Online vs. Offline Behavior: How to Design Strategic Agents for Distributed Coordinator-Free Environments

Highly distributed coordinator-free systems with many agents, such as peer-to-peer systems, are receiving much interest in research and practice. In such systems, free riding continues to be a severe and difficult problem. To reach a high degree of cooperation, researchers have proposed numerous incentive mechanisms against free riding. Our concern in turn is a more profound question: Under which circumstances do human individuals indeed resort to effective strategies in those settings? This question is important as humans control the agents. In economics, mimicking the behavior of humans is an accepted method for strategy design. In particular, this holds for complex settings where formal analyses must rely on simplifying assumptions that do not hold in reality. By means of extensive human experiments in one specific setup, namely structured P2P systems, this paper provides evidence that strategies in the settings under investigation often are the result of a non-strategic perspective. This perspective lets participants overlook obvious strategies that are effective. Further, our experiments reveal the following: Online players, i.e., individuals taking part in the system directly, intuitively tend to find better strategies than offline players, i.e., individuals who just implement agents. Offline players have difficulties predicting the strategies of others and overestimate the quality of their strategies. We conclude that a combination of 'online' and 'offline' strategy design is a cost-efficient and effective solution.