Social Norm Recommendation for Virtual Agent Societies

Norms express expectations about behaviours of interacting entities in both human and software agent societies. While humans naturally possess the ability to recognize existing norms and learn new ones, software agents representing human users (e.g. avatars in virtual worlds) have to be endowed with such capabilities. Such a norm recognizing and learning agent, using observed actions and interactions of other agents in a situated environment, can be considered a norm-aware agent. This paper contributes to the agenda of creating a norm-aware agent, by proposing an architecture for social norm recommendation, comprising norm identification, norm classification, norm life-stage detection, and finally norm recommendation. These recommendations can then either be provided to a human user (e.g. the user of an embodied virtual agent in a new environment) or can be used by the agent itself for choosing appropriate actions. We use a simulation-based study to demonstrate how the four phases of the norm recommendation system work. The contributions of this paper are: (i) a comprehensive account of norm recommendation: identification, classification, life-stage detection and how to combine them into a recommendation, and (ii) a first evaluation of the approach, which advances the state of the art for this problem.