In multiagent interactions, such as e-commerce and peer-to-peer file sharing, being able to accurately assess the trustworthiness of other agents is important for agents to protect themselves from losing utility. We focus on an agent's discount factor (time preference of utility) as a direct measure of the agent's trustworthiness in a number of settings. We prove that an agent's discount factor, when in context of the agent's valuations and capabilities, is isomorphic to its trustworthiness for a set of reasonably general assumptions and definitions. Further, we propose a general list of desiderata for trust systems and show how discount factors as trustworthiness meet these desiderata. We also show how discount factors are a robust measure of trustworthiness when entering commitments with adverse selection and moral hazards.;When agents can significantly increase each other's utility at a moderate cost, the socially optimal outcome is for the agents to provide favors to each other. However, when agents cannot support or enforce a market system, the favor environment forms a situation similar to the repeated prisoner's dilemma because each agent can unilaterally improve its utility by refusing to help others. We present an adaptive tit-for-tat strategy that provides a mutually beneficial equilibrium when agents may have differing private discount factors and when favor costs and benefits are stochastic and asymmetric. This strategy enables agents to treat previously unencountered agents with caution, communicate about the trustworthiness of other agents, and evaluate past communication for deception. We discuss the details of our simulation results and the impact of various parameterizations and communication.;Building from the favor model, we examine more complex transactions with private discount factors as a model for trustworthiness. We closely examine the case of simultaneous favors, which comprise a single market transaction where two parties perform an exchange. Further, we investigate more complex market models, where agents directly compete on price and quality. We derive a number of methods that agents can use to obtain and aggregate information of other agents' discount factors and valuations.;Despite the large body of work in reputation and trust in dynamic multiagent environments, no metrics exist to directly and quantitatively evaluate and compare reputation systems. We present a common conceptual interface for reputation systems and a set of four measurable desiderata, inspired by dynamical systems theory, that are broadly applicable across multiple domains. We discuss the implications, strengths, and limitations of our desiderata. Our discount factor as trustworthiness model performs well across the desiderata when measured against other established reputation models from the literature. We apply our desiderata to empirically evaluate the Amazon reputation mechanism in terms of actual ratings data obtained by sellers on Amazon's marketplace.
展开▼
