Collaborative agent planning using a subjective logic based trust model

In multi-agent systems, most of the time an agent will participate in some team activities, based on shared plans and aimed at achieving some complex collective goals. The complex goals are decomposed into basic activities to be performed by different individuals in the team; thus an agent might have to fullfill her obligations as a team member by pursuing particular team sub-goals. The agent will have to tackle the problem related to uncertainty about the knowledge, performances and risks involved by the decision to help her peers. This paper presents “Belief Recipe Trees”, an improved solution for the problem of deciding how to engage in a collaborative activity for helping a partner. It extends the approach based on SharedPlans and probabilistic recipe trees by using a trust model based on opinions (subjective probabilities) for the likelihood of success of basic actions performed by particular agents (instead of general probabilities describing the action as a whole). It offers the opportunity to estimate the probabilities even if just a small number of interactions have been previously observed and to judge about whether to seek for more information (and pay a price) or just to rely on the available one is suggested.