Trust-based Incentive Mechanisms for Community-based Multiagent Systems

Abstract

In this thesis we study peer-based communities which are online communities whose services are provided by their participant agents. In order to improve the services an agent enjoys in these communities, we need to improve the services other agents offer. Towards this goal, we propose a novel solution which allows communities to share the experience of their members with other communities. The experience of a community with an agent is captured in the evaluation rating of the agent within the community, which can either represent the trustworthiness or the reputation of the agent. We argue that exchanging this information is the right way to improve the services the agent offers since it: i) exploits the information that each community accumulates to allow other communities to decide whether to accept the agent while it also puts pressure on the agent to behave well, since it is aware that any misbehaviour will be spread to the communities it might wish to join in the future, ii) can prevent the agent from overstretching itself among many communities, since this may lead the agent to provide very limited services to each of these communities due to its limited resources, and thus its trustworthiness and reputation might be compromised.

We study mechanisms that can be used to facilitate the exchange of trust or reputation information between communities. We make two key contributions. First, we propose a graph-based model which allows a particular community to determine which other communities to ask information from. We leverage consistency of past information and provide an equilibrium analysis showing that communities are best-off when they truthfully report the requested information, and describe how payments should be made to support the equilibrium. Our second contribution is a promise-based trust model where agents are judged based on the contributions they promise and deliver to the community. We outline a set of desirable properties such a model must exhibit, provide an instantiation, and an empirical evaluation.