Applying the repeated game framework to multiparty networked applications

Abstract

This thesis presents repeated game analysis as an important and practical tool for networked application and protocol designers. Incentives are a potential concern for a large number of networked applications. Well-studied examples include routing and peer-to-peer networks. To the extent that incentives significantly impact the outcome of a system, system designers require tools and frameworks to better understand how their design decisions impact these incentive concerns. Repetition is a prevalent and critical aspect of many networking applications and protocols. Most networked protocols and architectures seek to optimize performance over a longer timescale and many have explicit support for repetition. Similarly, most players in networked applications are interested in longer horizons, whether they be firms building a business or typical individuals trying to use a system. Fortunately, the study of repeated interaction between multiple self-interested parties, repeated games, is a well-understood and developed area of economic and game theoretic research. A key conclusion from that literature is that the outcome of the repeated game can differ qualitatively from that of the one-shot game. Nonetheless, the tools of repeated games have rarely if ever been brought to bear on networking problems. Our work presents the descriptive and prescriptive power of repeated game analysis by making specific contributions to several relevant networking problems.

(cont.) The applications considered are inherently repeated in practice, yet our research is the first to consider the repeated model for each particular problem. In the case of interdomain routing, we first show that user-directed routing (e.g., overlays) transforms routing into a meaningfully repeated game. This motivates us to consider protocols that integrate incentives into routing systems. In designing such a routing protocol, we again use repeated games to identify important properties including the protocol period and the format of certain protocol fields. Leveraging this insight, we show how it is possible to address the problem of the repeated dynamic and arrive at a more desirable outcome. In the case of multicast overlay networks, we show how repeated games can be used to explain the paradox of cooperative user behavior. In contrast to prior models, our repeated model explains the scaling properties of these networks in an endogenous fashion. This enables meaningful examination of the impact architecture and protocol design decisions have on the system outcome. We therefore use this model, with simulation, to descry system parameters and properties important in building robust networks. These examples demonstrate the important and practical insights that repeated game analysis can yield. Further, we argue that the results obtained in the particular problems stem from properties fundamental to networked applications - and their natural relationship with properties of repeated games.

(cont.) This strongly suggests that the tools and techniques of this research can be applied more generally. Indeed, we hope that these results represent the beginning of an increased use of repeated games for the study and design of networked applications.