Professional Documents
Culture Documents
Yanis Varoufakis Game Theory
Yanis Varoufakis Game Theory
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
Springer is collaborating with JSTOR to digitize, preserve and extend access to Erkenntnis
(1975-)
ABSTRACT. Equilibrium game theory borrows from neoclassical economics its rational?
ity concept which it immediately puts to work in order to produce the basic results it
needs for building an elaborate narrative of social interaction. This paper focuses on
some recent objections to game theory's use of rationality assumptions in general, and
of backward induction and subgame perfection in particular, and interprets them in the
light of the postmodern critique of the grand meta-narratives which social theorists often
rely on for social explanation. The paper presents a defence of game theory which seeks
to accommodate the postmodern critique. However, it goes on to show that such a
defence is illegitimate and claims that the problem lies with the faulty conceptualisation
of the main concept on which game theory rests: that of Reason. Having established the
nature of the problem, it considers three alternative interpretations (Humean, postmod?
ern and Hegelian) of why the problem resists logical solutions and of its significance for
social theory.
INTRODUCTION
The battlelines in social theory have frequently been drawn along two
familiar views of human agency. First, there is the perception of th
sovereign agent whose autonomous desires forge the social structures
that will fulfil them. Society is, therefore, seen as the means by whic
the agents' ends will be instrumentally realised. Second, there is the
view of an individual whose desires are the product of social structure
Even if the agent rationally pursues her objectives, she is still a playthi
of social forces which she cannot control.
In this paper the above controversy is bypassed in favour of a deepe
controversy between the dominant variant of modernity and its foes
To accomplish this, I focus on a well-known game in which agent
have given payoffs and a unique equilibrium strategy. That a fier
controversy is engendered in a simple framework is evidence that on
does not need complex social interactions in order to end up wi
complex social phenomena. That this controversy also has the potenti
of inciting clashes between Humeans, postmodernists and Hegelians,
is an indication that game theory ought to be more than a search fo
clever strategies. Indeed, as I will argue, it should be primarily con?
cerned with the meaning of rationality in social settings.
split.2 Once this result is obtained, game theory relaxes its assumptions
progressively and tackles more complex versions of the same problem.
First, it allows for differences in attitudes toward risk and rewards the
(relative) risk-taker with a greater payoff, and, second, it introduces
asymmetric information in order to show that the possession of more
information is advantageous (see Harsanyi 1973).
Why is the above example representative of contemporary mod?
ernity? Two individuals facing each other in an instance of pure antag?
onism develop trains of thought which swiftly terminate the cacophony
that would have arisen in a pre-modern narration of their situation.
Instead of the drama of equally intelligent belligerents duelling to the
last for personal gain, Reason is called upon to furnish harmony and
efficacy. The fact that their rationality is common knowledge is pre?
sented as the bedrock of a uniquely rational train of thought that each
will, if rational, latch on to. The clear separation of Reason from
Unreason allows the theorist to view agents as identical computers
running the same software with identical initial information (input)
guaranteed by the assumption of perfect information. It is not surprising
that they will inevitably come to the same conclusion (output) and thus
agree without delay.
An immediate postmodern concern is that the computer metaphor
inhibits understanding of human responses since our reasoning can be
neither unique nor transparent. A more serious postmodern objection
is that to talk of Reason is to talk of a term that has no concrete
equivalent in social reality. It is not that agents are irrational, but that
it is unclear what it means to be rational in social interactions. If this
is so, the analogy with the numerical algorithm is misleading.3 Whereas
game theory treats simple social interactions with given objectives in a
way that the outcome can be assumed and used later for explaining
more complex situations (the analytic-synthetic road to explanation),
postmodernity claims that the only chance of defensible choices, even
in simple situations, materialises when we recognise the impossibility
of understanding our reasoning by means of metaphors that devalue
and oversimplify.
In a manner reminiscent of Parmenides' definition of nothingness
(i.e., a radical absence of reality4) game theory identifies Reason with
the residual left behind once Unreason has been expelled. By contrast,
postmodernity claims that human thoughts are irreducible to a field
where Reason is dominant. Critics of the modernity that lies behind
Table 1.
(1,0) (0,30) (50,29)
Up/ Up/ Up/
A -> B -> A -> (40,40)
Down Down Down
t= 1 ? = 2 i=3
Assumptions:
(a) AbB is rational
(b) AbBbA is rational
Fundamental conjectures:
(c)i=3 Ci: A^UP
(d) t = 2 C2: BbCl thus B -? UP
(e) f = 1 C3: AbC2 thus A -> UP
A's composite conjecture inducing the equilibrium outcome:
B will play UP at t = 2 if he expects that his payoff from doing so, i.e. 30, is greater than
what he can rationally anticipate at t = 3. At t = 3 I am the one who does the choosing
and I will clearly play UP leaving B with payoff 29. Since this is less than what he will
get from ending the game at t = 2, it is silly of me to expect him to play anything other
than UP. Thus, the conclusion that t = 3 will not be reached leads me to the conviction
that I am better off by playing UP at t = 1.
herself in his shoes by pretending that his payoffs are her own. This is
what allows game theorists to assume that the division game described
earlier is trivial and also that the passage from (c) to (e) and (f) in
Table 2 ought to be automatically accepted.
To sum up, Table 2 is a good example of how the dominant modernity
lurking behind game theory analyses a simple interaction between two
agents, of how it breaks beliefs down to their elemental components
and uses induction in order to put back together a string of conjectures
leading to an equilibrium result.
2. THE CHALLENGE
roles are finally assigned, whoever gets the part of A plays UP instantly.
CKR does for equilibrium game theory's view of Reason what the veil
of ignorance does for John Rawls' concept of justice: it defines it via a
process of de-personalisation. The second point to note is that the
adoption of this program requires that Reason is a means by which
agents (as well as game theorists) convert an expectation into a convic?
tion. For example, at stage 1 player A is facing a choice between a
certain reward (payoff 1 if she plays UP) and a conjecture concerning
what she will end up with if she plays DOWN. Backward induction,
faithfully reproduced in Tables 2 and 3, turns this conjecture into the
conviction that, were she to play DOWN, her payoff would be 0.
It is now time to explain why the above is highly problematic. The
critique of backward induction which follows has been around for some
time8 but its impact has not been felt outside the narrow circles that
produced it. Nevertheless, it is an important critique with repercussions
for the way social theorists incorporate game theory in their models
but also because it allows us to place the debates between game theorists
within the larger debates in social theory. It begins with a devious
thought that may cross A's mind:
I understand Table 2 well and I agree with assumptions (a) and (b). Therefore, I see
why its logic should lead me to the conclusion that UP at t = 1 is the only sensible strategy
for me. However, what if I choose to defy it?
Assumptions:
(a) AbB is rational with probability 1 - w = 1
BbA is rational with probability 1 - p = 1
(b) A and B know (a)
(c) If at t = 2 p were to equal 1 then: BbA -* DOWN at t = 3
Definitions:
(d) Let p' > 0 be the probability belief of B at time t = 2 that would induce B to play
DOWN at t = 2
(e) Let q be A's probability belief that p exceeds p', i.e. q = Pr(p > p')
(f ) Let q' > 0 be the probability belief of A at time t = 1 that would induce A to play
DOWN at t = 1
Fundamental conjectures:
(g) A believes that if she defies the logic of backward induction and plays DOWN at
t = 1, then B will revise p upwards at t = 2
(h) q > q' at t = 1 and, therefore, A ^ DOWN
where b and ?> denote the verbs believes and plays respectively
goes against the equilibrium prescription of game theory but which may
or may not be irrational. For example, in our game (see Table 1), if
A ever played DOWN at t = 3, we conclude that A is (instrumentally)
irrational. However, if A plays DOWN at t = 1, then she may or may
not be irrational depending on her ability to justify her choice in terms
of her objectives and beliefs. If she can justify her belief in the superior?
ity of playing DOWN as a strategy for reaching the (50, 29) outcome,
then her strategy is deviant albeit not irrational.
The logic of Table 4 is this: At t = 1 player A contemplates playing
DOWN instead of her equilibrium strategy UP for a simple reason: she
is hoping that by so doing t = 3 will be reached. Why? She thinks to
herself:
anticipate what B will think, agents are forced to stop operating like
automata, to ditch the program in Table 3, and to start thinking as
opposed to following formulae.9 A continues her reflection.
Since my choice deviates from that of the Table 2 recommendations, he will be forced
to find an explanation. There are two possibilities. One is that he will think that I am
irrational for not doing as Table 2 prescribes. If this is so, he will change his game plan
and play DOWN at t = 2 expecting my irrationality to overcome my senses so that at t =
3 I will choose DOWN. Of course, there is the other possibility that I must reckon with.
Player B may realise that this is exactly what I am thinking and refuse to believe that I
am irrational simply because I have chosen irrationally. Nevertheless, all I need in order
to consider playing DOWN is that B assigns a relatively low probability that I am
irrational; not that he is convinced of my irrationality. Let p be the non-zero probability
that he assigns to this prospect after observing my deviant choice at t = 2. If p > 1/11 (in
terms of part (d) in Table 4,p' = l/ll), then his expected return at t = 2 from playing
DOWN exceeds that from UP, therefore giving him a strong incentive to deviate from
his equilibrium strategy too, i.e. play DOWN at t = 2. So, I conclude that if my defiance
of the logic of Table 2 makes him think with probability 1/11 that I am irrational then
it may, after all, make sense for me to play DOWN at t = 1 since there is now a realistic
chance of getting 50 at t = 3 rather than 1 at t = 1.
A's behaviour, how can we say that A's deviant strategy is rational at
t = 1 when she cannot know how B will be thinking at t = 2? This is a
good point. It proves beyond doubt that our players face risky decisions
once CKR and the safety of the Table 2 logic are abandoned. This is,
however, no proof of the irrationality of deviance; it is merely confir?
mation that neither the equilibrium nor the deviant strategies are
uniquely rational.
In effect, when A contemplates the deviant strategy she is hoping
that she can deceive her opponent. Is this rational? The answer must
be that it is certainly not irrational. There is nothing in the structure
of this game to suggest that an instrumentally rational agent ought to
assume that she cannot out-manoeuvre her opponent. By the same
token, it is also rational to think that she cannot do this. The problem
with game theory and its CKR foundation is that it instils in agents'
minds the belief that deception can never work. It is unclear what
institution or psychological mechanism performs the same role in soci?
ety.
The result is that Table 2 cannot represent the uniquely rational train
of thought. This is a familiar postmodern view regarding modernist
theories which are accused of mistaking analogies for concepts.19 In
our case, game theory mistakes the consistency which results from
analogous behaviour (such as that prescribed by Table 3) for rationality.
A positive defence of game theoretic orthodoxy must begin with a
humble admission that Table 2 is only an embarkation point and not
the destination of the equilibrium narrative. The CKR assumption
should then be interpreted as an initial assumption to be relaxed soon
after the theory has got off the ground. The refined game theorist must
admit that there are different ways of conceptualising the game of Table
1 and that it is unwise to assume that a rational player A will never
play DOWN at t = 1. However, the positive defence must insist that,
if we are to understand what happens when two equally intelligent
players participate in this game, we must utilise the tools of equilibrium
analysis even if we respect its earlier conclusion based on quite restric?
tive rationality assumptions. Kreps et al (1982) offer a good basis on
which to build such a defence.
The first leg of a positive defence would be to modify the CKR
assumption as stated in parts (a) and (b) of Table 2. In its stead, it would
place the assumption that players may now suspect their opponent to
be irrational (see (a) and (b) in Table 5). Furthermore, it allows for
more than one kind of rationality so that player A may be rational and
yet not conceptualise the game according to Table 2; let me label the
latter the Reason of Backward Induction RBI. After making these
concessions, the positive defence procures its rationalisation of the
deviant strategy on game theoretical grounds. Central to it is the diver?
sity of logics which a rational player may adopt as well as the possibility
of flagrant irrationality. In Table 5 player B expects player A to be
irrational with probability p and is convinced that an irrational A will
always choose DOWN at t = 3.20 Also, if he thinks she is rational he
does not immediately assume that she will adopt RBI and the Table 2
logic. He expects a rational A to deviate from RBI with probability
1 - r (see part (h)) due to the adoption of an alternative mode of
reasoning. What kind of reasoning is that? It depends on how open
minded the theorist is. In the most stark of interpretations, to shun
RBI is identified with irrationality and 1 - r becomes the probability
of behaving irrationally with a view to confusing player B. Alterna?
tively, game theorists may wish to allow 1 - r to be the probability with
Table 5. The positive defence: a game theoretic explanation of the deviant strategy
Assumptions:
(a) AbB is rational with probability 1 - w ^ 1
(b) BbA is rational with probability 1 - p ^ 1
(c) A and B know (a) and (b)
(d) If at t = 2 p were to equal 1 then: B expects with certainty A to play DOWN at
t = 3 (see footnotes 8 and 15)
Definitions:
(e) Let p' > 0 be the probability belief of B at t = 2 that would induce B to play
DOWN at t = 2
(f) Let q be A's probability belief that, at t = 2, p >/?', i.e. q = Pr(p >/?') if she plays
DOWN at t = 1
(g) Let q' be the probability belief of A at time t = 1 that would induce A to play
DOWN at t = 1
(h) Let 1 - r be the probability belief of B that A, if rational, will adopt the RBI
reasoning (i.e. the reasoning of Table 2)
(i) Let s be the probability belief of B that A ?? DOWN at t = 1 when A is irrational
Fundamental conjectures:
(j) A rational player A believes that if she defies the logic of backward induction
(RBI) and plays DOWN at t = 1, then B will revise p upwards at t = 2 using
Bayes' rule (see Equation (1))
(k) q > q' at t = 1 and, therefore, A -> DOWN
where b and ?? denote the verbs believes and plays respectively
Table 6.
Player B's conjecture about player A at t=\
r UP V-P?r
Pi Rational
/ ^\-r DOWN (l-Pl)(l-r)
\ 1-5 UP Pi(l~s)
P\ Irrational /
DOWN P\S
A's choice at B's subj. prob, belief
t=l for each outcome at
t=l
(l-px)(l-r)+Pls
Hence, given values for r and s, both players can work out how
initial belief that A is irrational will be updated if A plays DOWN a
t= 1. Moreover, they both know the value of p' (i.e.. the degree
conviction at / = 2 that A is irrational so that B wishes to play DOW
at t = 2) and they can work out whether it would make sense f
rational A to play DOWN in order to ensure that p reaches p' (i.
that deviant behaviour at t = 1 pushes p2 up to the level of p'). In o
particular game in Table l,p' equals 1/11. Given an initial probabilist
belief by A that B is irrational equal to pu I re-write (1) as:
erts (see Kreps et al., 1982). It claims that once the CKR assumption
is dropped (i.e. once/?i > 0), player A may attempt creatively to exploit
the fact that her opponent does not know if she is rational at all or, if
rational, what kind of rationality she espouses. In addition, she may
have an incentive to defy RBI even though she initially conceived of the
game in terms of RBP. Turning the tables on the challenge of Section
2, game theory seeks to explain internally the logic of Table 4 by means
of Table 5.
The only striking difference between this sophisticated narrative and
that of Table 4 is that the latter begins with an assumption in common
and absolute belief in each other's rationality, whereas the former
requires that B experiences at least a little bit of uncertainty concerning
A's thoughts. One could construct a claim that Table 4 is superior to
the above as an explanation of deviant play at t = 1 because it allows
deviant thoughts even when both players are convinced that they share
the same reasoning. However, on its own this would be a thin claim.
For the defenders of game theory could retort that, in view of the
conclusions of Table 4, no rational player can be certain that she knows
the reasoning that her opponent will employ. Thus, assumptions (a)
and (b) in Table 4 are not the assumptions that rational players would
wish to make and, consequently, it makes more sense to relax the CKR
assumption instead.
The standard reply by game theorists which the above criticism shall
occasion is that if we want agents to entertain different expectations,
then there is no real problem. We assume that this is the case, equip
them with probability density functions which capture their uncertainty
about each other's subjective beliefs and we derive a complex asymmet?
ric information model that addresses the above concerns. However,
this would muddy the waters unnecessarily. There is no gain to be had
if a problem is elevated to a higher level of complexity without being
solved. If an equilibrium model with asymmetric subjective beliefs is
to work, we must assume that the probability density functions of one
player are known with certainty by another. So, instead of demanding
that agents use the same value of pl9 r and s, the positive defence now
demands that they are certain that they have the same probability
density functions over different values of pl5 r and s. But why would
they feel confident that this is so? Such confidence would be unaccept?
able on the grounds of rationality.26
In order to avoid the danger of getting bogged down in a pointless
argument about higher order probabilistic conjectures, and whether or
not they should be in equilibrium, let me make a concession that I do
not have to make and yet demonstrate that the common knowledge
the positive defence depends upon is implausible. Suppose that A and
B are equally rational and have at their disposal exactly the same values
of px, r and s, as the positive defence assumes (i.e. for the moment
disregard Criticism 1). Some unspecified process leads them both to
the conclusion that 1 - r = 1/4, s = 3/4 and px = 1/31. The question
then is: what will A and B do given these shared beliefs? Will they
have an incentive to move away from them? According to the equilib?
rium story, if player A is rational she must randomise at t = l.27 If the
outcome of this randomisation is DOWN then Equation (1) yields p2 =
1/11 and player B is forced to randomise at t = 2 too. This is a knife
edge situation where neither has an equilibrium pure strategy and
where each has to resort to an equilibrium mixed strategy - i.e. to
randomising. Is this what they will do?
Player A may believe that B will stick to the above scenario. If she
does, she has no overwhelming reason for playing DOWN, UP or for
And so on.
Since the actual outcome of the game will depend on which thought
each player terminates his or her climb up the ladder of conjectures,
we cannot predict what either of them will do. There is no optimal
stopping rule when one enters deviant trains of thought and for this
reason any equilibrium story (including those postulating probability
expectations over the two strategies of each player) is inappropriate.28
All we can safely say is that if A stops at DT3, then Table 2 applies.
If, on the other hand, she reaches DT5, then Table 4 aptly describes
her thoughts.
I predict two objections to Criticism 2. First, DTI may be denied on
the grounds that there is no reason why it is more likely to develop
than, say, DTI': A sets r= 1. However, in this case player B may
anticipate DTI' and move directly to DT4 setting p2 = 1. The point is
that at t ? 1 anything goes whichever deviant thought arises first. The
second objection is that Criticism 2 applies only when pl9 r and s are
such that/?2 = 1/11. Although this is correct, this case is too important
to dismiss as an exception that confirms a rule. Since A's choice at t =
1 will depend on whether she expects or not /?2 > 1/11, it is crucial for
the positive defence that there exists one combination of pu r and s
such that p2 = 1/11 so that A is made indifferent between UP and
DOWN at t = 1. Otherwise there is no clear demarcation between the
case where A will rationally play the deviant strategy and the case
where she will not. And without this demarcation, there can be no
equilibrium defence of game theory from the challenge of Section 2.
In conclusion, Criticism 2 reinforces the claim of Criticism 1 that there
can be no tenable equilibrium theory of what will happen if A and B
are gifted with equal amounts of Reason.29
sition to the Humean and the postmodern views. The former evokes
the image of a static Reason which, due to the inability of desires to
provide it with enough information, stays on the sidelines and refuses
to engage until t = 3 is reached, whereas the latter agrees that Reason
is absent at t = 1 but claims that this is due to its non-existence. By
contrast, Hegel would argue that Reason jumps into the fray at t = 1
and generates contradictory thought processes like those in Tables 2
and 4. And this is the rub. For it is these inconsistencies that give
Reason the opportunity to enrich itself with elements of fundamentally
opposed reasonings. Hegel views Reason as an evolving concept that
affects the agents' experience and, in contradistinction to Hume, is
affected by it.
Looking at our little game again, Hegel's dialectics suggest that at
t = 1 Reason generates two contradictory logics (Tables 2 and 4) which
are equally powerful; they are the thesis and the antithesis. The out?
come is only describable in historical (as opposed to logical) time
because of the logical equivalence of the two types of reasoning. How?
ever, once the game is played (and here Hegel would agree with Hume
ans and postmodernists that there is no way of predicting what will
happen if all the information we have is in Table 1), the Reason of
agents, as well as of theorists, emerges superior to what it was before
they encountered this game. As it absorbs both logics (Tables 2 and
4), Reason endows us with an understanding of the game that is inde?
scribable by one of the two tables although it is comprehensible by a
synthesis of the two. Put differently, our rationality was of a lower
order of development before we stumbled on this game. Generally, the
more complex the social phenomena to which men and women are
exposed the more advanced their Reason. Reason develops as rational
agents struggle to come to grips with the maze of conjectures that social
interaction (of which our game is a simple example) creates. To use
Hegelian language, rationality is not to be defined axiomatically but is
to be understood as a process.33 If we wish to follow modernity in
picturing Reason as a totality, we may still do so. However, it is not a
static totality but one whose aspects are in contradiction with each
other. And through this internal feud, the aspects of the totality (e.g.
the conjectures in Table 2 or 4) transform not only the totality but also
each other.34
It is quite obvious that Hegel and Hume are on modernity's side.
Excepting their disparate language, what is the significant difference
CONCLUSION
NOTES
spent arguing about the postmodern condition, and to Joseph Halevi for
intransigence. Nonetheless, this paper should be blamed entirely on me.
1 Here I am referring to a specific model of bargaining that has come t
literature: the so-called Nash program. Also note that in game theory, a
omic analyses, to be rational is to know how to deploy your means effec
to achieve your ends. Rationality is exclusively instrumental.
2 See, for instance, Rubinstein (1982). In his model the distribution w
the 50-50 division to the extent that one player issues her demand befo
the delay between demands vanishes, the equilibrium outcome tends to b
3 One can, perhaps, accommodate the postmodern view in terms of the co
Before the theory of chaos, one expected the same algorithm to give ide
fed the same initial values twice. Since the study of non-linear models ha
because the input can never be exactly the same twice, the output ma
different. So, why should we expect our two agents to come to the same
they espouse slightly different conventions by which to predict the thou
their train of beliefs may lead them to seriously different conclusions and,
ment.
4 See Finelli (1990) who traces the debate on the nature and role of
oppositions to the Sophists.
5 Recall that game theory does exactly this. It starts with simple gam
division game described earlier, and assumes solutions for them. Once th
it then looks at more complex situations (e.g. asymmetric informatio
earlier assumptions to obtain explanation. This is what I call the analytic-s
of game theory.
6 This is a variant of a game that appears quite often in discussions of ga
for instance, Binmore (1987) and Sugden (1989, 1991). For the purpose
sition, I assume that A is female and B is male.
7 To be precise this is the so-called subgame-perfect Nash equilibrium. A
rium is an outcome brought about by strategies which are chosen on the
which are ex post confirmed by the outcome. The equilibrium is subgam
game comprises more than one stage and such a coordination of strate
(i.e. an equilibrium) is achieved not only for the whole game, but, also, in
8 Binmore (1987) criticises over-reliance on backward induction, Sugd
that it is possible to have a game theory without this kind of induction
less ambitious and Pettit and Sugden (1990) cement the arguments aga
Tables 2 and 3. More recently, Sugden (1991) provides a good summ
against backward induction.
9 Thinking about the possibility of defying the theory that is supposed
behaviour, is a uniquely human capacity. It is also a capacity that makes
social scientist inordinately demanding. To disallow counterf actuals w
(which is what Table 2 does) is to ask for serious trouble since human rat
bad habit of instructing agents to ask, 'what if I do not obey the the
Chapter 6 of Varoufakis (1991), I argue that counterfactual reasoning is, a
and incompatible with equilibrium game theory.
10 The reader may notice that I have made a rather strong assumption c
B expects an irrational A to do. Indeed, I assume that an irrational A
opposite of what is good for her. This has allowed us to assume that if p = 1 - i.e. if B
is convinced that A is irrational - then he expects her to play DOWN at t = 3 with
certainty. This is, of course, too restrictive. Nonetheless, the main point I am making is
not lost if the assumption is relaxed. Suppose, for instance, that an irrational A chooses
as if by randomisation. Then, at t = 3 an irrational A plays UP or DOWN with probability
1/2. In this case, p and q can be re-computed fairly easily and the argument remains
intact.
11 Those familiar with game theory may protest that game theorists recognise the legit?
imacy of a logic such as that in Table 4 without abandoning game theory's tenets - for
example see Kreps et al. (1982). This is correct. However Kreps et al (1982) can only
do this after they assume right at the start that agents have some doubt about the
rationality of their opponents. Table 4 by contrast does not require such a dilution of
the common knowledge of rationality assumption: non-equilibrium strategies are rational?
ised even when everyone is (at the beginning) absolutely sure that all others are perfectly
rational. The ideas in Kreps et al. (1982) become relevant in section 4 in which they help
construct a defence of game theoretical orthodoxy.
12 The reader may ponder the generality of my conclusion in view of the fact that I have
focused on a single game. Is it fair to discuss the whole project of game theory on the
basis of one example? I think it is. For this is an example that contains a unique
Nash equilibrium (subgame perfect) which should, if game theoretical thinking is to be
vindicated, produce an unequivocal rational strategy (due to the uniqueness of the
equilibrium). If the logic of Table 4 is compatible with full rationality, then we have
evidence that the existence of a unique equilibrium does not necessarily tell us what
agents will do. Since game theory trades on the thought that it ought to, one example
where this is untrue is as good as a thousand.
13 The Harsanyi doctrine occupies a central role in game theory since on it rest a very
large number of solutions that would otherwise break down. In the present context, the
negative defence draws on it heavily. However there are game theorists who do not
accept this doctrine and who, therefore, would not invoke the negative defence. Perhaps
they will be inclined to adopt the positive defence of the next section.
14 Suppose for instance that q > 1/50 and A plays DOWN at t = 1 but that B sets p at
1/20 and plays UP at t = 2. Alternatively, suppose that q > 1/50, A plays DOWN at t =
1, B sets p equal to 1/8 thus playing DOWN at t = 2 and, finally, A plays at t = 3. In
both these cases one of the two has formed expectations that are proven erroneous ex
post.
15 This is effectively the thesis in Bernheim (1984).
16 Recall the earlier argument that the equilibrium logic is perfectly legitimate even if
not uniquely so. Thus, a player may still choose to be prudent and assume that, since
her opponent is equally rational, there is nothing she can do to confuse her.
17 Table 2, for example, depends on the unique projectibility of traits established at t -
3 onto t = 2 and t?\.
18 The game of Table 1 is truly interactive in that what player A does at t = 1 depends
entirely on what A thinks that B will think if... It is in such a game that the enigma of
human reasoning becomes pertinent and wrecks the certainty of backward induction. In
other cases, where the choice of one player can be made independently of conjectures
concerning the actions of another, then of course induction is straightforward. Consider
the following ten dot game. There are ten dots which two players take
The first player begins and may erase either one or two dots. Then
player's turn to either erase one or two dots. The player who crosses
wins. Working backwards, it is clear that the player who plays first has a
strategy: to erase only one dot at the beginning. In this way, she can be
out the 4th, 7th and, finally, the 10th dot whatever player B's choices. Bac
works impeccably in this game because A does not need to consider wha
A plays in one way rather than in another. Then, the trait identified at
the game is uncontentiously projectible to the very first stage when the g
19 Postmodernity actually rejects the very possibility of a concept
purposes, it may argue that analogies are often mistaken for concep
demonstrate the vacuousness of concepts.
20 The reader who would like to leave open the possibility that an irra
in an unpredictable manner will protest that this is too stringent an assum
the analysis will not change significantly if we envision an irrational pla
who chooses between UP and DOWN as if by randomisation. Footnot
with equal force.
21 I assume that A believes B to be rational with probability 1 ? w =
be easily extended to allow for two-sided uncertainty concerning ration
w>0.
22 For the updating mechanism to make intuitive sense, 1 - r > s, i.e. the probability
that A will adopt some logic different to that of Table 2 (RBI) if rational and thus choose
the deviant strategy must exceed the probability that an irrational A will choose the
deviant strategy. This is very sensible since otherwise there would be no reason for B to
believe that DOWN at t = 1 enhances the prospects that A is irrational.
23 Naturally, part (k) of Table 6 is tantamount to the condition px > 1/21.
24 Binmore (1987, 1988) has also voiced concern about the undiscriminating use of Bayes'
rule.
25. There is an interesting parallel here with Foucault's (1967) critique of 'modernity's
monologue'. Foucault claims that before the triumph of modernity, there used to be a
dialogue between rationality and madness. Later, this dialogue broke down and left us
with a monologue of rationality on madness. And yet, he goes on, there are dimensions
of sense in madness that are missing in what we tend to think of as Reason, or to put it
differently, there is a great deal of Reason in madness. Any attempt to evict madness
altogether in order to procure pure Reason is, therefore, ill-conceived. The reader who
is so inclined may interpret the assumption that there exists a uniquely rational estimate
of s as a technical manifestation of illegitimate attempts to cement this monologue.
26 A player's conceptualisation of her opponent's conjectures is, in itself, a theory. To
argue that one attaches, via induction, probabilities to different such theories and, in
addition, to insist that these probabilities are common property, is philosophically absurd.
Peirce (1932) draws the important distinction between the probability of a hypothesis
and the probability derived from a hypothesis. He writes:
of them the law held good . . . What induction does... is infinitely more to the
purpose.
27 The reason is that if A goes DOWN at t= 1, then Equation (1) will update B's
probabilistic assessment that A is irrational to p2 = 1/11. This posterior belief makes B
indifferent between UP and DOWN at t = 2. Thus, A anticipates that DOWN at t = 1
will make B randomise at t = 2, a thought that makes her unsure as to whether she ought
to play DOWN at t = 2. Consequently, she also randomises at t = 1.
28 See Skyrms (1990) for a discussion of deliberational disequilibrium.
29 Table 4 has presented this critique of equilibrium theory implicitly. Let p1 = 0. Then,
if player A played DOWN at t = 1, Equation (1) cannot be defined: an event occurred
that B had attached a zero probability to. So, what should B do in such a situation?
According to the equilibrium story, there is no answer. Can we speculate that, in the
absence of advice by the theory, player B may still revise p upwards (i.e. p2 > 0). If A
expects this to happen (and there is no reason why she should not), then she may
rationally choose DOWN at t= 1. Of course, there can be no equilibrium account of
what has happened. Therefore, equilibrium theory is inferior to the account of table 4
because rational agents may have an incentive to violate it.
30 Nietzsche wrote: "What therefore, is truth? A mobile army of metaphors, metonymies,
anthropomorphisms; truths are illusions of which one has forgotten they are il?
lusions . . . coins which have their obverse effaced and now are no longer of account as
coins but merely as metal" On Truth and Falsity on their ultramoral Sense in Levy (1964).
However, Heidegger successfully exposed holes in his arguments by demonstrating that
Nietzsche needs truth as a concept in order to argue against its meaning. Interestingly,
this is also a problem for Heidegger. Finelli (1990) claims that Being is denied by
Heidegger and his contemporary postmodernist followers, but that in their philosophy it
returns to determine human reality through its loss and emptiness.
31 Sugden (1991) illustrates this point in the context of (i) a critique of Savage's expected
utility theory and (ii) the theory of games.
32 Gerhard Adler writes: "The enigma of creativeness rooted in the irrational, indefinable
matrix of man's timeless psyche has held eternal fascination for him and has helped
produce the most memorable justification of his status as man" - see the Foreword in
Kirsch (1966).
33 The social anthropologist Levi-Strauss (1966) defines analytical Reason as the type of
logic that develops when humans try to understand natural phenomena of a low order
(e.g. hydrodynamics as opposed to the concept of time). He thinks that such logic is
frustrated when it is called upon to explain social phenomena. The result of this failure
is a new kind of Reason which, in Hegelian fashion, he terms dialectical. "... [Dialectical
reason thus covers the perpetual efforts analytical reason must take to reform itself if it
aspires to account for language, society and thoughts; and the distinction between the
two forms of reason in my view lies on the temporary gap separating analytical reason
from the understanding of life. Sartre calls analytical reason reason in repose; I call the
same reason dialectical when it is roused by action, tensed by the effort to transcent
itself."
34 Of course, postmodernity is eager to attack Hegel in the same way that it disparaged
Hume. The contradiction on which Hegel bases the sublation of Reason is seen as
both unresolvable and as unreal. It is unresolvable because Reason is meaningless and,
therefore, hardly capable of improving itself. It is unreal because when we talk of the
contradiction, we fall victims to the inferiority of our language. The latter is forced,
through its imprecision, to contrive false categories (such as Reason and Unreason) when,
in reality, the contradiction is, like truth in Nietzsche, an illusion that we have forgotten
that it is an illusion.
References
Aristotle: 1987, Nicomachean Ethics, transi, by J. Welson, Prometheus, New York.
Bernheim, D.: 1984, 'Rationalisable Strategic Behaviour', Econometrica 52, 1007-28.
Binmore, K.: 1987, 1988, 'Modeling Rational Players: Parts I and II', Economics and
Philosophy 3, 179-214 and 4, 9-55.
Derrida, J.: 1978, Writing and Difference, Routledge and Kegan Paul, London.
Finelli, R.: 1990, 'Production of Commodities and Production of Images: Reflection on
Modernism and Postmodernism', mimeo, Faculty of Philosophy, University of Rome.
Foucault, M.: 1967, Madness and Civilisation, Tavistock, London.
Harsanyi, J.: 1973, 'Games with Randomly Disturbed Payoffs: a New Rationale for
Mixed Strategies', International Journal of Game Theory 2, 1-23.
Hegel, G. W. F.: 1931, The Phenomenology of Mind, translated by J. Baillie, London.
Kirsch, J.: 1966, Shakespeare's Royal Self, Putnam and Sons, New York.
Kreps, D., P. Milgrom, J. Roberts, and R. Wilson: 1982, 'Rational Cooperation in the
Finitely Repeated Prisoner's Dilemma', Journal of Economic Theory 27, 245-52.
L?vi-Strauss, C: 1966, The Savage Mind, Weidenfeld and Nicholson, London.
Lyotard, J.-F.: 1984, The Postmodern Condition: A Report on Knowledge, Manchester
University Press, Manchester.
Norris, C: 1985, The Contest of Faculties: Philosophy and Theory After Deconstruction,
Meuthen, London.
Nietzsche, F.: On Truth and Falsity in their Ultramoral Sense, in Levy, O.: 1964, The
Complete Works of Friedrich Nietzsche, New York.
Peirce, C. S.: 1932, Collected Papers Vol. 2, Harvard University Press, Cambridge Mass.
Pettit, F. and R. Sugden: 1989, 'The Paradox of Backward Induction', Journal of Philos?
ophy, LXXXVI, 169-82.
Skyrms, B.: 1990, The Dynamics of Rational Deliberation, Harvard University Press,
Cambridge Mass.
Sugden, R.: 1989, 'Game Theory without Backward Induction', mimeo, University of
East Anglia.
Sugden, R.: 1989b, 'Spontaneous Order', Journal of Economic Perspectives, 3, 118-25.
Sugden, R.: 1991, 'Rational Choice: A Survey of Contributions from Economics and
Philosophy', The Economic Journal 101, 751-85.
Varoufakis, Y.: 1991, Rational Conflict, Basil Blackwell, Oxford.
Department of Economics
University of Sydney
Sydney 2006
Australia