Comments on the Interpretation of Game Theory

Author(s): Ariel Rubinstein

Source: Econometrica , Jul., 1991, Vol. 59, No. 4 (Jul., 1991), pp. 909-924
Published by: The Econometric Society

Stable URL: https://www.jstor.org/stable/2938166

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to
Econometrica

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 Econometrica, Vol. 59, No. 4 (July, 1991), 909-924

COMMENTS ON THE INTERPRETATION OF GAME THEORY

BY ARIEL RUBINSTEIN1

The paper is a discussion of the interpretation of game theory. Game theory is viewed
as an abstract inquiry into the concepts used in social reasoning when dealing with
situations of conflict and not as an attempt to predict behavior. The first half of the paper
deals with the notion of "strategy." Its principal claim is that the conventional interpreta-
tion of a "strategy" is not consistent with the manner in which it is applied, and that this
inconsistency frequently results in confusion and misunderstanding. In order to prove this
point, the term "strategy" is discussed in three contexts: extensive games in which players
have to act more than once in some prespecified order, games normally analyzed using
mixed strategies, and games with limited memory. The paper endorses the view that
equilibrium strategy describes a player's plan of action, as well as those considerations
which support the optimality of his plan rather than being merely a description of a "plan
of action." Deviation from the perception of a strategy as a mere "plan of action" fits in
well with the interpretation of the notion "game" which is discussed in the second half of
this paper. It is argued that a good model in game theory has to be realistic in the sense
that it provides a model for the perception of real life social phenomena. It should
incorporate a description of the relevant factors involved, as perceived by the decision
makers. These need not necessarily represent the physical rules of the world. It is not
meant to be isomorphic with respect to "reality" but rather with respect to our perception
of regular phenomena in reality.

KEYWORDS: Game theory, interpretation, strategy, game form, perception, relevance.

1. INTRODUCTION

I APPROACH THIS PAPER with the view that game theory is not simply a matter of
abstract mathematics but concerns the real world. This does not mean that the
object of game theory is to predict behavior in the same sense as the sciences
do, or indeed, that it is capable of such a function. I view game theory as an
analysis of the concepts used in social reasoning when dealing with situations of
conflict. It is an abstract inquiry into the function and logic of social institutions
and patterns of behavior. As game theory is at once abstract and concrete, we
must build a bridge between the abstract formal concepts of the theory and
reality. An interpretation is a mapping which links a formal theory with
everyday language.
Despite the vital need to provide a plausible interpretation for the basic
primitives of game theory, especially when game theory is applied to economics,
I have found only a few discussions of this issue. Noteworthy exceptions are
Aumann (1987a) and Binmore (1983, 1987b). I believe that discussion or
application of game theory is utterly meaningless without a proper interpreta-
tion. This task cannot be left entirely to philosophers of science, for it consti-

1 This paper is based on my Walras-Bowley Lecture delivered in the North American Economet-
ric Society meeting at the University of Minnesota, Minneapolis, June 1988.
I am deeply grateful to my friends Dilip Abreu, Ken Binmore, John Moore, Martin Osborne,
David Pearce, Avner Shaked, and Asher Wolinsky for their comments, suggestions, criticism, and
most importantly their encouragement during the period in which this paper was written. Three
referees provided exceptionally thorough comments on the first version of the paper.

909

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
910 ARIEL RUBINSTEIN

tutes the very essence of the theory. In this paper I will bring together various
comments on the interpretation of game theory.
How should we interpret game theory's two most basic primitives, "game
form" and "strategy?" The standard interpretation of a game form is that it
represents an exact and full description of the physical rules of a given of a
situation. It includes a list of decision problems for each of the participants in
the game. The information sets itemize the information available to the decision
makers concerning random events and other players' previous moves. The game
tree presents a chronology of potential events. The standard view is that this
description is exhaustive. To quote Kohlberg and Mertens (1986, fn. 3): "We
adhere to the classical point of view that the game under consideration fully
describes the real situation-that any (pre)commitment possibilities, any repeti-
tive aspect, any probabilities of error, or any possibility of jointly observing some
random event, have already been modelled in the game tree." A strategy is
usually interpreted as "a plan of action." M. Shubik refers to a strategy as "a
complete description of how a player intends to play a game, from beginning to
end." J. Friedman uses the phrase, "a set of instructions." These descriptions
are linguistically consistent with the Oxford Dictionary, which defines the word
"strategy" as "a general plan of action."
The first half of the paper deals with the notion of "strategy." Its principal
claim is that the conventional interpretation of a "strategy" is not consistent
with the manner in which it is applied, and that this inconsistency frequently
results in confusion and misunderstanding. In order to prove this point, I
discuss the use of the term "strategy" in three contexts: extensive games in
which players have to act more than once in some prespecified order (Section
2), games normally analyzed using mixed strategies (Section 3), and games with
limited memory (Section 4). I aim to endorse the view that equilibrium strategy
describes a player's plan of action, as well as those considerations which support
the optimality of his plan (i.e. preconceived ideas concerning the other players'
plans) rather than being merely a description of a "plan of action."
Deviation from the perception of a strategy as a mere "plan of action" fits in
well with the interpretation of the notion "game" which is discussed in the
second half of this paper. It is argued that a good model in game theory has to
be realistic in th. sense that it provides a model for the perception of real life
social phenomena (Section 5). It should incorporate a description of the
relevant factors involved, as perceived by the decision makers. These need not
necessarily represent the physical rules of the world (Section 6). It is not meant
to be isomorphic with respect to "reality" but rather with respect to our
perception of regular phenomena in reality (Section 7).

2. ON THE NOTION OF "STRATEGY" IN SEQUENTIAL GAMES

In an extensive game, a player's strategy is required to specify an action for

each node in the game tree at which the player has to move. In other words, a
player has to specify an action for every sequence of events which is consistent

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 INTERPRETATION OF GAME THEORY 911

C C C
1 2 1

S s s

FIGURE 1

with the rules of the game. In games which require a player to make at least two
consecutive moves (and most of the games which have been analyzed recently in
economic theory fall into this category), a strategy must specify his actions even
after histories which are inconsistent with the player's own strategy. For
illustration, consider the two player game form shown in Figure 1. According to
the natural definition of "strategy" as a complete "plan of action," player 1 is
required to specify his behavior, "Continue" or "Stop," at the initial node, and
if he plans to "Continue," to make provisional plans for his second decision
node in the event that player 2 chooses C. However, the game theoretic
definition of strategy requires player 1 to specify his action at the second
decision node, even if he plans to "Stop" the game at the first node.
Why does the notion of strategy as used by game theorists differ from a "plan
of action?"2 Could we not narrow the formal definition of a "strategy" to
specify an action only at the decision nodes which are not excluded from being
reached by the strategy? If we were only investigating Nash equilibria of
extensive games, then the game-theoretic definition would indeed be unneces-
sarily broad. The broad definition is, however, necessary for testing the rational-
ity of a player's plan, both at the beginning of the game and at the point where
he must consider the possibility of response to an opponent's potential deviation
(the subgame perfect idea). Going back to the example of Figure 1, assume that
each player plans to choose Stop at his first decision node. Testing the
optimality of player 2's plan following player l's deviation requires player 2 to
specify his expectations regarding player l's plan at his second decision node.
The component of player l's strategy after both players have chosen C provides
these expectations. Player l's strategy at his second decision node must be
interpreted as what would be player 2's (as opposed to player l's) belief
regarding player l's planned future play, should player 1 decide to deviate from
what was believed to be his original plan of action. Thus, a strategy encom-
passes not only the player's plan but also his opponents' beliefs in the event that
he does not follow that plan.3

2 In some circumstances it is natural to require a player to plan his behavior after his own
deviations. If the player is worried that he will make mistakes in the implementation of his
decisions, then a comprehensive plan of action should include a move for each history including
those which are the result of an error. This is the motivation behind Selten's trembling-hand perfect
equilibrium concept. However, if the possibility of error is significant enough to enter into the
player's considerations, then the technology of mistakes should be modeled explicitly as part of the
game.
3This interpretation was mentioned in Rubinstien (1982) for the perfect equilibrium in a
sequential bargaining model. Recently, Reny (1987) has used this interpretation to motivate his
explicable equilibrium notion.

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
912 ARIEL RUBINSTEIN

Interpreting a player's strategy after a deviation as the expectations of the

other players about his future behavior makes it problematic to speak of
"choice of a strategy." Player 1 does not choose player 2's belief. Game-theo-
retic models which use the notion of a subgame-perfect equilibrium (or any
other solution concept which includes a sequential rationality requirement)
therefore require re-examination. Consider, for example, the sequential bargain-
ing literature in which the authors assume (rather than conclude) that strategies
are stationary. That is to say, a player is confined by hypothesis to the use of
offers and response patterns (response to offers made by the other player) that
are independent of the history of the game. This literature presents this
stationarity assumption as an assumption of simplicity of behavior. Consider, for
example, player l's strategy: "Demand 50% of the surplus and reject any offer
which gives you less than 50%, independent of what has happened in the past."
Simplicity of behavior implies that player 1 plans to make the same offer and
make the same responses independently of how player 2 has reacted in the past.
However, this strategy also implies that player 2 believes that player 1 would
demand 50% of the surplus even if player 1 demanded 60% of the surplus in
the first, let us say, 17 periods of bargaining. Thus, stationarity implies not only
the simplicity of player l's behavior but also the passivity of player 2's beliefs.
This is unusual, especially if we assume simplicity of behavior. If player 2
believes that player 1 is constrained to choose a stationary plan of action, then 2
should believe (after 17 repetitions of the demand of 60%) that player 1 will
continue to demand 60%. Thus, assuming passivity of beliefs eliminates a great
deal of what sequential games are intended to model: namely, the changing
pattern in players' behavior and beliefs, as they accumulate experience.
Note also, that in games with more than two players, the interpretation of a
player's strategy following a deviation as the belief of the other players regard-
ing his future behavior, implies that players 2 and 3 hold the same belief about
player l's future behavior, not only on the equilibrium path, but also after
player 1 has deviated. This problem also arises when interpreting the sequential
equilibrium notion. Assume that player 1 has to choose from (a, b, c) and that
players 2 and 3 learn whether he chose within (a) or (b, c). If player 1 was
supposed to choose a, but did not, then the beliefs of players 2 and 3 about
player l's actual move must (by the sequential equilibrium definition) coincide.

3. MIXED STRATEGIES

The second context in which I would like to examine the interpretation of the
notion "strategy" is that of "mixed strategies." The concept of mixed strategy
has often come under heavy fire. To quote Aumann (1987a):4 "Mixed stra-
tegy equilibria have always been intuitively problematic... ", and Radner and
Rosenthal (1982): "One of the reasons why game-theoretic ideas have not found

4 I will do injustice to Aumann if I do not add that in the same paper he also describes Harsanyi's
justification of the mixed strategy concept as one which "rings so true."

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 INTERPRETATION OF GAME THEORY 913

more widespread application is that randomization, which plays a major role in

game theory, seems to have limited appeal in many practical situations." The
reason for the criticism is that the naive interpretation of a mixed strategy as an
action which is conditional on the outcome of a lottery executed by the player
before the game, goes against our intuition. We are reluctant to believe that our
decisions are made at random. We prefer to be able to point to a reason for
each action we take. Outside of Las Vegas we do not spin roulettes.
One must, however, be cautious. There are obviously cases in which players
choose random actions. Consider, for example, an employer who has time to
monitor only one of his two workers. The workers have to choose a level of
effort, either High or Low. If a worker is caught being lazy, his salary is cut. A
worker's tradeoff between effort and reduction in salary is such that if he faces a
probability 1/2 of being audited, he prefers to choose H rather than L. The
employer is interested in maximizing the number of agents who choose H.
Denote by p the probability with which the employer monitors worker A. His
real choice is not between monitoring A and monitoring B, but among the
possible numbers p. A suitable description of the situation is a leader-follower
game in which the employer first commits himself to a number p after which the
workers choose their level of effort. The policy p = 1/2 is the best policy for the
employer and, using a random device, is strictly optimal. However, this choice is
not a mixed strategy in which the two pure actions "monitoring A" and
"monitoring B" are the pure strategies. The employer is not indifferent be-
tween p = 1 (auditing only worker A) and p = 1/2. He strictly prefers the
latter. In contrast, in mixed strategy equilibrium, the players are indifferent
among all lotteries which consist of the actions taken from the support of the
mixed strategy.
It should also be mentioned that the use of mixed strategies is particularly
problematic in any situation where their execution is costly in terms of devoting
attention or time.S If implementing a mixed strategy is costly for the player,
then he will strictly prefer to use any of the pure strategies which appear in the
support of the mixed strategy rather than to waste the resources associated with
implementing the lottery.
Having rejected the naive interpretation of mixed strategies let us discuss two
other interpretations of the concept.
First, we will examine the large population case. One can think about a game
as an interaction between large populations. Each occurrence of the game takes
place after a random draw of the players from these populations.6 Alternatively,
a player stands for a large set of players, each of which chooses an action where
the payoff depends linearly on the fraction of the population which selects each
of the actions. In this context, a mixed strategy is viewed as the distribution of
the pure choices in the population. However, this is a limited context which

5 See the discussion in Abreu and Rubinstein (1988).

6 See Rosenthal (1979).

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 914 ARIEL RUBINSTEIN

does not justify the use of the notion in the case where a unique agent stands
behind each of the players in the game.
The second interpretation is the purification idea. A player's mixed strategy is
thought of as a plan of action which is dependent on private information which
is not specified in the model. Although the player's behavior appears to be
random, it is actually deterministic. If we add this information structure to the
model, the mixed strategy becomes a pure strategy in which the action depends
on the extraneous information. There are two major problems with interpreting
a mixed strategy as a rule of behavior which depends on exogenous factors.
First, it is hard to accept the assumption that behavior depends on factors which
are payoff irrelevant. There are reasons behind the choices people make and
their behavior is not motivated by what they consider to be irrelevant factors. If
the additional factors are payoff relevant (as Harsanyi (1973) insists), then the
model is incomplete in the sense that there are factors the players perceive as
relevant (compare with Section 7) which are excluded from the model. Second,
the predicted equilibrium behavior is highly fragile. If the manager's behavior is
determined by the type of breakfast he eats, then factors outside the scope of
the model, such as a change in diet or the price of eggs, may change the
frequency with which players choose their actions, thus inducing changes in the
beliefs of other players and causing instability.
Nevertheless, I agree that the purification idea provides a consistent interpre-
tation of a mixed strategy as a plan of action. However, in applying this idea to a
particular economic problem the economist needs to indicate which are the
"real life" exogenous variables on which players base their behavior. Thus, if an
industrial organization theoretician describes an equilibrium in price competi-
tion as an equilibrium with mixed strategies relying upon the purification
concept, he must specify the unmodelled factors which actually serve as the
basis for the firms' pricing policy. One has to show that the information
structure is rich enough to span the set of all mixed strategy equilibria.7 I am
not familiar with a single case in which the use of mixed strategies in economics
is justified by specifying a dependence of behavior on specific exogenous factors.
In fact the mixed strategy equilibrium is widely used as a technique for
calculating a solution in cases where the pure strategy equilibrium does not
exist.

7 Radner and Rosenthal (1982) investigate the following question: Given an information structure
about some random elements (which may or may not be payoff relevant), can the mixed strategy
equilibrium of the original equilibrium be attained as a pure Nash equilibrium of the perturbed
game? Conditions are provided under which this is the case. These conditions are not mild and
there are simple examples of informational structures by which a mixed strategy equilibrium could
not be purified. Another important related reference is Harsanyi (1973) who shows that under wide
conditions every mixed strategy equilibrium is a limit of some sequence of pure strategy equilibria in
some sequence of games in which all payoffs are "slightly" randomly perturbed and each player gets
exact information only about his own payoffs. From the perspective of looking for an interpretation
of mixed strategy, this result is rather weak. At most we can conclude that there is some sequence of
perturbations for which the mixed strategy equilibrium is an approximation.

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 INTERPRETATION OF GAME THEORY 915

Mixed strategy can alternatively be viewed8 as the belief held by all other
players concerning a player's actions. A mixed strategy equilibrium is then an
n-tuple of common knowledge9 expectations, which has the property that all the
actions to which a strictly positive probability is assigned are optimal, given the
beliefs. A player's behavior may be perceived by all the other players as
the outcome of a random device even though this is not the case. Adopting this
interpretation requires the reassessment of much of applied game theory. In
particular, it implies that an equilibrium does not lead to a prediction (statistical
or otherwise) of the players' behavior. Any player i's action which is a best
response given his expectation about the other players' behavior (the other
n - 1 strategies) is consistent as a prediction for i's action (this might include
actions which are outside the support of the mixed strategy). This renders
meaningless any comparative statics or welfare analysis of the mixed strategy
equilibrium and brings into question the enormous economic literature which
utilizes mixed strategy equilibrium.

4. GAMES WITH LIMITED MEMORY

The extensive form game allows us to discuss situations in which it is common

knowledge among the players that information concerning another player's
move, a random event, or previously available information (perhaps even about
his own actions) is to be withheld from a certain player.10 Let us now consider
situations in which a player knows that he might become doubtful about some
relevant information which he possessed in the past. Such fear of memory
confusion is clearly an important element in real life reasoning. In particular, it
is often the motivation for preferring modes of behavior which keep things
simple in order to avoid disasters resulting from this confusion.
Consider the following example which is illustrated in Figure 2. A driver is at
A and wishes to reach C. He could drive to C by the long route, which would
bring him to his destination directly without having to make any further
decisions, or he could use a short but unmarked road, in which case he would
have to make a turn at the second intersection. If he arrives at B or if he misses
the second intersection and reaches D he will be stuck in traffic jams and hence
waste several hours returning to C.
The driver knows that he is able to identify the turn to B but that when he
arrives at the turn to C, he will become confused and believe that there is a
10% chance that he has not yet passed the first intersection. The driver believes
that once he finds himself in this state of doubt, he will be alert and no longer
be confused. We usually assume that a player makes all possible inferences from
the information embedded in the game form; here, however, we assume that the

8 This view has been promoted in the literature by Aumann; see, for example, Aumann (1987b).
9 The important role of the common knowledge of expectations in this interpretation is empha-
sized in the discussion of rationalizability appearing in Bernheim (1984) and Pearce (1984).
10As in games of imperfect recall; see Alpern (1988).

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
916 ARIEL RUBINSTEIN

FIGURE 2

driver does not conclude from his state of doubt (which would occur only at the
second intersection) that he is actually at the second intersection.
Any presentation of the situation as an extensive game must allow a path of
length 4. It has to include a node, v1, in which the choice between the short and
long routes is made, a node, v2, which corresponds to the decision problem at
the turn to B, and an information set which corresponds to the state of doubt.
The two nodes in this information set, V3 and V4, must be different from v2,
since at v2 the decision maker does not have any doubts about his location. A
chance player node which precedes this information set enables us to model the
assumption that the decision maker believes with probability 0.9 that he is at the
turn to C (and will not have any more decisions to make) and with probability
0.1 that he is at the turn to B (and if he continues he will have one more
decision to make at the turn to C). When considering his action, at this
information set, he realizes that he may pass through the intersection to C as
well and thus the tree must include another successor node, v5.
Can a strategy in this game be interpreted as a plan of action? No. Looking at
the map, we observe that on any excursion the player could make at most 3
decisions (at A, at the turn to B, and at the turn to C). In the above extensive
form game, there is a path (v1, v2, v4, v5) in which the driver has to make 4

C via a Vi
long way T

B - V2

chance
0.9/ \0.1

C "V3 , ' B V4.W

D C -V5

FIGURE 3

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 INTERPRETATION OF GAME THEORY 917

decisions. Thus, in practice the node v5 is unreachable. The decision at v5 is not

a part of a plan of action made at A. It is added to the game form merely to
allow us to discuss the player's reasoning in the state of doubt.

5. TOWARDS A PERCEPTIVE INTERPRETATION OF GAME THEORY

In the previous three sections I have tried to explain the problems of

interpreting a strategy as a plan of action. In the case of sequential games, such
an interpretation does not apply to the part of a player's strategy which
supposedly describes a player's planned actions should he deviate from his
original plan. A mixed strategy can rarely be interpreted as a set of instructions.
In games with potential loss of memory, the game theoretic strategy has to be
interpreted as including hypothetical plans which can never be realized. I have
tried to argue that interpreting one's strategy as a description of the other
players' beliefs about his behavior makes more sense in these contexts. In this
section I argue that this interpretation fits neatly into a perceptive interpreta-
tion of the notion of a game.
Let me go back to the classical interpretation of the game form as a full
description of the physical events in the modeled situation. Notice that it is rare
that a situation involving a conflict of interests is described clearly and objec-
tively by a set of rules. The exceptions I can think of are "games" in the
colloquial sense. Unless the game instructions appear on the box bought at
"toys 'r' us," I cannot see how we can avoid the interpretation of a game form
as an abstract summary of the players' actual perceptions of the complicated
situations they are in.11 Players may be involved in a recurring situation but
might ignore the repetitive aspect of their position; players may not be aware of
some possibilities of error. They may consider only a few of the many available
actions; they may combine several actions into one option; they may decide to
ignore information which in principle they could take into account and so on.
As is often the case, we find an early reference to this point in John Nash's
work. When describing the strategic approach to bargaining, he states (Nash
(1953, p. 129)): "Of course one cannot represent all possible bargaining devices
as moves in the noncooperative game. The negotiation process must be formal-
ized and restricted, but in such a way that each participant is still able to utilize
all the essential strengths of his position."
I like to think of a game as being a comprehensive description of the relevant
factors involved in a specific situation as perceived by the players, rather than as
a presentation of the physical rules of the game. Let me demonstrate this view
while distinguishing between infinite and finite horizon games, particularly in
infinitely and finitely repeated games. My thinking on repeated games has been
influenced by the following quotation from Aumann (1959): "In the notion of a
supergame that will be used in this paper, each superplay consists of an infinite
number of plays of the original game G. On the face of it, this would seem to be

I See Selten (1978) for a discussion of "the level of imagination."

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
918 ARIEL RUBINSTEIN

unrealistic, but actually it is more realistic than the notion in which each
superplay consists of a fixed finite (large) number of plays of G.... Of course
when looked at in the large, nobody really expects an infinite number of plays to
take place; on the other hand, after each play we do expect that there will be
more. A. W. Tucker has pointed out that this condition is mathematically
equivalent to an infinite sequence of plays, so that is what our notion of
supergames will consist of."
This paragraph12 is an elegant statement of what in my opinion is the most
basic principle in the art of formal modelling. By using infinite horizon games
we do not assume that the real world is infinite. Models are not supposed to be
isomorphic with reality. An infinitely repeated game is meant to assist in
analyzing situations where players examine a long-term situation without assign-
ing a specific status to the end of the world. In contrast, the finitely repeated
game model corresponds to a situation in which the finite period enters
explicitly into the players' considerations. Using the terminology of formal logic,
we can say that finite horizon models are suitable only for modelling situations
in which the last period appears as an "individual constant" (a specified
element) in the players' model.
There is only a partial correspondence between the real length of a repeated
game (assuming that the term "real length" is well defined) and the selection of
a model to analyze it. Even short games may be better analyzed as infinite
horizon games.13 When subjects in a laboratory situation get instructions to p
the prisoner's dilemma twenty times with payoffs of between one and four cents,
it seems that the infinite horizon game captures their method of reasoning
better then a finitely repeated game.
Many game theoreticians have expressed their unhappiness with the sharp
discontinuities that exist between finite and infinite horizon game outcomes. I
have no difficulty with these discontinuities. I view infinite and finite horizon
models as representing two very different scenarios. I am equally unimpressed
with the fact that the limit of a sequence of long finite horizon models is an
infinite horizon game. For example, in the alternating bargaining model (see
Rubinstein (1982)) the perfect equilibria for the finite horizon model converge
to the infinite horizon unique perfect equilibrium (see Binmore (1987a)). How-
ever, the significance or insignificance of the infinite model does not depend on
this result. The convergence result is, in my opinion, nothing more than a
convenient tool for approximating the perfect equilibrium outcomes of finite
horizon models when the horizon is very long.
I take a similar position on some other game theoretic modelling issues. For
example, compare models with and without discounting. A model with discount-
ing is appropriate when considering situations in which the players are con-

12 Ignoring the problems associated with the term "mathematical equivalence" which I
understand.
13 This point of view weakens the significance of some game-theoretic results which are referred
to in the game-theoretic jargon as "paradoxes" (e.g., the finitely repeated prisoner's dilemma and
Selten's chain store paradox).

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 INTERPRETATION OF GAME THEORY 919

cerned with time and include time impatience in their strategic considerations.
A model without discounting is better suited to situations in which the players
ignore the timing component of an outcome. (This does not imply that if asked,
a player in the no-discounting model would be indifferent between getting one
million dollars today or getting the same amount in ten years time). Therefore, I
do not see the need to approximate a world without discounting using a world in
which the discount rate is close to 1. The similarity between the model with a
discount rate equal to 1, and models with discount rates of "almost" 1 is purely
mathematical. It is useful mainly as a technique to approximate the solutions for
games with discount rates close to 1.

6. RELEVANCE

If we adopt the view that a game is not a rigid description of the physical
rules of the world, then a game-theoretic model should include only those
factors which are perceived by the players to be relevant. Modelling requires
intuition, common sense, and empirical data in order to determine the relevant
factors entering into the players' strategic considerations and should thus be
included in the model. This requirement makes the application of game theory
more an art14 than a mechanical algorithm.
Adopting the view that a game should include only relevant factors entails
implications for many game-theoretic issues and in particular for the recent
discussion of "forward induction principles." This is demonstrated by one of the
most intriguing examples that I have ever seen in game theory. The example is
due to Eric van Damme (see van Damme (1989)15 and also Ben-Porath and
Dekel (1988) and Osborne (1990)) and is a development of ideas initiated in
"Kohlberg's example" (see Kohlberg and Mertens (1986)).

EXAMPLE: Two players 1 and 2 participate in a battle of the sexes in which

the nonzero outcomes are in dollar terms (3, 1) and (1, 3). Prior to their
engaging in the battle, player 1 (and only player 1), is able to dispose of one
dollar. Player 2 observes player 1 disposing or not disposing of the dollar. An
extensive form game which describes the above scenario is:

L R L R

T 3 1 0 0 T 2 1 -1 0
B 0 0 1 3 B -1 0 0 3

14 See Aumann (1987a).

IS van Damme offers the example as a demonstration of the importance of sunk costs and he
should not be indicted in any criticism of my position on the example.

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
920 ARIEL RUBINSTEIN

In the normal form of this game each player has 4 strategies:

LL LR RL RR

=T 3 1 3 1 0 0 0 0
=B 0 0 0 0 1 3 1 3
DT 2 1 -1 0 2 1 -1 0
DB -1 0 0 3 -1 0 0 3

van Damme invokes what is usually regarded as the rather weak solution
concept of the successive elimination of weakly dominated strategies: It is easy
to verify that successively, DB is dominated by = T, RR is dominated by RL,
LR is dominated by LL, = B is dominated by DT, RL is dominated by LL, DT
is dominated by = T, and we are left with (= T, LL). Thus, the mere fact that
player 1 could dispose of one dollar skewed the game in his favor.
As economists we should ask ourselves whether we can conceive of a conflict
of interests like this to be resolved by the identity of the player who has the
ability to dispose of resources. I very much doubt this to be the case.
How is "relevancy" related to this discussion? If disposing of the dollar were
a relevant consideration in the players' perception of the situation, then the
result would (probably) make sense. However, I cannot believe that any reason-
able person would consider a pre-game disposal of a dollar to be relevant in the
analysis of the battle of the sexes. It is my opinion that a formal description of
the situation should exclude the choice of disposal even in cases where a
description of the game is given by a referee who specifies the possibility of
disposing of the dollar (recall that there is rarely a referee).
We might think about "relevancy" as a matter of convention, which may be
explained by game-theoretic arguments. I prefer to view the determination of
relevancy as a stage in the game process which precedes the prediction of other
players' behavior and the calculation of the best plan of action. I prefer to leave
the explanation of the logic of perception to other branches of science (such as
evolutionary biology). Still, we can attempt to organize our intuition of irrele-
vancy by appealing to general principles.16 The following are three characteriza-
tions of the disposal of the dollar in van Damme's example which may be used
to form a general sufficient condition for irrelevancy:
(a) The disposal decision does not affect the payoffs of the players in the
battle of the sexes (when player 1 disposes of the dollar, the game is identical to
the game in which player 1 does not dispose of the dollar).
(b) The disposal decision does not reveal any unknown information (in
contrast to the role of spending money on advertising in the signalling litera-
ture). Even if disposal gives information about "the rationality of player 1," a
sensible conclusion might be that a player who throws the dollar out of the
window is just "crazy."

16See Harsanyi (1977, page 118).

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 INTERPRETATION OF GAME THEORY 921

(c) The disposal decision is not a part of a game which is identical or even
similar to the choice problem which player 1 has to confront in the battle of the
sexes. (In contrast, the problem which a player confronts in the first period of an
infinitely repeated game is identical to his problem in the second period.) Thus,
whether or not player 1 disposes of the dollar cannot provide information from
which player 2 learns about player l's behavior in the battle of the sexes.
The reader will probably feel that verbal statements made before a play of the
battle of the sexes satisfy the above three conditions for irrelevancy and yet they
are highly relevant. This point was made by Joe Farrell. Farrell (1985) suggests
that if the players in the battle of the sexes have a language and if only player 1
could (cheaply) talk and state (in English) that he is going to play UP, then he
has a serious advantage in determining the conflict's outcome. This sounds like
a plausible argument and Farrell correctly modelled communication as a part of
the solution concept rather then as a part of the game. However, this is not the
argument which is used in van Damme's analysis. If instead of being able to
throw a dollar out of the window, player 1 is allowed to throw a bill worth
nothing out of the window (or nod his head), or even state that "I am going to
play UP," then the process of successive elimination of weakly dominated
strategies is not powerful and all equilibria of the battle of the sexes would
survive. It is my impression that although language plays a crucial role in
resolving conflicts, game theory has so far been unable to capture this role.

7. REGULARITY

There is one additional issue to be raised in connection with the interpreta-

tion of game theory. Game theory deals with regularities.17 As Carnap (1966)
writes: " The observations we make in everyday life as well as the more
systematic observations of science reveal certain repetitions or regularities in
the world. ..." The laws of science are statements expressing these regularities
as precisely as possible. Events which are pure one-shot events, completely
disconnected and uncorrelated with other events are outside the scope of any
theory including game theory. Game theory should be employed as a descriptive
science only after the social scientist observes or expects some regularities in a
family of "similar" events. Accepting this view leads me to make the following
three comments:
a. Is a one-shot game an isolated event? Game theoreticians often emphasize
the one-shot character of the games analyzed. Careless interpretations of this
emphasis have led to a widespread confusion in the applied game theory
literature with respect to the distinction between one-shot and repeated games.
It is the repetitive nature of a situation which makes it possible to talk about
regularity. The distinction between a one-shot game and a repeated game is not
based solely on the naive criterion of whether the players repeat the conflict

17 See Lucas (1985) who expresses a similar view about economics.

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 922 ARIEL RUBINSTEIN

again and again, or whether they are involved in an isolated incident. The
proper criterion is concerned with whether the players take into account the
effect of their choice today on similar future games in which they will partici-
pate. If they ignore the effect of their behavior on future games, then the
framework of the game-theoretic one-shot game is appropriate. If the players
calculate the effect of their behavior on future games, the repeated game
framework is appropriate. If the players use their past experience to speculate
about other players' future behavior without taking into account the effect of
their own behavior, then we are dealing with dynamics.
b. The meaning of nonexistence. If what we are trying to model in game theory
are situations in which we expect regular behavior, then it is not true that all
descriptions of the world should have an equilibrium.18 The mere fact that a
game theoretician constructs a game does not mean that the game corresponds
to a regular mode of behavior. The modeller should not, in the case of
nonexistence, twist the analysis, but rather should check the adequacy of the
model as a description of conceived regularity.
This brings me back to the mixed strategies issue. One of the reasons that
mixed strategies are popular in both game and economic theory, in spite of
being so unintuitive, is that many models do not have an equilibrium with pure
strategies. However, the nonexistence of a solution concept in pure strategy
does not necessarily mean that we should look for stochastic explanations. It
means that the description of the game and the assumptions embedded in the
solution concept are not consistent with regularity. Expanding the model or
changing the basic assumptions are alternatives which the modeller should
consider at least as favorably as mixed strategies.
c. Regularity and language. The observed regularity depends on the language
employed. Behavior can be irregular in one language and regular in another. I
became aware of this simple point after a game I played with my baby daughter
when she was about one year old. Wanting to check her choice consistency and
knowing that she recognized colors, I put Blue, Red, and Green cubes in front
of her in an arbitrary order. I repeated this choice problem about a dozen times.
Her choices revealed a clear inconsistency. However, before I became "over-
joyed" with the idea that my cute baby daughter had refuted the basic postulate
of choice theory, I realized that the baby was actually amazingly consistent.
However, she was not consistent in her choice between Blue, Red, and Green.
She was consistent in her choice between Left, Center, and Right. She always
chose Left.... If my vocabulary did not include the words Left, Center, and
Right, I would be unable to describe this regularity. Given that we can use two
languages, we prefer to use the terminology in which we find the behavior to be
more regular.

18 My thinking on this matter was influenced by a statement made by my teacher, Menachem

Yaari, in commenting on a famous paper which showed nonexistence of equilibrium in an economy
with imperfect information. Yaari told the "shocked" economists that the problem lay neither in the
solution concept nor in the real world, but rather. . . in the model.

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 INTERPRETATION OF GAME THEORY 923

8. CONCLUSION

The comments I have offered in this paper were intended to emphasize the
inadequacy of the naive interpretation of game theory as a physical description
of the world and to promote the view that in game theoretic modelling, we must
regard as given the laws of perception, the bounds on rationality, and the
processes of reasoning employed by the players.
There exists a widespread myth in game theory, that it is possible to achieve a
miraculous prediction regarding the outcome of interaction among human
beings using only data on the order of events, combined with a description of
the players' preferences over the feasible outcomes of the situation. For forty
years, game theory has searched for the grand solution which would accomplish
this task. The mystical and vague word "rationality" is used to fuel our hopes of
achieving this goal. I fail to see any possibility of this being accomplished.
Overall, game theory accomplishes only two tasks: It builds models based on
intuition and uses deductive arguments based on mathematical knowledge.
Deductive arguments cannot by themselves be used to discover truths about the
world. Missing are data describing the processes of reasoning adopted by the
players when they analyze a game. Thus, if a game in the formal sense has any
coherent interpretation, it has to be understood to include explicit data on the
player's' reasoning processes. Alternatively, we should add more detail to the
description of these reasoning procedures. We are attracted to game theory
because it deals with the mind. Incorporating psychological elements which
distinguish our minds from machines will make game theory even more exciting
and certainly more meaningful.

Department of Economics, Tel Aviv University, Ramat Aviv, 69978 Tel Aviv,
Israel

Manuscript received August, 1988; final revision received August, 1990.

REFERENCES

ABREU, D., AND A. RUBINSTEIN (1988): "The Structure of Nash Equilibrium in Repeated Games
with Finite Automata," Econometrica, 56, 1259-1281.
ALPERN, S. (1988): "Games with Repeated Decisions," SL4M Journal of Control and Optimization,
26, 468-477.
AUMANN, R. (1959): "Acceptable Points in General Cooperative n-Person Games" in Contributions
to the Theory of Games IV, Annals of Mathematics Study, 40. Princeton: Princeton University
Press, pp. 287-324.
(1987a): "What Is Game Theory Trying to Accomplish?" in Frontiers of Economics, ed. by
K. J. Arrow and S. Honkapohja. Oxford: Blackwell.
(1987b): "Correlated Equilibrium as an Expression of Bayesean Rationality," Econometrica,
55, 1-18.
BEN-PORATH, E., AND E. DEKEL (1988): "Coordination and the Potential for Self Sacrifice,"
Graduate School of Business, Stanford University, Research Paper 984.
BERNHEIM, D. (1984): "Rationalizable Strategic Behavior," Econometrica, 52, 1007-1028.
BINMORE, K. (1983): "Aims and Scope of Game Theory," mimeo, LSE.
(1987a): "Perfect Equilibria in Bargaining Models," in The Economics of Bargaining, ed. by
K. Binmore and P. Dasgupta. Oxford: Blackwell.

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms
 924 ARIEL RUBINSTEIN

(1987b): "Modeling Rational Players I," Economics and Philosophy, 3, 179-214.

CARNAP, R. P. (1966): Introduction to Philosophy of Science. New York: Basic Books, Inc.
FARRELL, J. (1985): "Credible Neologisms in Games of Communication," MIT Working Paper 386.
HARSANYI, J. (1973): "Games with Randomly Distributed Payoffs: A New Rationale for Mixed
Strategy Equilibrium Points," International Journal of Game Theory, 3, 211-225.
(1977): Rational Behavior and Bargaining Equilibrium in Games and Social Situations.
Cambridge: Cambridge University Press.
KOHLBERG, E., AND J. F. MERTENS (1986): "On the Strategic Stability of Equilibria," Econometrica,
54, 1003-1038.
LUCAS, R. (1985): "Adaptive Behavior and Economic Theory," in Rational Choice, ed. by R. and
M. Reder, pp. 217-242.
NASH, J. (1953): "Two Person Cooperative Games," Econometrica, 21, 128-140.
OSBORNE, M. J. (1990): "Signalling, Forward Induction, and Stability in Finitely Repeated Games,"
Journal of Economic Theory, 50, 22-36.
PEARCE, D. (1984): "Rationalizable Strategic Behavior and the Problem of Perfection," Economet-
rica, 52, 1029-1050.
RADNER, R., AND R. ROSENTHAL (1982): "Private Information and Pure Strategy Equilibrium,"
Mathematics of Operations Research, 7, 401-409.
RENY, P. (1988): "Explicable Equilibria," mimeo, University of Western Ontario.
ROSENTHAL, R. W. (1979): "Sequences of Games with Varying Opponents," Econometrica, 47,
1353-1366.
RUBINSTEIN, A. (1982): "Perfect Equilibrium in a Bargaining Model," Econometrica, 50, 97-110
(and STICERD research report 80/13).
(1986): "Finite Automata Play the Repeated Prisoner's Dilemma," Journal of Economic
Theory, 39, 83-96.
SELTEN, R. (1978): "Chain Store Paradox," Theory and Decision, 9, 127-159.
VAN DAMME, E. (1989): "Stable Equilibria and Forward Induction," Journal of Economic Theory,
48, 476-496.

This content downloaded from

190.60.10.198 on Tue, 15 Mar 2022 16:36:12 UTC
All use subject to https://about.jstor.org/terms