Abstract

In multi-agent systems, agents must typically interact with others to achieve their goals. Since agents are assumed to be self-interested, it is important to choose reliable interaction partners to maximise the likelihood of success. Sub-standard and failed interactions can result from a poor selection. Effective partner selection requires information about how agents behave in varying situations, and such information can be obtained from others in the form of recommendations as well as through direct experience. In open and dynamic environments, agents face quick and unforeseen changes to the behaviour of others and the population itself. This paper presents a trust and reputation model which allows agents to adapt quickly to changes in their environment. Our approach combines components from several existing models to determine trust using direct experiences and recommendations from others. We build upon previous models by considering the multi-dimensionality of trust, recency of information, and dynamic selection of recommendation providers. Specifically, we take a multi-dimensional approach to evaluating both direct interactions and recommendations. Recommendation sharing includes information about the recency and nature of interactions, which allows an evaluator to assess relevance, and to select recommenders themselves based on trust.

Citations