done to investigate the possibility that
the formation subjects built good networks for the task, but either ran out of
time to reach unanimity, or included
subjects who behaved very stubbornly
because they had significant edge expenditures and thus strongly held out
for their preferred color.
Performance on the control experiments was even worse. The surprising
conclusion seems to be that despite
the fact that subjects clearly understood the task, and were now given
the opportunity to solve it not on an
arbitrary network, but one collectively
designed by the population in service of the task, they were unable to
do so. One candidate for a structural
property of the subject-built networks
that might account for their difficulty
in the biased voting task is (
betweenness) centrality, a standard measure of
a vertex’s importancee in a network.
Compared to the networks used in the
original, exogenous-network biased
voting experiments, the distribution
(across vertices) of centrality in the
subject-built networks is considerably
more skewed.
16 This means that in
the network formation experiments,
there was effectively more reliance on
a small number of high-centrality vertices or players, making performance
less robust to stubbornness or other
non-coordinating behaviors by these
players. Indeed, there was moderately
positive and highly significant correlation between centrality and earnings,
indicating that players with high centrality tended to use their position for
financial gain rather than global coordination and information aggregation.
Despite their demonstrated ability to solve a diverse range of computational problems on a diverse set of
networks, human subjects seem poor
at building networks, at least within the
limited confines of our experiments so
far. Further investigation of this phenomenon is clearly warranted.
Concluding Remarks
Despite their diversity, our experi-
ments have established a number of
rather consistent facts. At least in mod-
e The betweenness centrality of vertex v is aver-
age, over all pairs of other vertices u and w, of
the fraction of shortest paths between u and w
in which v appears.
erate population sizes, human subjects
can perform a computationally wide
range of tasks from only local interaction. Network structure has strong but
task-dependent effects. Notions of social fairness and inequality play important roles, despite the anonymity of our
networked setting. Behavioral traits
of individual subjects are revealed despite the highly simplified and stylized
interactions; with language removed,
subjects persistently try to invent signaling mechanisms.
There are a number of recent efforts related to the research described
here. Some compelling new coloring
experiments7, 20 have investigated the
conditions under which increased connectivity improves performance. Our
experimental approach has thus far
aimed for breadth, but studies such as
these are necessary to gain depth of understanding. We have also usually done
only the most basic statistical analyses
of our data, but others have begun to
attempt more sophisticated models.
6
Perhaps the greatest next frontier is
to conduct similar experiments on the
Web, where a necessary loss of control
over subjects and the experimental environment may be compensated by orders of magnitude greater scale, both
in population size and the number of
experimental conditions investigated.
Recent efforts using both the open web
and Amazon’s Mechanical Turk online
labor market have started down this
important path.
2, 19, 23
Acknowledgments
Many thanks to the stellar colleagues
who have been my coauthors on the
various papers summarized here: Tan-moy Chakraborty, Stephen Judd, Nick
Montfort, Sid Suri, Jinsong Tan, Jennifer Wortman Vaughan, and Eugene
Vorobeychik. I give especially warm
acknowledgments to Stephen Judd,
who has been my primary collaborator
throughout the project. Thanks also
to Colin Camerer and Duncan Watts,
who both encouraged me to start and
continue this line of work, and who
made a number of important conceptual and methodological suggestions
along the way.
References
1. Camerer, C. Behavioral Game Theory. Princeton
university Press, Princeton, nJ, 2003.
2. Centola, d. the spread of behavior in an online social
network experiment. Science 329, 5996 (2010),
1194–1197.
Michael Kearns ( mkearns@cis.upenn.edu) is a professor
in the Computer and Information science department
of the university of Pennsylvania. his research interests
include machine learning, social networks, algorithmic
game theory, and computational finance.