How Teams Benefit from Communication Policies: Information Flow in Human Peer-to-Peer Networks David Reitter Katia Sycara Christian Lebiere Yury Vinokurov Antonio Juarez Carnegie Mellon University Pittsburgh, PA 15213 reitter@cmu.edu, katia@cs.cmu.edu, cl@cmu.edu, jerryv@andrew.cmu.edu, ajuarez@andrew.cmu.edu Michael Lewis University of Pittsburgh Pittsburgh, 15260 ml@sis.pitt.edu Keywords: Communication Networks, Social Networks, Belief Propagation, Cognitive Modeling ABSTRACT In an experiment involving teams of humans playing a cooperative game, we study the effect of local communication policies on the efficiency and the performance of teams and of individuals in different positions within a network. This design provides an experimental model of human communities, where information may spread from peer to peer by word of mouth. With this model, we explain the realistic tradeoff between liberal information dumping and targeted information sharing by human peers. Human subjects exchanged natural- language messages with relevance to a task, thereby sharing knowledge across a community. Communication took place along the edges of a small-world graph. Cooperation and individual efforts were incentivized. In one condition, participants were asked to request specific information and only supply information that they knew was needed. In another condition, they were asked to supply and forward as much information as possible. We found that a targeted communication policy was successfully implemented by the participants, increased task success, shortened the time it took to get answers to questions, increased efficiency (task success per communication bandwidth), and may have done so selectively for nodes with fewer connections. 1. Introduction The basic tenet of teamwork is that joint problem-solving increases efficiency: the whole of all collaborators is stronger than the sum of their contributions. The alterna- tive view of collaboration is that “none of us is as dumb as all of us.” Such inefficiencies are commonly blamed on communication overheads. In many modern professional settings, access to information is increased (e.g., through real-time retrieval on the internet), or has been proposed to be increased, for instance with soldiers on battlefields having access to a range of real-time data (Owens, 1995). Two scenarios represent opposite approaches: information is either disseminated widely to all people regardless of its utility, or it is selected by the information consumer based on filtering by, e.g., keywords. Neither of these extremes is satisfactory: wide dissemination burdens the recipient with filtering, while filtering by the sender does not allow down- stream nodes to aggregate information of which they are unaware. To our knowledge there has not been any systematic explo- ration of system architectures that deliver information to the right recipient at the right time through a communica- tion network, or what the potential benefits and pitfalls may