Interaction-Based Trust Evaluation in a Team of Agents Using a Determination of Trust Model

Trust has been widely recognized as one of the most important factors influencing team performance. The ability to accurately evaluate the trustworthiness of team members (agents) is crucial for effective team performance. Interaction data among agents are suitable sources of information determining each agent's trustworthiness. However, the existing interaction-based trust models are usually task specific and are only applicable to some well-defined domains (tasks). This paper addresses the problem of accurate trust evaluation in a team of agents by proposing an interaction-based trust evaluation model - the Determination of Trust Model (DoTM), which is applicable to various team tasks. The DoTM maps the relationships between interaction records and the trustworthiness of an agent through a supervised learning algorithm. To take full advantage of the interaction data, before being fed into the machine learner, interaction data are pre-processed by three data processing methods, i.e., combining data from multiple runs, involving indirect interaction records and calculating relative data across agents. A series of experiments are conducted on a simulation platform which performs a cooperative food foraging task. Different types of flawed agents are introduced to distinguish between agents with different trustworthiness. The experimental results demonstrate that the DoTM achieves high accuracy and consistency in scenarios involving different types of flawed agents. The DoTM is compared with an existing interaction-based trust model - LogitTrust and achieved significantly better evaluation accuracy in all considered scenarios. Moreover, the impact of each data processing method is demonstrated through experimental investigations.