No communication or interaction of any kind outside that provided by the system was permitted. The system tabulated the total financial compensation earned by each subject throughout a session, and subjects were paid by check at a later date following the session. Compensation was strictly limited to the actual earnings of each individual subject according to their own play and the rules of the particular task or game; there was no compensation for mere participation. Following a session, subjects were given an exit survey in which they were asked to describe any strategies they employed and behaviors they observed during the experiments.

Within an individual experimen-
tal session, the overall collective task
or problem was fixed or varied only
slightly (for example, an entire session
on graph coloring), while the underly-
ing network structures mediating the
interaction would vary considerably.
Thus, the sessions were structured as
a series of short ( 1 to 5 minutes) ex-
periments, each with its own network
structure but on the same task. This is
the natural session format, since once
the task and incentives are explained
to the subjects, it is relatively easy for
them to engage in a series of experi-
ments on differing networks, whereas
explaining a new task is time-consum-
ing. Each experiment had a time limit
imposed by the system, in order to
ensure the subjects would not remain
stuck indefinitely on any single experi-
ment. In some sessions, there were
also conditions for early termination
of an experiment, typically when the
instance was “solved” (for example, a
proper coloring was found). A typical
session thus produced between 50 and
100 short experiments.

 

Summary of Experiments The accompanying table briefly summarizes the nature of the experiments conducted to date, describing the collective task, the network structures used, the individual incentives or mechanism employed, and some of the main findings that we detail below. Our first remark is on the diversity of these experiments along multiple dimensions. In terms of the

Figure 1. Sample screenshot of subject Gui for a biased-voting experiment; many other sessions involved similar Guis.

The central panel shows the subject’s
vertex (currently in the “blue” state)
with black edges to network neighbors
and their current states; red lines denote
edges between the subject’s neighbors.
The bottom action panel allows the subject
to change their current state any time,
while the top panel specifies their incentives
and elapsed time in the experiment.

Summary of experiments to date. ER stands for Erdös-Renyi, PA for preferential attachment.

task Description

graph coloring17

coloring and consensus10 networked trade13

networked bargaining3 independent set15

biased voting14

network formation16

networks

cycle+chords; PA clique chain w/rewiring eR; PA; structured; all bipartite assorted assorted eR and PA between types; minority power endogenous to the game

incentives/mechanism

differ with neighbors differ/agree with neighbors limit orders for trades for opposing good nash bargain on each edge kings and pawns with side payments consensus with competing individual preferences biased voting minus edge expenditures

Sample Findings
chords help; importance
of information view
opposite structure/task effects
comparison to equilibrium theory;
networked inequality aversion
behavioral price of obstinacy
side payments help;
conflict and fairness
well-connected
minority rules
poor collective
performance
oCTobeR 2012 | VoL. 55 | no. 10 | CommuniCAtionS oF thE ACm 59