Bayesian Cognitive Diagnosis in Believable Multiagent Systems

This paper considers how beliefs can be ascribed in socially intelligent multiagent systems by observing the agents' behaviour, to build an image of their mental state. When agents are 'believable', the actions they perform are a function of their mental pattern and their personality traits. These traits involve different behaviours, and therefore different orderings of beliefs to revise after each interaction step, to keep the mental state image consistent. We examine how a probabilistic ordering of beliefs can be assigned in the hypothesis of a 'closed' multiagent world, and how belief revision can be treated in a bayesian way. We show, in particular, that gradual updating of beliefs enables getting hints for a suspect of deceipt and suggestions on when and how to deceive. 1 Introduction A rational agent takes decisions about how to behave in a given circumstance by reasoning on its own mental state, that is on first order beliefs and goals. When it is programmed so as to show a 'so..