7.7. The previous delinition of a mixed strategy implies thai player 2
can ntndOlnil',e among any of his four pure strategies in the set S2 ={oo, of, fo, ffl.
It does not, however, allow him to choose strategies of the form "If player J plays 0
then l' II play f, whi le if he plays F then J'II mix and play f with probability 1."
To allow for strategies that let players randomize as the game unfolds we define a
lIew concept as follows:
Definition 7.6 A bclutvioral strategy speeifjes for each information set hi Hi
an independent probability distribution over A;(h
i
) and is denoted by (Ji : Hi ...-+
6A;Ch
i
), where (J;(0i(h
i
is the probability that player i plays action ai(h
i
) E Ai(h
i
)
in information set hi'
Arguably a behavioral strategy is more in tune with the dynamic nature of the
extensive.form game. When u::;ing sllch a strategy, a player mixes among his actions
whenever he is called to play. This differs from a mixed strategy, in which a player
mixes before playing the game but then remains loyal to the selected pure strategy.
Luce and Raiffa (1957) provide a nice analogy for the different strategy types we
have introduced. A pure strategy can be thought of as an instruction manual in which
each page tells the player which pure action to take at a particular information set, and
the number of pages is equal to the number of information sets the player has. The
set Si of pure strategies can therefore be treated like a library of such pure-strategy
manuals. A mixed strategy consists of choosing one of these manuals at random and
then following it precisely.
In contrast a behavioral strategy is a manual that prescribes possibly random
actions on each of the pages associated with play at particular information sets.
To see this, consider the sequential-move Battle of the Sexes game in Figure 7.9.
Player 2 has two information sets associated with the nodes XI and X2, which we
denote hf and hf, respectively, and in each he can choose between two actions
in A2 = {o, fl. A pure strategy would be an element from S2 {oo, of, fo, .tf}
A mixed strategy would be a probability distribution (Poo' Pof, Pro' Pff)' where
PS
2
2: 0 and P
oo
Pol Pro Pff = 1. Denote a behavioral strategy as four prob-
abilities, (J2(0(hf, (}2(f(hf, (J2(0(hf, and (}2(f (hf", where (J2(0(hf +
(J2CfChf = (J2(o(hf) (J2Cf(hD) = 1. In Figure 7.9 we have used (}2(0(hf =
1, (}2Cf(hf = ~ and (J2(o(hf = (J2Cf(hn) = i
Having defined two kinds of nonpure strategies, the obvious question is whether
we need to consider both kinds of strategies for a complete description of possible
behavior, or whether it suffices to consider only one kind or the other. We can
answer this question by answering two complementary questions. The first is, given
a mixed (not behavioral) strategy, can we find a behavioral strategy that leads to
the same outcomes? Using our example, can we replace a mixed strategy of the
7,2 Strategic:; :IIHI Nash Eqllilihriulll ..14')
()
2
() 2
FIGURE 7.9 Behavioral strategies in the sequential,move Battle of the Sexes game.
form (POII' Pol' PIlI' PJJ) with L'\'ES, P,I = ) with a behavioral strategy of the fonn
(o'2(o(hf, o'2(f(h2)), a;(f(h.n that leads to the same randomization
over outcomes?
The answer is yes, which is quite easy to sec. Conditional on reaching xI t.he
probability or playing () is Pr(oIO} :::::: p()o+- Pof" and conditional on reaching
XI the probability of playing f is Pr{fIO} Plo+ Pff' so that Pr{oIO}
PrUIO} I. Sirnilarly, conditional on reaching X2 the probability of playing
() isPr{olF} = Poo -+ Pro' and the probability of playing f is Pr{IIP'}:::c:::
Pof + Prt" so that Pr(ol F} + Pr{flF} = 1. 'I'hus we can define a behavioral strat,
egy (}2(a2(h)) Pr{a2Ih) that yields the same randomization as the mixed strategy
(Poo' Pol' Pfli' PfI)'
The complementary question is whether, given a behavioral strategy, we can
find a mjxed strategy that leads to the same outcomes. Again, turn to the example
as an illustration and consider the behavioral strategy shown in Figure 7.9, where
0"2(0(0 1, a2(f(0 = and a2(0(F a2(f(F Notice that if the
player lIses a mixed strategy (Poo' Pol' Pill' Plf) then conditional on player I choos-
ing 0, action 0 will be chosen with probability }J()O + Pol' and action f will be chosen
with probability Pro Po' Similarly, conditional on player I choosing action ()
will be chosen with probability Poo + Pro' and action f will be chosen with prob-
ability PoI VII' 'I'hus to replicate the behavioral strategy, four equalities must be
satisfied by the mixed strategy (1'00' Pol' Pro' Plf):
1
Pr {oIO} = P00
Pof = 3'
Pr{fIO} = Pro +PtJ =
1
Pr{oIF} = Poo + Pfo
2:
1
Pr{f IF) = Pot
PIt = 2'
It may seem like we have four equations with four unknowns, but this system
of equations is "underidentified" because these four equations are really only two
equations. By definition it must be true that
Poo + Pol + I' fo PIf = 1,
141 .. ('Iwplcr 7 Preliminaries
wllich lllal h:lvC oilly equations wilh four It follows that
/I/(/II.\, vahlcs ur (1'00' PoI' I) PO) wi II satisfy the equHl iOIlS, j Illpl ying that this par-
behavioral slrategy can be generated by all inllllile IlIlmber of different mixed
strategies. hll' exalllple, (1'00' Pol' Pro' PII) 0,0, /)' will lead 10 equivalent
outcollles as the bdwvionli slralegy, but so will (Poo, Pol' 1'/0,1'/1) - (A,
This example may suggest Illat any randomization over play can be represented
hy eilher l1lixed or belwviOr:lI slrategies. As it turns OUI, lhis is true under a rather
mild condition.
Definition 7.7 A gamc or I}('I:/,ect recall IS one 111 which no player ever forgets
information that previollsly knew.
Thai is, a game of perfect recall is one in which, jf a player is called upon to
move more Ihan ollce in a gnmc, then he must reITlember the moves that he chose in
his prcvious information sets. Practically all of the analysis in game theory, and in
applications of game theory to the social sciences, assurnes perfect recall, as will we
in Ihis text. For the class of perfect-recall games, Kuhn (J (53) proved that mixed and
behavioral strategies arc equivalent, in the sense that given strategies of i 's opponents,
the same distribution over outcomcs can be generated by either a mixed or a behavioral
strategy of player i.
Remark (The Absent-Minded Driver) Despite our focus on games of perfect recall,
it may be interesting to dwell for a momcnt on the example of the "absent-minded
driver" that was introduced by Piccione and Rubinstein (1997) to depict a simple game
(in fact, a single-person decision problem) with imperfect recalL The story goes as
follows. An absent-minded driver is driving home along the highway. The first exit
on the road, exit I, takes him to a bad neighborhood, yielding him a payoff of O. Exit
2, farther down the road, is the best route to his home, yielding him a payoff of 4.
If he misses exit 2 then he will eventually get home the long way, yielding him a
payofT of 1 . Imperfect recall means that he cannot distinguish between exits 1 and 2
and therefore cannot remember whether he has already passed one exit or not. This
decision problem is depicted in Figure 7.10. Define a "planning" mixed strategy of
the player as a probability p that he will exit at any exit he passes (both nodes in
his single information set)-that is, the probability that the driver will commit to exit
when be is at an intersection. His expected payoff from this strategy is Op +4p(1 -
p) 1(1 p)2 2p + I, which is maximized at p As Piccione and
Rubinstein argue, a puzzle emerges once the driver finds himself at an intersection.
He knows that with some probability q he is at Exit J, and with some probability
(1 - q) he is at exit 2. The driver's payoff for choosing to exit with probability p is
now, at the intersection, q[4p(1- p) + 1(1 - p)2] + (1 q)(4p 1(1 p)], which
is equivalent to the planning problem only when q = 1, that is, only when he believes
for sure that he is at the first exit. This leads to an interesting dynamic inconsistency
in that the driver would plan one thing in advance and then rationally change his mind
once he finds himself at an intersection.
4. This follows because Poo + Pol' = together with Poo + Pof + Pro + PIf 1 implies that
Pfo + Pfl and, at the same time, Poo + Pfo = together with }Joo + Pol' + Pfo + PfI = I
implies that Pol .+. Pff Therefore two of the five equations are redundant.
Strategies(lJHI N:1SIl <I> 143
()
4
FIGURE 7.10 Theabsent-mindeddriver.
o F
()
F
2 o o 2 o o I
o o 2 o 2
Sequential Sil11tlltaneOll.'>

FIGURE 7.11 TheBattleoftheSexesgame: two ven-dons.
7.2.3 Normal-Form Representation of Extensive-Form Games
Considerthe two variantsoftheBattleoftheSexes game presented in Figure 7.II.
Thesimultaneous-moveversion in therightpanel is onethatwehaveseenbeforein
its matrixform, asfollows:
Player2
o F
o 2, 1 0,0

Player I
F 0,0 1,2 I
Now considerthe sequential-moveBattleofthe Sexes game depicted in the left
panel. Recall that S] = {O, F} and S2 too, of, f0, fn,where f0, for example,
meansthatplayer2playsf afterplayer1plays 0,whileplayer2playsafterplayer
1plays F.Thisgamecanberepresentedbya2 x 4 matrixas follows:
Player2
00 of .fa .ff
o
Player1
F
As this matrixdemonstrates,eachofthefourpayoffsintheoriginalextensive-form
gameisreplicatedtwice.Thishappensbecauseforanypurestrategyofplayer I, two