See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/6378575

The emergence of simple language in an experimental coordination game

Article  in  Proceedings of the National Academy of Sciences · June 2007

DOI: 10.1073/pnas.0702077104 · Source: PubMed

CITATIONS READS
119 415

2 authors, including:

Massimo Warglien
Università Ca' Foscari Venezia
88 PUBLICATIONS   2,226 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Conceptual Spaces View project

COPE -- Center for Organizational Plasticity and Evolution View project

All content following this page was uploaded by Massimo Warglien on 22 May 2014.

The user has requested enhancement of the downloaded file.

The emergence of simple languages in an
experimental coordination game
Reinhard Selten*† and Massimo Warglien‡
*Laboratorium für Experimentelle Wirtschaftsforschung, University of Bonn, Adenauer Allee 24-42, D-53113 Bonn, Germany; and ‡Ca’ Foscari
University of Venezia, Scuola Superiore di Economia, Fondamenta San Giobbe 873, Cannaregio, 30121 Venice, Italy

Contributed by Reinhard Selten, March 6, 2007 (sent for review November 3, 2006)

We investigate in a series of laboratory experiments how costs and plain the properties of binary relations appearing in natural
benefits of linguistic communication affect the emergence of language. Blume (15) has modeled how learning efficiency is
simple languages in a coordination task when no common lan- favored by modularity of language. Weber and Camerer (16)
guage is available in the beginning. The experiment involved report an experimental study of the emergence of a common
pairwise computerized communication between 152 subjects in- code based on natural language under efficiency pressure.
volved in at least 60 rounds. The subjects had to develop a common The experiments described here try to explore how costs and
code referring to items in varying lists of geometrical figures benefits of linguistic expression affect the emergence of simple
distinguished by up to three features. A code had to be made of a languages or codes in a coordination task. At the beginning, the
limited repertoire of letters. Using letters had a cost. We are subjects do not yet have a common language; they must therefore
interested in the question of whether a common code is developed, develop it in the course of interaction.
and what enhances its emergence. Furthermore, we explore the We look at the emergence of languages in an extremely
emergence of compositional, protogrammatical structure in such simplified laboratory environment. We consider only the bene-

SOCIAL SCIENCES
codes. We compare environments that differ in terms of available fits of transferring meaning correctly and the cost of language
linguistic resources (number of letters available) and in terms of expression. Thus, factors emphasized by sociolinguists (17) are
stability of the task environment (variability in the set of figures). not present in our experimental setup. This permits us to isolate
Our experiments show that a too small repertoire of letters causes the two ‘‘economic’’ factors.
coordination failures. Cost efficiency and role asymmetry are In psycholinguistics, there is a tradition exploring how inter-
important factors enhancing communicative success. In stable acting individuals adapt natural language to coordinate in some
environments, grammars do not seem to matter much, and instead unfamiliar task (18, 19). Weber and Camerer (16) have success-
efficient arbitrary codes often do better. However, in an environ- fully extended such type of experimental design to the investi-
ment with novelty, compositional grammars offer considerable gation of economic issues. Our experiments are different be-
coordination advantages and therefore are more likely to arise. cause our subjects cannot use a natural language but have to
shape an artificial one. This permits us to put more in focus the
communication 兩 compositionality 兩 economics of language emergence of common codes and to explore their structural
properties.
Our experiments are also very different from those that inves-
A ccording to the linguist André Martinet, language is shaped
‘‘by the permanent conflict between man’s communicative
needs and his tendency to reduce to a minimum his mental and
tigate the acquisition of artificial grammar (20), because our focus
is on the emergence of an artificial language rather than on its
physical activity’’ (1). This means that the benefits of commu- acquisition. Moreover, we do not look at the evolution of linguistic
nication are compared with the memory and articulation costs competence. Of course, all our subjects are already endowed with
of linguistic expression. A tendency toward the optimization of this competence. We investigate how simple languages are created
the difference between benefits and costs is postulated. by linguistically competent subjects in a situation where no common
Even if many contemporary linguists would disagree with language is available in the beginning.
Martinet’s view, it is widely recognized that some sort of A recent paper by Galantucci (21) explores the emergence of
economizing principle is at work in shaping human language: for graphical signs in the environment of a video game that requires
example, Chomsky’s (2) ‘‘minimalist program’’ relies on assump- cooperation of two subjects. This paper offers some interesting
tions of economy of derivation and representation, and on insights based mostly on a qualitative discussion of the behavior
principles of least effort. However, he maintains that ‘‘questions of 10 pairs of subjects. Our experiment was developed indepen-
concerning abstract computational mechanisms are distinct dently and differs in many respects from Galantucci’s. First, our
from those concerning communication’’ (3). Another interesting focus is more on the structure of language rather than the nature
approach to language structure is optimality theory (4), which of signs. In our experiments, signs are predefined and have to be
postulates a process of constraint satisfaction that guides the combined. Moreover, our communication protocol is more
shaping of linguistic outputs. constrained, communication is costly (unlike in Galantucci’s
A recent, growing body of literature has exploited principles experiment), and the payoff structure is quite different. Finally,
of communication effectiveness and efficiency to model, math- our experiment is based on a much broader sample of subjects,
ematically or by computer simulation, the emergence of gram- permitting a deeper statistical analysis.
In our experiments, subjects see a list of geometrical figures on
matical languages from initial no communication (5–10). As will
a computer screen. These figures differ by shape (circle, square,
become apparent later, the basic models used in this literature
have some resemblance with the structure of our experiments.
Moreover, we should mention a long tradition of using prin- Author contributions: R.S. and M.W. designed research; R.S. and M.W. performed research;
ciples of least effort to explain and model language statistical R.S. and M.W. analyzed data; and R.S. and M.W. wrote the paper.
regularities like Zipf’s law in word distribution (11–13). The authors declare no conflict of interest.
There is also a small but growing literature on the economics †To whom correspondence should be addressed. E-mail: [email protected].
of language that relates language structure to cost–benefit This article contains supporting information online at www.pnas.org/cgi/content/full/
considerations. On the theoretical side, Rubinstein (14) has 0702077104/DC1.
shown how optimization considerations can contribute to ex- © 2007 by The National Academy of Sciences of the USA

www.pnas.org兾cgi兾doi兾10.1073兾pnas.0702077104 PNAS 兩 May 1, 2007 兩 vol. 104 兩 no. 18 兩 7361–7366

 periods figures letters: I letters:
(cost) II,III, IV
(cost)
1-10 R,S(2) R,S(2)

11-30 + + R,S(2) R,S (2),

M,Z (3)

+ +
31-60 R,S(2), R,S,V(2),
Z(3) M,Z,F(3)
+
* * *
1a Experiments I,II,II (session I), IV (session I)

periods figures letters (cost:2)

Fig. 2. Figures in experiment IV’s second session.

1-6 Z,H,N
(referred to as experiments I, II, III, and IV). It shows which
figures were presented in which periods and which repertoire of
7-16 Z,H,N, U,Y,T
letters was available. We use the name ‘‘block’’ for a section of
an experiment in which figures and repertoire of letters remain
unchanged. Each of the experiments I and II had only one
session of 60 periods. In experiments III and IV, the participants
17-26 Z,H,N, A,M,Q,J
stayed for a second session, which began after a short pause in
which they could not communicate. The first and the second
6 blocks of 6 figures.
session of experiment III were identical with experiment II. The
Z,H,N,U,Y,T, same is true for the first session of experiment IV, but not for the
27-62 The 3 shapes, 3 inserts, 3 colors A,M,Q,J
introduced before: second one. There, first, sufficiently many features are intro-
randomization of order within block duced, to induce some familiarity with the structure of the
environment. Then, 36 new combinations of three features
1b Experiment IV (session II) (shapes, inserts, and colors) are presented in blocks of six and
selected for communication just once (Fig. 2). This ‘‘novelty’’
Fig. 1. The experimental treatments. structure was explained to the subjects just before the beginning
of the second session. Also, in experiment III, subjects were not
informed about the conditions of the second session before the
etc.), inserts (star, dot, etc.), and sometimes by color. To each of end of the first one.
the figures in the screen, a subject must assign a message, a string
of permissible letters. The subjects interact anonymously in Results
pairs. They both always face the same list of figures and the same The Impact of the Size of the Repertoire. In our first experiment (I),
repertoire of permissible letters. In each period, one subject and only two (and then three) letters were permissible. After running
one figure are randomly chosen, and then the message specified
the experiment, we observed less convergence to common code
by the code is transmitted to the other player. The transmission
then we had expected. In the literature on language evolution
is successful if and only if the messages specified by both codes
(22), it is common to emphasize the role of the human capability
are the same. A payoff is obtained for a communication success,
to produce a great variety of phonetic signals for the develop-
but the letters have costs that must be borne by the sender. After
each period, both players receive feedback on the chosen figure ment of linguistic competence. Therefore, we supplied a larger
and the messages specified by the code of the receiver. After repertoire of permissible letters in all further experiments.
receiving the feedback, they can change their codes. Table 1 reports results of experiments I on the one hand and
We are interested in whether a common code is developed, experiments II and the first sessions of III and IV on the other
and what enhances its emergence. We also explore the structure hand. The latter three sessions are identical and can therefore be
of common codes, which may or may not have a grammatical pooled. We refer to them as the ‘‘pool.’’
structure. We compare environments that differ with respect to The table shows in how many cases a common code was
the size of the repertoire of permissible letters and with respect reached. We say that a pair has reached a ‘‘common code’’ if the
to novelty (i.e., the frequency with which new figures occur). codes of the two players for the last period agree with respect to
We conducted four versions of our experiment. The rules all figures chosen for communication in the course of the game.
apply to all of them. The versions differ with respect to the (Because figures are chosen randomly, some of them may never
figures presented and the repertoire of permissible letters and be chosen for a pair. Of course, one cannot expect code
their costs. All experiments begin with a small number of figures agreement for such figures.)
and small repertoires (two or three figures and the same number As Fig. 1 shows, experiment I and the pool differ only with
of letters). The set of figures and the repertoire of letters expand respect to the size of the repertoire of the letters. This difference
over time in different ways. is important for the emergence of a common code. This effect
Fig. 1 gives an overview of the four versions of our experiment is significant on the 10% level (P ⫽ 0.069) according to a Fisher’s

7362 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0702077104 Selten and Warglien

Table 1. Pairs reaching a common code in experiments I and
II–III–IV (pool)
Common code No common code
+ +
Pool 28 29
Experiment I 5 14 a list of 6 figures

exact test, one-tailed. This suggests that a greater repertoire is in none +

fact facilitating the emergence of linguistic coordination.
From a logical point of view, two symbols are sufficient for the R RM RZ
construction of a code. In principle, communication could be
based on a binary code. However, it seems to be the case that the S SM SZ
availability of a sufficiently large variety of letters not only makes
it easier to achieve communication efficiency but also has a
cognitive effect that facilitates linguistic coordination. Later, in
connection with our discussion of grammars, we shall come back a compositional grammar
to this point.

The Influence of Role Differentiation and Efficiency on Payoffs. In this none +

section, we restrict ourselves to the analysis of the pooled
experiments (II, and the first session of III and IV). We shall look R RR RS
at rank correlations among a number of aggregate variables

SOCIAL SCIENCES
defined for pairs as follows: S SS SR
Agreement. Average number of figures with the same code for the
two players, over all periods.
Changes. The sum of all code changes for all periods. A code
a noncompositional grammar
change is a change of a message attached to a figure (e.g., if, in
a period, the messages for three figures are changed, this counts
as three code changes).
Asymmetry. The fraction of all code changes made by the player
with more code changes.
none +
Code costs. The sum of code costs for all figures and both players
R M SS
averaged over periods.
Payoff. The sum of all net payoffs for both players in all period.
S RZ Z
Fig. 3 shows a graph whose nodes stand for the variables
defined above and whose edges indicate Spearman rank corre-
lations significant at least at the 1% level, one-tailed. The
number at an edge is the Spearman rank correlation coefficient a nongrammatical code
of the variables associated to the points connected by the edge.
Fig. 4. Types of code.
A causal interpretation of Fig. 3 from left to right suggests
itself: code costs reflect the efficiency of coordination within a
pair. The higher the code cost, the lower its payoff. A high degree confusion by simultaneous changes: if, after a period, both
of asymmetry results in few changes and much agreement. players adjust to the code of the other, differences in codes are
Payoffs in turn are higher the more agreement and the fewer
not reduced. Asymmetry ensures that any change reduces code
changes there are. The negative correlation of changes and
differences.
agreement is a consequence of the common causal influence of
It is good for the payoff of the pair if they choose low-cost
asymmetry.
codes and define asymmetric roles as early as possible. The two
Total asymmetry means that one of the players acts as the
leader (or the teacher) and the other as a follower who passively factors seem to be independent: there is no significant correla-
adapts to the leader’s code. This role differentiation avoids tion between asymmetry and code cost, notwithstanding the
relatively high number of 57 observations.

changes Language Structure: Forms of Grammar. It is a further purpose of

-.64 this study to investigate factors affecting the structure of com-
mon codes. We distinguish between common codes with and
-.78
-.58 without a grammar. A code has a compositional grammar if it is
asymmetry obtained by mapping instances of features to letters or strings of
letters, called graphemes, and by arranging them in a fixed order
+.82 payoff of features. Thus, there may be two features, shapes and inserts,
+.45 two instances of ‘‘shape,’’ circle and triangle, and two instances
agreement
of ‘‘insert,’’ dot and cross. The compositional grammar shown in
-.54 Fig. 4 uses the graphemes ‘‘R’’ and ‘‘S’’ for circle and triangle,
and ‘‘M’’ and ‘‘Z’’ for dot and cross. In the fixed order, shape
code cost
precedes insert. Thus, our use of the word ‘‘compositional’’
Fig. 3. Spearman rank correlations among aggregates for pairs significant at conforms to the tradition that begins with Frege (23), according
least on the 1% level, one-tailed. to which a language is compositional when the meaning of a

Selten and Warglien PNAS 兩 May 1, 2007 兩 vol. 104 兩 no. 18 兩 7363
 Table 2. Communication achievement of pairs in the last period
Common code

No common Noncompositional Compositional Number of pairs,

code Ungrammatical grammar grammar all experiments

66 (58%) 32 (28%) 1 (1%) 14 (12%) 113

In experiments III and IV, the two sessions are counted separately.

complex expression can be derived uniquely from the meaning mars seems to be related to the concepts of productivity and
of its constituents and the rules for combining them. This systematicity, usually associated to compositionality in the se-
definition survives in modern semantics (24). mantic literature (26); however, we do not want to go into more
One also finds noncompositional grammars and ungrammat- details in this respect here.
ical codes in the data. Admittedly, we cannot offer an exact Actually, there is no necessity to use a different letter for every
general definition of the boundary line between these two instance of every feature. A fixed position for shapes, positions,
categories. We speak of a noncompositional grammar if a and colors has to be fixed anyhow, given the constraints of the
principle suggests itself that generates the relationship. experiment. Thus, the same letter could be used for a circle, a
The noncompositional grammar of Fig. 4 maps circles and star, and blue. However, it seems to be very difficult for the
triangles to the letters ‘‘R’’ and ‘‘S,’’ respectively, and maps the subjects to devise positional codes that rely heavily on positions
instances of inserts into the number of letters. Obviously, of letters to convey meaning. This is probably an important
the code has a regular structure but not a compositional one. The reason why less common codes are found in experiment I, where
nongrammatical code of Fig. 4 also follows a principle. The first only one additional letter is made available for two new instances
letter stands for the shape, a star is indicated by two equal letters, of features in the transition from block two to block three.
and the cross by two different letters. Therefore, no compositional grammar was ever adopted by a
Of course, our concept of grammar does not capture the pair during experiment I. In this respect, it is interesting that a
intricacies of the use of the term in linguistics. It does not go full positional system of expressing numbers evolved only rela-
beyond what is called a ‘‘protolanguage’’ by Jackendoff (25) (i.e., tively late in history (27). Less efficient systems like the roman
a combination of linear order and semantic roles). However, it numbers lasted for a long time before positional notations were
is sufficient within the frame of our experiments. introduced.
Table 2 shows what kind of status, with respect to a common
code, pairs have achieved in the last period. As one can see, no Emergence of Grammar in Response to Expressive Needs. An advan-
common code is reached in the majority of all cases. Among the tage of grammar is the possibility to express new contents never
common codes, approximately two of three are ungrammatical. expressed before, but nevertheless having them understood by
Noncompositional grammars are rare, and grammars, if they are other speakers of the same language. Bickerton (28) and Pinker
present in the last period, tend to be compositional. and Bloom (29) discuss this point from an evolutionary stand-
point (although from different perspectives). This advantage
Extendability and the Persistence of Grammars. At the end of the counts in a sufficiently rich environment.
experiments, most grammars are compositional, but many non- Experiment IV provides such an environment in the second
compositional grammars can be found at the end of block two. session. After introducing sufficiently many combinations of two
In the course of our experiments (with the exception of the features, 36 new combinations of three features are presented
second session of experiment IV), new figures and new letters just once. This is explained in the instructions before the session
are introduced from one block to the next. The old figures begins [see supporting information (SI) Text].
remain in the list, and new ones are added. We speak of an In the second session of experiment IV, only 5 of 17 pairs
‘‘extension’’ of an earlier code if the figures that were present developed grammars, all of them compositional, and reached
before are associated with the same messages as before. We say agreement soon after the beginning of the last phase.
that a common code of a pair at the end of block two is Table 4 shows the average payoffs for pairs with and without
‘‘persistent’’ if, at the end of block three, this pair has a common grammars in experiment IV’s second session. All five pairs
code that is an extension of its earlier code. with grammar had higher payoffs than each of the 12 pairs
Table 3 shows the persistence of different types of grammars without grammar (the Mann–Whitney U test yields a significance
in the sessions of the pool. The compositional grammars are all of 0.001 for this).
persistent, whereas all noncompositional ones are nonpersistent Compositional grammars are highly successful in the novelty-
(Fisher’s exact test yields a significance of 1%, two-tailed). providing environment of experiment IV and languages without
Moreover, all pairs with noncompositional grammars at the end such grammars are far less successful. The relationships between
of block two have no grammar at the end of the session. grammar and the expressive needs created by novelty are further
The results show that compositional grammars are easily illustrated by the comparison between experiment III and ex-
extendable, whereas the same is not true for noncompositional periment IV.
grammars. The easy extendability of compositional grammars is The first sessions in experiment III and IV follow the same
facilitated by the availability of one new letter for every new rules. In experiment III, the same rules apply also to the second
instance of a feature. The extendability of compositional gram- session. The comparison of experiment III and IV aims at the

Table 3. Persistence of grammar types present at the end Table 4. Average payoff per pair in experiment IV,
of block two second session
Persistence No persistence Pairs in IV, second session Average payoff

Compositional 5 0 With grammar 635 talers

Noncompositional 0 5 Without grammar 211 talers

7364 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0702077104 Selten and Warglien

Table 5. Code structure by independent groups, experiments transmission contexts (e.g., from parents to children, from
III and IV natives to immigrants), such status differences are pervasive.
With grammar Without grammar (iii) In stable environments as those considered in point ii,
grammar does not matter much, and efficient arbitrary codes
First session often do better. However, compositional grammars have the
Experiment III 2 7 advantage of being more easily extendable to broader environ-
Experiment IV 1 5 mental demands. Noncompositional grammars are more fragile
Second session and are easily lost if new conditions have to be met.
Experiment III 1 8 (iv) In an environment with novelty, in the sense that often the
Experiment IV 4 2 need arises to express something that never has been expressed
before, compositional grammars offer considerable coordination
advantages. Therefore, under such circumstances, compositional
exclusion of mere experience effects as an explanation of the grammars are more likely to arise. In this respect, our findings
emergence of grammars in the novelty treatment of the second parallel and complement hypotheses proposed in the literature on
session of experiment IV. language evolution (28, 29). In our experiments, all subjects have
At the end of the first session of both experiments III and IV, grammatical competence but they make relatively little use of it
subjects had to be reassigned to new pairs for the second session. unless pressure of novelty gives them an incentive to do so.
Thus, in comparing such experiments, we will consider indepen-
Methods
dent subject groups as a unit of observation (i.e., groups of
participants such that there is no matching across groups in the In our experiments, the subjects sit at computer terminals and
first and second sessions). cannot communicate verbally with each other. The interaction is
We are primarily interested in numbers for the second session anonymous and only by formal messages. Subjects are randomly
because, there, a comparison of the effects of novelty and mere paired at the beginning of a session and the pairs remain the same

SOCIAL SCIENCES
repetition is possible. Fisher’s exact test yields a significance of for all periods of the session. During a session, the subjects in a pair
2%, one-tailed, for the four-field table for the second session always interact only with each other. We now shortly explain the
(Table 5, session 2). Significantly more groups with a grammar general features common to all of our experiments. Detailed
emerge in the second session of the novelty condition. We can instructions to subjects are reported in the SI Text.
exclude the possibility that the increase of the number of
Players. There are two players.
grammars is due to a learning effect that should also be there in
the second session of the ‘‘repetitive’’ experiment III.
Periods. The experiment runs over 60 (or 62) periods.
Discussion
Figures. In every period, each of the two players sees the same
Our experiments present a very simplistic and constrained envi-
figures on the screen, listed in different random order.
ronment for the evolution of common codes. Nevertheless, the
results surprisingly reproduce many features of natural languages,
Structure of Figures. Figures have shapes, inserts, and, in one
and may throw light on interactive factors affecting the evolution of
experiment, colors. These ‘‘features’’ have different ‘‘instances’’
a language. In the following, we discuss some broader implications (shapes: circles, squares, etc; inserts: dots, crosses, etc.; colors: red,
of our findings, which of course runs the risk of exaggerating the yellow, etc.).
transferability of our laboratory results to broader contexts.
(i) The size of the repertoire of elementary linguistic symbols Codes. In every period, each of the two players has the same
seems to be important in facilitating linguistic coordination. In repertoire of permissible letters (e.g., ‘‘R,’’ ‘‘S,’’ and ‘‘T’’) to be
the experiments of Galantucci (21), there seems to be a high rate used in messages that take the form of strings of permissible
of success in creating communication systems. We also find letters. Each player has to fix a code assigning a message to every
successful coordination (although at somehow lower rates), but figure. The code can be changed in every period.
our experiment I shows that a too small size of the repertoire may
be a serious obstacle for the attainment of a common code. This Communication. In every period, one player, the sender, and a
is not only due to the fact that more symbols permit a higher figure are randomly chosen. The message corresponding to this
degree of cost efficiency by shorter expressions but also to some figure in the code of the sender is transferred to the other player,
properties of the human cognitive system. If the number of the receiver. The message is understood if the codes of the
elementary symbols is too small, one has to rely more on sender and the receiver agree with respect to the message
positional structure. Positional systems of expressing meaning, assigned to the chosen figure.
like in the Arabic number notation, are hard to devise for the
human mind, although they can be easily used once they are Period Payoff. Each player receives 10 units of the experimental
available, as shown by the history of mathematics. currency (taler), if the message has been understood, and zero
(ii) In an environment in which the same messages occur many talers otherwise. However, the sender bears the cost of the message,
times, cost efficiency and role asymmetry are factors enhancing composed of costs for each letter in the string. The costs of letters
communicative success, whereas grammars do not offer partic- (e.g., three talers for ‘‘R’’ and two talers for ‘‘S’’ and ‘‘T’’) do not
ular advantages under such circumstances. Role asymmetry change over time and are known to the players when they fix their
between a leader and an imitator avoids mismatches by simul- codes.
taneous adjustments to the code of the other. In dialogue theory
(19), the role of imitation is also emphasized as conductive to the Feedback. After each period, both players are informed about
conversational alignment of interlocutors. Our results throw whether the message was understood. If the message was not
additional light on this phenomenon and suggest looking more understood, feedback is supplied to the receiver about the chosen
closely to how role asymmetry might facilitate alignment pro- figure and to the sender about the message attached to the figure
cesses. The importance of role asymmetry (30) is clearly visible in the receiver code.
in the experiment despite the fact that there are no natural or The experiments were run at the Experimental Economics
social status differences among the subjects. In natural language Laboratory of Trento University (Trento, Italy). In an experimental

Selten and Warglien PNAS 兩 May 1, 2007 兩 vol. 104 兩 no. 18 兩 7365
session, 34–40 participants were organized in fixed pairs interacting Marco Tecilla gave invaluable help writing the code of the experi-
anonymously by computer terminals. Communication was re- mental software. Phil Johnson-Laird provided helpful comments. We
stricted to the exchange of formal messages according to the rules. thank the Rovereto Cognitive Science Laboratory and the Trento
Computable and Experimental Economics Laboratory for their or-
An experiment lasted 1–2.5 h. Subjects received monetary payoffs ganizational and financial support in running the experiments. Min-
proportional to their earnings in talers. The conversion rate was two istero dell’Università e della Ricerca Grant 2005139342 and Fondo per
euro cents per taler, plus a fixed sum of five euros for their gli Investimenti della Ricerca di Base Grant RBNE03A9A7 contrib-
participation. uted to the project funding.

1. Martinet A (1964) Elements of General Lingusitics (Faber and Faber, London). 16. Weber RA, Camerer C (2003) Manage Sci 49:400–415.
2. Chomsky N (1995) The Minimalist Program (MIT Press, Cambridge, MA). 17. Labov W (2001) Principles of Linguistic Change (Blackwell, Oxford), Vol 2.
3. Hauser MD, Chomsky N, Fitch WT (2002) Science 298:1569–1579. 18. Clark HH, Wilkes-Gibbs D (1987) Cognition 22:1–39.
4. Prince A, Smolensky P (2004) Optimality Theory: Constraint Interaction in 19. Pickering MJ, Garrod S (2004) Behav Brain Sci 27:169–190.
Generative Grammar (Blackwell, Oxford). 20. Reber AS (1993) Implicit Learning and Tacit Knowledge (Oxford Univ Press,
5. Nowak MA, Krakauer D (1999) Proc Natl Acad Sci USA 96:8028–8033. Oxford).
6. Nowak MA, Plotkin JB, Jansen VAA (2000) Nature 404:495–498. 21. Galantucci B (2005) Cogn Sci 29:737–767.
7. Nowak MA, Komarova NL, Niyogi P (2001) Science 291:114–118. 22. Lieberman P (1984) The Biology and Evolution of Language (Harvard Univ
8. Batali J (1998) in Approaches to the Evolution of Language, eds Hurford JR, Press, Cambridge, MA).
Studdert-Kennedy M, Knight C (Cambridge Univ Press, Cambridge, UK), pp 23. Frege G (1882) Zeitschrift fur Philosophie und Philosophische Kritik 100:25–30.
405–426. 24. Cann R (1993) Formal Semantics: An Introduction (Cambridge Univ Press,
9. Steels L (2000) in Proceedings of ECAI 2000, ed Horn W (IOS Publishing: Cambridge, UK).
Amsterdam). 25. Jackendoff R (2002) Foundations of Language (Oxford Univ Press, Oxford).
10. Kirby S (2001) IEEE Trans on Evol Computation 5:102–110. 26. Fodor J, Lepore E (2002) The Compositionality Papers (Oxford Univ Press:
11. Zipf GK (1949 Human Behaviour and the Principle of Least-Effort (Addison– Oxford).
Wesley: Cambridge, MA). 27. Boyer CB (1968) A History of Mathematics (Wiley, New York).
12. Mandelbrot BB (1953) in Proceedings of the Symposium on Applications of 28. Bickerton D (1990) Language and Species (University of Chicago Press,
Communications Theory (Butterworth, London), pp 486–502. Chicago).
13. Ferre I, Cancho R, Solé RV (2003) Proc Natl Acad Sci USA 100:788–791. 29. Pinker S, Bloom P (1990) Behav Brain Sci 13:707–726.
14. Rubinstein A (1996) Econometrica 64:343–356. 30. Tomasello M (1999) The Cultural Origins of Human Cognition (Harvard Univ
15. Blume A (2000) J Econ Theory 95:1–36. Press, Cambridge, MA).

7366 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0702077104 Selten and Warglien

View publication stats