Learning Collective Behaviour from Local Interactions

In this paper, we propose a reasoning model for extracting collective behaviours in Multi-Agent Systems (MAS) from regularities in interaction streams. Using reflexive structures to describe the actions and knowledge of the system, called views, we build chronicles that catch and register the actual situations related to events and actions occurrences. These chronicles are then intensionalised in order to extract regularities in a single autonomous agent's behaviour, thus defining a local behaviour. We then propose to extend this model for extracting collective behaviours in MAS using machine learning algorithms. Finally, we try to show that this dynamic analysis of an agent's runtime can lead to organisations in MAS.