Agent-Based Trust and Reputation in the Context of Inaccurate Information Sources

Abstract

Trust is a prevalent concept in human society that, in essence, concerns our reliance on the actions of other entities within our environment. For example, we may rely on our car starting to get to work on time, and on our fellow drivers, so that we may get there safely. For similar reasons, trust is becoming increasingly important in computing, as systems, such as the Grid, require integration of computing resources, across organisational boundaries. In this context, the reliability of resources in one organisation cannot be assumed from the point of view of another, as certain resources may fail more often than others. For this reason, we argue that software systems must be able to assess the reliability of different resources, so that they may choose which of them to rely on. With this in mind, our goal is to develop mechanisms, or models, to aid decision making by an autonomous agent (the truster), when the consequences of its decisions depend on the actions of other agents (the trustees). To achieve this, we have developed a probabilistic framework for assessing trust based on a trustee's past behaviour, which we have instantiated through the creation of two novel trust models (TRAVOS and TRAVOS-C). These facilitate decision making in two different contexts with regard to trustee behaviour. First, using TRAVOS, a truster can make decisions in contexts where a trustee can only act in one of two ways: either it can cooperate, acting to the truster's advantage; or it can defect, thereby acting against the truster's interests. Second, using TRAVOS-C, a truster can make decisions about trustees that can act in a continuous range of ways, for example, taking into account the delivery time of a service. These models share an ability to account for observations of a trustee's behaviour, made either directly by the truster, or by a third party (reputation source). In the latter case, both models can cope with third party information that is unreliable, either because the sender is lying, or because it has a different world view. In addition, TRAVOS-C can assess a trustee for which there is little or no direct or reported experience, using information about other agents that share characteristics with the trustee. This is achieved using a probabilistic mechanism, which automatically accounts for the amount of correlation observed between agents' behaviour, in a truster's environment.

More information

Identifiers

Catalogue record

Export record