Trust Assessment Through Continous Behaviour Recognition

Computational models of trust typically assume that an assessment of the trustworthiness of an individual can be formed from learning from the outcomes of a sequence of atomic tasks, as well as other evidence such as reports from third parties. Further, they assume that an agent’s trustworthiness can be modelled by a single probability distribution. In this paper we explore alternative mechanisms that allow these assumptions to be relaxed. We propose a trust assessment model based on Markov Switching Regimes, where direct and third-party observations about an agent’s behaviour follow interrelated Autoregressive processes. We argue that this offers a richer model of trustworthiness and a means to combine trust assessment with within-task monitoring.